Long-Term Owner's Analysis of Huawei HiSilicon Ascend

Business Understanding and Industry Structure

By official definition, HiSilicon is a semiconductor and device design company whose capabilities span connectivity, sensing, video and audio processing, intelligent computing, chip architecture and process technology, high-performance circuit design, security, and other fields. "Ascend" is one of HiSilicon/Huawei's core series in artificial intelligence computing, providing AI computing power for digital centers, edge, consumer terminals, and IoT scenarios. Put differently, HiSilicon is the platform chip-design entity, and Ascend is one of its most important AI computing product families.

How does Ascend make money? The official framing already gives a fairly clear outline. Huawei's enterprise business disclosures state that Ascend Computing builds Atlas solutions on Ascend AI processors and foundational software, with product forms including modules, cards, edge stations, servers, and clusters, covering the full process from inference to training. It also provides solutions for large-model training, inference, and post-training. Huawei's 2025 annual report adds a more complete commercial loop: Ascend does not merely sell chips; it forms integrated sales across "chips, servers, supernodes, software stack, cloud services, and industry solutions." In other words, there is more than one charging point: hardware sales, system delivery, software enablement, cloud computing power, and integrated industry-solution monetization.

Who are the customers? At least based on official disclosures, Ascend and Huawei Cloud computing infrastructure have already landed in internet, finance, telecom, power, government, healthcare, education, manufacturing, oil and gas, and other industries. Huawei's 2025 annual report says its 384-NPU SuperPoD has been deployed at scale in internet, finance, telecom, and power industries. Huawei Cloud covers more than 30 industries and more than 500 scenarios, and AICS had more than 1,800 customers by year-end 2025. This customer profile means Ascend is heavily tilted toward government and enterprise customers, large internet companies, regulated industries, and computing centers, not a standardized and fragmented consumer-products market.

That also means the recurrence and predictability of revenue are above average, but not especially high. If you sell an AI accelerator card or a supernode system, revenue naturally leans toward project-based, batch-delivery, capital-expenditure spending. If you sell cloud computing power, developer tools, and industry solutions, repurchase and subscription-like attributes are stronger. Ascend is clearly now in a stage of "hardware volume ramp + ecosystem expansion + cloud-service absorption," so the revenue mix is likely to include both one-off large orders and recurring platform revenue. Because Ascend does not disclose stand-alone financials, outside investors cannot know the proportions.

On cost structure, this is not an easy business. HiSilicon explicitly describes itself as a chip design company, which means it still needs to rely on external manufacturing, packaging and testing, materials, and equipment systems. Huawei Group's 2025 R&D spending reached RMB 192.3 billion, equal to 21.8% of revenue; it had 114,000 R&D employees, or 53.7% of total headcount, and 165,000 active granted patents. Such a high R&D weight does not reflect a business form that has matured into rent collection. It reflects continued high-intensity investment and an industrial offensive racing global leaders.

Dependencies are also high and concentrated in a few key areas: advanced process and packaging, procurement by large domestic customers, policy support, and software-ecosystem migration. Reuters, citing IDC, reported that China's AI accelerator market shipped about 4 million units in 2025, with NVIDIA still holding 55% and Chinese vendors together holding 41%; among domestic vendors, Huawei shipped about 812,000 units, Alibaba 265,000 units, and Baidu and Cambricon 116,000 units each. This structure shows two things: China's domestic-substitution window is real, and NVIDIA remains very strong, while Ascend is nowhere near a stage where "the competition is over".

If the only question is whether this is a business I can understand, my answer is: yes, but it is not a simple business. It is not selling a can of Coca-Cola, nor is it selling a bank card. It is a highly engineered, capital-intensive infrastructure business with strong coupling to policy and ecosystems. As a long-term owner, I can understand how it creates value; as an outside investor, I still cannot see clearly how, at what pace, and at what rate of return it converts that value into distributable cash flow. I therefore score "business understandability" at 3/5.

If we widen the question to the industry, the picture becomes more interesting. AI computing infrastructure is almost certainly still in a long-term growth stage, not a mature or declining stage. NVIDIA generated USD 215.938 billion of revenue in fiscal 2026, including USD 193.7 billion from data center revenue, up 68% year over year. That proves at a global level that AI infrastructure spending remains very strong. In China, domestic substitution adds another layer of structural demand for Ascend. Strong demand is almost beyond dispute.

At the same time, this is an industry whose ranking can be easily disrupted by technology paths, regulatory rules, export controls, and supply-chain constraints. A U.S. Commerce Department official publicly said in March 2026 that Huawei could produce "no more than 200,000" advanced AI chips in 2025. Huawei, meanwhile, continued in its 2025 annual report and subsequent public communications to emphasize growth in the Computation/Ascend ecosystem, scaled supernode deployments, and next-generation cluster advances. For outside investors, this shows that the key industry variable is not demand alone, but whether demand can be converted on time through constrained capacity, a mature software stack, and stable yield.

So the more accurate characterization is: a good industry, but not an easy one; a hard battle inside a large opportunity, not a cash cow whose moat forms naturally. If asked whether it is a good company in a good industry or an excellent company in a poor industry, my judgment is closer to the former, with one qualifier: in the context of China's domestic substitution, it is a strong company in a good industry; in an unconstrained global competitive context, it remains a strong challenger still catching up. I score "industry attractiveness" at 4/5.

If the stock market closed for five years, would I want to own this business? If I were an industrial owner who could fully control the enterprise and accept a long period of high investment, yes. If I were an outside investor requiring transparent pricing and clear shareholder rights, no, because I cannot obtain that ownership vehicle.

Moat and Management

Ascend's most credible moat is not a single chip specification. It is "a full-stack AI infrastructure that is locally usable in China". This moat has several layers. The first is chip and systems-engineering capability. Huawei officially disclosed that the Atlas 800T A3 supernode server can support up to 384 NPUs in high-speed interconnection, 48TB of on-chip memory with unified addressing, and 784GB/s bidirectional interconnect bandwidth, while more than 300 Atlas 900 supernodes had been cumulatively deployed, serving more than 20 customers. The second layer is the software stack and developer ecosystem. Huawei disclosed 4 million Ascend developers as of year-end 2025, and has open-sourced or opened software such as CANN and the Mind series. The third layer is cloud absorption capability. Huawei Cloud has connected CloudMatrix AI Infra, AICS, and enterprise-cloud scenarios, with more than 1,800 customers. The real competitive barrier comes from the stacking of these three layers, not from a single chip model.

If assessed item by item across the ten moat types you requested, my judgment is as follows. Brand advantage: medium to strong in Chinese government, enterprise, and regulated industries, because Huawei's brand asset of being able to deliver, take responsibility, and execute systems engineering is strong; in the global high-end AI training market, this brand advantage is far smaller than NVIDIA's. Cost advantage: insufficient evidence, so I do not classify it as a strong moat. Ascend may approach or locally optimize system-level cost per token through supernode interconnection and full-machine engineering, but public evidence is not enough to support a clear cost advantage in single-chip performance, mature manufacturing, and yield. Scale advantage: medium to strong, because of developers, partner count, deployment scale, and Huawei Group's sales system. Network effects: medium. The Ascend/CANN chain of "developers, partners, solutions, and customers" has formed positive feedback, but its strength still cannot match CUDA's global developer lock-in. Switching costs: medium to strong, especially for large Chinese customers that have already migrated models, toolchains, and supply chains to domestic stacks. Channel advantage: strong. Huawei already has a massive sales and delivery network across carriers, government and enterprise, cloud, energy, power, and manufacturing. Patents, licenses, and regulatory barriers: strong, especially under domestic compliance, security controllability, and policy direction. Data advantage: medium, reflected more in the combination of cloud and industry solutions than in the chip business alone. Corporate culture and operating capability: strong, with extremely high R&D density and long-term investment. Capital allocation capability: medium; the direction is rational, but results still need a longer period of verification.

In aggregate, I score "moat strength" at 4/5, but this 4 must carry a geographic qualifier: it is a 4 inside China's locally controllable computing ecosystem, and should be discounted in the unconstrained global market. Put differently, the real boundary of this moat is not "the global first tier of AI chips," but "one of the primary options for China's large-scale AI infrastructure under geopolitics and domestic substitution." This moat has generally been widening in recent years because developers, partners, supernode deployments, and cloud absorption have all been expanding together. It is not yet wide enough to ignore hard constraints such as process technology, capacity, performance, and application migration costs.

How long and how much capital would competitors need to replicate it? If they only replicate a single point product, the time may be 2 to 4 years and the capital may be several billion to more than RMB 10 billion. If they have to replicate the full system of "chips + full machines + supernodes + software stack + cloud + industry solutions + sales and delivery," the difficulty rises sharply. On one hand, Huawei's 2025 R&D spending of RMB 192.3 billion and cumulative RMB 1.382 trillion over the past ten years show that the group is continuously feeding this system at the group level. On the other hand, after a developer and partner ecosystem forms, what must be replicated is not only technology, but also training, migration, commercialization, and delivery organization. For many chip-design companies, that threshold is not realistic.

As for whether Ascend can raise prices in an inflationary environment, my answer is more conservative than many optimistic views. Ascend is not a luxury good or a strong consumer brand. It faces large-customer procurement and computing-investment decisions, where customers watch performance, supply, and total cost of ownership rather than list price. Ascend is more likely to preserve bargaining power through performance, delivery certainty, and integration efficiency than through pure price increases like Moutai. It has some pricing power, but not strong pricing power in the classic sense.

Can it remain profitable in an economic downturn? Looking at Huawei Group, the answer leans yes. In 2025, Huawei had operating profit of RMB 96.937 billion, an operating margin of 11%, a liability ratio of 55%, cash and short-term investments of RMB 361.426 billion, and total borrowings of RMB 239.284 billion, implying roughly RMB 122.1 billion of net cash cushion. Looking only at Ascend stand-alone, I do not have enough evidence to make that conclusion, because its separate statements are not disclosed, and public domestic AI-chip comparable Cambricon still had negative operating cash flow even after turning profitable in 2025. In other words, the group can absorb pressure, but that does not mean the stand-alone business has already become a cash machine that can cross cycles.

Now consider management and capital allocation. Huawei's equity structure is quite unusual. As of year-end 2025, Huawei Investment & Holding was owned by 169,054 employees and retired beneficiaries, while Ren Zhengfei's ownership was close to 0.59%. This means it was never a governance framework in which a listed company answers to outside minority shareholders. It is closer to an "employee community + management's long-term investment" framework. The advantage is very strong strategic persistence, with R&D investment less likely to be held hostage by quarterly results. The drawback is that outside investors do not have a standardized shareholder oversight and price-discovery mechanism.

My basic assessment of management is: integrity and long-term orientation broadly pass; verifiability for outside shareholders does not. Why broadly pass? Huawei continued to invest in R&D at high intensity over the past five years, and after sanctions it still strengthened chips, operating systems, manufacturing tools, and the developer ecosystem. Capital allocation is clearly tilted toward long-term survival and technological autonomy rather than short-term financial cosmetology. The parent company still planned to distribute RMB 77.095 billion of dividends to employee shareholders in 2024, which also shows cash return from operating results. Why does outside-shareholder verifiability fail? Because you cannot examine it the way you would a listed company: whether buybacks occurred when shares were undervalued, whether equity incentives diluted holders, whether acquisitions created per-share value, or whether management fully reviewed mistakes externally. For a private company, these questions are institutionally hard to verify from the outside.

I therefore score "management and capital allocation" at 3/5. The direction is right and execution is strong, but from the perspective of a conservative long-term value investor like you, insufficient transparency is itself a deduction.

Financial Quality and Owner Earnings

The scope must be clear here. HiSilicon/Ascend does not independently disclose 5 to 10 years of financial statements, so the core table below uses Huawei Group's consolidated statements to observe whether the platform behind it has the ability to keep funding and absorb pressure. For the cash-flow and working-capital realities of the AI-chip industry, I also add one public comparable, Cambricon, as a side reference. Any data that cannot be broken down to Ascend stand-alone is explicitly labeled.

The table below comes from Huawei's five-year financial summary in its 2025 annual report, on an IFRS consolidated basis. Revenue growth, net margin, CFO/net profit, approximate ROE, approximate ROA, and approximate net cash are rough calculations I made from the annual-report table.

This table gives several important conclusions. First, Huawei Group's 2021 to 2025 revenue CAGR was about 8.5%. It was not a straight upward line, but it has recovered from the trough for four consecutive years. Second, accounting profit quality improved meaningfully in 2024 and 2025, because operating cash flow exceeded net profit for two consecutive years, and 2025 CFO/net profit reached a rough 1.87x. Third, the balance sheet is generally sound. The 2025 liability ratio was 55%, and cash plus short-term investments exceeded total borrowings, so there is no obvious financial-leverage pressure at the group level. Fourth, these conclusions only show that "the Huawei platform can absorb pressure, invest, and provide funding." They do not mean Ascend stand-alone has proven itself to be a high-ROIC, high-FCF independent good business.

If we further break down the group structure, Huawei's 2025 revenue included RMB 375.014 billion from ICT Infrastructure, RMB 344.473 billion from Consumer, RMB 32.161 billion from Cloud Computing, RMB 77.312 billion from Digital Power, and RMB 45.018 billion from Intelligent Automotive Solution. Huawei also noted that its 2025 "cloud computing business scope, including revenue inside other Huawei segments" reached RMB 72.075 billion. These numbers show that Ascend is more likely embedded inside ICT Infrastructure, Cloud, and some industry solutions as a computing base. Its value is large, but its financial aggregation is not separately visible.

For many financial metrics you care about, I must answer in three categories. Metrics that can be roughly viewed with group data: revenue growth, operating margin, net margin, operating cash flow, ROE, ROA, liability ratio, and approximate net cash, all shown in the table above. Metrics that can be judged at the group level but have limited meaning for Ascend: dividends and buybacks. Huawei pays cash dividends to employee shareholders, but there is no public-share buyback logic for minority shareholders to evaluate. Metrics that cannot be reliably answered: HiSilicon/Ascend stand-alone free cash flow, free-cash-flow conversion, ROIC, net debt/EBITDA, interest coverage, share-count changes, Ascend stand-alone inventory/receivables/payables changes, and maintenance capital expenditure. For these, my answer is: unknown; segment financials are required.

To avoid misjudging Ascend's economics by only looking at Huawei's platform, public comparable Cambricon is worth a look. Cambricon's 2025 revenue was RMB 6.497 billion, up 453.21% year over year; net profit attributable to the parent was RMB 2.059 billion; R&D spending was RMB 1.169 billion. On the surface, this looks very bright. But the same annual report also shows 2025 net operating cash flow was still RMB -498 million, net investing cash flow was RMB -4.530 billion, sales to the top five customers accounted for 88.66% of annual sales, purchases from the largest supplier accounted for 55.34% of annual procurement, inventory rose 178.67% year over year, and accounts receivable rose 120.17%. This is very important because it reminds us that China's domestic AI-chip industry can grow rapidly and report high profits, while still having cash-flow lag, customer concentration, rising working capital, and heavy dependence on a few upstream and downstream counterparties.

So on whether profits are real cash profits or accounting profits, my judgment is: at the Huawei Group level, 2024 and 2025 profit quality was acceptable, with operating cash flow at least supporting reported profit; but Ascend stand-alone profit quality is unknown, and public industry comparables show that AI-chip profits do not naturally equal cash profits. On whether growth requires large capital investment, the answer is clear: yes. Huawei Group's RMB 192.3 billion R&D spending and Cambricon's continued high-intensity R&D plus negative operating cash flow both show this is not a model that becomes lighter as it grows. It is a model that requires continuous investment as it grows.

On accounting risk, I have not seen direct signals of financial fraud or major audit qualifications in public materials. Huawei's annual report states that its consolidated financial statements are prepared under IFRS. Cambricon's 2025 annual report received a standard unqualified opinion from Pan-China. The real risk is not "obvious signs of fraud," but the boundary of disclosure: for outside investors, the opacity of Ascend's statements is itself the largest accounting and governance risk, because you cannot perform segment-level cash-flow attribution.

Next is Owner Earnings. Strictly following Buffett's approach, owner earnings should start from net income, add back non-cash charges such as depreciation and amortization, then deduct maintenance capital expenditure and necessary working-capital additions. For HiSilicon/Ascend stand-alone, this model cannot be implemented, because revenue, profit, depreciation, maintenance capex, working capital, and taxes are not independently disclosed. For Huawei Group, one can barely make an extremely conservative lower-bound estimate: directly treat "operating cash flow minus net investing cash outflow" as a lower-bound owner earnings figure. Under this very conservative basis, which clearly mixes in strategic investments, Huawei's 2025 figure was about RMB 52.59 billion, 2024 about RMB 38.66 billion, and 2023 RMB -28.95 billion. I emphasize that this is not strict free cash flow, because investing activities include more than maintenance capital expenditure. It is only a lower-bound anchor showing that if most investment is treated as necessary, distributable cash would be compressed very low.

Therefore, the most important conclusion in this section is not a precise number, but a disciplined conclusion: a reliable Owner Earnings valuation cannot currently be performed for Huawei HiSilicon Ascend. We can only confirm that the platform behind it has cash flow and R&D capability; we cannot confirm that Ascend stand-alone already has long-term stable, distributable, high-quality cash flow.

Valuation and Margin of Safety

I follow the three methods you requested in this section, but frankly: none of the three can produce a high-confidence numerical range for Ascend under currently available, externally verifiable information. This is not an excuse. It is the most basic honesty in value investing: without verifiable inputs, do not pretend to have precise outputs.

Method one, discounted owner earnings. For listed companies, this is usually the preferred method; for Ascend today, it is unusable. There are three reasons: no independent revenue, no independent profit, and no independent capital-expenditure or working-capital data. Even if I forced a model, it would be built on guesses about revenue, gross margin, capex, and R&D capitalization, followed by guesses about growth and discount rates. The result would only create an illusion that looks rigorous but cannot be verified. I therefore will not give you a fake DCF result. If Ascend is independently listed in the future and discloses positive owner earnings for more than three consecutive years, I would then model it using conservative growth of 5% to 10%, a discount rate of 10% to 12%, and terminal growth of 2% to 3%. Current conclusion: not executable.

Method two, relative valuation. This method is actually more educational today, because it tells you that the comparables themselves show why this sector is hard to invest in by saying "peers are expensive, so this should also be expensive." Domestic pure-play public AI-chip comparable Cambricon had a delayed quote of about RMB 1,250.36 per share as of early June 2026; Reuters/LSEG pages showed median-type metrics of roughly 289x PE, 94.99x PS, and 61.04x PB. This is a typical "high expectations, low tolerance for error" valuation. On the other side, global leader NVIDIA's current share price was about USD 208.64, with PE of about 31.8x. But NVIDIA's fiscal 2026 revenue had already reached USD 215.938 billion, with data center revenue of USD 193.7 billion and GAAP gross margin of 71.1%. It is a mature leader that has already delivered scale, profit, and cash flow. Putting the two together, you can see that comparable valuations in this industry are extremely split, depending on whether you are valuing an already proven global oligopolist or a domestic-substitution company still priced around theme and runway. For an unlisted, non-carved-out business like Ascend, neither side's multiples can be directly transplanted.

This leads to a clear disciplinary judgment: if Ascend is independently listed in the future and the market prices it like Cambricon, with high PS/PB and multiple years of prepaid growth, I would not touch it. That would be expectation and policy trading, not value investing. Conversely, I would only begin to look seriously if it could disclose stable owner earnings, reduce customer concentration, improve operating cash flow, and trade at least 30% to 40% below conservative intrinsic value.

Method three, asset or liquidation value. For traditional manufacturing, real estate, or financial-asset holding companies, this method is often useful. For a business like Ascend, whose value mainly resides in IP, R&D teams, software stacks, ecosystem, customer validation, and systems-solution capability, book assets likely both understate real value and cannot be directly realized as liquidation value. More fatally, Ascend has no separate balance sheet, so even book net assets cannot be verified. In other words, the conclusion of this method here is not "undervalued," but "not applicable".

Therefore, I must answer the valuation outputs you requested according to reality: conservative intrinsic value range: cannot be reliably estimated; reasonable intrinsic value range: cannot be reliably estimated; optimistic intrinsic value range: cannot be reliably estimated; current price discount or premium to intrinsic value: not applicable; required margin of safety: currently unverifiable; ideal buy-price range: not applicable; acceptable hold-price range: not applicable; clearly overvalued price range: not applicable. These answers may not feel satisfying, but this is where qualified value-investing analysis is more useful than inventing a precise number.

Still, margin of safety can be answered from a methodological requirement perspective. If Ascend becomes tradable equity in the future, I would require at least three conditions at the same time. First, three consecutive years of positive operating cash flow and positive, or nearly positive, free cash flow on stand-alone statements. Second, customer concentration materially better than the extreme state of current domestic comparables, meaning the top five customers cannot consume more than half of revenue. Third, the buy valuation should be no higher than 70% of conservative owner-earnings value, or no higher than 15 to 20 times sustainable owner earnings. Before those three conditions are met, discussing "margin of safety" is only wordplay.

If you want a more direct answer to the seven margin-of-safety questions, my conclusions are as follows. First, whether the current price is cheap enough: impossible to judge, because there is no public price. Second, the most fragile assumption is not demand, but whether demand can steadily turn into mature, replicable, distributable economic profit. Third, if growth falls short of expectations, Ascend may still exist at a strategic level, but capital returns could deteriorate quickly. Fourth, if margins decline, Huawei Group may still endure, but Ascend's stand-alone investment case could fail outright. Fifth, if valuation multiples compress after a future listing, permanent loss is entirely possible; high-expectation peers like Cambricon are the warning example. Sixth, a "good company, bad price" outcome is highly likely because the AI sector is often thematic first and tested by cash flow later. Seventh, the most rational action today is not to chase a "price without a price," but to wait for better verifiability.

The conclusion is very clear: the current margin of safety is insufficient, not because the price is definitely high, but because price and value cannot be reliably connected from an outside shareholder's perspective.

Risks, Bear Case, and Opportunity Cost

If permanent capital loss is placed first, I would rank the risks of Huawei HiSilicon Ascend as follows. First is availability risk: can you own its economic interest in a compliant, transparent, and exit-able way? Today, the answer is basically no. Second is disclosure risk: without stand-alone statements, outside investors cannot identify whether growth comes from real profitability, internal group resource transfer, or a temporary policy and supply-demand gap. Third is supply-chain and technology risk: advanced processes, packaging, materials, and capacity remain constrained, and U.S. officials' views on production ceilings show that outsiders still have substantive concerns about manufacturing delivery. Fourth is customer and industry risk: large-customer dominance, project-based delivery, government and enterprise procurement, and cloud resource allocation all imply revenue concentration and timing volatility. Fifth is valuation risk: if the business is listed in the future and the market prices it feverishly as a "core domestic AI asset," even a good business can become a bad investment.

More specifically, competitive risk has not disappeared because of domestic substitution. NVIDIA remains the world's strongest AI infrastructure company, with fiscal 2026 data center revenue of USD 193.7 billion and companywide revenue of USD 215.9 billion. What makes it formidable is not only chip performance, but also development tools, framework adaptation, customer habits, and software lock-in. China's market window exists more because of export restrictions plus demand for local controllability. It should not be simplified into "global competition has reversed."

Technology substitution risk is also real. AI computing is not a single-product market with a settled end state. It is a market in which GPUs, custom ASICs, training-inference separation, and system-level interconnection paths keep changing. Huawei's current approach clearly leans toward "full-stack systems engineering + supernode interconnection + cloud absorption." That is an advantage, and also a path dependence. If customers regain access to a more mature global technology stack in the future, or if other domestic players achieve lower cost and faster iteration in some scenarios, Ascend's local advantages may be eroded.

Regulatory and policy risks are two sides of the same coin. On one hand, policy encouragement for localization supports demand for Ascend. On the other hand, that demand also means the business is heavily affected by policy cadence, procurement standards, and industry compliance requirements. Reuters reported at the end of 2025 that China had required new data-center projects receiving state funding to prioritize domestic AI chips. This policy tilt is certainly positive for Huawei, but this demand attribute means the business model is not fully spontaneous and market-driven. It carries exogenous rules. Exogenous variables can build a business, and they can also change its pace.

Financial leverage risk is low at the Huawei Group level. In 2025, Huawei had cash and short-term investments of RMB 361.426 billion, total borrowings of RMB 239.284 billion, approximate net cash of RMB 122.1 billion, and a liability ratio of 55%. The safety cushion is not thin. The real issue is not balance-sheet leverage, but whether continued ultra-high R&D and computing-infrastructure investment can form clearer capital returns over the next few years.

The strongest bear case is this: Ascend may be a strategically important business, but not necessarily one that can deliver high long-term returns to outside shareholders. The reason is that the AI-chip industry naturally requires high investment, iterates quickly, and overlays demand cycles with policy cycles. Even if demand is very strong, process technology, capacity, customer concentration, price competition, and ecosystem migration costs may keep profits and cash flow unstable for a long time. Public comparable Cambricon gave a clear signal in 2025: explosive growth in revenue and profit does not prevent operating cash flow from remaining negative and customer concentration from staying extremely high. Put differently, bears are not really bearish on "AI computing demand." They are bearish on whether that demand can be returned to outside shareholders through healthy cash flow.

What facts would make me admit the judgment was wrong? From a stance of "bullish on the business but avoiding the investment," the following facts would make me upgrade the assessment: first, Ascend stand-alone disclosures begin and show positive free cash flow for three consecutive years; second, the customer structure diversifies and reliance on the top five customers falls meaningfully; third, developers, partners, supernode deployments, and Huawei Cloud AICS customers continue to rise steadily; fourth, supply-chain delivery improves and the business is no longer constrained for a long time by production-ceiling debates. Conversely, the following facts would make me downgrade the business assessment: first, developer and partner ecosystems stagnate; second, Atlas 900/CloudMatrix deployment expansion slows significantly; third, domestic market share declines; fourth, group-level operating cash flow weakens while R&D/capital investment stays high; fifth, key large customers clearly return to overseas stacks across the board.

The conclusion becomes clearer when we compare it with other opportunities. Compared with NVIDIA, the strongest competitor in the same industry, Ascend's industrial position is "strong domestic strategic substitute," while NVIDIA is "a global economic oligopolist that has already delivered." The former has high business value but is not investable; the latter has a visible daily price, complete financial statements, and verified economics. Compared with domestic public proxy Cambricon, Cambricon gives you tradability, but also extremely high valuation, extremely high expectations, and customer-concentration risk. Compared with CSI 300, the latter had a trailing PE of about 14.23x as of 2026-06-08; it may not be very cheap, but it at least gives you public liquidity, broad diversification, and transparent statements. Compared with China's 10-year government bond, the latter's current yield was roughly around 1.7%; the return is low, but measurable, comparable, and holdable. For a balanced but conservative investor, spending current capital to "study a good business that cannot be bought or valued" carries a fairly high opportunity cost.

So here are direct answers to your four comparison questions. Is buying it clearly better than buying an index? Not today, because you cannot buy transparent, value-able equity in it at all. Is the expected return enough to compensate for the risk? This cannot currently be proven. Is it worth occupying capital? For ordinary public-market capital, no. If you could hold only 5 assets, does it deserve a place in the portfolio? As a "business research object," yes; as a "currently executable asset," no.

Investment Checklist and Final Conclusion

First, here is the Checklist you requested. The judgments strictly distinguish between the business itself and current investability.

【Final Rating】 Avoid

【One-Sentence Investment Thesis】 Ascend is an AI infrastructure business with high strategic value and a strengthening domestic moat in China, but it is not currently a value-investing target that outside investors can verify, value, or buy.

【Core Bull Case】

• Huawei has built Ascend into a full-stack computing base across "chips, servers, supernodes, software stack, cloud services, and industry solutions," rather than a single-point chip business.

• China's domestic AI accelerator market has a real domestic-substitution window, and Huawei is near the front of the first tier among domestic vendors.

• Developers, partners, industry solutions, and supernode deployments are all expanding, showing that the ecosystem is not just a paper story.

• Huawei Group has a sound financial base, with RMB 127.384 billion of operating cash flow in 2025 and cash plus short-term investments above borrowings, giving it long-term investment capacity.

• HiSilicon/Ascend benefits from Huawei's channel, delivery system, and industry relationships, which matter especially in government, enterprise, and large-customer markets.

【Core Bear Case】

• It is not an independently listed company, has no public price, and cannot support a margin-of-safety calculation.

• HiSilicon/Ascend has no separate financial statements, so outsiders cannot verify real free cash flow or return on capital.

• The industry is capital intensive, supply-chain constrained, and fast evolving. It is not naturally a high-cash-flow business.

• Domestic public comparables show that even with high growth and high profits, the AI-chip industry can still have customer concentration and negative operating cash flow.

• If it is independently listed in the future and priced by the market at high multiples as a "core domestic AI asset," a "good company, bad price" outcome is highly likely.

【Key Assumptions】

• Domestic substitution demand continues for many years, rather than being a short-term policy pulse.

• Ascend's software ecosystem keeps expanding, with continued growth in developers, partners, solutions, and cloud customers.

• Supply-chain delivery capability continues to improve and is not trapped for a long time by capacity or yield bottlenecks.

• Huawei Group remains willing to support this business with group resources, rather than treating it as a short-term tactical project.

【Fair Buy Price】 There is currently no applicable public buy price. If Ascend is independently listed in the future, I would only consider it under the following conditions: three consecutive years of positive free cash flow, significantly improved customer concentration, and a purchase valuation no higher than 70% of conservative owner-earnings value, or no higher than 15 to 20 times sustainable owner earnings. This price framework is a methodology, not a current quote.

【Target Holding Period】 Only if it becomes an investable target in the future would it qualify for discussion of a 10-year-plus holding period. Until then, it is more suitable for a "long-term watchlist" than a "buy list."

【Expected Annualized Return】 Currently impossible to estimate. This is not because the business lacks prospects, but because there is no public buy price, no independent financials, and no boundary for outside minority shareholder rights. Any specific annualized return number would be false precision.

【Maximum Loss Risk】 For ordinary public-market investors, there is currently no standardized buyable path, so "how much could be lost after buying" is temporarily not applicable. If one participates at a high valuation in any hype-driven vehicle related to this business in the future, the worst case could be a permanent loss of more than 50% from valuation mean reversion. If one obtains exposure through opaque, non-standardized channels, the worst case could approach a full principal loss in liquidity and valuation terms. This is the core reason I insist on "Avoid."

【Tracking Indicators】

• Whether the number of Ascend developers continues to grow above 4 million.

• Whether Kunpeng/Ascend partner count and industry-solution count continue to rise.

• Whether Atlas 900/CloudMatrix deployments continue to expand.

• Whether Huawei Cloud AICS customer count continues to exceed and expand from the 1,800-customer base.

• Whether Huawei ICT Infrastructure and Cloud revenue growth improves.

• Whether Huawei Group operating cash flow continues to exceed net profit.

• Changes in domestic-vendor share and Huawei share within China's AI accelerator market.

• Whether public signals related to supply-chain constraints and production ceilings improve.

• Whether domestic public comparables move from "profit growth" to "cash-flow growth."

【Signals That Trigger Reassessment】

• Huawei/HiSilicon begins to disclose Ascend financials separately.

• Key ecosystem metrics stagnate or decline.

• Supply-chain bottlenecks fail to improve for a long time, and mass-production capability falls short of expectations.

• Large customers clearly shift back to overseas technology stacks.

• If Ascend is independently listed in the future, valuation materially detaches from cash-flow capacity.

【Final Recommendation】 Calmly viewed, Ascend is a business worthy of respect, study, and long-term tracking. But under Buffett-style value-investing discipline, what one should truly buy is not "an exciting story," but "cash flow you can calculate, hold, and buy at the right price." For Huawei HiSilicon Ascend today, the first two conditions partly hold, while the last one, price and margin of safety, does not. My recommendation is not to chase, but to keep observing and wait for verifiability before discussing valuation and buying.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.