브로드컴: 심층 리서치 보고서

위대한 성장주 가운데 10년 5배를 찾아 — 상방을 묻는다: "훨씬 더 커질 수 있는가?"

• QHow high is its market ceiling? Is it growing a bigger slice of an existing pie, or creating an entirely new market?8/10

  A

  Conclusion first: it is doing both — carving custom share out of the "existing pie" of Nvidia general-purpose GPUs, while turning "hyperscaler in-house custom silicon" into a new category that essentially did not exist before. Viewed through Baillie Gifford's upside lens, the ceiling on this first question is indeed very high: Broadcom's own management draws the 2027 fiscal-year serviceable market (SAM) for custom XPUs + networking at 60 billion~90 billion dollars, plus over 100 billion dollars for networking silicon; and that is only the figure contributed by "three hyperscale customers," with another four customers in contact not counted in (see io-fund's breakdown of Broadcom's SAM and Tom's Hardware on the million-XPU clusters in 2027). In the report's framing, this line corresponds to management's forward visibility of "already having line of sight on 2027 AI chip revenue, at a scale over 100 billion dollars," and to Hock Tan's PSU incentive that ties the performance threshold to FY2030 AI revenue of 60 billion~120 billion dollars.

  How high is the ceiling (upside lens). Pull the lens out to around 2030, and external third parties estimate: the entire AI accelerator TAM is heading toward over 600 billion dollars (2033, Bloomberg Intelligence); within that, the custom ASIC piece climbs at roughly 27% CAGR to about 118 billion dollars by 2033, with its share rising from 8% in 2024 to 19%. The nearer-term SemiEngineering (Geoff Tate) estimate: assuming hyperscalers + OpenAI buy two-thirds of accelerators and split between ASICs and GPUs, custom AI accelerators could reach 100 billion~130 billion dollars by 2030 — directionally consistent with Hock Tan's 2027 figure of 60 billion~90 billion, just further out. And Broadcom currently holds about 80% of custom ASIC share, which even when diluted by new entrants would "very likely stay above 50%." In other words, this pie itself is growing through the rapid upward shift of the ASIC share, and Broadcom is the entrenched leader of that piece — in the upside script, this is a large pool that can still expand rapidly over a ten-year horizon, not a stock market that has already peaked.

  Carving the pie or making the pie? Both, but with different weights.

  • The part carving the existing pie: custom XPUs grab "the portion of hyperscaler compute budget that would otherwise have flowed to Nvidia general-purpose GPUs." This is a substitution relationship — Google TPU, Meta MTIA, Anthropic TPU capacity, all essentially letting big customers escape dependence on a single GPU platform. But to be honest: this "carving" has a ceiling, because GPUs will not be relegated to a supporting role. The external consensus is that GPUs will still hold about 81% of the accelerator market by 2033, and Nvidia will still defend about 70%–75% share before 2030 (Bloomberg Intelligence / SemiEngineering); for flexibility, the CUDA ecosystem, and risk diversification, CSPs will continue to buy GPUs heavily. The report also captures Broadcom's niche precisely: it does not go head-to-head with Nvidia on general-purpose GPUs, but acts as "the preferred collaborator when customers want to take the custom route."
  • The part making a new pie: what truly "creates a new market" is turning "hyperscaler in-house custom silicon" from what used to be near-nonexistent, scattered internal projects into a new industry layer with scale, multi-year long-term orders, and a co-engineering paradigm. That line in the report — "Broadcom went from the person who sold the rebar and pipes when constructing a building to the person who participates in the building's design" — together with Google's long-term order locked to 2031, Anthropic's roughly 3.5GW from 2027 onward, and Meta's first phase over 1GW, is exactly about this: it is not selling one more standard part on an existing shelf, but turning "custom silicon + network interconnect + system co-design" into a new, repeatable, recurring-revenue category.

  But the upside lens does not mean you should buy now — that is the YMYL bottom line. This report's rating is Avoid (downgraded from Watch), and the reason is precisely not that the ceiling is too low, but that the price has already bought in the optimistic realization of that high ceiling ahead of time: at a current price of about $481.57 and a market cap of about 2.28 trillion dollars, the corresponding TTM free-cash-flow yield is only about 1.3%, well below the roughly 4.46% 10-year Treasury over the same period. In the report's owner-earnings scenarios, even the optimistic case (FY2027 revenue about 125 billion, 41x OE) only pushes fair value to about $423, below the current price. So the answer to the first question is: the ceiling is indeed high and real, and this pie is indeed being both carved and grown by Broadcom; but "the market is big enough" answers "can this business grow for a long time," not "is the odds good for entering at this price" — the report judges that the ideal buy must wait below $173, and that does not change because of the TAM story. The height of the ceiling is exactly what explains why the market is willing to give it such demanding pricing, and is also exactly where that pricing is most fragile: what it needs is no longer "the market being big enough," but "that big-enough market being realized into per-share owner earnings with high precision and without missteps."

  2026년 6월 3일
• QCan its revenue at least double over the next five years? Is the growth driven mainly by volume, price, or new businesses?7/10

  A

  Conclusion first: doubling in five years is almost a "low bar," and the driver is highly concentrated in one place — AI XPU/networking volume ramp (volume), not VMware repricing (price), and even less new customers (new business). Starting from FY2025's 63.887 billion, doubling to about 128 billion does not even require a full five years on management's terms. But to tell this upside lens honestly: this report's rating is Avoid — whether it can double, and whether the current price of $481.57 is worth buying, are two entirely different things, addressed separately below.

  Layer one: the arithmetic of doubling is actually quite loose. What management gave in March 2026 was 2027 AI chip revenue "line of sight" exceeding 100 billion dollars, backed by roughly 73 billion dollars of AI backlog, which per the company's framing delivers over roughly six quarters (≈18 months). Take this line of sight at face value: AI alone heads toward 100 billion in 2027, and adding non-AI semiconductors (networking/storage/broadband/wireless; within FY2025 total semiconductor revenue of 36.858 billion, AI is the incremental driver while the rest still has a base of several billion) and software of about 27 billion, total revenue in 2027 could land in the 170–200 billion range — meaning "doubling in five years to 128 billion" is achieved around FY2027 on management's line of sight, and the remaining two or three years are just consolidation. Even with a steep discount, using the report's own far more conservative scenarios — [FY2027 revenue 90 billion (conservative) / 105 billion (neutral) / 125 billion (optimistic)] — the slope to double to 128 billion by FY2030 in the neutral case is not demanding. So the answer to "can it double in five years" is most likely yes; the real suspense was never whether revenue can double, but whether this pile of revenue can convert proportionally into per-share owner earnings.

  Layer two: breaking down the drivers — overwhelmingly "volume."

  • Volume (AI XPU + AI networking) = the absolute main engine. Q1 FY2026 AI revenue was 8.4 billion, up +106% year over year, with Q2 guidance of 10.7 billion (implying about +140%). The increment is not unit-price gains, but customer system-level expansion: six confirmed custom XPU customers — Google, Meta, OpenAI, Anthropic, Apple, and others, a single Anthropic order of about 21 billion dollars delivering from late 2026, Google's long-term contract locked to 2031, and Meta's first phase over 1GW. This is the classic "volume" — customers build out cluster scale and number of generations, and Broadcom's revenue follows the deployed gigawatts and chip count. For doubling, more than 80% must be realized through this one line.
  • Price (VMware repricing) = a one-time step, already largely realized and now exiting. VMware's price hikes/bundling were key to growing the software book in FY2024–2025, but this is a one-time repricing, not sustainable compounding. By Q1 FY2026, infrastructure software revenue growth had already slowed to about 1% (report framing puts software quarterly revenue at about 6.8 billion). In other words, "price" contributed a large chunk in the past, but for "doubling again over the next five years" it contributes almost nothing more, and the roughly 1% low growth is actually a drag — a 27 billion software book climbing only in single digits will at most add several billion over the next five years, a marginal contribution to doubling far smaller than AI. Layering on the European CISPE antitrust complaint and partner-ecosystem friction, the software line is a "cash-flow backstop" for doubling, not a "growth engine."
  • New business (new customers/new categories) = an option, not a locked-in cornerstone for doubling. OpenAI (the sixth, ramping from 2027) and names like Apple provide upside options — whether they can turn from "signing/first phase" into stable revenue at the several-billion scale will only become clear in 2026–2027. They are optional items that further raise the 100-billion line of sight, but should not be counted as a "necessary condition for doubling," or that would mean pre-spending the not-yet-realized into the baseline.

  Layer three (honestly laying out the facts): doubling can happen, which does not mean you can buy now. Baillie Gifford's lens presses on "why hasn't the market realized" — but with Broadcom, the market has, if anything, realized it too fully and already priced in doubling and even more optimistic distant visions ahead of time. The report gives exactly this judgment: the current price of $481.57 corresponds to about 2.28 trillion market cap and a TTM free-cash-flow yield of only about 1.3%, below the roughly 4.46% 10-year Treasury; in the report's owner-earnings scenarios, even the optimistic scenario (FY2027 revenue about 125 billion) gives a fair value of only about $423, still below the current price, so the rating is downgraded from Watch to Avoid, with the ideal buy price set below $173. Two more points must be nailed down: ① as of the benchmark date 2026-06-03, FY2026 Q2 earnings had not yet been disclosed (after the US-Eastern close on 6/3, early morning Tokyo 6/4), so the most critical validation point of 10.7 billion had not yet landed, and the market is "front-running confirmation"; ② the board's PSU for Hock Tan writes AI revenue directly as the performance threshold (any four consecutive quarters in FY2028–2030 with an AI revenue floor of 60 billion, top tier 120 billion), showing that even the incentive is betting on AI volume — this reinforces the judgment that "doubling depends mainly on AI volume," but it also means that once big customers retreat from super-linear expansion back to linear, the entire valuation paradigm switches with it.

  One sentence to close: doubling to about 128 billion over the next five years is highly feasible, and almost entirely depends on the "volume" of AI custom ASICs/networking — VMware's "price" is an already-realized one-time step that has slowed to about 1%, and new customers are an option not a cornerstone. But doubling is a revenue story, and the report's "Avoid" targets the price: revenue can double, yet the current price has already priced in doubling and more, so per this report's framework you must wait for a return to around $300 and see Q2/Q3 nail down AI volume and margins before the odds reopen.

  2026년 6월 3일
• QFive years out, what will take over as the next growth engine? Does this "second curve" exist today?6/10

  A

  Conclusion first. Using Baillie Gifford's upside lens to find Broadcom's "third curve," the most concrete answer is: it does not need to wait five years to invent a new engine, because the two things that most resemble a second/third curve — inference-side custom silicon, and optical interconnect / co-packaged optics (CPO) — are no longer slideware today, but have respectively secured formal orders and entered volume production. But just as honestly: however beautifully this curve is realized, it does not change this report's landing on "Avoid": the optimistic scenario in the body also caps at only $423/share, while the current price of $481.57 (corresponding to about 2.28 trillion market cap and a TTM free-cash-flow yield of about 1.3%, below the roughly 4.46% 10-year Treasury) has already bought in these long-dated tailwinds ahead of time. In other words, the third curve is a "real sprout," but not a "cheap call option."

  Rank four candidates by "maturity that exists today" — this is the core of the question.

  • ① Inference-side custom ASIC: already exists, and already under contract (the hardest). This is not imagination. Broadcom is co-developing a proprietary accelerator with OpenAI, which market reports describe as a 10-billion-dollar-class project positioned for inference, targeting volume production before the end of 2026; OpenAI's overall 10GW framework uses 3nm + 2nm two-generation designs, with the first batch deployed in the second half of 2026; and Meta's MTIA (the I stands for Inference) roadmap is called "alive and well" by Hock Tan. The significance: Broadcom's first curve relies on the expansion of training supercomputers, while inference is the demand side that, once AI truly enters everyday applications, is larger in volume, more sustained, and closer to a "tollbooth" nature. It does not take over after five years, but grows layered on top of the current ASIC curve — and that is exactly the most powerful part of the upside thesis.

  • ② Optical interconnect / co-packaged optics (CPO): already in volume production, but not yet the profit mainstay (second hardest). Broadcom's Bailly platform is the industry's first 51.2T CPO Ethernet switch to enter volume production, with power efficiency about 70% lower than pluggable optical modules and 8x better silicon area efficiency, and has already previewed third-generation 200G/lane technology and signed a data-center optics integration agreement with Corning. Logically CPO is the bottleneck link unavoidable for million-card clusters (electrical signals "can't keep up" at higher bandwidths), and Broadcom, stuck at the switch chip + silicon photonics position, is hard to replace. But to be honest: CPO is still early penetration layered into the switch business, far from a scale that can support a standalone revenue curve, and its more realistic role is to widen Broadcom's networking moat rather than immediately become a third profit pool.

  • ③ More hyperscaler in-house silicon: sprouting, and accelerating (medium). The customer list has expanded from "Google/Meta" to six confirmed XPU customers (Google, OpenAI, Meta, ByteDance, Fujitsu, plus a billion-dollar-class "fifth customer" that analysts guess is Apple or xAI), and disclosed about 73 billion dollars of AI backlog; Arm/SoftBank are also separately building a companion CPU for OpenAI's Broadcom accelerator. This is consistent with the body — this report has confirmed Google's long-term contract to 2031, Anthropic's roughly 3.5GW from 2027 onward, and Meta's first phase over 1GW. It is more like "the same curve widening its customer base" than a brand-new engine, but customers going from 2–3 to 6+ itself reduces single-customer concentration risk (the body also flags the worry of one customer accounting for 32% of FY2025 revenue).

  • ④ Automotive/industrial Ethernet: closer to imagination, at least not a near-term growth engine (the weakest). Broadcom does have positioning in automotive/industrial Ethernet, but this currently belongs to non-AI semiconductors, which management describes as a "U-shaped recovery": enterprise networking, server storage, and industrial remain sluggish, with non-AI semiconductors at about 4 billion dollars quarterly and roughly flat sequentially (management's FY2026 earnings-call framing). Treating it as the third curve five years out is for now mostly "imagination," lacking evidence that it can form high-slope growth.

  Finally back to the honest side. Under the upside lens, Broadcom's biggest "point the market may not have fully realized" is that it has not just one AI curve, but a depth of training ASIC → inference ASIC → optical interconnect, three stages staggered in time and each already landed, which is far steadier than betting on a single generation of products, and is also where it differs from a pure cyclical chip stock. But the report's "Avoid" is not denying this depth, but saying the value of this depth has already been overdrawn by the price: even counting both inference and CPO into the most optimistic scenario, the cap of the body's owner-earnings framework is only $423/share, while the current price of $481.57 and market cap of about 2.28 trillion are higher (Q2 earnings revealed precisely after the close on the benchmark date June 3). So the complete answer to this question is: the second/third curve has indeed sprouted or even landed, which is a real positive; but at the odds of a 1.3% free-cash-flow yield against a 4.46% long bond, "the curve exists" does not automatically equal "you should buy now" — the body gives an ideal buy zone below $173 and a fair holding zone of $263–355, and this question's findings can reinforce the long-term quality thesis but cannot overturn the current valuation discipline.

  This is research analysis, not investment advice. Markets carry risk; invest prudently.

  2026년 6월 3일
• QWhat is its core competitive advantage? Will this moat widen or narrow over the next three to five years?6/10

  A

  Conclusion first: the moat itself is real, but the trend over the next three to five years is not "linear widening," but "binding deeper while being deliberately diluted" — the net effect is most likely flat to marginally wider, not the kind of sustained steepening that Baillie Gifford's upside narrative requires. To capture the tension in one sentence: engineering collaboration binds Broadcom and big customers ever more tightly (the moat's depth is increasing), but the same customers are systematically introducing second and third design sources and pushing up the share of thinner-margin custom ASICs (the moat's "toll premium" is being ground down). These two forces exist simultaneously, so whoever says "Broadcom's moat will only keep widening" ignores what the customers are actively doing on their side.

  What the moat actually is: four lines, but with different weights. The report breaks it down clearly, and I re-rank by "will it change over the next three to five years": ① Design collaboration + switching costs of custom XPUs — this is the hardest line. Google TPU, Meta MTIA, and Anthropic capacity agreements are all multi-year engineering that co-designs the customer roadmap, packaging, networking, system, and software stack together — not standard parts; the April 2026 agreement with Google is written directly to 2031 (report). ② Network interconnect share — Ethernet switching/connectivity is Broadcom's position in AI data centers to "collect tolls even without customization," with elasticity coming from interconnect penetration. ③ Capital-light — FY2025 capex was only about 1% of revenue, with an FCF rate of about 42% (report); this line will hardly change and is structural. ④ Expense discipline — FY2025 gross margin was 68%, partly from the lower software labor costs after the VMware integration (report); this line is management capability, sustainable but not a binding force on customers. What truly determines "widen/narrow" is line ①, because it is both the deepest moat and the place customers most want to loosen.

  The widening force: order depth + engineering stickiness thickening. Don't just look at the status quo — look at the trend; this side is indeed widening: Broadcom's AI backlog (XPU + AI components) has already exceeded 73 billion dollars, about half of its 162 billion total backlog, expected to deliver over roughly 18 months; XPU customers expanded from 3 to 6, and Anthropic added about another 11 billion dollars of orders. In the report's framing Q1 AI revenue was 8.4 billion dollars, Q2 guidance 10.7 billion dollars, and management has "line of sight" on over 100 billion dollars for 2027, and the board ties Hock Tan's PSU reward directly to AI revenue (any four consecutive quarters in FY2028–30 with AI revenue starting at 60 billion, top tier 120 billion). Baillie Gifford would say "what the market hasn't realized is: every additional generation of chips bound and every additional customer raises migration costs exponentially, and the backlog locks in toll rights for the next three years" — this thesis holds at the engineering level. The VMware leg is likewise deepening its lock-in, except the stickiness comes from "migration is too painful" rather than "customers love it" (report) — switching cost, not brand affection.

  The narrowing force: customers actively introducing a second source + custom share diluting margins — this is the real reverse force over three to five years. This side must be laid out honestly: Google is in talks with Marvell over two new chips (a memory-processing unit and an inference-optimized TPU), adding another design partner to its custom-silicon supply chain alongside Broadcom and MediaTek (The Next Web); even more notable, Google has kept the high-value training-side TPU v8AX (Sunfish) with Broadcom while handing the inference-side TPU v8x (Zebrafish) to MediaTek. Although those two Marvell chips are only in talks and not yet under contract, the direction is clear: big customers are actively building "multiple design sources" to hedge dependence on a single partner for pricing/capacity/roadmap. This is exactly the real-world version of the report's pre-mortem scenario where "a weighty customer offloads part of the generations/design work to a second supplier, and Broadcom goes from core co-developer to important but replaceable executor" — what it grinds down is not Broadcom's absolute revenue, but its bargaining rank in the customer roadmap. The second source of erosion comes from the margin structure: custom AI hardware has a gross margin of about 45%–55%, below Broadcom's traditional business, and Mizuho estimates FY2026 is thereby diluted by about 166 basis points of gross margin (Investing.com citing Mizuho); the report also explicitly lists "rising custom-chip share dragging down margins" as a key risk in the optimistic scenario. In other words, the higher this AI revenue curve climbs, the more the "moat gold content" per unit of revenue (gross margin, pricing power) is pulled down instead. Layer on the VMware-side CISPE antitrust complaint and European partner-ecosystem friction (report), and the software leg's moat is "stickiness real, commercialization frictional."

  The net three-to-five-year judgment: the moat "will not narrow but will not significantly widen either," which is not enough for Baillie Gifford's upside narrative. Putting the two forces together, my judgment is — the depth (switching costs, backlog lock-in) is increasing, but the width (irreplaceability to customers, pricing-power premium) is being actively flattened by customers and the margin structure, with the net effect close to flat to marginally wider. Broadcom will not collapse instantly (report: the software side first takes the regulatory/partner hit, the semiconductor side first takes the hit from top-customer allocation and margins), but the Baillie-Gifford-style upside of "the moat keeps widening, driving a five-bagger over ten years" requires Broadcom after 2027 to truly precipitate engineering collaboration into steadier customer stickiness and platform toll rights (the report's own words), whereas the customer behavior observable today (introducing a second source) and the margin trend (custom dilution) are precisely the early signals in the opposite direction. And the valuation reality must be put on the table: this report's rating is "Avoid" (downgraded from Watch), with the current price of about $481.57 corresponding to about 2.28 trillion market cap and a TTM FCF yield of only about 1.3%, far below the roughly 4.46% 10-year Treasury (report). Even if the moat stays the status quo and does not narrow, this moat is already priced by the market on the ultra-optimistic case of "continuing to widen significantly" — the "quality" of the moat and the "price" of buying at this moment are two different things. Baillie Gifford's lens asks "why hasn't the market realized," but on this moat question, the market has not underestimated Broadcom's moat; if anything, it has priced in the not-yet-happened "widening" ahead of time.

  2026년 6월 3일
• QIf its core business were disrupted, does it have the gene for self-reinvention? How does it treat mistakes and bad news?6/10

  A

  Conclusion first: Broadcom has a strong "self-reinvention" gene, but it is a specific kind of reinvention — shifting tracks through M&A + repricing through cost discipline, not through internal disruptive innovation. Its way of treating bad news is "honestly admit + quickly switch direction + but never back down on already-set pricing strategy." This gene has been extremely effective over the past two decades, but applied to Baillie Gifford's question of "if the core business were disrupted, can it save itself," the answer is a qualified "yes."

  One: the history of shifting tracks is real, but in essence it is capital allocation, not technological disruption. The report breaks Broadcom's evolution into four stages — the asset cleanup of the Avago era, acquiring the old Broadcom to become a networking-semiconductor platform, loading in software to become a "semiconductor + software" dual engine, and then on to AI custom ASICs. This migration path does exist: the origin traces to HP/Agilent semiconductors, with KKR/Silver Lake buying it for about 2.7 billion dollars in 2005 to grow Avago, acquiring the old Broadcom Corporation in 2015 at an enterprise value of about 77 billion dollars, and after 2018 bolting on CA, enterprise security, and finally VMware. But the key point is the line the report repeatedly stresses — what Broadcom is good at is not "assembling the map" but "efficiency repricing at an already-mature infrastructure layer." It is not that an analog-device company internally incubated communication chips and then incubated software; it buys mature assets, cuts non-core R&D, and grips the toll rights of high-margin product lines. Third-party summaries of Hock Tan's playbook are consistent: identify key technologies, cut non-core R&D, focus on the world's top 1,000 customers. So your judgment is right: this is "self-reinvention," but reinvention of the M&A-integration type, not the organic-disruption type.

  Two: the best sample of how it treats "its own misjudgments / external bad news" is the 2018 blocked Qualcomm acquisition. This is the most persuasive evidence of Broadcom's adaptability: in March 2018 the US government halted, on national-security grounds, its roughly 117 billion dollar hostile takeover of Qualcomm — a thorough strategic failure. Broadcom's reaction was not to dwell on it or stall, but to turn around within a few months, that November acquiring CA Technologies for 18.9 billion dollars, pivoting the whole company's M&A focus from pure semiconductors to infrastructure software — the very starting point of today's VMware dual engine. In other words, when one path is sealed off, it can extremely quickly transplant the same set of "buy a franchise + cut costs + collect cash" capabilities to a brand-new track. This is exactly what Baillie Gifford wants to see in "does it have the gene to save itself when the core business is disrupted" — Broadcom's answer is: it does not live on by any single generation of products, it lives on by the capital-allocation engine itself, and the engine can point at any high-margin, high-switching-cost mature asset.

  Three: how it treats bad news — verbally very honest, but strategically unyielding — a double-edged sword. Here we must look at two categories, and be honest, because this report's rating is "Avoid":

  • Its own misjudgment (custom-chip margins being squeezed): Broadcom's handling is transparent admission, no whitewashing. The report notes a drop of over 11% after the December 2025 earnings; the actual market reaction was more violent, dropping about 17% over three days. But the reason it dropped was precisely that management proactively put the bad news on the table — the CFO directly said some AI systems would have "lower" gross margins (because more components must be sourced to assemble racks), and Hock Tan added that 2026 AI demand was "hard to judge precisely." This is a management team that does not rely on accounting cosmetics and dares to say unpleasant things, corresponding to the report's rating of "management credibility: high." This is actually a good gene in the Baillie Gifford sense: facing bad news rather than hiding it.

  • External bad news (VMware regulatory backlash): here Broadcom exposes the boundary of its adaptability. Facing CISPE's 2026 antitrust complaint to the EU (alleging price hikes as high as nearly 10x, with some customers at 900%, forced bundling, and charging by forecast rather than actual usage), Broadcom's response is hardline confrontation rather than appeasement: publicly "strongly disagreeing with CISPE's allegations" and calling it "an organization funded by hyperscalers", while still cutting European partners and maintaining price hikes. This is exactly what the report worries about: cost/pricing discipline is its core engine, so it will not soften commercialization just to calm a dispute — for shareholders that is pricing power, for regulators it is an ever-growing tail risk. The report lists this as a "real regulatory risk" but has not yet judged it "triggered."

  Four: the honest landing back on Baillie Gifford's question. If one day the core curve of AI custom ASICs were disrupted (for example, customers pulling design back in-house at scale, or a second supplier demoting it from "core co-developer" to "replaceable executor" — exactly the second scenario in the report's pre-mortem), Broadcom could most likely still transplant the engine to the next high-margin asset as it did in 2018, which is its genuine moat-like adaptability. But two things must be kept clear: first, this kind of reinvention depends on a continued supply of suitable M&A targets + balance-sheet room, and the report has pointed out that tangible net assets are thin while goodwill and intangibles are thick (FY2025 net intangibles of 32.273 billion dollars), so the downside protection is in essence future cash flows rather than asset liquidation; second, its gene is "cost-cutting repricing" rather than "zero-to-one organic innovation" — facing a disruption that needs a true technological-paradigm leap rather than something acquisition + integration can solve, Broadcom's history has not proven it can internally grow a second growth pole. So Baillie Gifford's "upside lens" can give Broadcom a not-low score on adaptability, but this does not change the report's "Avoid" conclusion at all — what the report opposes was never company quality or its adaptability, but the price of $481.57, corresponding to about 1.3% TTM free-cash-flow yield (against the roughly 4.46% 10-year Treasury over the same period), which has already bought in all the optimism of "it will keep doing well." A good self-reinvention gene does not equal a good buy point now.

  2026년 6월 3일
• QDoes management (especially the founder) have a long-term vision, with interests deeply aligned with the company? Are they willing to sacrifice current profit for five-to-ten-years-out?5/10

  A

  Conclusion first: on long-term vision and capability, Hock Tan is almost textbook "qualified"; the direction of interest alignment is right, and the contract locks him to 2030 — but on "willing to sacrifice current profit for five-to-ten-years-out," he is precisely the opposite. Broadcom's cultural core is "high cash returns + cost discipline + buybacks," not the Baillie-Gifford-favored school of "sacrificing current profit at all costs for the long term to reinvest." So measuring by the ruler of Baillie Gifford's question 6, Broadcom scores high on "long-term + aligned" and clearly mismatches on "sacrificing current for the long term." This layer of tension is exactly a profile of a company of extremely high quality that this report nonetheless rates "Avoid."

  Long-term vision and interest alignment: hard alignment, but see both sides clearly

  The positive evidence is solid. Hock Tan has been CEO since March 2006, with almost no deviation over twenty years, assembling a pile of "originally unsexy" semiconductor assets into a semiconductor + software dual-engine platform — this kind of cross-cycle capital-allocation resolve is itself what Baillie Gifford values most as "long-termism." The alignment has been further reinforced recently: in September 2025 the board gave him a new PSU whose core performance metric is not the stock price but AI revenue, locking his service period to FY2030. The threshold the report gives is "any four consecutive quarters in FY2028–FY2030 with AI revenue over 60 billion, top tier 120 billion." Adding finer tiers (not expanded in the report): this award's target share count is about 610,521 shares, with zero vesting below 60 billion, full vesting at 90 billion, double at 105 billion, and a cap of 300% above 120 billion, and Broadcom states about 96% of CEO pay is variable, the vast majority tied to long-term stock price/performance. Directionally, management and shareholders are firmly tied to the single rope of "AI revenue realization" — the report's direct "high" score on "management credibility" holds up.

  But there is a side the report did not lay out, which is very key to Baillie Gifford's question: Tan is a long-term, consistent net seller, not adding to his position. His actual holdings, depending on the framing, are roughly 908,000 shares, about 418 million dollars, in January 2026, a tiny fraction in a company with a 2.28 trillion dollar market cap. And the direction is downward: in June 2025 he sold 40,000 shares for about 10.73 million dollars, in September 148,514 shares for about 50 million dollars, again in late September 100,000 shares (about $339–340), in December 130,000 shares for about 42.38 million dollars, with direct holdings dropping to about 595,638 shares, and in January 2026 about another 70,000 shares. That is, over more than half a year the disposals run on the order of ten billion dollars in value. Such disposals likely follow pre-set plans and are normal for executives, not equivalent to bearishness; but under Baillie Gifford's ruler of "deep interest alignment / willing to hold for the long term," what truly binds him is the contractualized PSU, not his own stock that he keeps reducing — these two have different gold content and must be honestly distinguished.

  Willing to sacrifice current profit for five-to-ten-years-out? — this is precisely the opposite of Broadcom's culture

  This is the tension this question most needs to clarify. Baillie Gifford's ideal picture is the Amazon-style "long-term sacrifice of current profit for the moat at all costs." Broadcom is another species: the report repeatedly stresses that one of its moats is "management's extreme emphasis on capital allocation and expense discipline," relying on "ruthless" cost control, investing R&D only where it best maintains toll rights, FY2025 capex of only about 1% of revenue (623 million dollars), a free-cash-flow rate of about 42%, plus large buybacks (a substantial part of which is merely hedging the dilution from 7.568 billion dollars of SBC). It is a "high-cash-return tollbooth," not a "heavy-reinvestment machine that burns cash for the long term."

  So for question 6, the honest answer is: Tan has a long-term vision, and the PSU hard-ties interests to AI revenue out to 2030 — but he most likely will not actually "sacrifice current profit for five-to-ten-years-out," because sacrificing current profit to reinvest heavily is simply not Broadcom's playbook. For a true Baillie-Gifford-style upside investor, this is not necessarily a demerit (high FCF + strong discipline can also compound), but it means Broadcom's upside does not come from "enduring pain to reinvest heavily and eventually become a giant tree," but from "whether this one string of AI revenue can stay taut." And the report's reason for "Avoid" echoes this: management bets the incentive on the single variable of AI revenue, so the market will fixate on just this one variable — layered with a current price of about $481, a TTM free-cash-flow yield of only about 1.3% (against the roughly 4.46% 10-year Treasury), good management cannot backstop this price. Baillie Gifford's upside lens can appreciate Tan's long-term alignment, but this report's judgment is: at this price, however good the alignment, it does not constitute a reason to buy.

  2026년 6월 3일
• QIf it disappeared tomorrow, how much would customers miss it? Is its growth model sustainable and not dependent on harming society and regulation?6/10

  A

  Conclusion first: "whether it would be missed" and "how that missing came to be" are two entirely different things and must be separated. On the semiconductor side (Google TPU, Meta MTIA, Anthropic), if Broadcom disappeared tomorrow, customers would miss it enormously, near-irreplaceable in the short term; but the "being missed" on the VMware side is in essence "wanting to leave but being unable to" bound by switching costs + price hikes, not "can't do without you." The report itself states the latter plainly: "many customers keep paying because migration is too painful, not because the overhaul is welcome — that's called switching cost, not brand affection," and on that basis downgrades the rating to Avoid. So for Broadcom this question is actually "half a real moat, half overdrawing the customer relationship."

  ① Customer dependence — the semiconductor side is genuinely "much missed." Custom TPU/MTIA are not standard parts of "order today, ship next week," but co-engineering that aligns the customer roadmap, packaging, networking, system design, and software stack over many years. Broadcom used an 8-K to lock its TPU/AI rack agreement with Google to 2031, wrote Anthropic's roughly 3.5GW TPU compute from 2027 onward into formal disclosure, and Meta committed to over 1GW in its first phase. Such deep co-development is extremely hard to switch within 12–18 months, so on the AI side, "disappearing tomorrow would be much missed" holds — this is Broadcom's hardest moat. But to be honest about the risk: the report's pre-mortem lists "a weighty customer offloading more generations/part of the design to a second supplier, with Broadcom retreating from 'core co-developer' to 'important but replaceable executor'" as one of the most likely halving scenarios. That is, "much missed" ≠ "permanently irreplaceable" — hyperscalers naturally have an incentive to cultivate a second supplier. Layer on a fact: a single customer contributed 32% of FY2025 revenue and 44% of receivables — the dependence is mutual, and Broadcom's dependence on big customers is just as high.

  ② Sustainability of the growth model — the VMware line, the manner is the problem. Much of Broadcom's growth after acquiring VMware relies on price hikes + channel cuts + bundling + prepayments to "squeeze cash." CISPE filed an antitrust complaint with the EU in March 2026 and requested interim measures, saying these arrangements pushed cumulative costs up by over 1,000%, with the trigger being Broadcom's termination in January 2026 of the European VMware Cloud Service Provider (VCSP) partner program, kicking the vast majority of European CSPs out of the sales system (CISPE directly called it a "death sentence"). This is not an isolated case: earlier in the AT&T lawsuit, VMware quoted a 1,050% annual price increase. This "forcibly collecting tolls via switching cost" growth is, by definition, overdrawing the customer relationship and actively inviting regulation — it is the opposite kind of "dependence" to the Google/Anthropic "customers actively binding for many years."

  Regulatory facts (verified online to 2026-06-03): the risk is real, but has not yet landed as a substantive remedy. As of the benchmark date, the European Commission's status is still "complaint received, being assessed under standard procedure" — no interim measures granted, no formal case opened, no ruling, and Broadcom's Q1 10-Q also made no material legal-loss provision. The report judges this line as "a real regulatory risk, but it cannot yet be judged triggered," consistent with the latest public status, and it should not be exaggerated into "already disrupted by regulation." Another interesting counterpoint is the AT&T side: its court filings show the cost of migrating off VMware was about 40–50 million dollars, but "payback is fast and IRR is very high, so it decided to prioritize migrating away" — this precisely shows that VMware's switching cost is real but not infinite: when the price hike is absurdly large, the switching cost gets paid off all at once instead. This is exactly the inherent fragility of the "squeezing cash via price hikes" growth model.

  The honest landing under the Baillie Gifford lens (upside lens, but not bullish for the sake of being bullish): Standing on the bullish probe of "why hasn't the market realized," the customer stickiness on Broadcom's AI co-development side has actually been fully priced by the market and even pre-spent — the current price of $481.57 corresponds to about 2.28 trillion market cap and a TTM free-cash-flow yield of only about 1.3% (against the roughly 4.46% 10-year Treasury), on which basis the report gives "Avoid." So for question 7, what truly has not been fully realized is not "how much Broadcom would be missed," but how much of this "being missed" is healthy and voluntary, and how much is squeezed out via price hikes and switching costs and will eventually have to be paid back: the missing on the AI side is healthy and sustainable, while the missing on the VMware side comes with a tail of regulatory cost and customer backlash. Lumping the two together and giving valuation as if they were the same kind of "irreplaceable" is precisely the biggest expectation mismatch in this stock right now.

  2026년 6월 3일
• QWhat are the unit economics of this business (gross margin, incremental returns)? Do they get better or worse as it scales? Where is the money it earns spent?8/10

  A

  Conclusion first: Broadcom's unit economics are top-tier among large-cap tech — FY2025 overall gross margin about 68%, free-cash-flow rate about 42%, capex only about 1% of revenue, almost a "tollbooth" that burns no heavy assets. But incremental returns are getting marginally worse: the custom ASICs driving growth have lower gross margins than legacy chips and software, dragging the blended gross margin down. The money earned mainly flows back to shareholders (buybacks + dividends) and M&A, and a substantial part of buybacks is merely hedging the dilution from equity incentives. The report scores business quality as "strong moat, high cash-flow quality," but the rating is "Avoid" — good business, bad price.

  How good the unit economics are (genuinely very good). In the report's framing: FY2025 revenue 63.887 billion dollars, GAAP net profit 23.126 billion, operating cash flow 27.537 billion, free cash flow 26.914 billion (OCF/net profit about 1.19x, cash thicker than book profit), capex only 623 million, about 1% of revenue, depreciation 574 million — maintenance capex and total capex almost coincide, so owner earnings ≈ FCF, with essentially no accounting trap of "earning money that doesn't come back as cash." Of the two segments, software is the profit engine: FY2025 software revenue 27.029 billion, operating profit 20.765 billion, nearly level with semiconductors' 21.232 billion. Software gross margin is especially high; under Broadcom's official quarterly framing, the infrastructure software segment gross margin is about 93%; while the overall 68% is the result of blending semiconductors at sixty-some-percent gross margin with software at about ninety-percent gross margin (the 68% in the report body uses the GAAP framing; official FY2025 GAAP gross profit 43.294 billion / revenue 63.887 billion ≈ 67.8%, with a higher non-GAAP gross margin of about 78.6%). The root of the extremely high incremental returns is exactly: more revenue requires almost no more plants, and the bulk of marginal profit can fall into cash.

  As it scales: overall still getting better (operating leverage + software backstop), but the gross-margin line is getting marginally worse. The tension comes from product mix — the custom ASICs and AI rack-level systems driving growth most strongly have lower gross margins than legacy chips and software, and the higher the AI share, the more the blended gross margin is pulled down. This is not bearish speculation, but what management itself says: on the Q4 FY2026 call Hock Tan made clear the AI business gross margin is lower than the rest of the company (including software), guiding Q1 FY2026 blended gross margin down about 100 basis points sequentially (mainly the rising AI share, partly also passing through third-party component costs), while stressing that "the gross-margin percentage trends down, but the absolute amounts of gross profit and operating profit still rise," recovering the account at the operating-profit level via operating leverage. The sell side has also cut accordingly: Bank of America lowered its FY26/27/28 gross-margin estimates to about 72.1%/70.4%/70.9% (non-GAAP framing), on the grounds that legacy semiconductors are more profitable than custom ASICs. So the more precise statement is: scaling lets the "absolute amount" of operating profit keep getting better, but the unit gross margin is "getting worse" — exactly the failure scenario in the report's pre-mortem of "rising custom-chip share dragging down margins, larger revenue but thinner."

  Where the money earned is spent (buybacks + dividends + M&A, with the bulk of buybacks filling SBC dilution). Capex costs almost nothing, and free cash flow mainly goes to three places. One, shareholder returns: per Broadcom's official FY2025 cash flow statement, for the year after interest dividends paid 11.142 billion and open-market buybacks 2.45 billion, totaling about 13.6 billion dollars returned to shareholders; the dividend is still growing — the Q1 FY2026 quarterly dividend was raised 10% to $0.65 per share, the 15th consecutive year of growth, and the board approved another 7.5 billion dollar buyback authorization through calendar 2026. Two, there is a key "water" here: on the same statement, another 3.86 billion dollars of stock was bought back to hedge the withholding tax on equity-incentive vesting — this is even larger than the 2.45 billion of open-market buybacks, meaning a large part of buybacks is not actually shrinking the share count but offsetting SBC dilution (FY2025 SBC was as high as 7.568 billion, which the report lists as a risk eroding the "true per-share cost"). Three, M&A and deleveraging: Broadcom's cash has historically served Hock Tan's capital-allocation model over the long term (Avago acquiring the old Broadcom at 77 billion enterprise value, then loading in CA / Symantec enterprise security / VMware), layered with debt repayment — at FY2025 year-end net debt was about 51.883 billion, with net debt/adjusted EBITDA still at over 1x.

  One sentence to close: unit economics and capital allocation are both Broadcom's strengths — the cash-machine character is real, shareholder returns are generous, and the cost-cutting is ruthless; but "incremental returns" are being marginally diluted by the low margins of AI custom chips, and a substantial part of buybacks is merely filling the SBC hole. The report acknowledges the high quality of this business yet rates it "Avoid"; the disagreement is not over whether the business is good, but that the current price (about $481.57, with a TTM FCF yield of only about 1.3% against the 4.46% 10-year Treasury) has already bought this quality cash flow too expensively. This passage is an objective survey of the report and public materials, not investment advice.

  2026년 6월 3일
• QFor it to 5x over ten years, what conditions need to hold simultaneously? Are these realistic? What expectations does today's stock price imply?2/10

  A

  Conclusion first: on the ten-year five-bagger question, Broadcom almost cannot pass. In Baillie Gifford's (LTGG) way, "5x over ten years" requires the market cap to grow from the current about 2.28 trillion dollars (current price $481.57) to about 11.4 trillion dollars — that is not an ordinary growth stock's target, but one that would surpass any company in the world today and requires several extremely favorable things to hold simultaneously. Breaking the conditions down one by one, each "holding alone" is aggressive, and the joint probability of "holding simultaneously" is extremely low. This is exactly the fundamental reason the report gives "Avoid" and presses the ideal buy price below $173.

  The four conditions that must hold simultaneously for a ten-year five-bagger, and the reality of each:

  • ① AI custom-silicon revenue realized to an extremely high scale, and able to compound sustainably. Management's framing is 2027 AI chip revenue "line of sight" over 100 billion dollars (Q1 already at 8.4 billion, Q2 guidance 10.7 billion, with the about-to-be-disclosed Q2 results corresponding per market expectations to about 140% year-over-year growth). But a "ten-year five-bagger" requires not just clearing 100 billion, but stacking another ten years of high compounding on that base, with Google/Meta/Anthropic long-term orders continuously expanding and not switching to a second supplier — the scenario the report's pre-mortem worries about most (deployment retreating from super-linear to linear, stalling at 50–60 billion in 2027). Directionally somewhat positive, but "sustained ten-year high compounding" far exceeds the realized evidence. Reality: medium-low.
  • ② Margins not dragged down by custom silicon. Broadcom's cash quality is a genuine strength (FY2025 FCF rate about 42%, capex only about 1% of revenue, gross margin 68%). But custom ASICs themselves have lower margins than software, and both the report and the market worry the rising AI share will dilute overall margins; in December 2025 it dropped over 11% after earnings precisely because of "lower custom-chip margins." To 5x over ten years, you must keep the owner-earnings rate almost flat while revenue surges — reality: medium, and in inherent tension with condition ① (the harder it pushes into custom silicon, the more it pressures margins).
  • ③ The valuation multiple not contracting. This is the hardest line. Today's price is itself built on "optimistic mega-cap AI platform valuation": trailing P/E about 94x, TTM FCF yield only about 1.3%. For a ten-year five-bagger relying on earnings 5x with the multiple unchanged, owner earnings must also 5x and hold above 40x; if earnings grow less, the multiple must not contract or must even expand further — but the current price is already a historical high, with almost no room for "the multiple to re-expand," and the report's pre-mortem of a "valuation-paradigm switch" compressing from 40x to 25–30x is more realistic instead. Reality: low.
  • ④ Customer concentration and governance risks not blowing up. FY2025 a single customer contributed 32% of revenue and 44% of receivables; VMware faces a CISPE antitrust complaint in Europe, and software growth has slowed to about 1%; SBC is on the high side (FY2025 7.568 billion dollars), continuously diluting per-share returns. None of these going wrong over ten years is a somewhat-optimistic assumption. Reality: medium.

  What expectations does today's stock price already imply? In one sentence: it has already pre-spent "near-perfect AI realization" into the price, and the room left for a ten-year five-bagger has been overdrawn ahead of time by itself. In the report's three owner-earnings scenarios, even the optimistic case (FY2027 revenue about 125 billion, owner earnings about 50 billion, applying 41x) gives a fair value of only about $423, already below the current price of 481.57; the neutral case is about $309, the conservative case about $216. That is, the current price is not betting "Broadcom will do well," but betting "Broadcom will do extremely well and disappoint the market almost not at all after Q2" — already a premium on top of the report's optimistic assumption. As a comparison, the lockable shareholder return at the current price (about 1.3% FCF yield + 0.54% dividend ≈ 1.8%) is significantly below the roughly 4.46% 10-year Treasury risk-free rate, no margin of safety. A starting point that must 5x further on top of an optimistic valuation is mathematically extremely strenuous.

  Honest close: Broadcom is indeed a high-quality compounding company (strong moat, thick cash flow, credible management), and Baillie Gifford's framework naturally wants to find five-baggers among such companies. But "high company quality" does not equal "it can 5x at this price." Under the dual constraint of a 2.28 trillion base + a current price already above the report's optimistic-valuation upper bound, the reality of the four conditions above holding simultaneously is very low, and any one of them slipping (especially valuation-multiple contraction, which is almost the norm at a high base) would turn the five-bagger to nothing. This is why the report rates this "good company" as "Avoid," and explicitly waits for a better price (ideal below $173) before discussing the odds. (The latest stock price/market cap was verified online on 2026-06-03, consistent with the report; data source: stockanalysis.com.)

  2026년 6월 3일
• QWhy hasn't the market realized all this? Is it that it can't understand, looks down on it, or can't see far enough? What will become the "narrative inflection point"?3/10

  A

  Conclusion first: this question is a "reverse" problem for Broadcom. Baillie Gifford's tenth question presupposes "the market hasn't realized, so it's an underpriced five-bagger"; but Broadcom is a star stock at 2.28 trillion dollars market cap, with the sell side almost unanimously "Strong Buy" (S&P Global framing of 47 analysts rating Strong Buy, 48 buys, 0 sells), a 12-month average target price of about $472–487, nearly level with the June 2 close of $481.57. In other words, the market has not "failed to realize," but has highly realized it, and may even have bought all the good news ahead of time. So the honest answer is not to force-fit "a cheap thing it can't understand," but: the market understands the value, but may not have fully priced the risk — this is exactly the reason this report downgrades the rating from "Watch" to "Avoid".

  So it is not "can't understand, looks down on it," but closer to "sees the value, but hasn't fully discounted the risk." Translating into Baillie Gifford's trichotomy: Broadcom neither "can't be understood" (the dual-engine platform, AI custom ASICs + VMware cash flow, is thoroughly researched by the sell side), nor can it be said to be "looked down on" (the valuation is already the top-tier pricing of a mega-cap AI platform). If there must be a component of "not seeing far enough," it is that both directions haven't seen far enough: the bulls haven't fully factored in the downside risk, with the pricing leaving almost no room for error on Q2 realization pressure, customer multi-supplier offloading, custom ASICs dragging down margins, or VMware European regulation — the report computes it bluntly: the current price corresponds to a TTM free-cash-flow yield of only about 1.3%, while the same-period 10-year Treasury is about 4.46%, and adding about 0.54% dividend yield, the return shareholders can lock in remains significantly below the risk-free rate, this buy price has no margin of safety. This is not "the market underestimated Broadcom," but "the market may have overestimated certainty." The real expectation gap more likely hides on the downside.

  On the "narrative inflection point": it could go up or down, and the report judges the downside odds to be less favorable.

  • The upside inflection: Q2 big beat + 2027 line of sight nailed down. Broadcom has pulled its Q2 AI revenue guidance up to 10.7 billion dollars (about +140% year over year, up from 8.4 billion last quarter), with management saying it has "line of sight, over 100 billion dollars" on 2027 AI chip revenue, and wrote the Google TPU long-term contract to 2031, Anthropic's roughly 3.5GW from 2027 onward, and Meta's first phase over 1GW into formal disclosure. If Q2/Q3 deliver consecutively and margins don't collapse, the narrative could upgrade from "AI option value" to "a confirmed compounding platform."
  • The downside inflection: Q2 falling short of 10.7 billion, or "good but not stunning enough." The report repeatedly stresses that Broadcom is currently at the position "most dangerous on expectations" rather than "most dangerous on fundamentals" — with extremely low tolerance for error on any single variable. Important note: as of the report's benchmark date (June 3), Q2 earnings had not yet been disclosed (the call was scheduled after the US-Eastern close on June 3), and the options market implied a post-earnings single-day move of about ±10.65%, showing the market itself knows this is a high-odds validation point. Other downside triggers: Marvell and other second suppliers grabbing part of the generations/design work of a weighty customer (Marvell says its custom-chip revenue will exceed 10 billion dollars in FY2029, with sector sentiment linkage), the CISPE antitrust complaint escalating from "under assessment" to a formal remedy measure by the European Commission, and the rising custom ASIC share structurally dragging down margins.

  This is exactly where the honest tension lies. Force-fitting the Baillie Gifford template would want to say "the market can't understand, there's five-bagger room"; but the fact on the table is the market already prices it as a top-tier AI platform, and the report's worry is precisely overestimation rather than underestimation. So this report gives "Avoid": the ideal buy must wait below $173, the holdable range is $263–355, and above $465 is clearly overvalued, with the current price falling into clearly overvalued. In the report's scenario valuation, even the optimistic case pushes fair value to only about $423, below the current price; the most realistic failure scenario is not the business collapsing, but expectations retreating — 2027 AI revenue stalling at 50–60 billion rather than clearing 100 billion, and the valuation compressing from over 40x owner earnings to 25–30x, so even without profits collapsing the stock could halve.

  One point to make clear: the above is the report's judgment at the pre-market moment of June 3; once Q2 earnings are revealed after the close, whether confirmed upward or refuted downward, it will immediately rewrite the direction of this "narrative inflection point." This is YMYL content, research analysis only, not investment advice; please rely on the latest disclosures and your own judgment.

  2026년 6월 3일

매수 전 반드시 물어야 할 질문 — "좋은 사업"을 찾고, 핵심은 "해자는 누구의 것인가?"

• QCan you explain this company's business model in one sentence?

  A

  In one sentence: Broadcom is a "dual-engine tollbooth"—on one side it sells deeply co-developed custom AI chips (ASICs), Ethernet switching, and connectivity/storage semiconductors to hyperscale customers like Google, Meta, and Anthropic; on the other it charges enterprises for "the systems they don't dare turn off" through VMware-centric infrastructure software. It makes its money in the infrastructure layer that customers find hard to replace yet must keep upgrading, collecting high-margin design fees, chip fees, software fees, and maintenance fees, while carrying almost no heavy-asset depreciation of its own.

  Splitting the two engines apart makes it clear. The first is the semiconductor engine: FY2025 revenue of about $36.9 billion, operating profit of $21.2 billion, with the profit core rapidly concentrating into AI custom accelerators and AI networking—Q1 FY2026 AI revenue already reached $8.4 billion in a single quarter, doubling year over year (+106%), with Google's TPU long-term contract locked through 2031 and Anthropic's roughly 3.5GW of compute from 2027 onward all written into formal 8-Ks. The essence here is not "a networking chip vendor riding the AI wave," but a co-developer of custom solutions that help large customers break free of dependence on a single GPU platform.

  The second is the infrastructure software engine: FY2025 revenue of about $27 billion, operating profit of $20.765 billion, nearly even with semiconductors, with VMware as the main body plus mainframe software, enterprise security, and the like. It makes money not because "customers love it," but because of switching costs—migration is too painful and production environments can't be stopped. The report puts this bluntly: Broadcom collects money for "the systems you don't dare turn off." Worth flagging: this engine's role is retreating from "growth story" back to "cash-flow chassis"—by Q1 FY2026, software revenue growth had slowed to about 1%, and the European CISPE antitrust complaint is heating up.

  Putting the two engines together is what sets Broadcom apart from a pure chip stock or a pure software stock: FY2025 capex was only about 1% of revenue, free cash flow margin about 40%—almost a tollbooth that keeps spitting out cash without needing heavy-asset capacity expansion. And precisely because the business itself is of such high quality, the report rates it "Avoid" this time, attributing the conflict entirely to "price" rather than "business model"—but that falls under "is the current price reasonable" (question seven), which is a different matter from "can this business be explained in one sentence." On the business model itself, Broadcom explains clearly and holds up.

  (Note: all segment and single-quarter figures in this item are consistent with the report's framing; the FY2026 Q2 earnings had not yet been disclosed as of the research baseline date of 2026-06-03, so the above are all confirmed FY2025 annual report and Q1 FY2026 quarterly data.)

  2026년 6월 3일
• QIs this market large enough? Is there still room for growth over the next 10–20 years?

  A

  The conclusion first: the market is large enough, and the runway over the next 10–20 years is long enough—and this is precisely not why the report is bearish. The rating is "Avoid," and the point of disagreement was never "TAM isn't big enough," but "the current price has already bought in the rosier outcomes within this large TAM ahead of time." So this question splits into two layers: Broadcom does stand on two still-expanding major tracks, but "the market is big" and "you should buy at the current price" are two different things.

  Track one: AI custom ASIC + networking interconnect—the engine with the largest space and highest visibility

  This is the main axis of the Broadcom story. Stack a few framings together and the TAM is anything but small:

  • Broadcom's own "Serviceable Available Market (SAM)": management frames the FY2027 AI revenue SAM for XPU + networking interconnect at $60 billion to $90 billion. The origin of this framing is that it has three hyperscale customers each with multi-generation XPU roadmaps, expected to deploy 1-million-XPU clusters each on a single fabric in 2027; beyond these three, about four more hyperscalers are already in deep talks to have Broadcom build their self-developed accelerators, meaning the potential customer pool is moving from 3 toward 7. The report's main text gives a more forward revenue framing—management says it has line of sight to "chips only" AI revenue >$100 billion for 2027; in other words, the company reported both a more TAM-oriented $60–90 billion SAM and a more self-capture-oriented >$100 billion framing, both pointing to "this curve can still ratchet up several more steps."
  • The upstream funding pool: the report's framing is that the combined AI infrastructure capex of several major firms that year was at least $630 billion, and was revised up in May to over $700 billion. Broadcom's revenue doesn't come from selling more phones to end users, but from extracting supply-chain profit from this multi-trillion-scale capex—the reservoir itself is getting bigger.
  • Third-party industry TAM also lines up: Omdia estimates the AI data center chip market at about $286 billion by 2030, and explicitly notes that the custom ASIC share is rising; the more optimistic IDTechEx and similar framings push data center/cloud AI chips above $400 billion by 2030. Note that Omdia simultaneously gives a caveat unfavorable to bulls: the overall AI chip incremental growth rate may peak around 2026—that is, the track is still large, but "super-linear expansion" may not last forever, which is exactly the industry footnote to the report's pre-mortem failure script of "2027 AI revenue stalling at $50–60 billion rather than blowing past $100 billion."

  To sum up this layer in one sentence: over the next 10–20 years, AI compute expanding from training to inference and from general-purpose GPU to customer-self-developed ASIC is a structural rather than one-off wave, and Broadcom is wedged into the position of "helping large customers escape dependence on a single GPU platform," so long-term space is not lacking.

  Track two: infrastructure software (VMware-centric)—large but mature, with growth already clearly downshifted

  This piece is the profit chassis, not a high-growth engine. FY2025 software revenue of $27.029 billion, operating profit of $20.765 billion, nearly even with semiconductors—absolutely large enough in scale, and the installed-base market for private cloud/virtualization is itself in the hundreds of billions (the private cloud market framing reaches nearly $200 billion by 2030, virtualization software on the order of $364.8 billion by 2030). But "the market is big" needs a discount here: Broadcom's play is to raise prices and collect maintenance fees on top of the installed infrastructure via switching costs, relying on "migration is too painful, production environments can't be stopped," rather than penetrating a new incremental market. The report already notes that Q1 FY2026 infrastructure software revenue growth had slowed to about 1%, compounded by the European CISPE antitrust complaint and partner-ecosystem friction. So this track's contribution to long-term space is "stable thick cash flow + a price-increase ceiling," not "recreating an incremental curve like AI"—it provides certainty, not slope.

  Finally, bringing this question back to the rating framing

  The market is big enough and the long-term space is there—on this point the report and external data agree, and corporate governance votes with its feet too: the board ties Hock Tan's PSU incentive directly to AI revenue, with FY2028–2030 thresholds of $60 billion for any consecutive four quarters of AI revenue and a top tier of $120 billion, which amounts to the company itself acknowledging that this TAM can support a forward revenue range of more than double.

  But please separate "big space" from "reasonable current price": the report gives "Avoid" precisely on the premise of acknowledging that the TAM is very large and AI is still accelerating, with the reason entirely in price—the current price of about $481.57 corresponds to a market cap of about $2.28 trillion and a TTM free cash flow yield of only about 1.3%, below the roughly 4.46% 10-year Treasury yield. In other words, by the report's judgment, such a large market has already been fully (even excessively) priced in; its bearishness is not a bet that the track will shrink, but a bet that "even if the track is very large, squaring the current valuation requires near-perfect, high-precision delivery." The answer to this TAM question is yes—large enough and long enough; but the report's stance is: a big TAM ≠ now being a good buy point, with the ideal buy waiting for a return to below $173 and a holdable range of $263–355. (The above is a restatement of the report's judgment and public facts, and does not constitute separate buy or sell advice.)

  2026년 6월 3일
• QIs its moat deep enough? Is it hard for competitors to replicate?

  A

  The conclusion first: Broadcom's moat is real, but it is not a wide, brimming moat—it is a set of "narrow and deep control points that people are now trying to bypass." Its hardest barrier is not VMware, but the "customer collaboration + architectural positioning + extremely light capital" triad on the custom ASIC side; VMware is more of a "real switching cost with an overstated commercialization narrative." For competitors to replicate it in place is very hard, but eroding its share piece by piece at the edges has already begun—and this is exactly the key tension behind a rating landing on "Avoid" while the moat is still graded "strong" by the report yet does not equal "should buy now."

  Three moats, ranked by real hardness

  I basically agree with the three moats the report names, but I would re-rank them:

  • Hardest: capital-light yet irreplaceable architectural positioning. This is Broadcom's scarcest asset. FY2025 capex was only $623 million, about 1% of revenue, with depreciation of $574 million, almost stuck together; yet it can charge across three high-value links in AI data centers—switching, interconnect, and custom accelerators—simultaneously. In other words, it doesn't carry tens of billions in data-center and GPU depreciation, yet it takes the fattest supply-chain profit pool from customers' capex—this is what the report calls the "tollbooth." This positioning is hard to replicate because it is built on underlying IP that requires long-term accumulation, such as SerDes, packaging, and network system-level design; external industry data generally puts Broadcom's high-end switching chip share above 80% and custom AI accelerator share at about 70% (note: shares are third-party estimates, not Broadcom official disclosures).
  • Second hardest: customer collaboration + switching cost. Google TPU, Meta MTIA, and the Anthropic compute agreement are not "order today, ship next week" standard parts, but engineering that takes years to coordinate the customer roadmap, packaging, networking, and software stack. The 8-K facts in the report hold up: in April 2026 Broadcom signed its TPU/AI rack agreement with Google through 2031, with Anthropic at about 3.5GW from 2027 onward and Meta's first phase exceeding 1GW. Why don't customers switch easily? Because custom chips can deliver 30%–50% lower total cost of ownership than general-purpose GPUs on hyperscale workloads, and once a design is baked into a customer's entire system generation, migrating means starting over. This is a real switching cost.
  • Third hardest: cost discipline. FY2025 gross margin rose to 68%, and Broadcom itself notes in the 10-K that part of the improvement came from lower software labor costs after VMware integration. This is less a moat than the execution moat of Hock Tan's twenty years of "knifework"—it lets the first two moats keep being efficiently monetized, but it is itself a capability, not a structural barrier.

  VMware: real switching cost, overstated "love"

  I completely agree with this line in the report, and would put it even more bluntly: VMware's stickiness is real, but what's sticky is "migration is too painful," not "the changes are welcomed." Customers keep paying because moving the virtualization base away means stopping production, re-testing, and taking on extreme risk—not because Broadcom raised prices to everyone's satisfaction. The biggest problem with this "moat" is that it is being eroded by regulation and the partner ecosystem: CISPE has formally filed an antitrust complaint with the EU and requested interim measures in March 2026, alleging that Broadcom terminated the European VMware cloud service provider (VCSP) program and, through bundling and prepayment commitments, raised costs by over 1000%; the European Commission confirmed receipt and is assessing it under standard procedure, but as of the baseline date no interim measures had been imposed. Compounded by Q1 FY2026 infrastructure software revenue growth already slowing to about 1%, my judgment is: VMware is a moat, but not a growth moat—it should be treated as the chassis for profit and cash flow, not given an extra growth premium. Pricing it as a "love-type brand moat" is the overstated, pseudo-moat part.

  Versus Nvidia / Marvell / AMD: hard to replicate, but already being "bypassed"

  • vs Nvidia: the two are fundamentally not the same kind of moat. Nvidia relies on full-stack CUDA ecosystem lock-in (FY2026 Q4 revenue $68.127 billion, GAAP gross margin 75.0%), which is "platform dominance"; Broadcom is precisely the party helping large customers escape platform dependence. This means Broadcom's moat doesn't conflict with Nvidia's, and may even be complementary, but it also means Broadcom's fate hinges on the premise that "hyperscale customers are willing to keep self-developing"—if one day general-purpose GPU price/performance overtakes and customers pull back on self-development, demand elasticity on this Broadcom line will be more fragile than Nvidia's.
  • vs Marvell: this is Broadcom's most direct listed rival on the "custom ASIC + interconnect" track, with the two together controlling about 95% of the custom ASIC co-design market. Broadcom's advantages are scale, depth of customer relationships, and a software cash-flow buffer; Marvell is purer, higher-elasticity, and more volatile (focused on AWS Trainium and Microsoft Maia). But here there is a key new signal that directly shakes the report's "Google long contract = deep moat" narrative: just days after Broadcom locked the agreement through 2031, Google was reported to be in talks with Marvell to co-develop two new AI chips—a memory processing unit (MPU) to pair with existing TPUs, and a new TPU dedicated to inference—amounting to bringing in a third design partner beyond Broadcom and MediaTek (note: the negotiations have not yet been signed). This shows that even Broadcom's deepest customer is actively leaving itself "a second supplier + bargaining leverage." Counterpoint accordingly expects Broadcom's custom ASIC share to fall back from about 70% in 2026 to about 60% in 2027, with Marvell rising to about 25%. The moat isn't broken, but "irreplaceable" is being downgraded to "important but partially replaceable."
  • vs AMD: AMD is strong in standard CPU/GPU (FY2025 data center revenue $16.6 billion, +32% year over year), both competing with and complementing Broadcom—customers going the "self-developed ASIC + Broadcom co-development" route will reduce general-purpose GPU dependence, but in hybrid deployments Broadcom networking may also coexist with AMD. Broadcom's moat was never "beating AMD," but "becoming the preferred collaborator when customers want to take the custom route."

  To close in one sentence: the moat is deep enough, and in-place replication by rivals is extremely hard (SerDes/packaging/system collaboration + twenty years of customer relationships + a software cash-flow buffer), which is the basis for the report grading the moat "strong." But "deep enough" doesn't equal "impossible to bypass"—the VMware leg is being gnawed by regulation and the partner ecosystem, and the custom ASIC leg is being diluted by customers actively cultivating second suppliers. This is exactly the report's sharpest point: it doesn't deny the moat, but reminds you that the market has priced in above $481 (corresponding to a roughly 1.3% TTM free cash flow yield vs the 4.46% 10-year Treasury) on the assumption of "the moat is never bypassed + perfect long-term AI delivery," pricing a moat "whose boundaries are being tested" as one that is "always full." The factual judgment on the moat (strong) and the judgment on buy price (Avoid, ideal buy point below $173) are not contradictory here—the above is a restatement of the report's judgment and public facts, and does not constitute investment advice.

  2026년 6월 3일
• QWhere does its growth come from? (Industry growth / market share / price increases / capital allocation)

  A

  The conclusion first: the absolute mainstay of this round of Broadcom's growth is the pairing of "industry growth × market share," and the two are highly superimposed—the AI capex pie itself is exploding in size, and Broadcom is eating an ever-larger share on the non-Nvidia "custom ASIC + AI networking" route. Price increases (VMware repricing) are second-tier, having already retreated from "growth engine" to "cash-flow chassis"; capital allocation is more about "locking in + thickening per-share" the cash earned above, rather than recreating a new growth source. The four sources are not used with equal force, with weights roughly: AI volume (industry + share) ≫ VMware price increases > capital allocation.

  Below I break it down by the four sources this question asks for; I've tagged FY/quarter on figures from the report's framing, and attached source links to everything outside the report.

  1) Industry growth—this is the largest sail. Broadcom's revenue elasticity doesn't come from selling one more phone at the end, but from several hyperscale customers rewriting capex structure at the system level. The report cites: 2026 major-firm AI infrastructure spending is expected to total at least $630 billion (Reuters, March), and by May Marvell's public statement had revised it up to over $700 billion. Broadcom itself doesn't carry the data centers and GPU depreciation, yet it can take supply-chain profit from this capex—FY2025 capex was only about 1% of revenue, a textbook case of "others build the building, I sell the rebar and pipes." This layer is the "rising tide"—it determines how high the ceiling is.

  2) Market share—Broadcom's true alpha is here, and it's accelerating. This is the piece most worth heeding among the four sources. Per the report: Q1 FY2026 AI revenue $8.4 billion, +106% year over year, Q2 guidance $10.7 billion (the earnings itself had not been disclosed as of the report baseline of 6/3, so this is still guidance, not actuals). Adding external disclosures makes it clearer—management-guided Q2 AI revenue is about +140% year over year, +27% sequentially, with the slope not flattening but steepening. The source of share splits into two parts:

  • Custom ASIC / XPU: Broadcom builds the self-developed accelerators for major customers like Google TPU, Meta MTIA, Anthropic, and OpenAI, essentially helping customers "escape Nvidia platform dependence." The hardest share evidence is the order book—management disclosed AI orders with a backlog exceeding $73 billion, covering about the next 18 months, roughly half of total combined backlog, and recently signed a fifth XPU customer (about $1 billion in orders, delivery by end of 2026); of which Google, Meta, and Anthropic together account for about 70% of AI revenue and the 2027 pipeline. The report also confirms customer long-termization: the April 8-K locked the Google TPU agreement to 2031, with Anthropic at about 3.5GW from 2027 onward and Meta's first phase already exceeding 1GW.
  • AI networking interconnect: this piece is often overlooked, but share is rising fast. Tomahawk/Jericho switching + optical interconnect eats the attached budget of "every $1 of XPU typically pairs with $0.4–0.6 of networking," with management framing AI networking climbing to about 40% of Q2 AI revenue. That is, in this AI layer Broadcom eats two shares—"compute + connectivity"—rather than only selling accelerators.

  Putting these two together: the 10-K in the report explicitly names FY2025 semiconductor growth as "mainly driven by custom AI accelerators and AI networking products"—so for Broadcom, industry growth and market share are superimposed in the same direction, very hard to cleanly separate, which is the root reason its valuation has been pushed so high.

  3) Price increases—VMware repricing, but its positioning has slid from "growth" to "chassis." After acquiring VMware, Broadcom did textbook commercialization repricing: switching to pure subscription, per-core licensing with a 72-core minimum purchase, and converging product lines into VCF/vSphere Foundation big bundles, with quite a few small-to-mid customers and universities facing annual cost increases of 200% to over 1200%, the strategy clearly being "prefer fewer customers paying more each, take high margin over market share." This play did contribute high software growth + gross margin lift in FY2025 (report: FY2025 software revenue $27.029 billion, software operating profit $20.765 billion, nearly even with semiconductors; gross margin rose to 68%, part of it from lower software labor costs after integration). But the key turning point is: by Q1 FY2026, infrastructure software revenue growth had slowed to about 1% (about $6.8 billion)—the one-off dividend of price increases is being digested. So VMware now looks more like a stabilizer "underpinning" overall cash flow and margins, rather than the main engine of marginal increment; and it carries the tail risk of the European CISPE antitrust complaint (report: filed in March, being assessed by the European Commission, but no formal remedy imposed as of the baseline date).

  4) Capital allocation—lock in cash + thicken per-share, rather than recreate growth. Hock Tan's twenty-year play is "buy mature, critical infrastructure assets → cut costs → wring contract cash flows into a steeper return curve" (Avago → old Broadcom $77 billion, CA, Symantec enterprise security, VMware, pieced together into today's dual engine). But at the current juncture, capital allocation mainly takes the form of shareholder returns rather than new acquisitions: FY2026 Q1 newly authorized $10 billion buyback, quarterly dividend raised to $0.65 (+10%), FY2025 full-year return to shareholders of $13.6 billion (of which $11.1 billion dividends, $2.5 billion buybacks). Two honest "discounts" to the report's framing here: one, a substantial part of the buyback only hedges equity-incentive dilution—FY2025 SBC was as high as $7.568 billion, Q1 FY2026 still $2.176 billion, not pure share shrinkage; two, the company hasn't publicly disclosed its next major acquisition plan, so M&A integration's contribution to "future increment" is currently more about historical momentum (VMware profit has already been wrung out) than a landed new engine.

  Finally, back to this report's judgment (faithfully restated, not drawing the opposite conclusion for it): the report acknowledges the above growth narrative—especially the AI-volume share expansion—as extremely high quality, with the latest data "accelerating rather than stalling"; it gives "Avoid," ideal buy price below $173, with the conflict not in growth itself but in price having already bought in this growth ahead of time (including Q2, this yet-to-be-disclosed validation point, and the >$100 billion forward line of sight for 2027)—the current price of about $481.57 corresponds to a TTM free cash flow yield of only about 1.3%, below the roughly 4.46% 10-year Treasury. In other words: the growth sources are very hard, but the report holds that "a good business" doesn't equal "the current price is a good buy point." The answer to question 4 itself is clear—growth comes mainly from AI industry expansion and Broadcom's share lift in custom ASIC/networking (the two superimposed are the absolute mainstay), VMware price increases are a now-flattened second tier, and capital allocation mainly serves to lock in and thicken per-share.

  2026년 6월 3일
• QIs management reliable? Are they honest and rational?

  A

  Conclusion first: on the management gate, Broadcom is the highest-scoring of the seven questions—Hock Tan's twenty-year capital allocation record is real "honest and rational," and the board using AI-revenue-bound PSUs to keep him through FY2030 does align interests with shareholders. But "rational" doesn't equal "paying up for the current price," and high SBC plus a CFO parachuting in from the largest customer's camp right around the Q2 validation point are two flaws you must look at with eyes open. The report gives this one "management credibility: high," and I agree with that direction; below I lay out the facts in full and fill in what the report didn't expand on.

  One, capital allocation record: this is the source of Broadcom's moat that needs the least questioning. Hock Tan has been CEO since March 2006, with cumulative total shareholder return exceeding 3500% since the 2009 IPO. His play has barely deviated in twenty years: buy mature but critical infrastructure assets (the roughly $77 billion enterprise-value acquisition of old Broadcom in 2015, CA in 2018, Symantec enterprise security in 2019, the $69 billion VMware in 2023), cut redundancy, press resources into the highest margin, charge with price discipline, and roughly split free cash flow half into dividends, half into the next deal. The result is that set of numbers in the report: FY2025 revenue $63.887 billion, capex only about 1% of revenue, FCF $26.914 billion (FCF margin about 42%). Measured by Buffett's yardstick—"honest and rational" first looks at how the CEO spends shareholders' money—giving Tan full marks on this item is not excessive.

  Two, incentive design: direction aligned, but bets the whole company on one variable. The report already notes the key fact; I'll complete all tiers of the PSU. On September 3, 2025, the board's independent directors granted Tan a PSU with a target of 610,521 shares, the performance metric being not stock price but AI Revenue: the measurement window is any consecutive four quarters from FY2028 to FY2030, with AI revenue below $60 billion vesting zero, reaching $90 billion vesting in full, $105 billion doubling, and over $120 billion earning 300%, and requiring Tan to remain employed through the end of FY2030 (the report's framing only gives the two endpoints of $60 billion/$120 billion; I filled in the two middle tiers from the proxy statement disclosure). The good side: this welds the CEO's wealth directly to AI delivery, far more honest than schemes that dress up short-term EPS, and it answers the biggest succession concern for a company with a single helmsman of twenty years—"retention." The side to be wary of: when incentives are extremely focused on a single variable, management framing has an inherent motive to "talk it up"—management's March 2026 forward-looking statement that it has "line of sight to over $100 billion" in 2027 AI revenue is in essence visibility rather than realized revenue, and that is precisely where the PSU-unlock money sits; the same AI number is both guidance to shareholders and the CEO's bonus line, so read it with a discount.

  Three, CFO transition: both the timing and the provenance warrant a couple of extra looks. This is the point the report tagged "neutral, leaning cautious," but which I think can be spelled out one more layer. Broadcom announced on April 2, 2026 that Amie Thuener would become CFO, effective June 12; the old CFO Kirsten Spears retired after twelve years at the company and stays on as an advisor for nine months to ensure a smooth transition—the handover arrangement itself is orderly, not a crisis-style departure. But there are two details the report didn't expand on: first, the timing—changing the finance head at the same time as the most critical validation point, Q2 on June 3/4, is not a plus for narrative credibility at a company priced by the market as an "AI high-elasticity cash-flow platform"; second, and more important—the new CFO comes from Alphabet, having served since 2018 as VP, corporate controller, and chief accounting officer at Google's parent, with a résumé including PwC and FASB. This means Broadcom's new CFO comes directly from its single largest customer's camp (the report discloses that in FY2025 one customer contributed 32% of revenue and 44% of receivables, with the market broadly pointing to Google/TPU). This double-edged sword cuts both ways: on the upside, it brings the largest customer's financial language and depth of relationship into its own boardroom; on the downside, a revolving door of finance executives between the largest customer and its supplier is a signal where the governance boundary of interests bears watching, not a neutral personnel change to be skipped. Her base salary is $700,000, target bonus 100%, signing bonus $1 million.

  Four, high SBC: this is the most concrete deduction within "rational." Per the report: FY2025 stock-based compensation of $7.568 billion, Q1 FY2026 still $2.176 billion. Broadcom does buy back, but a substantial part of the buyback only hedges incentive dilution rather than being pure share shrinkage, meaning that within that pretty "surface high cash flow," one piece is first handed to employees and executives, then bought back with shareholders' money. It hasn't faked cash flow—operating cash flow/net income is about 1.19x, and owner earnings nearly hug FCF—but "the owner earnings truly left for ordinary shareholders per share" should take a discount versus headline FCF, which is exactly why the report grades financial soundness only "upper-middle" and lists SBC as one of the four bearish points.

  To close in one sentence: on the management question, Broadcom passes, and passes beautifully—discipline, alignment, and retention are all in place, and a large part of the answer to "who owns the moat" is Tan's own twenty-year-flawless capital allocation capability. But please separate two things: management being credible doesn't equal the current price being buyable. Tan is a famous believer in price discipline, never chasing up on his own asset purchases; when it's the ordinary investor's turn to buy this stock, the report gives precisely "good company, bad price, rating Avoid, ideal buy below $173." Applying this high regard for management as-is to the buy timing is exactly the misjudgment the report most wants to stop you from making.

  2026년 6월 3일
• QWill it be stronger in 10 years? (Will users, profits, and brand strengthen?)

  A

  The conclusion first: ten years out, Broadcom will most likely be "stronger, but not with a wider moat." Its scale, profit pool, and engineering ties with hyperscale customers will all grow thicker; but the moat's "exclusivity"—that is, "only Broadcom can do this"—will likely be slowly diluted from today's near-monopoly. This is a question I'm answering as the 6th (the "will it be stronger in 10 years" among Buffett's seven good-business questions); breaking it down with the report's framework, the most honest judgment is: profit and scale strengthen, position doesn't weaken, but that wall of relative monopoly will narrow.

  First, the forces that make it "stronger in 10 years"—these the report explains very clearly, and they hold up.

  • Customer collaboration has upgraded from "supply" to "co-build." The report notes that Broadcom used an 8-K to lock its TPU/rack agreement with Google to 2031, with Anthropic at about 3.5GW of TPU compute from 2027 onward and Meta MTIA's first phase exceeding 1GW. These are not standard-component orders, but multi-year engineering that welds the customer roadmap, packaging, networking, and system design together across several generations. Beyond the report, two more orders of magnitude can be added: behind these collaborations is, per management's framing, about $73 billion of AI backlog, and identified XPU customers have expanded to six—Google, Meta, ByteDance, Anthropic, Fujitsu, OpenAI, and the like. The more customers and the deeper the ties, the stronger the platform effect.
  • The real chokepoint is networking and SerDes, not just logic design. This is the physical bedrock of the report's "capital-light but irreplaceable architectural position" moat: Broadcom's Tomahawk switching ASIC and leading SerDes IP are the backbone that connects tens of thousands of custom accelerators into training clusters—"without Broadcom's networking silicon, hyperscale customers can't run their own ASICs at scale." The TCO economics of doubling bandwidth per generation and replacing multiple old chips with a single chip will keep pinning customers to the same supplier. This leg only gets thicker ten years out.
  • The incentive mechanism welds management to "scaling up AI." The report mentions the PSU the board gave Hock Tan in September 2025, with the performance metric being AI revenue directly, aiming to keep him through FY2030. The external public framing is more specific: for any consecutive four quarters of AI revenue in FY2028–FY2030, below $60 billion vests not a single share, blowing past $120 billion earns the full 300%. This answers the question "will it try harder to thicken customers and profit in ten years" with real money: "yes."
  • Software cash flow is the chassis. VMware can underpin profit even without growth (FY2025 software operating profit $20.765 billion, nearly even with semiconductors), giving Broadcom the surplus to outlast customers across multiple generations. So across the three dimensions of customer count, profit pool, and brand (the preferred co-developer for "de-Nvidia"), being "stronger" ten years out is basically certain.

  But whether the moat gets "wider," I differ from the simple bulls—I think it narrows, and the most threatening force the report itself names in its pre-mortem.

  • Customer self-development moving outward and second suppliers have shifted from "risk assumption" to "something already happening." The report's second pre-mortem, "weighty customers move more generations/partial designs outward to second suppliers, and Broadcom is downgraded from core co-developer to replaceable execution vendor"—this isn't ten years away, it's happening now. The hardest evidence is Broadcom's largest customer, Google: its TPU now has Broadcom + MediaTek dual design partners, and it is in talks with Marvell on an inference TPU and memory processing chip, pulling a third party into the supply chain. Marvell's own earnings also explicitly list "customers developing fully internal design capabilities" as a structural threat that holds for both first and second suppliers. Putting these together, Counterpoint expects Broadcom's custom ASIC share to slide from today's 70%+ toward about 60% in 2027. Share falling and customers betting across multiple lines means the bargaining tier is, on a ten-year horizon, more likely to be ground down than fortified.
  • Single-customer concentration is an amplifier. The report discloses that in FY2025 one customer accounted for 32% of revenue and 44% of receivables, and lists this as a hard signal that "triggers re-assessment." The real test of whether the moat is "wide enough" is how far this behemoth customer goes on self-development and multi-sourcing—each step it moves outward thins the "irreplaceable" quality of Broadcom by one notch.
  • Regulatory and software-side friction cuts the other leg. VMware's CISPE antitrust complaint is heating up, software revenue growth has slowed to about 1%, and the report puts it well: that is "switching cost," not "brand love." Stickiness maintained by migration pain is, the longer it lasts, the more likely to be replaced by cloud-native and forced loose by regulation—ten years out, the software moat is more likely flat or slightly narrowed, not wider.

  So my net judgment on question 6: ten years out, Broadcom will be a larger company with thicker profit and more customers—"stronger" holds; but today's relative monopoly of "near sole dominance in custom ASIC + networking-silicon chokepoint" will be diluted under the three forces of customer self-development, second/third suppliers, and regulation, with the moat evolving from a "deep well of exclusivity" into a "very wide moat that others come to divert water from." In other words, the moat's absolute thickness (the engineering barrier of networking/SerDes/system collaboration) will be preserved or even deepened, but its relative exclusivity will narrow. This dovetails precisely with the report's "Avoid" rating logic: the report doesn't deny that this is a high-quality compounder (moat graded "strong," stronger in ten years), it denies the current price of $481.57, a roughly 1.3% TTM free cash flow yield (against the 4.46% 10-year Treasury)—the market has bought in ahead of time the rosier script of "the moat will only widen over ten years and barely be shared."

  A timing note: this report's baseline date is 2026-06-03, and Broadcom's FY2026 Q2 earnings were only disclosed that evening (after the U.S. East market close), so they were not yet public at the time of writing and the above does not include actual Q2 values; what to watch in the Q2/Q3 delivery are the two moat-quality indicators the report lists—whether AI revenue delivers about $10.7 billion as guided, whether the adjusted EBITDA margin holds at 67%–68%, and whether the major customer shows further signs of moving outward.

  2026년 6월 3일
• QIs the current price reasonable? Is there a margin of safety?

  A

  The conclusion first: by the last of Buffett's "seven good-business questions," the current price is not reasonable, and there is almost no margin of safety. The report classifies the current price as "clearly overvalued" and rates it "Avoid," and after checking the data I agree with this closing judgment—good business, bad price. Worth noting: the core of this question is price, so I specifically checked the latest market data: as of the close on June 2, 2026, AVGO was at $481.57 (up 4.70% from the prior close of $459.97), exactly consistent with the report's baseline price, with a market cap of about $2.28 trillion, still among the top ten companies globally by market cap. In other words, the price hasn't drifted away, so the conclusions below can be used directly.

  Why there's no margin of safety—in one sentence: shareholders' actual return can't beat Treasuries.

  • The current price of $481.57 corresponds to a TTM free cash flow yield of only about 1.3% (FY2025 FCF $26.914 billion, the trailing four quarters of about $28.9 billion spread over the $2.28 trillion market cap), while over the same period the 10-year Treasury yield is about 4.46%.
  • Add the FY2026 target dividend of $2.60/share, a current-price dividend yield of about 0.54%, and what shareholders can "lock in" is only around 1.8%, markedly below the risk-free rate. In other words: if after buying, profit grows zero over the next three years and the valuation doesn't expand either, the return you get is worse than just buying Treasuries. The essence of a margin of safety is "you don't lose even if assumptions fall through," and this price level is exactly the opposite—it requires the assumptions to be delivered for you not to lose.

  Compared against the three owner-earnings scenarios, the current price is a premium, not a discount. The report did three tiers of absolute valuation using owner earnings (because capex is only about 1% of revenue and maintenance capex ≈ total capex, owner earnings nearly equals FCF):

  • Conservative $216 (30x OE, FY2027 revenue about $90 billion)—about -55% relative to current price;
  • Neutral $309 (35x OE, revenue about $105 billion)—about -36%;
  • Optimistic $423 (41x OE, revenue about $125 billion, AI sustaining ultra-high prosperity)—about -12%.

  The key is not some point estimate, but the structure: even the most optimistic scenario only gets to about $423, while the current price is already above $481. This means the market's current pricing is betting not "Broadcom will keep doing well," but "Broadcom will keep doing extremely well, and disappoint almost no one after Q2." The report accordingly gives three price signals: ideal buy ≤$173, holdable $263–355, clearly overvalued ≥$465, with the current price landing in the clearly overvalued zone, and the multiple-choice answer to "is there a margin of safety" being: no.

  This dovetails precisely with the first six questions of the Buffett framework. The picture from the preceding questions is: clear business model (dual engine: custom ASIC/networking semiconductors + VMware infrastructure software cash flow), large enough market (hyperscale customer AI capex as the main engine), strong moat (customer co-development switching cost + extremely light-capital architectural positioning + Hock Tan's cost and capital allocation discipline), visible growth (Q1 AI revenue $8.4 billion, Q2 guidance $10.7 billion), credible management (CEO incentive tied directly to AI revenue). These all hold—but precisely because the company is this good, the price has prepaid the good news too fully, and the margin for error has been compressed to extremely low. And question 7 is the closing question: no matter how good the business, buying it expensive is not a good investment. Buffett's "buy an excellent company at a reasonable price" only satisfies "excellent company" at $481, not "reasonable price."

  One more timing note for you to gauge the odds. The most critical validation point—FY2026 Q2 earnings—had not been disclosed as of the report baseline date (and my verification point), scheduled to be released after the U.S. East market close on June 3; the market broadly expects Q2 AI revenue of about $10.7 billion, about +140% year over year, and total revenue of about $22 billion. That is, the current price is "buying the good news in full before confirmation." The report's stance is to wait: if it returns to the $300–350 range while Q2/Q3 delivers AI revenue, margins don't collapse, and the VMware regulatory dispute doesn't escalate, re-rating the odds would be much better. To be clear, this is the report's judgment framing, which I faithfully restate while laying out the above facts; for specific buying or selling, please decide based on your own risk tolerance and the latest earnings results (once Q2 is released, both data and price may change).

  2026년 6월 3일

"가치 포착점"을 찾아 — 핵심은 "미래 최대 이익은 어느 단계에서 병목이 되는가?"

• QWhere does this company sit in the value chain?

  A

  Conclusion first: Broadcom is not a single point on the value chain. It spans an entire "tollgate corridor" cutting across custom compute-chip design + data-center network interconnect + connectivity/storage semiconductors + infrastructure software. In the language of Serenity's first question, the starting point for positioning it on the entire AI compute chain is this: it sits at the throat of the route where hyperscale customers "bypass Nvidia and build their own compute" — it helps you design the compute chip, it connects the network between chips, and once connected, the private-cloud software running on top is still its own. It barely touches fabs and carries no GPU depreciation (FY2025 capex was only about 1% of revenue), yet it can skim off the fattest design fees, chip fees, and software maintenance fees from other people's hundreds of billions of dollars of AI capex.

  Below, we break down each of its four chokepoint positions in the chain, to see clearly "which pipes the money flows through it via."

  One, custom ASIC / XPU design — "not building GPUs, building your own GPU." This is Broadcom's segment closest to the AI compute core, and the main engine of its valuation re-rating. It does not fight Nvidia head-on in general-purpose GPUs. Instead, it does co-design and tape-out of custom accelerators (XPUs) for those hyperscale customers who do not want to be tied to Nvidia forever (Google TPU, Meta MTIA, Anthropic, and reportedly OpenAI). Its role in the value chain is fabless design + IP (especially SerDes high-speed interfaces) + packaging/system co-engineering, translating a customer's chip requirements into mass-producible silicon, with manufacturing outsourced to TSMC. How hard is this chokepoint? Look at share: Broadcom holds about 70% of the design-services market for custom AI ASICs, Bloomberg Intelligence's earlier range was 60–80%, Marvell about 20–25%, and together the two prop up more than 80% of hyperscaler self-designed AI silicon. By the report's framing, Q1 FY2026 AI revenue was 8.4 billion, up +106% year-over-year, and with Google's long-term contract signed through 2031 and Anthropic at about 3.5GW starting in 2027, both show it is a "co-builder" in customers' roadmaps rather than a "short-order supplier."

  Two, Ethernet switching and interconnect — "the network that links chips into clusters is essentially its alone to call." This segment is even more monopolistic than the first, and it is where Broadcom's true "throat" lies. An AI cluster is not just a pile of chips; tens of thousands of XPUs/GPUs must be connected by a network into one "supercomputer," and the core switch chips of that network (the Tomahawk series, and the Jericho routing series used across data centers) are basically taken by Broadcom: it holds about more than half of revenue share in the Ethernet switch-chip market, and in merchant-silicon deployment terms its share is as high as a "dominant" level of about 86%–90%. The latest generation Tomahawk 6 is claimed to be the world's first 102.4 Tbps switch chip, Jericho4 is used to connect AI compute across data centers, and by external reporting the order backlog for AI switch chips alone is over 10 billion dollars. Worth noting: AI networking is becoming an ever-larger slice of Broadcom's AI revenue (Futurum says management positions AI networking as a steadily rising share of AI revenue) — meaning it earns not just the "designing chips" money but also, in passing, the "connecting these chips together" money, two toll booths stacked on the same set of customers.

  Three, connectivity and storage semiconductors — "the cash cushion of the old core business." Going further to the more foundational layer of the chain is the connectivity (optical components, PCIe/retimer, NIC-related) and server/storage semiconductors (including Fibre Channel SAN) that Broadcom was built on. This part is not the protagonist of the AI story, but it is a high-margin base that has been embedded in customers' supply chains for decades, and together with the two segments above forms the "Semiconductor Solutions" division — by the report's framing, FY2025 segment revenue was 36.858 billion and operating profit 21.232 billion. Its significance: it gives Broadcom a near-complete hand at the data center's physical layer (optics, electrical, interconnect); when a customer builds an AI system, from chip to NIC to switch to long-distance interconnect, it is very hard to route around it.

  Four, VMware infrastructure software — "the operating-system tax sitting above compute." In the final segment Broadcom stepped out of hardware and lodged itself at the data center's software control point. VMware (private cloud/virtualization) + mainframe software + enterprise security make up the "Infrastructure Software" division, with FY2025 revenue of 27.029 billion and operating profit of 20.765 billion by the report's framing, nearly even with semiconductors on profit — this is where it differs most from Marvell and Nvidia: beyond hardware, it also collects a fee on "the system you dare not shut down." Its position in the chain is to abstract CPU/GPU/storage/network into a private cloud and then charge by subscription; the moat is migration cost rather than customer preference (this is also the source of friction such as the European CISPE antitrust complaint and software revenue growth already slowing to about 1% in Q1). The value of this layer is cash-flow backstopping, making Broadcom's chain position less fragile than a pure cyclical chip stock.

  Connecting the four segments to see the essence of its "chokepoint position": Broadcom holds not Nvidia's kind of "platform dominance," but every tollgate on the alternative route of "helping big customers escape the Nvidia platform" — it helps you design the compute chip, it connects the network between chips, the physical connectivity components are its own, and the private-cloud software running on top is also its own. It is a two-in-one: "the infrastructure co-development platform for hyperscale customers in the AI era + the tollgate of legacy enterprise IT." This is precisely the true starting point for the later value-capture questions (where profit ultimately lodges, whether the market has already discovered it): it stands simultaneously at the most expensive design end and the most irreplaceable interconnect end of AI capex, and is itself extremely capital-light — a very large slice of the excess profit on this chain does indeed flow through its pipe. (The report's judgment on this "dual-engine platform" is that the fundamentals are extremely strong and the moat is strong, but it assigns an "Avoid" rating, with the reason lying not in the chokepoint position but in the price — the current price of about $481.57 corresponds to about a 2.28 trillion market cap, with a TTM free-cash-flow yield of only about 1.3%, below the about 4.46% on the 10-year Treasury; that is a matter of judgment, relayed here faithfully.)

  2026년 6월 3일
• QWhat does it actually sell? And what actually makes the money?

  A

  Conclusion first: what Broadcom sells is "the infrastructure control point you dare not easily switch off, yet must keep upgrading" — on the hardware side this is custom ASICs (accelerators built for Google TPU, Meta MTIA, Anthropic), Ethernet switching, and connectivity/storage chips; on the software side it is the private cloud, mainframe, security, and enterprise software centered on VMware. But "what it sells" and "what actually makes the money" are not the same thing: by volume, semiconductors are the breadwinner; by how much profit each dollar squeezes out, software is more vicious. The report places the two segments' operating profits side by side (semiconductors 21.232 billion / software 20.765 billion, nearly even), and that framing is not wrong, but it hides the real contrast — software earns nearly the same profit on less revenue.

  Break the two down to the gross-margin level, and the gap finally shows. Using the raw numbers from Broadcom's FY2025 10-K segment footnotes:

  • Semiconductors: revenue 36.858 billion, cost of goods sold 11.74 billion → gross profit about 25.1 billion (gross margin about 68%), segment operating profit 21.232 billion (operating margin about 58%).
  • Software: revenue 27.029 billion, cost of goods sold only 1.902 billion → gross profit likewise about 25.1 billion (gross margin about 93%), segment operating profit 20.765 billion (operating margin about 77%).

  In other words, the gross-profit amounts of the two are nearly identical (about 25.1 billion each), but software achieves it on nearly 10 billion less revenue than semiconductors. Each dollar of software revenue drops about $0.77 in operating profit, semiconductors only about $0.58; that 1.9 billion of cost of goods sold for software means VMware as a business has almost no marginal manufacturing cost — customers keep paying mainly because migration is too painful and the production environment cannot go down (switching cost), which is exactly what the report calls "collecting the fee on the system you dare not shut off." So on the question "what actually makes the money," the software segment is the margin champion, and semiconductors are the absolute profit amount and the growth engine — one provides thickness, the other provides slope.

  One trap on the report's framing that must be called out: those two segment operating profits (21.232 + 20.765 ≈ 42 billion) do not add up to the company's actual operating profit. The 10-K has a large chunk of "unallocated expenses" not counted in segments: amortization of acquisition-related intangibles 8.062 billion, stock-based compensation 7.568 billion, restructuring 0.667 billion, acquisition costs 0.216 billion, totaling about 16.5 billion — mostly the accounting baggage left by the VMware acquisition. After deducting it, FY2025 company-wide GAAP operating profit is actually only 25.484 billion (not 42 billion). These two large unallocated items are almost all tied to that software acquisition, so although the software segment's 93% gross margin is real, from an ordinary shareholder's standpoint the profit VMware actually leaves you gets eaten down a notch first by this amortization + SBC corridor — which is why the report repeatedly stresses that "high cash flow does not mean per-share returns are equally excellent."

  Finally landing on value capture: what the market is repricing Broadcom on today is not this 93%-gross-margin software base (that is a stable cash cushion), but whether that custom-AI-accelerator curve within semiconductors can keep climbing super-linearly — Q1 FY2026 AI revenue already reached 8.4 billion, up +106% year-over-year, and management raised the Q2 guidance to 10.7 billion. So closing with "what it sells vs. what actually makes the money": software is the most efficient profit outlet today (the margin winner), and semiconductors are the absolute bulk of profit and the sole source of future incremental growth (the growth winner); the report's "Avoid" is precisely because the price has already bought in advance a near-perfect realization of that AI curve ahead, not because the business itself does not make money.

  2026년 6월 3일
• QWhy do customers buy it? And who provides these capabilities of its?

  A

  Conclusion first: what customers buy from Broadcom is not an off-the-shelf standard part, but "that dedicated chip co-developed over many years + the high-speed interconnect and advanced-packaging integration capability that others cannot stack up in the short term" — once you are on Broadcom's roadmap, switching suppliers means overturning several years of system design. But conversely, a fair part of Broadcom's own underlying ability to "build these things" is not in its own hands: advanced process and advanced packaging are bet on TSMC, the EDA tool stack is bet on Synopsys/Cadence, and HBM is bet on three memory makers. Its truly proprietary moat is design IP (especially SerDes) and the engineering ability to integrate these external elements into one mass-producible large chip. Below we break it down.

  One, why customers buy it: stickiness comes from "co-development + switching cost," not a single component

  The report covers this layer very well and can be cited directly. Broadcom's custom ASICs are "not an order-today, ship-next-week standard-component business, but a multi-year engineering effort that co-orchestrates the customer roadmap, packaging, network, system design, and software stack together." This shows specifically in three classes of major customers:

  • Google TPU: Broadcom has been a design co-builder of Google TPU since the early generations, and in April 2026 locked the partnership through 2031 via an 8-K;
  • Meta MTIA: in April 2026 announced an expanded partnership, with the first phase committing more than 1GW, pointing to multiple chip generations and multi-gigawatt deployment;
  • Anthropic: a TPU compute plan of about 3.5GW starting in 2027 has been written into Broadcom's formal disclosure text.

  Once such a relationship is established, the customer's switching cost is extremely high — not "swap a chip," but redoing the entire set of validation, packaging, network topology, and software stack. The report also points out that Broadcom has the traits of "capital-light, thick cash flow, and high customer switching cost rarely seen in semiconductors." This is the first source of stickiness: not that Broadcom's components are so unique, but that the very act of "getting that customer's dedicated chip built" requires years of coordination.

  But one point the report does not develop should be added: why do these hyperscale customers, who have their own chip teams, still have to go to Broadcom? The answer lies in two pieces of hard IP that Broadcom owns and that others cannot make up in the short term —

  • SerDes (high-speed serial transceivers) is Broadcom's signature strength. It has cumulatively shipped over 350 million embedded SerDes channels, and pushed per-lane rates to 200G and evolving toward 224G. In an AI cluster, the bandwidth between chips and between racks is the lifeline, and the maturity of SerDes directly determines whether compute can be stacked up; this is the part that even Google and Meta would rather buy externally.
  • The integration know-how of advanced packaging. Broadcom launched the industry's first 3.5D F2F (face-to-face) XPU platform, XDSiP, where a single package can stack multiple compute dies + up to 12 HBM, with volume shipment starting February 2026. Integrating multiple dies, HBM, and I/O in one package to the point of mass production is a task with a high barrier of engineering experience.

  Put another way, what customers buy is the whole set of "dedicated design + SerDes + packaging integration," and the stickiness = the multi-year run-in cost of this set.

  Two, who provides its capabilities: above its own IP, manufacturing and tools are highly dependent on upstream

  This is the side the report basically does not unpack head-on, but it is most critical to "value capture." Broadcom's capex is only about 1% of revenue and it carries almost no heavy-asset depreciation — the report treats this as a "tollgate"-style moat. But the other side of this coin is: Broadcom is a fabless (no-fab) design company, and its physical ability to "make chips" is essentially rented. It depends specifically on three upstream lines:

  1. TSMC — advanced process + advanced packaging, neither can be missing. Broadcom's custom XPUs are made on TSMC's 3nm process, and its 3.5D XDSiP is explicitly built on TSMC's CoWoS-L advanced packaging. The key point: the real capacity bottleneck is not at Broadcom, but at TSMC's CoWoS packaging lines — CoWoS has been the primary bottleneck limiting AI-chip shipments since 2024, and capacity expansion still cannot keep up with demand. That is, whether Broadcom can truly turn 8.4 billion or 10.7 billion of AI revenue into shipments partly depends on how much quota it can grab on TSMC's advanced process and CoWoS, and that link is not its call.
  2. Synopsys / Cadence — the EDA tool stack. Every XPU must be designed with EDA tools, and Broadcom, like all ASIC designers, runs on the Synopsys/Cadence tool chains; by estimate, Broadcom's ASIC division's total annual spend on EDA tools + IP licensing + emulation hardware is about 200–500 million dollars. EDA is a highly oligopolistic chokepoint that Broadcom cannot route around.
  3. HBM memory — Samsung / SK Hynix / Micron. XDSiP packs up to 12 HBM3/HBM4 in one package, all from external memory giants, and HBM is likewise a hot commodity in the AI cycle.

  So Broadcom's capability stack can be cut this way: what it owns is "design IP (SerDes, etc.) + system/packaging integration engineering + customer relationships," and what it buys in is "process capacity + packaging capacity + EDA tools + HBM." What is truly irreplaceable about it is the position standing between the customer and TSMC, designing and integrating one workable custom large chip; and its own biggest external dependency and most realistic physical ceiling is TSMC's advanced process and CoWoS packaging quota.

  Putting these two sides together, the implication for "value capture" is clear: Broadcom's stickiness toward downstream customers is real (co-development + switching cost), but it does not have equal pricing dominance toward upstream (TSMC, EDA) — it is one link on this chain whose profit is very fat yet not the only chokepoint. Layer on the high concentration disclosed in the report of a single customer accounting for 32% of FY2025 revenue and 44% of accounts receivable, and the answer to this "source of capability" question actually implies Broadcom's two-ended risk: downstream relies on the expansion pace of one or two large customers, upstream relies on TSMC's capacity ramp — this is precisely one of the underlying reasons the report assigns an "Avoid" rating and stresses that the current stock price has extremely low margin for error.

  2026년 6월 3일
• QOver the next 3–5 years, where will demand growth come from?

  A

  Conclusion first: over the next 3–5 years, more than ninety percent of Broadcom's incremental growth comes from one thing — a handful of hyperscale customers pushing the deployment of custom AI accelerators (XPU/ASIC) + supporting Ethernet interconnect from "pilot production" toward "multi-gigawatt ramp." The software (VMware) line has already receded from "growth engine" to "cash-flow base," with growth slowing to about 1%, basically no longer contributing incremental growth. Break the drivers apart from the parts that have already slowed:

  Main engine one: hyperscale-customer custom-XPU expansion, already locked in by long-term contracts. By the report's framing, Q1 FY2026 AI revenue already reached 8.4 billion dollars, up +106% year-over-year, and management raised the Q2 guidance to 10.7 billion dollars. One granular point the report does not state should be added: that 10.7 billion for Q2 corresponds to about +140% year-over-year, while the entire Q2 total-revenue guidance is about 22 billion dollars, up +47% year-over-year — meaning AI is not just growing but accelerating, and the acceleration already far exceeds the company's overall base. Supporting this curve are three sets of formally disclosed long-term contracts: Google's TPU and AI-rack agreement signed through 2031; Anthropic taking 1GW in 2026 + about 3.5GW starting in 2027 of TPU compute via Broadcom (note this 3.5GW is an elastic commitment whose "usage depends on its commercial performance," not an unconditional delivery); Meta's MTIA self-designed chip first-phase commitment exceeding 1GW, and it is the first phase of a multi-gigawatt deployment. The report names three; one customer-structure fact can be added: the market generally counts Broadcom's AI customers at six — besides the above, it also includes OpenAI, plus media-speculated Fujitsu and ByteDance. Of these, the OpenAI line warrants caution: external reporting says Broadcom will ship its first self-designed chip for it in 2027, providing 1GW+, but Broadcom has not made a same-level confirmation with an 8-K as it did for Google/Anthropic, so it looks more like "potential incremental" than "locked-in incremental."

  Main engine two: penetration of AI network interconnect, two legs of the same order as XPU. What Broadcom sells has never been just the accelerator itself, but also the Ethernet switching, interconnect, and NICs that connect these accelerators within the data center. The 10-K itself attributes FY2025 semiconductor growth to two lines pulling together: custom AI accelerators and AI networking products. This is key to Question 4: every additional gigawatt of compute a customer deploys nearly synchronously amplifies the interconnect budget, so the network leg lets Broadcom "take two bites" in the same wave of capex. And how big is this wave of capex? The four hyperscalers' 2026 AI infrastructure spending combined is estimated by Reuters at at least 63 billion dollars, and Marvell by May revised the cloud-provider figure up to over 70 billion dollars — Broadcom's demand elasticity comes not from end users buying more phones, but from these few players changing how they spend at the system level.

  Far-out view: 2027 AI-chip revenue >100 billion dollars — that is "visibility," not "confirmed revenue." In March 2026 management gave a framing of having a "line of sight" to 2027 AI-chip revenue "significantly exceeding" 100 billion dollars. I must, per the report and the facts, label its nature faithfully: this is forward-looking management visibility, not orders fully converted into realized revenue. The board even tied Hock Tan's PSU retention incentive directly to AI revenue — over FY2028-FY2030, any four consecutive quarters with an AI-revenue threshold of 60 billion and a top tier of 120 billion dollars. Incentive alignment is a plus, but it also means this growth narrative is highly bet on a single variable, and the realization slope itself is a risk point (this is exactly one of the core concerns behind the report downgrading the rating to "Avoid"; relayed here only, no separate conclusion drawn).

  The part that has already slowed: VMware / infrastructure software, receding from a growth story back to a cash-flow base. The report says Q1 FY2026 infrastructure-software growth has slowed to about 1%, and this is accurate; two external granular points can be added to complete the picture: the precise figure is Q1 software revenue about 6.796 billion dollars, up +1.4% year-over-year, which is a cliff-like drop versus Q4 FY2025's +19.2% and Q1 FY2025's +46.7%. But this 1% is easily misread — in the same disclosure, VMware's own revenue was up +13% year-over-year, annual recurring revenue (ARR) up +19%, and total contract value of bookings 9.2 billion dollars, and the company also guided Q2 software back to about 7.2 billion dollars, up +9% year-over-year. In other words, 1% is the consolidated figure for the entire software segment (including mainframe, security, and other legacy businesses) dragged by a high base and restructuring, and the renewal stickiness of the VMware core has not collapsed. Its role for the next 3–5 years is therefore clear: no longer a source of incremental growth, but a stable base backstopping the high-elasticity AI cash flow; the real risk is the European CISPE antitrust complaint and partner-ecosystem friction, not the software suddenly stalling.

  A timing note: this report's base date is 2026-06-03, and Broadcom's FY2026 Q2 earnings call is scheduled for that very day at 17:00 ET after market close, so the 10.7 billion / 22 billion above are all guidance values, not already-published actuals; as of my verification, the Q2 actual figures had not yet entered public search and should be taken from the earnings release itself. For Question 4 this does not affect the conclusion — the demand-side base (who is buying, what they are buying, how much they are buying) has already been written very clearly by the long-term contracts and capex plans.

  2026년 6월 3일
• QIf industry demand grew 5x, which link would run short first?

  A

  Conclusion first. The report itself does not develop this question (it focuses on valuation discipline, not the supply-side bottleneck), so the vast majority below relies on external verification. After checking link by link along the chain, the answer is: if demand rose another 5x, what runs short first is not a single link, but two chokepoints one upstream and one downstream — upstream is CoWoS advanced packaging (including the HBM it ties up), downstream is power and electrical equipment (transformers/grid interconnection). The position that truly jams depends on which end you stand at: to make the chip, CoWoS runs short first; to power the chip up and run it, power runs short first. If I must single out the hardest one, I pick high-voltage transformers and grid interconnection on the power side — because its expansion cycle is the longest, the most irreplaceable, and the hardest to unlock with money and process within two or three years. Custom-XPU design capability (that is, what Broadcom itself does) is instead the least tight link on this chain.

  Link-by-link check (from highest to lowest tightness):

  • CoWoS / advanced packaging — the most urgent chokepoint on the silicon side. This is the currently acknowledged first link to run short; it sits upstream of everything: every custom ASIC, to integrate HBM onto it, must pass through CoWoS (or CoWoS-L), and TSMC's packaging capacity is now more of a constraint than wafer fabrication itself. TSMC is expanding CoWoS from about 35,000 wafers/month at the end of 2024 to about 130,000 wafers/month by the end of 2026 (nearly quadrupling), but utilization still hugs 100%, Nvidia has reserved the lion's share, and others can only go to OSATs such as ASE/Amkor to queue for overflow capacity. The most intuitive evidence: the CoWoS crunch directly forced analysts to cut Google's 2026 AI-chip output forecast from 4 million to 3 million units — and that is precisely the customer on Broadcom's main battlefield. If demand rose another 5x, this is the first to blow.

  • HBM — already sold out, but it "grows" inside the package. SK Hynix's and Micron's 2026 HBM capacity is basically sold out, with some customers' reservations already queued into 2027, and 2026 HBM/DRAM prices are expected to rise high double digits to 20%. It is tight enough that OpenAI publicly listed memory as the primary bottleneck risk for training/inference. But note: the HBM shortage and the CoWoS shortage are two segments on the same stacking chain (memory → interposer → packaging throughput → substrate → yield), so it is more of a compound chokepoint bound together with CoWoS than an independent track.

  • Optical interconnect / Ethernet — the second tier, and being locked up early by Nvidia. The switch ASIC itself is not the scarcest (Broadcom supplies it), and what is truly tight is the lasers upstream of optical modules: McKinsey estimates 800G optical modules will still fall short of demand by 40-60% by 2027, with the 1.6T gap of 30-40% persisting all the way to 2029; EML lasers have a gap of about 36% in 2026 and about 48% in 2027 (TrendForce). The most persuasive is the hard-cash vote: on March 2, 2026 Nvidia poured 2 billion dollars each into Lumentum and Coherent to lock priority EML laser capacity — squeezing rivals past 2027. This is a real bottleneck, but its magnitude and irreplaceability rank below packaging and power.

  • Power / data center — the longest cycle, the hardest to solve, the "hardest" one in my mind. Do not be misled by "the chip is the bottleneck": since 2025, the chokepoint has migrated from the rack to the substation, and power availability has formally become the primary constraint on new builds. The hard numbers are striking — GSU step-up transformer lead times of 144 weeks, power transformers 128 weeks, and high-end models already stretched to 5 years; of the 12-16GW of data-center capacity planned for 2026, only about one-third is actually under construction, the rest waiting on electrical equipment; Gartner expects power shortages to jam 40% of AI data centers by 2027; in the U.S. about 2,300GW is backlogged in the grid-interconnection queue, exceeding the nation's entire existing installed capacity. Why I rank it the hardest: packaging/HBM at TSMC and Hynix can multiply capacity in 18-24 months, while transmission infrastructure takes 5-10 years from planning to energization, and money cannot buy it faster — which is exactly what has forced the "BYOP" play of bringing your own gas turbines / small nuclear.

  • Custom-XPU design capability — the least tight link on this chain. Broadcom + Marvell together hold about 95% of custom-ASIC co-design share, Broadcom itself holds about 73 billion dollars of AI order backlog, and its capex is extremely light — showing that the design end can take on the orders, and is not yet queued past capacity. Design is a high-barrier scarce capability (which is exactly what the next question chases — "is Broadcom this chokepoint"), but it is not the first physical bottleneck to "run out of stock" under 5x demand; what truly runs out is the packaging capacity downstream of it and the power upstream of it.

  To close in one line: if demand rose 5x, the first to jam is CoWoS advanced packaging (and HBM along with it) — the gate closest to Broadcom that decides how much it can ship; and across the entire compute-deployment chain, power and high-voltage electrical equipment are the ultimate chokepoint with the longest cycle, the hardest to pry open with money and process. One silicon and one electrical, they jam at the two ends of "can be built" and "can be run."

  2026년 6월 3일
• QIs this company that link which runs short first?

  A

  Conclusion first: Broadcom is the link on this AI compute chain that is "near the bottleneck, but not the hardest bottleneck." If custom-compute demand really rose another 5x, the first to jam and the hardest to route around is not Broadcom, but two physical capacity nodes behind it — TSMC's CoWoS advanced packaging and the Korean duopoly's HBM. What Broadcom holds is the "design/IP + network interconnect" layer, and its chokepoint position is indeed hard, but the hardness is in "customers find it troublesome to switch," not in "physically there is no second path." These two kinds of hardness differ greatly in odds.

  First describe what the true hard bottleneck looks like, as a yardstick. On the CoWoS side, TSMC's capacity is basically booked out and sold out through the end of 2026, with lead times stretched to 50+ weeks, and monthly capacity has to race from about 35,000 wafers at the end of 2024 to about 130,000 wafers by the end of 2026 (nearly 4x expansion still not filling demand), while Nvidia has already reserved the lion's share and others are scrambling for the scraps; organic substrates simply cannot achieve the interconnect density required to connect logic die and HBM, and this is a physical manufacturing limit, not a design choice, so it cannot be routed around. The HBM side is even more extreme: the two Korean players (SK Hynix + Samsung) hold about 79% of global capacity, SK Hynix alone takes 57% share, and Micron's 2026 HBM4 capacity is entirely sold out, able to meet only 50%–65% of even core customers' requested volumes. These two are the textbook-grade chokepoints: single geography, oligopoly, and expanding capacity means pouring money into fabs over several years; if demand quadruples — quintuples, they blow up on the spot first. Broadcom itself knows this too — it publicly warned in early 2026 that TSMC's capacity is tight and will transmit into a squeeze on its chip supply. Note: Broadcom's own XPUs must also queue for CoWoS, so it is "jammed upstream of the bottleneck, getting jammed along with it," not the bottleneck itself.

  Now to Broadcom's chokepoint position in this link — hard, but hard in a different way. The report characterizes its moat as "customer co-development + switching cost" and "capital-light but architecturally irreplaceable in position," and this judgment is correct, with the external data even brighter: Broadcom holds about 70%+ share of custom-AI-accelerator design services, plus Marvell's about 20%–25%, the two combined about 95%, and holds about 73 billion dollars of AI backlog (corresponding to the report's Q1 8.4 billion, Q2 guidance 10.7 billion, 2027 sight-line >100 billion); its SerDes/interconnect IP (Tomahawk 6 already does 102.4Tb/s, paired with 200G PAM4 and co-packaged optics) is an IP library of high-speed interfaces, HBM integration, TCAM/SRAM, and ARM cores built over a dozen-plus years, with no short-term substitute. The report's Question 7-style colloquial test of "what if it shut down tomorrow," applied here, gives the answer: the next-generation TPU/MTIA roadmaps of major customers would be genuinely slowed — and that is one of the features of a true bottleneck.

  But "the degree to which it can be routed around" is precisely the watershed between Broadcom and CoWoS/HBM, and the core of this question. The difference: CoWoS/HBM jam capacity, while Broadcom jams "whether the customer is willing to do the design itself / find someone else to do it" — and the latter is being actively diluted by customers. The hardest evidence is Broadcom's largest and oldest customer Google: it has already gone dual-supplier on the seventh-generation TPU, keeping the training chip with Broadcom and handing the inference chip (v7e/Zebrafish) to MediaTek, with the openly stated purpose of using MediaTek's low price to press down Broadcom's premium, and by estimate the per-chip cost can drop by up to about 30%. MediaTek, Marvell, and even Alchip (cutting into the AWS chain) are all squeezing in, and the SerDes layer has already become a three-way melee among Broadcom/Marvell/MediaTek. In other words: Broadcom is the link that is "the most expensive, the strongest, the hardest to fully replace, but by no means the only one"; customers cannot swap out all of it, but they can, at the margin (a certain generation, a certain inference line, the I/O part), replace it with a cheaper second source. CoWoS has no such "30%-cheaper second source" option, and neither does HBM — this is the true gap in chokepoint hardness.

  Landing on the investment judgment (which also echoes the report's "Avoid"): pricing Broadcom as "an unroutable hard bottleneck on the AI chain" is one of the implicit premises of the current optimistic valuation of $481, about a 2.28 trillion market cap, and a TTM FCF yield of only about 1.3%; whereas the fact is it is more like "a very hard but marginally erodible tollgate," not "a physically irreplaceable gate." If demand really rose 5x, Broadcom would almost certainly keep growing fast (it is in the immediate upstream of the first link to run short, and is the de facto first choice at the design layer), but its profit elasticity would be squeezed on both ends: upstream, CoWoS/HBM are the true rent-extracting bottleneck, while downstream, customers keep pressing down its design premium with second sources and in-house design. The report's "Avoid" and ideal buy below 173 dollars are essentially saying — the price paid for the "hardest bottleneck" buys the link that is "second-hardest, and being partly routed around," and the odds are not worthwhile. This is a judgment; the objective facts are laid out here, and each of you can weigh them yourself.

  2026년 6월 3일
• QIf this company shut down tomorrow, what would happen to the industry chain?

  A

  Conclusion first: If Broadcom suddenly disappeared tomorrow, AI infrastructure would go through an "industry-wide emergency brake"—the sheer scale of the short-term shock is itself reverse proof that it is today the single most load-bearing node on the AI compute chain. But stretch the timeline to the medium term, and whether it's custom XPU/ASIC, high-end Ethernet, or the VMware private-cloud stack, real alternatives do exist; switching just comes at the cost of "months to a year or two of stoppage." This is exactly its true position of "strongly irreplaceable, but not an absolute monopoly," and it's also the underlying logic behind why this report gives "Avoid" rather than "bearish on the business": the business is extremely hard to replace, but that doesn't make the current price reasonable.

  Block one: custom XPU/ASIC and high-end networking—short-term instant shock, medium-term replaceable but very expensive and very slow

  The short-term shock is certain and severe. The report names Broadcom's positioning on the customer side: the Google TPU long-term contract runs to 2031, Anthropic at about 3.5GW of TPU compute from 2027, Meta's first-phase MTIA deployment of over 1GW, Q1 FY2026 AI revenue already at $8.4 billion, and Q2 guidance of $10.7 billion. What the report doesn't unpack is its "share scale" in the industry chain—by external accounts, Broadcom holds about 70% of the custom AI ASIC share and is called "dominant" in the data-center switch-chip market (share is often cited around 90%), and AI switching alone has a backlog of over $10 billion. What it supplies isn't a "swap-it-next-week" standard part, but multi-year engineering deeply co-developed with customer roadmaps, packaging, networking, and software stacks. So "shut down tomorrow" means: the generation of TPU/MTIA that Google/Meta/Anthropic are ramping or mass-producing is cut off on the spot, the next-generation designs in development are left hanging, and the Tomahawk 6 (102.4T) / Jericho 4 fabric clusters connecting tens of thousands of accelerators instantly lose their primary supply—which is equivalent to directly disrupting the cadence of the combined $630 billion to $700 billion-level AI capex plans of several hyperscalers that year. This "pinch it and the whole chain stops" property is itself the proof of the report's "capital-light but architecturally irreplaceable position" moat.

  But in the medium term it is not irreplaceable, and the alternatives are already on the field. The key fact is: even Broadcom's most loyal customers are actively building a "second supplier." Google is building a multi-supplier architecture of Broadcom + MediaTek + Marvell—Broadcom does the training chips (next generation codenamed Sunfish), MediaTek takes the low-cost inference variant of TPU v8, and Marvell is in talks on inference TPUs and memory processing units. Counterpoint predicts that by 2027 Broadcom will have about 60% and Marvell about 25% of the custom accelerator share, showing the "single source" landscape is being diluted. On the high-end networking side, replacements are even more ready-made: Nvidia Spectrum-X, Cisco's newly launched Silicon One G300 (102.4T), and Marvell Teralynx can all benchmark against Tomahawk. The cost is time and money: switching ASIC suppliers means re-taping-out, and silicon generally takes 18–24 months from design to mature mass production. So the medium-term picture is: the chain won't break, but you have to endure one to two years of capacity vacuum and rework cost—enough to slow the whole industry, but not to permanently paralyze AI infrastructure.

  Block two: the VMware private-cloud stack—enterprise IT would "freeze instantly," but the underlying layer is migratable, and migration is already happening

  The short-term risk is a "frozen runtime state," not "data loss." The report puts it clearly: VMware's stickiness comes from migration being too painful and production environments not being allowed to stop, not from customers loving it—this is switching cost, not brand loyalty. External data confirms it is still the absolute foundation: in 2026 VMware still holds the largest share of on-prem server virtualization (around the 70% level). If it disappeared tomorrow, the massive enterprise workloads running on ESXi/vSphere would lose updates, patches, and official support, which would be especially hard on regulated industries (banking, healthcare, government)—but because virtual machines are inherently portable, this is more like a "forced acceleration of moving house" than the data center going dark.

  And the report already flags that this "move" is underway, with external data giving the magnitude: Broadcom cut perpetual licenses and raised subscription prices by about 300%, triggering the largest migration wave in a decade. Nutanix disclosed that 40% of FY2025 orders came from VMware replacement, Proxmox enterprise subscriptions rose 60% in 2024, and Hyper-V/Azure Stack and Red Hat OpenShift Virtualization are also mature options; some long-term forecasts hold that VMware's share will fall from about 70% in 2024 to 40% in 2029. So medium-term replaceability is certain, and the bottleneck is likewise time—a large enterprise's one-time virtualization-platform migration easily takes years and requires re-testing the entire production environment. This also ties back to the report's judgment: VMware revenue growth has already slowed to about 1%, and the European CISPE antitrust complaint is heating up—its "lock-in" on enterprise IT is real, but that lock is slowly being pried open, and the risk that its long-term value-capture power is overestimated is more worth watching than the "stoppage risk."

  One-line close: Broadcom is one of the few key nodes in AI infrastructure where "the whole chain emergency-brakes the moment it disappears," and on this point the report and external data align closely; but both businesses have alternatives actively landing, and the supply-cut damage is a "capacity and migration vacuum measured in quarters to a year or two," not permanent paralysis. This "extremely hard to replace, yet not an absolute monopoly" is exactly what supports the report's core stance—the business's scarcity is real, but the current price (about $481.57, corresponding to about $2.28 trillion market cap, with a TTM free-cash-flow yield of only about 1.3%) prices that scarcity to a level that allows almost no room for error, so the question has never been "will the business be replaced," but "how much margin of safety is left at this price." This is the report author's judgment, offered for your own assessment.

  2026년 6월 3일
• QCan customers replace it? How long would it take? And how many years would it take a new competitor to enter?

  A

  Conclusion first: these are two completely different time scales. For a customer to "replace" Broadcom—even just shifting part of a design to a second supplier—the realistic window is 2–4 years (a chip generation from kickoff to volume production alone takes 12–24 months, and what gets switched is mostly "adding a second source" rather than "kicking out Broadcom"); whereas for a new player to go from zero to Broadcom's level of "custom XPU + network co-design," what must be accumulated is a decade-scale body of IP, packaging, and customer-engineering relationships—this is precisely what the report means by "high barriers, heavy customer co-development."

  Can customers replace it: they can "divert," but it's very hard to "replace," with a 2–4 year window

  The report records a key fact in its falsification conditions: in April 2026, Broadcom signed long-term TPU and rack contracts with Google running as far as 2031. The mention in the question that Google is using Marvell as a second source is also true—but you have to be clear which kind of "switch" it is. According to The Information, Google is in talks with Marvell on two new chips (a memory processing unit MPU and an inference-oriented TPU), bringing Marvell in as a "design services" partner. But that report is very blunt: the negotiation is not yet signed, the timing falls just days after Broadcom locked in the 2031 long-term contract, and it's "adding a third design partner, not replacing Broadcom"—in the supply chain Broadcom still makes the high-performance flagship variant, while MediaTek makes the "e" variant that is 20–30% cheaper; the strategy is diversification, not replacement. So even if it closes, Broadcom is still designing Ironwood and the TPU 8t training chip, and holds >70% of the custom AI accelerator share.

  Why can't it be switched quickly? Because every XPU is deeply co-developed with the customer's roadmap, packaging, SerDes, networking, and software stack—it's not an off-the-shelf standard part. A custom ASIC project takes about 12–24 months to go through design → tape-out → product validation → mass-production qualification before it enters production, with a single NRE running from tens of millions to over a hundred million dollars. So the realistic source-switching cadence is: Google and Marvell are aiming to finalize the memory chip design next year (2027) first, then enter trial production—meaning from "start talking" to "the switched-out silicon actually ramps in volume" takes at least 2–3 years, and most likely lands first on the new inference/cheaper variants, with Broadcom's flagship training generation untouched in the short term. This is also the basis for Broadcom's confidence in amassing about $73 billion in custom-customer backlog and giving multi-year revenue visibility: the hyperscalers are doing "multi-supplier diversified procurement," not "defecting overnight." The report calling this moat "customer co-development and switching cost" is accurate—the stickiness is real, but one must honestly add: share can be slowly diluted, and the second source isn't noise but real long-term marginal pressure.

  How many years for a new competitor: decade-scale IP + packaging + trust accumulation

  The time scale on the entry side is much longer. Broadcom making TPUs for Google is itself a decade-long relationship running from the first-generation TPU in 2014 to today; what it supplies isn't just "drawing the layout," but a whole set of silicon implementation capabilities—high-speed SerDes interfaces, ASIC design, power management, packaging—covering every generation. Broadcom itself says it has poured in over $3 billion to support these customer collaborations. The barrier is specifically lodged in three places: ① SerDes and network IP—this is the family fortune Broadcom has accumulated from its veteran networking/switching business, and it's exactly where the report's "network co-design" sits; ② advanced process + CoWoS packaging capacity—Broadcom and its rivals all have to compete with Nvidia for TSMC's advanced nodes and packaging capacity, and the packaging link is already tighter than the wafer itself; ③ customer engineering trust—hyperscalers will only entrust hundreds of millions in NRE and years of roadmap to the few who can "turn architecture into yield-able mass production."

  The result is that this track is extremely concentrated: Broadcom about 60%, Marvell about 35%, the two combined about 95%. Worth noting: even Marvell, the acknowledged number two, isn't a "new entrant"—it's already a listed, comparable veteran, yet it's still early in its AI ramp and still carrying heavy R&D for new tape-outs. A player truly starting from zero, to assemble SerDes/network IP, secure packaging capacity, and then win the trust of hyperscalers to entrust core generations, would need 5 years just to land edge orders, and 7–10 years to enter core co-development—and during that time Broadcom's customers and generations keep moving forward.

  Bringing it back to this rating

  Stacking the two scales together: the entry barrier (decade-scale) won't be breached in the short term, and the moat is real; but the source-switching/diversion window (2–4 years) is opening—Google's move to Marvell and the hyperscalers' in-house efforts are real footnotes to the script in the report's pre-mortem where "weighty customers move part of the design out to a second supplier, and Broadcom shifts from core co-developer to an important but replaceable executor." This is exactly the inner logic behind the report maintaining "Avoid": the barrier protects the business from collapsing, but it can't protect the valuation from being compressed—when the current price (about $481.57, about $2.28 trillion market cap, TTM FCF yield of only about 1.3%) has already priced in "Broadcom will forever be the irreplaceable exclusive core party," then even if the second source merely nibbles marginal share or shaves pricing, that's enough to shrink this "near-perfect" pricing. The report's ideal buy price is below $173; here this is offered purely as objective relay, with no contrary conclusion of its own.

  2026년 6월 3일
• QCan supply expand? What conditions does it require?

  A

  Conclusion first: Whether the supply of Broadcom's business "can expand" has to be split into two completely different things. Broadcom itself (the design + customer co-development layer) actually expands quickly and lightly—fabless, with capex only about 1% of revenue (the report's figure); what expands is the engineering team and customer-R&D bandwidth, not factory buildings. But what really gates shipments isn't in Broadcom's hands—it's in the TSMC advanced process, CoWoS packaging, and HBM above it; these three are the hard constraints right now, won't loosen in 2026, and won't see substantial relief until 2027. So the correct depiction is: Broadcom is the "porter of the bottleneck," not the bottleneck itself—it can fill orders, but it can't conjure this upstream capacity by itself.

  Layer one: how fast Broadcom itself can expand—very fast, but with a hidden ceiling Broadcom doesn't build fabs; for it, expansion is mainly two things: ① grabbing more upstream capacity quotas; ② lining up enough design engineers and project bandwidth to deeply co-develop with customers. It's already doing the former—according to TrendForce, Hock Tan has already "locked up capacity through 2026–2028" across advanced wafers, HBM, substrates, lasers, PCBs, and other key links—this is the most important talisman for a fabless player during a capacity shortage: it doesn't expand capacity itself, but pre-books others' capacity. The latter is Broadcom's true "soft bottleneck": custom XPUs aren't standard off-the-shelf chips; each one requires years of coordinating the customer's roadmap, packaging, networking, system, and software stack (the report also stresses this is its most core moat and source of switching cost). The problem is that such coordination heavily consumes scarce senior SoC design engineers, and Broadcom is now running custom projects for about 6 hyperscale customers simultaneously (Tom's Hardware lists Google, OpenAI, Meta, ByteDance, Fujitsu plus an unnamed "$10-billion mystery customer"). Running engineering coordination across so many generations and so many customers in parallel is itself a ceiling on expansion speed—it's not capacity you can buy within a quarter by spending money; it relies on an engineering organization built up over many years. In other words, Broadcom's expansion condition at this layer is: grab the quota + keep engineering bandwidth from being diluted by the number of projects, both of which it currently still holds; Broadcom and Marvell together take about 95% of the custom ASIC co-design market, showing this capability is extremely scarce.

  Layer two: upstream hard constraints—this is where "demand up 5x first runs short," and it won't loosen in 2026 Broadcom's XPUs are all riding TSMC's most advanced process (3nm, with the first 2nm compute SoC already shipping in February 2026, using the 3.5D XDSiP platform stacking 3nm/2nm multi-die + 6 HBM), and the three gates on this chain are all full right now:

  • Advanced process (N3/N2): TSMC's advanced-process wafer demand is about 3x the available supply, all three 2nm fabs are fully booked, delivery lead times of 78–104 weeks already stretch to 2028, and customers have to queue more than a year ahead to get capacity.
  • CoWoS advanced packaging: this is the acknowledged chokepoint for AI chips. TSMC is expanding monthly capacity from about 35,000 wafers at end-2024 and about 75,000–80,000 at end-2025, to a target of about 130,000 by end-2026 (nearly quadrupling)—but even so, CoWoS-L/S is currently reported fully booked, and Nvidia alone has locked up a substantial chunk of advanced-packaging quota, squeezing the rest of the players.
  • HBM: SK Hynix, Samsung, and Micron's 2026 HBM4/HBM3E capacity is almost 100% locked up by hyperscaler long-term contracts, with new contracts having to wait until 2027.

  These three plus substrates, lasers/optical modules, and PCBs are exactly the list Broadcom itself named—Broadcom executive Natarajan Ramachandran directly called TSMC capacity in 2026 "effectively a bottleneck, pinching and to some degree choking the entire supply chain," and made clear that capacity expansion won't keep up until 2027.

  What this means for Broadcom: expansion is possible, but the cadence isn't set by it—it can sign the Google long-term contract through 2031 and articulate the ">$100 billion AI revenue line of sight" for 2027 (the report's figure), but cashing these into shipments has to queue behind the expansion timetable of TSMC advanced process + CoWoS + HBM, and that timetable reads "tight in 2026, loosening only in 2027." So the real conditions for easing supply aren't on Broadcom's balance sheet, but on three things at other companies: TSMC advanced packaging ramping on schedule to ~130,000 wafers/month, N2 capacity continuing to scale, and HBM makers bringing 2027 new capacity online. The implication for investors is two-way: in the short term, tight capacity actually protects Broadcom's pricing and order visibility (scarcity = bargaining power), but it also shows Broadcom's growth ceiling is gated upstream—which precisely echoes the point the report repeatedly stresses: Broadcom's real risk isn't in the factory or the lab, but in "whether it can smoothly cash already-locked-in forward demand into per-share owner earnings." The slower the upstream ramp and the fiercer the quota competition, the more likely this cash-conversion curve gets ground from "super-linear" down to "linear," and that's exactly the share-price-halving script in the report's pre-mortem.

  2026년 6월 3일
• QWhere will profit ultimately flow in the industry chain?

  A

  Conclusion first: On the AI-compute chain, profit won't go to one link to swallow alone—it gets divided among three "chokepoint + toll" nodes: HBM memory (SK Hynix), upstream process/advanced packaging (TSMC), and the accelerator platform (Nvidia GPUs and Broadcom custom XPUs); the cloud providers who ultimately monetize this compute are instead the link that spends the most and has the worst return on capital per unit. Broadcom's role isn't "the one taking the fattest profit pool," but a rare toll booth on this capital-heavy chain that is extremely capital-light and extremely cash-rich—what it captures isn't the largest absolute chunk of profit, but the cleanest chunk with the highest return per unit of input.

  Spreading the profit pools along the chain (current figures for each link)

  • HBM (SK Hynix) = currently the strongest in unit profitability and the most underestimated link. The report body barely unpacks this, so external data must fill it in: SK Hynix's Q1 2026 operating margin surged to about 72% and gross margin to about 79%, leaving Nvidia (about 68%) and TSMC (about 58%) all behind, with HBM share around 59%, and customer orders already exceeding the next three years of capacity planning. In the bill of materials of a single GPU, HBM is often the most expensive single item, so the link that "first runs short when demand goes up 5x" is actually memory, and its pricing power is currently harder than the GPU's. The cost is that it is capital-heavy and strongly cyclical—it expands capacity via ten-billion-class new fabs like M15X, and it falls the hardest in a downturn.

  • TSMC = irreplaceable but the most capital-heavy chokepoint. The advanced process and CoWoS packaging for the whole industry (Nvidia GPUs and Broadcom XPUs alike) can almost only be taped out at TSMC, and CoWoS capacity is fully booked into 2026 and expanding at double the pace. But it trades about $40–42 billion of capex per year for about 60% gross margin and about 51% operating margin—high profitability, but it has to keep pouring cash back into factories, with a return on capital per unit far below a fabless company like Broadcom.

  • Nvidia = the platform with the largest absolute profit pool. FY2026 revenue about $216 billion, data center about $194 billion, gross margin about 75%, collecting a toll on "general-purpose compute" via its CUDA hardware-software integrated ecosystem; in absolute terms it still eats the most GPU dollars.

  • Broadcom's custom XPU + networking = a toll booth with no worse margins but the lightest capital input per unit. This is where the report and external data together support Broadcom most: in Q1 FY2026 Broadcom's overall gross margin was about 77%, semiconductor gross margin about 68%, operating margin about 66.4%, adjusted EBITDA margin 68%, and FCF margin about 41%, consistent with the report's figure (FCF margin about 42%); and its capex is only about 1% of revenue (just $623 million in FY2025), throwing all the capital-heavy process work to TSMC. In other words, Broadcom's gross margin level isn't an order of magnitude below Nvidia's or SK Hynix's, but it doesn't have to carry factory depreciation—on the "input-output ratio/ROIC" dimension, the fabless designer is instead the best-positioned on the whole chain.

  How much can Broadcom capture, and is it sustainable?

  By scale, what Broadcom captures is a high-quality but not-the-largest-in-absolute-terms slice of profit on the AI-compute chain: in FY2026 its AI semiconductor revenue is expected to be in the $3-billion-per-quarter level and up (report: Q1 already $8.4 billion/quarter, Q2 guidance $10.7 billion, management citing a 2027 line of sight above $100 billion); against Nvidia's nearly $200 billion data-center pool, it is the main contractor for "the in-house route outside Nvidia," with share depending on how fast the Google TPU, Meta MTIA, and Anthropic lines expand, rather than taking all.

  On sustainability, two points must be stated honestly, and they decide whether the profit Broadcom captures is a "toll booth" or a "fabrication fee": First, custom ASICs are co-fabrication done to customer demand, and the report explicitly warns that a rising share of custom chips in the future will pull margins down; after the December 2025 earnings, the stock fell over 11% precisely on this margin concern—so whether its high FCF margin can hold is itself the key variable for the later question (profit elasticity) in this logic chain. Second, the link on this chain that ultimately turns compute into money is the cloud providers who spend the most, not Broadcom: Google achieved a roughly 78% reduction in Gemini inference cost via its in-house TPU, and the vast majority of this "saved/earned" end value falls into Google's own pocket, while Broadcom takes a toll on the design and tape-out link. In other words, Broadcom helped its customers wrest back the "pricing power" of compute from Nvidia, but the bulk of the profit wrested back belongs to the customers; Broadcom's share is the stable, capital-light, but capped segment.

  So back to "where does profit ultimately flow": in the short term, the fattest and hardest-pricing-power links are HBM and the irreplaceable TSMC process, and the largest in absolute terms is the Nvidia platform; in the medium-to-long-term "deployment phase," profit is migrating from the shovel-selling chip side toward "money-collecting" parties like cloud platforms with their own in-house silicon. Broadcom's uniqueness is that it doesn't fight for food in the fattest pool, but holds the toll gate on the whole capital-heavy chain that is the lightest in capital and cleanest in cash conversion—and this is the foundation of the report's "high-quality compounding" depiction of it. But the value capture of this logic chain stops here: what it captures is a segment of high-quality toll, not the largest profit pool of end-compute monetization; and the real controversy around Broadcom right now has never been "can it capture profit," but "at the current price of $481, with a TTM FCF yield of only about 1.3%, has it already front-loaded all the next few years of tailwind for this toll segment"—this is the close that the later "has the market found it" question will address, and on this basis the report gives an "Avoid" rating (base date 2026-06-03, Q2 earnings not yet disclosed).

  2026년 6월 3일
• QHow large is the company's profit elasticity? If revenue grows 10%, how much will profit grow?

  A

  Conclusion first: Broadcom is a company whose "average economics are gorgeous, but whose marginal economics are dragged down by its growth mix." A rough projection: if revenue grows 10%, operating profit (EBIT) grows roughly 10%–15%—that is, a moderate positive operating leverage of 1.0–1.5x, not the explosive 2x-plus leverage of a pure software company; going one step further down to the "per-share owner earnings" layer, the leverage gets further ground down by high SBC, approaching or even slightly below the operating-profit growth rate. This is a projected range, not a precise multiple; below I make clear how I computed it and why I don't dare report it on the high side.

  The "numerator" of leverage: extremely capital-light + high gross margin, giving a base for positive leverage. In the report, this company's cost structure is very well-suited to amplifying profit: FY2025 revenue $63.887 billion, gross margin about 68%, the two segments of semiconductor + software combined operating profit about $42 billion (semiconductor $21.232 billion + software $20.765 billion), while capex was only $623 million and depreciation $574 million, the two almost pinned together. This means two things: one, at the company level there's almost no heavy-asset depreciation cost that surges with capacity, so new revenue doesn't first require a big fixed-asset outlay; two, the operating expenses below the gross-margin line (R&D, SG&A) are relatively fixed—the Hock Tan team is most famous for expense discipline. Putting these two together, every additional dollar of revenue can, in theory, sink a higher proportion than "average gross margin" into operating profit—this is the source of operating leverage. If incremental revenue can also run at ~68% gross margin with operating expenses largely not scaling proportionally, then operating-profit growth reaching 1.3–1.7x the revenue growth is within the range of common sense.

  But the "denominator" of leverage is pushed back by two reverse factors, and this is the key. First, what drives this 10% growth is precisely the lower-margin custom ASIC. The report already notes that after the December 2025 earnings the stock fell over 11% on "lower custom-chip margins and possible future gross-margin pressure"; external tracking gives a more specific magnitude—BofA cut its Broadcom FY2026/FY2027 gross-margin forecast to 73% and 71% (previously 75.4%/73.6%), because the rising share of XPU/custom silicon dilutes the mix that originally had about 93% software gross margin. In other words, the gross margin of that marginal slice of revenue is markedly below the 68% average (the incremental gross margin on ASIC hardware may be only in the 40-to-50% band), so the profit that incremental revenue can sink down is significantly thinned—this pushes the optimistic 1.3–1.7x leverage above back down to 1.0–1.3x. The more growth leans on AI custom chips, the more obvious this "gross-margin dilution," and this is exactly the "bigger but thinner revenue" the report repeatedly stresses.

  Second, going from operating profit to per-share return, high SBC takes another bite. Broadcom's FY2025 stock-based compensation (SBC) was about $7.568 billion, equivalent to about 12% of revenue and about 18% of combined operating profit, and Q1 FY2026 still had $2.176 billion (annualized ~$8.7 billion, rising rather than falling). SBC is a huge quasi-fixed cost that scales with talent retention/expansion, and the full-year diluted share count of ~4.85 billion shares is ~140 million more than basic shares (about 3% potential dilution per year). The report is right: although Broadcom is doing large buybacks, a sizable part of the buybacks merely offsets incentive dilution rather than purely shrinking the share count. The result is—even if GAAP operating profit is amplified with revenue, the leverage that lands at the "per-share owner earnings" layer gets ground flat once more by SBC and dilution offset, and EPS growth is likely only on par with, or even slightly below, operating-profit growth.

  So how to give the range. Putting together the positive base of "capital-light + high gross margin" with the two reverse factors of "ASIC dragging down incremental gross margin + high SBC eroding per-share":

  • At the operating-profit (EBIT) level: revenue +10% → roughly +10% to +15% (leverage factor ~1.0–1.5x), and the more growth tilts to custom ASIC, the closer to the lower bound;
  • At the per-share owner-earnings level: leverage is weaker, roughly in line with or slightly below operating profit, because the quasi-fixed cost of SBC and the dilution offset eat part of the amplification effect.

  In one sentence: Broadcom's "average" margins (68% gross margin, about 42% FCF margin, 1% capex) look like a perfect toll booth, but the portion of growth the market is now paying up for has a structurally lower marginal margin, so the operating leverage that actually materializes is moderate, not exponential. This precisely echoes the core of the report's "Avoid" rating—the market has front-loaded the near-perfect combination of "high growth + high cash flow both maxed out," but on the operating-leverage account, marginally it isn't actually as generous as it looks. (The above is a projection based on the report's financial structure, not a precise forecast; a change in any assumption—incremental gross margin, opex elasticity, SBC trajectory—would change the result.)

  2026년 6월 3일
• QHas the market already discovered this company? Or has it not yet realized all of this?

  A

  Conclusion first: this "company," Broadcom, was discovered long ago, even over-discovered; but what Serenity's twelfth question really asks—"is there still an unrecognized expectation gap at the value-capture point"—the answer is: yes, and its direction stands opposite to the stock price. The market repeatedly prices Broadcom as an "AI toll booth," but systematically underestimates two things: large customers are breaking the "toll booth" into multiple suppliers, and custom ASIC is pulling this machine's margins down. These two points are exactly the root of why the report downgraded the rating from "Watch" to "Avoid."

  That "the company is discovered" leaves no suspense at all. The current price is about $481.57 (the report's figure, base date 2026-06-03), corresponding to about $2.28 trillion market cap, with a TTM free-cash-flow yield of only about 1.3%, far below the roughly 4.46% 10-year Treasury of the same period—this itself is "discovered to the point it can't be discovered more" pricing. The sell side is even more one-sided: per S&P Global's survey of 47 analysts, Broadcom is a "Strong Buy" consensus with an average target price of $482.31, and for the month 32 Buy, 3 Hold, 0 Sell, with Morgan Stanley raising its target from 470 to 485. Note one detail: the consensus target price basically sits right on the current price—meaning Wall Street isn't "expecting it to keep rising," but has already bought the forward-expansion narrative of Google/Meta/Anthropic into today's price. In the report's scenario table, even the optimistic-case fair price is only about $423, yet below the current price, which is the same thing as this "target price ≈ current price" market state: the good news is basically already bought up.

  One under-priced expectation gap: customers are actively turning the "toll booth" into "multi-lane tolling." The market defaults to Broadcom being a near-irreplaceable exclusive co-development party on custom ASIC (Broadcom does hold about 70% of the custom AI accelerator market), but this "exclusive" narrative was personally torn open by its largest customer in April 2026: Google split the eighth-generation TPU v8 into two chips—the training chip Sunfish designed by Broadcom, and the inference chip Zebrafish designed by MediaTek, plus Marvell and Intel together forming a "four-way supply chain" to challenge Nvidia, with MediaTek's inference chip having a 20–30% lower unit cost. This is exactly the second script named in the report's pre-mortem—"weighty customers move part of the generation/design work out to a second supplier, and Broadcom shifts from 'core co-developer' to 'an important but replaceable executor'"—except the report wrote it as a risk assumption, while reality has already written it into Google's product roadmap. It hasn't dropped Broadcom out of the supply chain (the most profitable training-chip work is still in Broadcom's hands), but it shakes from the root the premise of "Broadcom = exclusive toll booth" that the market has priced in.

  The second under-priced expectation gap: this "high-cash-flow machine's" margins are being diluted by its own growth engine. The market is used to imagining the cash-flow elasticity after AI revenue doubles using Broadcom's overall gross margin of about 68% and FCF margin of about 42%, but AI revenue is precisely the lowest-margin slice: Broadcom's AI chip gross margin is about 65%, and as the share of system-level customization rises, semiconductor gross margin may slide from 65% toward 60%, while the VMware software gross margin of 80%+ is what truly props up the overall margin. That is—the more fiercely AI revenue grows, the faster the share of high-margin software in the revenue mix gets thinned, and "revenue grows beautifully, but per-share owner earnings can't keep up with revenue's slope." This is the same logic chain as the report's "revenue may become 'bigger but thinner'" and "software growth has already slowed to about 1%, with its support role retreating from a 'growth story' back to a 'profit foundation.'" The market is paying for "high growth + high margins" simultaneously, and these two have inherent tension in a custom-ASIC-dominated structure.

  So how to reconcile these two sides: At the company level—already discovered, and most likely over-priced; a 1.3% FCF yield against a 4.46% risk-free rate, even the report's optimistic scenario can't justify $481. At the expectation-gap level—the market prices the "AI toll booth" narrative as an exclusive, perpetual, high-margin compounding asset, yet underestimates the two structural variables of "customers diversifying suppliers" and "custom ASIC dragging down margins" that would discount the tolling right. This is exactly where the "Avoid" rating lands: not betting the business will collapse, but holding that the market has paid full price for an idealized expectation-capture point, while the real capture point is quietly narrowing and thinning. In one sentence—everyone has seen this company; but what many haven't seen is that the "toll booth" they think is ironclad is being gradually rebuilt, bit by bit, by its own customers and product mix. A reminder is needed: this question lands on the report's base date (before the FY2026 Q2 earnings disclosure), and if Q2 delivers and cements the 2027 line of sight, the "materialization timing" of this expectation gap will be pushed back; but the direction of the two slow-moving variables—supplier diversification and the margin structure—won't reverse because of a single-quarter beat.

  2026년 6월 3일