Cerebras Systems: A Long-Term Business Owner's Perspective

Hunting ten-year five-baggers among great growth stocks — pressing the upside question: "Can it get much bigger?"

• QHow high is its market ceiling? Is it growing a slice of an existing pie, or creating an entirely new market?6/10

  A

  Bottom line up front: Cerebras's ceiling is high, but it is fundamentally competing for a "low-latency inference" slice inside a large, already-proven, rapidly expanding pie (AI accelerators and inference infrastructure), rather than creating a brand-new market from nothing. Its scarcity comes from architectural differentiation (wafer-scale), not from opening up unmet demand.

  First, how big is the pie itself. AI accelerators and AI infrastructure are still growing fast, and the report judges that "long-term industry demand is very likely strong," giving it an industry-attractiveness score of 4/5. Objective data supports this: in Nvidia's fiscal 2026 (ended January 2026), full-year data-center revenue was already at hundred-billion-dollar scale, with quarterly data-center revenue exceeding $40 billion, showing the total demand pool for AI compute is large enough to hold multiple winners. In other words, Cerebras is not short of market space — it faces "an existing pie that is still growing fast."

  But it is not carving out the whole pie, only the "latency-sensitive inference" slice of it. The report is explicit: the company is shifting its narrative center of gravity "from training gradually toward latency-sensitive inference," and its website claims inference performance "up to 15x faster than GPUs"; the AWS partnership confirms this positioning too — in March 2026 the two companies announced pairing Cerebras's CS-3 system with AWS Trainium3, with Trainium handling prefill and Cerebras handling decode, aiming to push cloud inference speed to 5x existing solutions, available only through Amazon Bedrock. This is a strategy of "using differentiation to win a high-end niche within an existing inference market," not creating demand that did not previously exist.

  What determines the ceiling of this niche? It is determined by "how much inference workload will actually pay for latency." OpenAI's compute-purchase agreement worth more than $20 billion and 750 megawatts of low-latency inference capacity (expandable to 2 gigawatts through 2030) proves this niche demand is real and can be large; the report also cites $24.6 billion in remaining performance obligations (RPO) as of the end of 2025 as supporting evidence. So the "absolute height" of the ceiling is not low.

  The real constraint is not the height of the ceiling, but how much share Cerebras can capture, and for how long. The report repeatedly emphasizes that "the largest profit pool is clearly concentrated in mature platform players like Nvidia," with AMD, Broadcom, Marvell, AWS Trainium, Google TPU, and Groq all fighting over the same pool. In other words, it is pursuing high-end niche penetration in a large market "surrounded by giants, where the giants themselves are also developing their own inference chips" — the pie is large, but whether it can steadily eat its slice, without its low-latency ecosystem edge being quickly matched by giants, remains an open question.

  One-line characterization: ceiling = high (an existing large pie that is still growing fast); positioning = competing for the "low-latency inference" slice within that large pie, relying on architectural differentiation rather than creating new demand; constraint = share and durability, not market space. On the question Baillie Gifford's LTGG most cares about — "growing a slice of an existing pie or creating a new market" — Cerebras falls into the former category, and the most fiercely contested version of it.

  Jun 11, 2026
• QCan its revenue at least double over the next five years? Is growth driven mainly by volume, price, or new business?6/10

  A

  Bottom line up front: revenue "at least doubling" over the next five years is very likely — but not because the company's operations are especially stable, rather because it is sitting on a massive, already-signed backlog. Growth is driven mainly by "volume" (fulfilling large customers' compute contracts), with price and new business as secondary factors; the real question is not "can it double," but "is this kind of doubling high-quality, sustainable, and low-dilution."

  First, where does the confidence in doubling come from. The report body already sketches an extremely steep historical curve: revenue rose from $24.62 million in 2022 to $510 million in 2025 (hardware $358 million, cloud services $152 million), a roughly twenty-fold expansion over three years. More important, and something the report body doesn't fully unpack but which is decisive for this question: as of the end of 2025, the company disclosed remaining performance obligations (RPO) of $24.6 billion, made up mainly of the OpenAI multi-year compute agreement worth more than $20 billion and 750 megawatts. Relative to 2025 revenue of only $510 million, that is about 48 years' worth of revenue sitting in signed contracts. So "doubling to $1 billion+ within five years" is a low bar from a contract-coverage standpoint.

  But there is a timing mismatch here that the market easily overlooks, and it is exactly the crux of this question. Management estimates that of this $24.6 billion RPO, only about 15% will be recognized across 2026 and 2027 combined, with the vast majority not converting to revenue until 2028 and beyond. That 15% is roughly $3.7 billion spread over two years — layered on top of the $510 million base, meaning near-term growth in 2026-2027 is positive, but nowhere near as explosive as the headline backlog suggests; it is "back-loaded." This also explains reports that the company warned of slowing revenue growth in the first quarter of 2026, a matter of delivery/recognition timing rather than demand. So doubling — or several-fold growth — by five years out (through 2030/2031) is likely achievable, but the growth curve is "flat first, then steep," not linear.

  Breaking down the growth drivers — mainly "volume":

  • Volume (primary driver): essentially fulfilling large orders like OpenAI's 750-megawatt (expandable to 2-gigawatt) deal, plus the incremental cloud-inference volume from AWS going live on Amazon Bedrock in the second half of 2026. This is a story of "how many watts of compute get deployed."
  • Price (secondary, even under pressure): the report notes 2025 gross margin fell back from 42.3% to 39.0%, which public reporting attributes to rising data-center costs. This means growth is not currently price-driven — quite the opposite, the company is absorbing cost pressure from scaling delivery.
  • New business (adds quality, but still a small share): cloud services were $152 million in 2025, about 30% of revenue, representing the shift from "selling systems" to "selling recurring compute," but the report explicitly notes that this segment's "true retention, gross margin, and renewal quality still lack several consecutive quarters of post-IPO data to validate."

  Under Baillie Gifford's LTGG framework, the honest answer to this question is: the "doubling" answer is yes, and likely with room to spare, driven by the volume of contracts already in hand; but growth quality is discounted — customers are highly concentrated (discussed further below: 86% of 2025 revenue came from two UAE entities), near-term recognition is slow, gross margin is under pressure, and historically positive cash flow has depended heavily on prepayments. So this is a stock with "high growth certainty but growth quality still to be proven," not the ideal growth stock where "volume and price both rise, and it gets healthier the longer it grows."

  Jun 11, 2026
• QFive years from now, what will take over as the next growth engine? Does this "second curve" exist today?4/10

  A

  Bottom line up front: Cerebras's current "first curve" is actually two stages layered together — early on it was training-system sales, and now the primary engine has shifted to low-latency inference. Asked "does a second curve exist today," the honest answer is: a second curve (moving from "selling hardware systems" to "selling recurring cloud compute services") has already begun to show, but it is still small and far from proven; it looks more like a quality upgrade of the first curve than an independent new growth pole capable of taking over on its own five years from now.

  First, clarify the current engine. The report body clearly states the company is shifting its narrative center of gravity "from training gradually toward latency-sensitive inference" — this is a curve switch Cerebras made on its own. The real growth engine right now is fulfilling the OpenAI low-latency inference compute contract worth more than $20 billion and 750 megawatts, plus AWS connecting the CS-3 to Amazon Bedrock for inference decode. This inference curve itself is still in its steepest climbing phase.

  So what is the "second curve" candidate? The most visible one, and the direction the company is actively betting on, is shifting revenue structure from "project-based hardware delivery" to "recurring cloud/dedicated compute services." The evidence: 2025 cloud services revenue was $152 million, about 30% of total revenue (hardware was $358 million); the report also cites this share based on company disclosure and reads it as "the business model migrating toward more recurring cloud computing services." If this line can keep scaling, with high renewal rates and stable margins, it would be the key to re-rating Cerebras from "a capital-intensive hardware/compute integrator" to "a software-like AI infrastructure company" — precisely the watershed for whether the valuation can be supported.

  But calling it a second curve that "can take over five years from now" is not yet backed by enough evidence:

  • It is still only about 30% of revenue today, not large enough to take over independently;
  • The report explicitly warns that this segment's "true retention, gross margin, and renewal quality still lack several consecutive quarters of post-IPO data to validate" — meaning it currently looks more like "another delivery format for the same large customers (including OpenAI/AWS)" than an entirely new customer base or entirely new demand;
  • It is highly homologous with the first curve (inference compute) — same customers, same supply chain (the core chip still depends on the single foundry TSMC) — limiting how much it actually diversifies risk.

  What Baillie Gifford is asking with "does a second curve exist today" emphasizes: when the primary engine slows down, is there another engine already ignited that can take over? For Cerebras: today there is no mature new business independent of inference compute that could step up when the primary engine slows. Cloud-ification is a direction, a quality upgrade, but it still lives off the same handful of mega-customers and the same supply chain. This is also part of why the report gives "business understandability" only 3/5 and judges the long-term economics as "not yet verifiable" — the growth story is heavily concentrated in one main line: fulfilling existing large orders plus cloud-ifying delivery, lacking a second, independent safety cushion.

  Jun 11, 2026
• QWhat is its core competitive advantage? Will this moat widen or narrow over the next three to five years?3/10

  A

  Bottom line up front: Cerebras's core competitive advantage is "low-latency inference performance enabled by its wafer-scale architecture" — a real technology barrier, but not yet a wide and durable economic moat. Over the next three to five years, it more likely faces a two-sided outcome: "widening on some dimensions, narrowing on others as giants catch up," with the net effect depending on whether it can convert today's speed lead into ecosystem lock-in and sustainable pricing power. The report scores the moat 2/5, and I agree with this cautious assessment.

  First, the advantage is real. The differentiators the report lists hold up: the core is the Wafer-Scale Engine and CS-3 system — turning an entire wafer into a single processor. The company claims 4 trillion transistors and 900,000 AI cores per chip, and claims inference "up to 15x faster than GPUs." This architecture is not a slide deck: OpenAI signed more than $20 billion in compute, and AWS pairs the CS-3 with Trainium3 for inference decode, targeting 5x faster than existing cloud solutions — two top-tier buyers have validated with real money that it does lead in the "latency-sensitive inference" niche.

  But after breaking the moat down item by item, the report's judgment is sober-minded; viewed by time horizon:

  • What history/the present can prove: switching costs are "moderate" (once real-time applications are optimized onto the Cerebras platform for latency, switching creates friction); patent/process/engineering barriers are "moderate" (replicating a wafer-scale system requires substantial time, talent, and capital); engineering and sales execution are "moderately strong." These show it has established a real technical position "then/now."
  • Looking forward three to five years at the margin, several factors are clearly weak: brand is "weak to moderate," nowhere near becoming a de facto standard like Nvidia's CUDA; network effects are "weak" — OpenAI API compatibility lowers the trial barrier but does not create CUDA-style ecosystem lock-in; scale advantage is "weak" — against Nvidia's data-center revenue of more than $40 billion in a single quarter, Cerebras's full-year $510 million is still a rounding error; cost advantage is "unproven."

  Why the risk of "narrowing" over three to five years should not be understated: this is an industry the report calls "extremely brutal." Within the same inference profit pool, Nvidia, AMD, Broadcom, and Marvell are all pushing their own solutions, and even its own major customers are developing in-house alternatives — AWS has Trainium, Google has TPU. "Today's leading latency metric" is dynamic: once more mature ecosystems close the latency gap, Cerebras's differentiation premium could compress quickly. The report's own line cuts to the heart of it: "today's leading performance metric does not automatically translate into excess returns on capital a decade from now."

  Why it could also "widen": if the deployments by AWS (the first hyperscaler to commit to its hardware) and OpenAI turn "choosing Cerebras for latency" into a developer habit, switching costs could move from "moderate" toward "high," and combined with another generation of wafer-scale engineering accumulation, the moat could deepen further. But this needs time, and needs Cerebras to keep leading in the gap while giants develop in-house — this is "possible, but not yet realized."

  Baillie Gifford's honest answer on "will the moat widen or narrow over three to five years" for Cerebras is: it has a real barrier (wafer-scale engineering + latency lead + top-tier customer endorsement), but is currently a "narrow moat" whose direction is undetermined and highly dependent on factors outside its control (how fast giants catch up, whether the ecosystem can lock in). This is exactly why the report characterizes it as having "a technology barrier, but not yet proven to have a wide and durable economic moat," giving it 2/5 — and why it looks more like "a high-risk challenger in a good industry" than a moat-protected company worth holding at a high price for the long term with confidence.

  Jun 11, 2026
• QIf its core business were disrupted, does it have the genetic makeup to reinvent itself? How does it handle mistakes and bad news?4/10

  A

  Bottom line up front: Cerebras has already proven a capacity for "self-reinvention" once — it actively shifted its center of gravity from "selling training systems" to "low-latency inference," a genuine strategic pivot. But on "how it handles mistakes and bad news," its record is mixed: it has shown boldness in pivoting the business, but it also carries real historical baggage and unremediated present-day internal-control weaknesses on financial integrity and governance. For a company meant to be held long term, the latter will keep weighing on my trust.

  First, the self-reinvention genes (a positive). The product evolution the report describes is itself a reinvention: shifting the training narrative "gradually toward latency-sensitive inference," and restructuring outside partnerships around inference — landing more than $20 billion in OpenAI compute and an AWS inference integration split into prefill/decode. Being able to actively pivot away from the "world's largest training chip" banner toward the inference battleground that has more commercial traction today shows the team does not cling to its original path and has real adaptive flexibility. The report gives "corporate culture and execution" a moderately strong score, precisely because it sees the company "push a differentiated product into OpenAI, AWS, healthcare, and enterprise search use cases." So from a capability standpoint, "can it reinvent itself again if the core business is disrupted" leans positive.

  But the evidence on "how it handles mistakes and bad news" is not reassuring:

  First, the company proactively acknowledges in its prospectus that it has material weaknesses in internal control over financial reporting. Acknowledging it is honest, but the fact that the "weakness remains unremediated" is itself a present-day risk, and the report lists it as a governance risk that should not be glossed over.

  Second, and the heaviest historical baggage: CEO Andrew Feldman was involved in an accounting-controls case during his time at Riverstone Networks — according to public reporting, the SEC alleged he participated in a scheme that inflated revenue by nearly $30 million, and he personally pleaded guilty in 2007 to a charge of circumventing accounting controls. The report notes this history is "verifiable through SEC and litigation records." Putting "a CEO with a prior guilty plea for circumventing accounting controls" next to "the same company today disclosing its own material internal-control weakness" is exactly the reddest of red flags on this question of handling bad news — not that history determines the present, but that this combination demands greater transparency and a longer observation period to rebuild trust.

  Third, the governance structure weakens the external checks needed for correction: Class B shares carry 20 votes per share, significantly weakening common shareholders' ability to check management. On the positive side, the company added an independent director and an audit committee chair in 2024 (per the report), showing it is "catching up"; but under super-voting rights, the strength of external correction remains limited.

  Putting both sides together, Baillie Gifford is asking whether the company will "face adversity honestly, and have the capacity to reinvent." Cerebras's profile is: it has real reinvention genes at the business level (bold pivots, decent execution, similar to WPM's one successful transformation); but on financial integrity/governance, historical stains plus present-day unremediated internal-control weaknesses plus super-voting rights constitute a substantive deduction on "how it handles mistakes." The report accordingly judges "is management fully trustworthy" as a direct Fail, and gives both moat and management 2/5 — I agree: being bold enough to reinvent technologically does not equate to being reassuring on governance; the two must be evaluated separately.

  Jun 11, 2026
• QDoes management (especially the founder) have a long-term view, with interests deeply aligned with the company? Are they willing to sacrifice today's profit for five to ten years from now?5/10

  A

  Bottom line up front: on "long-term view" and "willingness to sacrifice today's profit for five to ten years from now," Cerebras clearly qualifies — the company keeps burning large sums to build capacity and absorbing cash up front for long-dated compute contracts, which is itself a long-term bet made at the expense of near-term profit; the founder is also deeply present. But on "interests being deeply and healthily aligned with the company and minority shareholders," it must be discounted — super-voting rights create a mismatch between the founding team's control and common shareholders' economic interests, and combined with the CEO's accounting-integrity history, "alignment" is more about control than the kind of interest alignment that reassures outside shareholders.

  First, "long-term view + willingness to sacrifice today's profit" (positive, and well supported). This is exactly what Cerebras has no shortage of:

  • Report data shows that in 2025, even as revenue reached $510 million, the company still posted an operating loss of about $146 million, free cash flow of about -$393 million, and capex as high as $383 million (about 75% of that year's revenue). This posture of "consuming cash first in order to grow" is essentially trading today's profit for a decade of compute territory.
  • It was bold enough to sign OpenAI's multi-year contract worth more than $20 billion and 750 megawatts (expandable to 2 gigawatts), and reportedly accepted a $1 billion, 6% annual-interest loan from OpenAI to build data centers — all textbook examples of "taking on current-period capital and financial burden in exchange for delivery five to ten years out." On "willingness to sacrifice the near term for the long term," management is qualified, even aggressive.

  Next, the founder's presence and technical commitment (positive). The report notes CEO Andrew Feldman is a co-founder, and the management and founding team have deep technical backgrounds. Founder-led, technology-driven — this fits the "founder still at the helm long term" profile Baillie Gifford favors.

  But the "deep interest alignment" side must honestly be discounted:

  • Governance mismatch: Class B shares carry 20 votes per share, meaning the founding team can lock in control with a relatively small economic stake. This is "control-alignment," but it simultaneously weakens common shareholders' voice — the report explicitly lists super-voting rights as a governance discount item. The direction of alignment favors the founder, not necessarily outside shareholders.
  • Integrity discount: the CEO pleaded guilty in 2007 to "circumventing accounting controls" during his Riverstone years, layered on top of the company's own present-day disclosure of a material internal-control weakness. What Baillie Gifford wants is "someone worth entrusting with a decade," and this combination of history plus current status directly lowers that sense of security.
  • Growth still highly dependent on outside transfusions rather than self-funding: the report traces a dense financing history (private rounds, Series H, revolving credit, then $5.55 billion raised in the IPO), and notes the company's growth model "remains, in essence, highly dependent on external financing rather than internal free-cash-flow compounding." For common shareholders, this means dilution — in 2025, diluted share count jumped from about 49 million to about 172 million, and the report notes "future per-share value must grow faster just to offset the dilution." Capital-allocation maturity (rational buybacks/dividends/stable returns on capital) has not yet been demonstrated.

  Baillie Gifford's honest conclusion on this question: long-term view = strong; willingness to sacrifice near-term profit for the long run = strong (even somewhat aggressive); but "interests deeply and healthily aligned with (minority) shareholders" = weaker — super-voting rights + integrity history + ongoing dilution mean alignment skews toward founder control rather than reassuring outside long-term owners. The report accordingly gives "management and capital allocation" 2/5 and judges "is management fully trustworthy" as Fail. I agree: this is a team "bold enough to bet on the long term, but whose governance credibility needs more time and a cleaner internal-control record to prove."

  Jun 11, 2026
• QIf it disappeared tomorrow, how much would customers miss it? Is its growth sustainable and free of dependence on harming society or inviting regulatory backlash?4/10

  A

  Bottom line up front (dual lens). Irreplaceability: in the "extreme low-latency inference" niche, a handful of top customers would miss it quite a lot in the short run; but zoomed out to the whole AI compute market, it is far from the point where "the industry stops if it disappears" — substitutability remains high. Social/regulatory sustainability: this is Cerebras's most fragile side — its growth is highly tied to UAE customers and a cross-border relationship under national-security scrutiny, which is growth that carries regulatory and geopolitical risk built in, not clean, sustainable growth.

  First, "how much would customers miss it" (irreplaceability, moderate but narrow). In latency-sensitive scenarios, it does give specific customers real reasons not to leave:

  • OpenAI locked in its low-latency inference compute with a multi-year contract worth more than $20 billion and 750 megawatts, and lent it $1 billion to build data centers, while also taking potential equity and warrants of up to 11% — a customer willing to pay, take equity, and lend money all at once shows Cerebras genuinely matters to it on "speed."
  • AWS chose to connect the CS-3 to Bedrock for inference decode, targeting 5x faster than existing cloud solutions specifically because of its unique latency characteristics.

  But the breadth and depth of that "missing it" are both limited: the report judges network effects "weak" and brand "weak to moderate," far from the kind of de facto standard lock-in CUDA has achieved; and its two flagship customers are themselves developing in-house alternatives — AWS has Trainium, Google has TPU, and Nvidia/AMD/Broadcom are all pushing inference solutions. In other words, if Cerebras disappeared tomorrow, the most latency-sensitive workloads would suffer a short-term shock and be forced back onto somewhat slower alternatives, but AI inference as a whole would not grind to a halt. Irreplaceability = moderate within the niche, weak globally.

  Next, social/regulatory sustainability (the real red line). Baillie Gifford wants to confirm "growth does not depend on harming society or inviting regulatory backlash," and Cerebras clearly loses points here:

  • Customers are extremely concentrated in entities linked to one country: the report and public disclosures agree — about 86% of 2025 revenue came from two UAE-related entities, with MBZUAI at about 62% and G42 at about 24%. Growth being tightly bound to a handful of buyers in a single region is itself a sign that "the growth mode is insufficiently diversified and insufficiently robust."
  • A real history of national-security review: the report explicitly notes the company's G42/UAE-related business previously triggered CFIUS review and delayed the IPO; and exporting advanced AI chips to the Middle East sits in a sensitive corner of export controls to begin with (the company says it has obtained export licenses for the CS-2/CS-3/future CS-4 to the UAE). In other words, its largest revenue engine sits right on the line of geopolitics and export controls — a shift in the regulatory stance could impede growth.

  Putting the dual lens together: customers would miss it (within the niche), but not to the point of being irreplaceable; more critically, its current growth mode is highly dependent on cross-border relationships with major customers under regulatory scrutiny — growth that carries a built-in sustainability discount. This is exactly why the report singles out "customer concentration risk" and "regulatory and geopolitical risk" as top risks, and one of the core reasons behind the Avoid rating. Against Baillie Gifford's dual standard of "irreplaceable + clean, sustainable growth," Cerebras barely passes the first test and clearly fails the second.

  Jun 11, 2026
• QWhat do this business's unit economics (gross margin, incremental returns) look like? Do they improve or worsen at greater scale? Where does the money earned go?3/10

  A

  Bottom line up front: Cerebras's unit economics are currently "moderate, and heading in an unclear direction" — gross margin is about 39% and still falling back due to rising data-center costs; more importantly, growing bigger has not made it "throw off more cash as it grows," but rather "burn more cash as it grows": incremental return is currently negative. Money earned (and money raised) has gone almost entirely into heavy capex to build capacity. This is the fundamental reason the report judges unit-economics-related items as a Fail.

  First, the gross-margin layer (moderate, under pressure). The report's own trajectory is one of improvement: gross margin rose from 11.7% in 2022 to 42.3% in 2024, proving it is "not just selling systems at a loss." But 2025 gross margin fell back to 39.0%, which public reporting attributes to rising data-center costs. This matters: a company with genuinely "software-like, scale-as-leverage" unit economics typically sees gross margin rise as it scales; Cerebras instead got pushed back down by costs while ramping volume, suggesting its economic nature today skews more toward "capital-intensive hardware/compute integration" than high-margin software. This is also the exact question the report poses rhetorically — "will it ultimately be a high-margin, software-like AI infrastructure company, or a capital-intensive, project-based hardware integrator dependent on large-customer prepayments?"

  Next, incremental returns and "does scale make it better or worse" (clearly worse). This is the sharpest part of this question:

  • In 2025 the company posted record revenue, but operating loss was about $146 million, operating cash flow was only about -$10.05 million, and free cash flow was about -$393 million.
  • On the surface 2025 had about $238 million in GAAP net income, but the report points out this was mainly driven by about $363.3 million in unusual items (Other Unusual Items, related to fair-value changes/extinguishment of the forward contract liability), not operating cash generation. Strip that out, and true operations are still unprofitable.
  • The impressive positive cash flow in 2024 (operating cash flow about $452 million, FCF about $429 million) was likewise an "illusion" — the report notes the Change in Unearned Revenue that year was as high as +$679 million, a point-in-time inflow driven by large-customer prepayments; by 2025 that line flipped to -$176 million, and cash flow turned negative right along with it. So "scale up → cash flow improves" does not hold for Cerebras; instead, as scale grew, prepayments pulled back and cash flow turned negative.

  Finally, "where does the money earned go" (almost entirely into capex). The report offers two especially telling ratios: 2025 capex was $383 million, about 75% of that year's revenue, and about 11x depreciation and amortization (about $34.45 million). These numbers make it plain: this is not "spinning off cash as it grows," but "consuming cash first in order to grow" — the money is going mainly into racks, data centers, and capacity. Layer on top of that accepting a $1 billion loan from OpenAI to build data centers, plus its own heavy equity and debt financing (the IPO alone raised $5.55 billion), and it's clear that capital today is going toward "building capacity, grabbing ground," not returning cash to shareholders.

  Baillie Gifford's honest conclusion on this question: unit economics = moderate and under pressure (39% gross margin, and falling); incremental return = currently negative, with greater scale temporarily worsening rather than improving cash flow; where the money goes = heavy capex to build capacity. This is the opposite of the "scale is leverage, bigger means more profitable, high-return incremental capital" profile Baillie Gifford favors — the report accordingly judges both "can it generate stable free cash flow" and "is its return on capital excellent" as Fail, and rates it Avoid on that basis. One caveat for balance: if the cloud-services share keeps rising and the backlog scales up post-2028, unit economics could improve; but that is a "direction yet to be proven," not "an already-established fact."

  Jun 11, 2026
• QWhat conditions would all need to hold simultaneously for it to rise five-fold in ten years? Are those conditions realistic? What expectations does today's price already imply?2/10

  A

  Bottom line up front: starting from the current level of about $237, roughly $52 billion in market cap (on about 219.6 million basic shares; on a fully diluted basis of about 305 million shares, roughly $72 billion in equity value), a five-fold rise in ten years would require a long chain of high-difficulty conditions to hold simultaneously — and today's price has already priced in a substantial share of that "flawless execution" in advance. The report judged the earlier $279.72 price as offering "no margin of safety, Avoid"; the current price, though down about 40% from its high of $386, is still significantly above the report's estimated intrinsic-value range.

  What conditions would need to hold "simultaneously" (none can be missing) for a ten-year five-fold gain:

  1. Revenue reaching platform-scale. The report's optimistic scenario only assumes 2035 revenue of $6-10 billion or more; supporting a five-fold gain basically requires landing at the upper end of that range or higher. This means high-quality fulfillment of the $24.6 billion RPO, plus continuing to win new orders beyond the existing contracts.
  2. A qualitative shift in unit economics. Owner Earnings margin would need to climb from today's "negative on a conservative basis" all the way to the roughly 20% assumed in the report's optimistic scenario, with gross margin stabilizing around 50% (the reality is 2025 gross margin was 39% and falling).
  3. De-risking the customer base. The 86% of 2025 revenue concentrated in two UAE entities would need to be substantially diversified, without being interrupted by export-control/CFIUS-type regulation.
  4. Capex intensity coming down sharply. From capex at about 75% of revenue in 2025, converging toward a level that can generate meaningful free cash flow — that is, flipping from "burning more cash the longer it grows" to "throwing off more cash the longer it grows."
  5. Dilution under control. Diluted share count already jumped from about 49 million to about 172 million (+251%) in 2025; future financing/lock-up expirations cannot dilute per-share value substantially further.
  6. The moat holding up against giants catching up. Under pressure from Nvidia, AMD, Broadcom, and in-house customer alternatives (AWS Trainium, Google TPU), the latency lead would need to hold for long enough (the report gives the moat only 2/5).
  7. Governance and internal controls fully remediated. The currently disclosed material internal-control weakness would need to be confirmed resolved in an audit sense, without recurring.

  Are these conditions realistic? Each one, viewed alone, is "possible," but requiring all seven to hold simultaneously and continuously for ten straight years without a single stumble is a textbook case of "stacking heroic premises." The report's own judgment is direct: justifying the current price "typically requires even more heroic premises... I believe this already goes beyond the bounds of conservative and rational." I agree — this is not "somewhat optimistic," it is "treating ten years of flawless execution as the base case."

  What expectations does today's price already imply? The report's own valuation anchors make this most concrete to quantify:

  • Absolute method: the report's three-tier intrinsic value is conservative $20-30, reasonable $30-50, optimistic $50-80 per share. Even the most generous optimistic tier ($50-80) is far below the current price of about $237. In other words, the market has not just priced in the optimistic scenario, but has layered roughly another 3-5x premium on top of it.
  • Cash-support method: the report estimates post-IPO net cash support of only about $19-21 per share. The great majority of the current price is not existing assets, but an extremely high expectation for a decade of operating success.
  • Relative method: the report put Cerebras's P/S at about 167x (fully diluted) when the price was $279.72; at the current price of about $237, fully diluted P/S is still roughly 140x; against Nvidia's P/S of about 20x (with ROIC above 100%) and Broadcom's about 30x — these already-profitable, already-proven-cash-flow leaders are nowhere near this expensive. The report's own line is the key takeaway: "peers being expensive does not mean Cerebras is cheap."

  Baillie Gifford's honest conclusion on this question: a ten-year five-fold gain is not impossible at the "absolute ceiling" level (the industry is large enough, the backlog is thick enough), but it requires seven high-difficulty conditions to be fulfilled continuously and simultaneously; and today's price of about $237 has already pre-spent several times over the report's optimistic scenario — the implied expectation is close to "flawlessly executing into the next platform-scale giant over the next decade." This means that even if the fundamentals keep improving, the valuation starting point has already used up most of the upside; conversely, if any single condition disappoints, the reversion space the report calculates (toward the $20-30 range) corresponds to possibly 80%-90% permanent capital loss risk. The conclusion matches the report: the price is the problem, not the company's lack of opportunity.

  Jun 11, 2026
• QWhy hasn't the market recognized all this yet? Is it a failure to understand, a failure to respect, or a failure to see far enough? What would become the "narrative inflection point"?3/10

  A

  Bottom line up front: this is not a case of "the market doesn't get it and it's being overlooked" — quite the opposite, the market understands the story very well, and has already priced it (even over-priced it) as "a future platform-scale giant." So the real question here is not "why hasn't the market recognized the value yet," but "why is the market willing to pay so much for a perfect future that hasn't been delivered yet." The narrative inflection point therefore runs both ways: on the upside it has largely already been used up; what would actually change the plot is the hard data that proves or disproves "high-quality recurring cash flow."

  First, responding to the "doesn't understand/doesn't respect/doesn't see far enough" framing — none of the three really holds:

  • Not "doesn't understand": this was one of the most closely watched tech IPOs of 2026 — it priced above the expected range at $185, raised $5.55 billion, and jumped about 108% on its first day, with demand oversubscribed more than 20 times. Sell-side coverage has also rolled out: within a week of the quiet period ending, Morgan Stanley set a $250 target (Overweight), Citi set a Street-high $340 (Buy), UBS and Needham both set $300, and Wedbush set $270. The market has not just understood it — it has collectively turned bullish.
  • Not "doesn't respect": OpenAI's more-than-$20-billion compute contract plus equity stake, AWS exclusively connecting the CS-3 to Bedrock, and the $24.6 billion RPO are all bullish anchors the market has weighted heavily.
  • More a case of "seeing too far ahead, and only one side of it": the market treats the ten-years-out backlog fulfillment and margin transformation as nearly certain, while relatively downplaying several "present-day facts" the report keeps emphasizing — 86% of revenue from two UAE entities, a real 2025 operating loss of about $146 million and free cash flow of about -$393 million, capex at about 75% of revenue, a material internal-control weakness, and the CEO's accounting-integrity history.

  So where is the real perception gap? Not in "hidden value the market has missed," but in the mismatch between two different pricing languages — the market prices off a growth/RPO narrative, while a value framework prices off already-verified cash flow. The report quantifies this gap: three-tier intrinsic value of $20-80 per share, against a current price of about $237 — even the optimistic tier is re-premiumed several times over. This means the narrative's "upside inflection" has essentially already happened and already been captured, leaving mostly downside risk and the difficulty of sustaining such a high valuation from here.

  What would become the narrative inflection point (bidirectional, with the following data points mattering most):

  • Downside (more likely to arrive first): ① order-fulfillment pace from major customers disappointing — management has already flagged that only about 15% of the $24.6 billion RPO will be recognized in 2026-2027, and growth is expected to slow in Q1 2026; any disappointing quarterly report could puncture the "high growth" expectation; ② customer concentration failing to come down over the long run, or the UAE relationship again triggering regulatory/export-control action; ③ large-scale lock-up expirations or refinancing bringing another round of dilution; ④ internal-control deficiencies recurring or remediation being delayed; ⑤ Nvidia/AMD/AWS closing the low-latency inference gap, collapsing the differentiation premium. The report notes that if any of these occurs, "the stock's reversion could be very severe," with a reversion toward the $20-30 range corresponding to possibly 80%-90% permanent capital loss.
  • Upside (a higher bar): several consecutive quarters proving "operating cash flow, after stripping out deferred-revenue swings, stays consistently positive + gross margin stabilizes above 50% + customers diversify + capex intensity declines" — in other words, turning the "story" into "verifiable, high-quality cash flow." The report is clear that only meeting this whole set would move it up from the watch list.

  Baillie Gifford's honest conclusion on this question: the market has not failed to notice — it has chosen a pricing language that is extremely favorable to Cerebras, treating a perfect future as the base case. The narrative inflection point is more likely to be triggered by negative surprises on fulfillment pace/cash-flow quality/regulation than by "the market finally discovering it was undervalued" — because under a value framework, its problem has never been being undervalued, but being over-extended. This is entirely consistent with the report's main judgment: a good company in embryo, at a bad price; fine as a high-priority watchlist name, not appropriate as a value-investing buy at the current price.

  Jun 11, 2026