NVIDIA Investment Memo

Cost structure and dependencies. NVIDIA runs a classic fabless model, but that does not mean "easy." Core manufacturing depends on TSMC and Samsung, HBM depends on SK hynix, Micron and Samsung, advanced packaging depends on CoWoS, and assembly and test depend on external partners such as Foxconn, Wistron and Fabrinet, leaving the supply chain heavily concentrated in Asia. At the end of FY2026 inventory was 21.403 billion dollars, and PwC listed "valuation of inventory write-downs and purchase obligations" as a key audit matter; at that time inventory plus outstanding inventory purchase, long-term supply and capacity obligations totaled 95.2 billion dollars. By FY2027 Q1 the company further disclosed 30 billion dollars of multi-year cloud-service commitments and 6 billion dollars of other supplier commitments; in addition, at the end of FY2026 the company expected to enter into roughly 22.7 billion dollars of future lease obligations during FY2027 to FY2030, mainly data center leases to support R&D. In other words, although capex as a share of revenue on the books is not high, the true capital being tied up is rising rapidly through inventory, prepayments, capacity lock-ins, cloud-compute contracts and lease commitments.

Customer, channel and key-person dependence. In FY2026 the company's two direct customers each accounted for 22% and 14% of total revenue; at period end, the top three direct customers represented 25%, 18% and 13% of accounts receivable. Compared with the FY2021 to FY2023 state of "no customer exceeding 10% of revenue," customer concentration has clearly risen. At the same time, the company depends heavily on Jensen Huang's product judgment, ecosystem drive and partner relationships. My view is: this business is now clear enough to understand; but it is not a "simple business," it is a business that is "clear in its model, complex in its technical execution." If the stock market were closed for 5 years and the only question were "do I want to own this company," my answer is yes; if you add "buying it new at today's price," my answer is in no hurry to buy.

Business understandability score: 4/5. The high-level logic is very clear: sell the compute platform, sell the ecosystem, sell the total-cost-of-ownership advantage; but the underlying cadence, competitive substitution and accounting volatility are far more complex than the "slow-variable businesses" that traditional value investing loves most.

Industry Landscape and Moat

What stage is the industry in? The AI server and accelerated-computing industry is still in a phase of rapid expansion. In public materials in April 2026, IDC stated that the global server market grew 52.4% year over year in the fourth quarter of 2025, driven by large-scale deployment of GPU servers, and cited a "five-year compound growth rate" of 24.1%; another IDC document showed that full-year 2025 server market revenue reached a record 444.1 billion dollars, up 80.4% year over year, with GPU server revenue rising 59.1% year over year in the fourth quarter and accounting for more than half of total server market revenue. Demand is undoubtedly strong, but this is not a stable, evenly growing industry; it is a high-growth, high-volatility, strongly cyclical capex industry shaped jointly by politics, technology, supply and power constraints.

How to view the competitive landscape? On the most direct GPU product competition, AMD remains the closest "head-on rival"; on the strongest substitute in an investment sense, Broadcom represents the route of large customers moving toward custom ASICs; add Google TPU, Amazon Trainium, Microsoft Maia and other in-house chips, and NVIDIA already faces not merely "another company selling GPUs" but customers themselves rebuilding the hardware stack. In May 2026 Reuters noted that the inference market is larger and more competitive than the training market, with AMD, Intel, large cloud providers and in-house chips all vying for it; in March of the same year, Reuters reported that Broadcom expected AI chip sales to exceed 100 billion dollars in 2027, underscoring that custom processors are posing a structural threat to NVIDIA's dominance. Even in May 2026, Reuters reported that Anthropic was discussing renting servers based on Microsoft's in-house chips, showing that "buy only NVIDIA GPUs" is not the sole answer for the future.

Where exactly is NVIDIA's moat? I think the moat is not merely "fast chips" but rather a complete platform, a deep ecosystem, fast delivery, thick software and strong networking coordination. The company repeatedly emphasizes in its 10-K that its data center platform is a co-design of GPU, CPU, DPU, NVLink, switches and the software stack; customers do not buy a card, they buy a system scaling to hundreds of thousands of nodes. More crucial is software: official materials say CUDA has crossed twenty years, and as of GTC 2026 the corresponding developer base exceeded 6 million; the official CUDA-X page discloses that more than one million developers use CUDA-X. This means NVIDIA's core barrier is not just the performance of one chip generation but a composite network of developer toolchains, framework compatibility, libraries, models, reference designs, partners and talent supply. Put differently, when a user switches to another platform, the cost is not only chip replacement but also engineering migration, model retuning, operations rebuild and talent retraining.

Moat checklist judgment. Brand advantage: yes, and in AI infrastructure procurement it is already one of the de facto standards. Cost advantage: yes, but not the lowest manufacturing cost; rather a lower total cost of ownership. Scale advantage: extremely strong, reflected in R&D, supply lock-in, ecosystem and customer coverage. Network effects: yes, mainly from the positive feedback among developers, frameworks, applications, hardware and cloud platforms. Switching costs: high, especially for large training and production-grade inference stacks. Channel advantage: fairly strong, with global coverage through OEM/ODM/cloud/system integrators. Patents and technical barriers: strong, but systems engineering and the software stack matter more. Data advantage: moderate, less important than the platform/ecosystem advantage. Corporate culture and operational capability: very strong, shown in the one-generation-per-year product cadence and the execution to scale GPUs into racks and networks. Capital allocation capability: broadly good, but not flawless. Most of the judgments above have a factual basis, but the "ranking of strength" is my inference.

Is the moat widening, stable or narrowing? My judgment is: on the whole still widening, but wear is starting to appear at the edges. Widening is reflected in the deepening of the software stack, networking, complete systems and developer ecosystem; wear is reflected in hyperscale customers trying to bypass general-purpose GPUs and turn to custom ASICs more tailored to their own workloads, while under export controls the company has itself defined the China market as "effectively hard to compete in." The FY2026 10-K even states explicitly that by the end of FY2026 the company was "effectively excluded" from China's data center compute market; the same document discloses that H20 export restrictions once caused 4.5 billion dollars of inventory- and purchase-obligation-related charges. Combine this with IDC data reported by Reuters in April 2026 -- domestic Chinese vendors had already taken roughly 41% of the local AI accelerated-server market -- and you can see the moat is not expanding without cost.

Several key judgments. For a competitor to replicate NVIDIA, it is not enough to "make a faster chip"; it must replicate the hardware-software coordination, developer ecosystem, channel relationships and supply chain, which usually takes years and enormous capital. But large customers do not necessarily need to "fully replicate NVIDIA"; they only need to build a good-enough substitute on a specific workload to potentially erode NVIDIA's fattest pool of profit. In an inflationary environment the company has very strong pricing power, as evidenced by gross margins around 70% or higher over the past three years and the FY2027 Q1 GAAP gross margin of 74.9%; nor is the company fragile in a downturn -- amid the FY2023 gaming-business adjustment headwind it still delivered 26.974 billion dollars of revenue, 4.368 billion dollars of net income and 5.641 billion dollars of operating cash flow -- but if AI capex enters a digestion period in the future, profits could still swing sharply. The high margins of the past come partly from a structural moat and partly, clearly, from the cyclical dividend of AI demand outstripping supply.

Industry attractiveness score: 4/5. Moat strength score: 4.5/5.

Management and Capital Allocation

Is management trustworthy? On governance and incentive alignment, I lean toward a high rating. Jensen Huang is the founder and has served as CEO since 1993; as of March 23, 2026 he beneficially owned roughly 870.6 million shares, about 3.58% of the company. The company sets share-ownership requirements for both executives and directors, and the disclosure notes that Huang and the other NEOs have all met the relevant requirements; the company also prohibits executives and employees from hedging, pledging and margin-trading the stock, and has a performance-pay clawback mechanism. These arrangements cannot guarantee that every decision is perfect, but they at least show that management and shareholder interests are broadly aligned, rather than the classic agency problem of "asset-light, high pay, low ownership, cashing out by telling stories."

Is management long-term oriented? Based on public disclosure, yes. The FY2026 proxy discloses that 96% of the CEO's target compensation is tied to company performance, and about 48% of other NEOs' target compensation is performance-linked; the company emphasizes using annual and multi-year performance targets to determine the number of PSUs earned, requires multi-year continued service, and offers no aggressive arrangements such as special change-of-control terms, tax gross-ups or automatic accelerated vesting for executives. Huang's FY2026 total compensation was 36.34 million dollars, which looks far from low, but is not outrageous for a company in the 5-trillion-dollar class; what is genuinely more interesting is that the company places the incentive emphasis on long-term equity and performance rather than on cash bonuses alone.

How well is capital allocated? This part must be taken apart. On reinvestment, over the past three years management has poured nearly all operating momentum into the product roadmap, supply-chain lock-in, networking, systems and software, objectively creating staggering increments of revenue and cash flow; by results, highly successful. On M&A, the most successful in the past was the earlier Mellanox, which moved NVIDIA from a GPU company toward a networking/systems company; recent equity investments also show aggressiveness, but their market value is not easy to assess. On buybacks and dividends, in FY2026 the company repurchased roughly 40.388 billion dollars of stock and paid 0.974 billion dollars in dividends; in FY2027 Q1 it bought back roughly 20.170 billion dollars more, while the board added an 80 billion dollar buyback authorization and raised the quarterly dividend from 0.01 dollars per share to 0.25 dollars. The catch is that these buybacks largely occurred at high valuations and cannot be called the textbook "counter-cyclical, buy heavily when undervalued" capital allocation. It looks more like: under strong cash flow, returning a large amount of cash to shareholders and steadily shrinking the share count. For long-standing shareholders this is not necessarily bad, but for a strict value investor it is not optimal either.

Is share dilution out of control? I think not. At the end of FY2026 shares issued and outstanding were 24.304 billion, below the 24.477 billion at the end of FY2025; by the end of FY2027 Q1 they fell further to 24.221 billion. That is, although the company has considerable SBC -- 6.386 billion dollars in FY2026, about 6.840 billion dollars TTM -- the scale of buybacks far exceeds the dilution, so per-share ownership is ultimately rising rather than diluting. Still, SBC is a real shareholder cost, so in my Owner Earnings estimate I treat it as an economic cost rather than simply ignoring it because it is "non-cash."

Does management candidly discuss mistakes and risks? Relatively candidly. In the 10-K the company does not dodge the 4.5 billion dollar H20 hit, states outright that it is "effectively excluded" from the China data center market, and repeatedly flags in its risk factors the faster product cadence, export restrictions, supply-demand mismatch, customer design shift-out and overseas competition. I take this as a positive signal: it is not "only talking up the TAM and never the cost of being wrong." Of course, when the competitive position is too strong, any management is more likely to look "honest and correct"; the real test will come when the industry enters its next trough.

Management and capital allocation score: 4/5. I withhold a 5 mainly for two reasons: first, the buybacks at the current stage are not obviously occurring at undervalued levels; second, the company is still in a "super tailwind period," and the reverse test of its capital-allocation ability is not yet sufficient.

Financial Quality and Owner Earnings

The conclusion first: NVIDIA's profit quality is generally very high and its cash-flow capability very strong, and in the headwind year of FY2023 it also proved it can earn money outside of a boom tailwind. But at the same time, the most recent quarter's GAAP net income was significantly lifted by investment gains, and the headline PE understates the true operating valuation. In FY2027 Q1, revenue was 81.615 billion dollars, operating income 53.536 billion dollars and net income 58.321 billion dollars, of which "other income (expense), net" was as high as 15.929 billion dollars; the cash flow statement also shows 15.936 billion dollars of equity-securities gain adjustments for the quarter. In other words, a large chunk of the quarter's net income is not cash profit created by the core operations but investment fair-value gains. A long-term owner cannot treat that part as a repeatable operating capability.

Key Financial Metrics Table

Unit: billion dollars; TTM is estimated as FY2026 minus Q1 FY26 plus Q1 FY27; free cash flow is operating cash flow minus purchases of property, equipment and intangibles. Figures for 2020 to 2023 are taken from the FY2022/FY2023 10-K, 2024 to 2026 from the FY2026 10-K, and the FY2027 Q1 data used for TTM are taken from the 10-Q for the quarter ended April 26, 2026.

Now to quality. From FY2024 to FY2026, revenue grew from 60.9 billion dollars to 215.9 billion dollars, operating cash flow from 28.1 billion dollars to 102.7 billion dollars, and free cash flow from 27.0 billion dollars to 96.7 billion dollars; from the angle of "does growth get more cash-hungry the longer it runs," the answer is clearly no. The company earns more the more it grows and generates stronger cash flow the more it grows. Capex as a share of revenue was only about 2.8% in FY2026, looking very "light" on the surface; but if you also count inventory, receivables, prepayments, supply agreements, cloud-service commitments and lease commitments, it is really "light on accounting capex, not light on true operating capital tied up." In FY2026 accounts receivable rose from 23.065 billion dollars to 38.466 billion dollars, and inventory from 10.080 billion dollars to 21.403 billion dollars; by FY2027 Q1 the two rose further to 40.710 billion dollars and 25.797 billion dollars. It is still consuming working capital at high speed, which shows that current growth is not entirely the "no investment required" SaaS-style kind.

Margins and returns. The gross margins for FY2024, FY2025 and FY2026 were roughly 72.7%, 75.0% and 71.1%; operating margins about 54.1%, 62.4% and 60.4%; net margins about 48.8%, 55.8% and 55.6%. The FY2027 Q1 GAAP gross margin was 74.9% and operating margin 65.6%, but the net margin was as high as 71.5%, clearly lifted by non-operating gains. On the balance sheet, by FY2027 Q1 the company held 13.237 billion dollars of cash, 37.098 billion dollars of marketable debt securities, 30.237 billion dollars of marketable equity securities and 43.364 billion dollars of non-marketable securities, while short-term and long-term debt combined was only about 8.47 billion dollars, leaving an extremely strong net-cash/net-financial-asset position and an absurdly high interest coverage ratio -- FY2026 operating income alone covered interest expense more than 500 times over. Conclusion: this is not a high ROE pried up by leverage but a high return pried up by gross margin, turnover and extremely low capital intensity. Only, headline ROE/ROA is further flattered by investment gains, so in judging it one should rely more on operating cash flow and operating income.

Accounting quality and latent risks. I see no clear sign of financial fraud or aggressive revenue recognition: PwC gave an unqualified opinion on its financial statements and internal controls and considered internal control effective in all material respects. What deserves attention is not "whether there are signs of large-scale fraud" but two more realistic issues: first, fair-value changes raise the volatility of net income; second, the scale of inventory and purchase commitments is enormous, involving judgments about future demand, export controls and product life cycle, which is also the key audit matter the auditor named. For a long-term investor, this means: profit is broadly real but must be put through a de-noising process; cash flow is more credible than EPS.

Owner Earnings estimate. On a very conservative basis, I am unwilling to simply annualize FY2027 Q1 GAAP EPS. My method is: start from TTM operating cash flow of roughly 125.6 billion dollars, subtract all capex of roughly 6.57 billion dollars, then deduct TTM SBC of roughly 6.84 billion dollars as an economic cost to shareholders, do not treat the recent large investment fair-value gains as sustainable operating profit, and do not count the potential appreciation of the 30.2 billion dollars of marketable equity and 43.4 billion dollars of non-marketable securities in the operating valuation. On this logic, conservative TTM Owner Earnings is roughly in the 110 billion to 115 billion dollar range; against the current 5.36 trillion dollar market cap, that is about 47 to 49 times Owner Earnings. Without deducting SBC, looking only at TTM FCF, it is about 45 times. My view is: for a balanced, conservative-leaning long-term investor, one should understand the current valuation as closer to 48 times rather than 33 times.

Intrinsic Value and Margin of Safety

First the latest market pricing. NVDA's latest share price is about 219.98 dollars, corresponding to a market cap of roughly 5.36 trillion dollars. Over the same period the US 10-year Treasury yield is about 4.57%. Against the TTM free cash flow of roughly 119.1 billion dollars estimated above, the current FCF yield is only about 2.2%; against conservative Owner Earnings of roughly 112 billion dollars, the current Owner Earnings yield is also only about 2.1%. This means: by buying today, you are not buying "a very thick current cash return" but buying the premise that "for many years to come growth must stay high and the moat must not narrow obviously." That premise may well hold, but the very fact of it shows the margin of safety is not wide.

Method one: Owner Earnings discounting. The following is entirely a set of valuation assumptions, not facts. To be conservative, I start from Owner Earnings of roughly 110 billion dollars, do not treat investment gains as an operating norm, and do not separately add option value for the investment portfolio.

• Conservative scenario: 8% annualized growth over the next 5 years, 4% for the following 5 years, a 10% discount rate, and 3% terminal growth. This corresponds to an equity value of roughly 1.8 trillion to 2.0 trillion dollars, or about 75 to 85 dollars per share.

• Neutral scenario: 12% annualized growth over the next 5 years, 6% for the following 5 years, a 9% discount rate, and 3.5% terminal growth. This corresponds to about 115 to 135 dollars per share.

• Optimistic scenario: 16% annualized growth over the next 5 years, 8% for the following 5 years, an 8% discount rate, and 4% terminal growth. This corresponds to about 190 to 220 dollars per share. The implication of these results is direct: the current price already sits near the upper edge of the optimistic scenario. If you are a conservative value investor, this is not a price with "a cushion still in place."

Method two: relative valuation. On headline market data, NVDA's current PE is about 33.4 times, below the roughly 68 times headline PE of AMD and Broadcom; on a conservative estimate, NVDA's P/FCF is about 45 times, close to Broadcom's roughly 44 times and below AMD's roughly 50 to 60 times range; NVDA's PB is about 27 times, far above AMD's roughly 5 times, while Broadcom's PB is of weaker reference value owing to VMware-acquisition accounting and large intangibles. Broadcom's FY2025 free cash flow was about 26.9 billion dollars, and the market is more willing to assign it a higher multiple, partly because of its high proportion of software revenue, but it also carries markedly heavier debt; AMD, meanwhile, still lags NVIDIA in data center growth. Conclusion: NVIDIA is not the most expensive among its peers, but that cannot be read as "cheap," because the peer group as a whole sits in a high-valuation zone. At most, relative valuation shows that the market considers NVIDIA's quality high enough to trade alongside the most expensive batch of AI-infrastructure assets.

Method three: asset/liquidation value. For a company like NVIDIA, the liquidation method is hardly a core approach, because its value lies not in plant and equipment but in the software stack, developer ecosystem, customer relationships, chip architecture, system design and future cash flows. Even so, the latest 10-Q shows the company still holds roughly 50.3 billion dollars of cash and marketable debt securities, 30.2 billion dollars of marketable equity securities and 43.4 billion dollars of non-marketable securities, against interest-bearing debt of only about 8.5 billion dollars. That is, it has a layer of real financial-asset cushion; but adding all of it together accounts for only a small fraction of the current market cap. The conclusion the asset method gives is not "it has great liquidation protection" but "even if it is not cheap, it is not propping up its valuation with high leverage." What truly underpins today's price is still its cash-generating ability over the next decade.

Intrinsic value range and price bands. Based on the three methods above, the ranges I give are:

• Conservative intrinsic value range: 75 to 95 dollars per share

• Fair intrinsic value range: 110 to 140 dollars per share

• Optimistic intrinsic value range: 180 to 220 dollars per share

On this basis, the current price of about 220 dollars carries roughly a 57% to 100% premium to fair intrinsic value; the premium to conservative value is higher; against optimistic value it is close to "broadly fully priced." For a balanced, conservative-leaning investor, I think:

• Required margin of safety: at least 25% to 30% below fair value

• Ideal buy price range: 80 to 110 dollars

• Acceptable holding price range: 110 to 160 dollars

• Clearly overvalued price range: above 180 dollars, especially above 200 dollars

Here I deliberately separate "acceptable to hold" from "suitable to buy": for those already holding at low cost, I do not advocate getting off lightly; for new money today, I do not think the odds here are good enough.

Margin-of-safety judgment. The most fragile assumption in the current valuation is not whether next year's EPS can beat, but whether the high growth rate of AI capex over the next 5 to 10 years can persist through competition, regulation and technical substitution. If growth comes in below expectations while the company still keeps high margins, then returns are likely only "ordinary" and hard to make "excellent"; if margins fall from the current 70%+ gross margin to around 60% while the valuation multiple compresses from 45 times FCF to 25 to 30 times, then even if the business is still good, shareholders could experience a long stretch of low returns or even a permanent capital loss. Yes, this is a very textbook case of "great company, bad price" risk.

Risks, Comparisons and Final Verdict

The most important risks. First, competition risk: AMD keeps catching up on the general-purpose AI GPU route; Broadcom helps customers build custom ASICs; Google, Amazon, Microsoft and other cloud providers' in-house chips are more cost-attractive on the inference side. Second, technical substitution risk: NVIDIA remains dominant in the training market, but the inference market is larger and more fragmented, and if framework compatibility and development tools gradually weaken the CUDA lock-in, the profit pool may be eroded before revenue is. Third, regulatory and geopolitical risk: the company has stated explicitly that it is effectively excluded from China's data center market, and the H20 episode has already caused a 4.5 billion dollar loss. Fourth, customer concentration risk: the two largest direct customers accounted for 36% of FY2026 revenue, and the top three receivable customers for 56% of period-end receivables. Fifth, supply chain risk: heavy dependence on advanced process nodes, HBM and advanced packaging. Sixth, overvaluation risk: the market has already discounted a great deal of the company's next-decade growth. Seventh, accounting noise risk: net income is distorted by fair-value swings in equity investments. Eighth, operating-capital commitment risk: inventory, long-term purchases, cloud services and data center lease commitments are all expanding rapidly.

The strongest counterargument. The strongest bear case is actually straightforward: the market is treating NVIDIA as the "tollbooth of AI-era infrastructure," but it may instead be just a super hardware platform earning excess profit during a period of scarcity. Once cloud providers and model companies migrate part of their workloads to custom ASICs, in-house inference chips or alternative software stacks, NVIDIA may still be a winner, but no longer the winner with "the fattest profit, the strongest pricing power and the highest multiple" it is today. At that point, revenue growth, gross margin and the valuation multiple would step down together. Investors who hold this view usually focus more on "the cost curve of inference, the marginal return of customers' in-house efforts, and the process by which GPUs go from a luxury to standardized infrastructure once supply expands."

What facts would overturn my positive judgment? If the following facts emerge, I would conclude that I must admit my judgment was wrong:

• Data center revenue growth, after excluding the 53-week factor, falls quickly below 20% and deteriorates over several consecutive quarters;

• Gross margin stays below about 65% without an obvious one-time factor;

• Major frameworks and customer workloads show clear "de-CUDA-fication" on the training or inference side;

• Hyperscale customers shift more and more key clusters to in-house/custom chips;

• Export controls expand further to a broader range of products, or "design de-Americanization" emerges in other markets too;

• Management starts using large, low-return acquisitions to sustain the narrative. None of these is a price signal; they are signals of deteriorating business quality.

Comparison with other opportunities. Compared with Broadcom, NVIDIA has a cleaner balance sheet, a stronger ecosystem and higher quality in its single business, but Broadcom offers a more diversified AI-infrastructure exposure through custom ASICs and software/subscription revenue; compared with AMD, NVIDIA's current profitability, cash flow and ecosystem advantages are all markedly stronger. Compared with the S&P 500, NVIDIA's company quality is most likely superior, but today's buying odds are not necessarily superior; compared with the 10-year Treasury at 4.57%, NVIDIA's current Owner Earnings/FCF yield of about 2.1% to 2.2% is markedly lower, which means almost all of your excess return must be delivered by future growth. To the question "if you could hold only 5 assets," my answer is: on business quality it qualifies; on today's price, I would rather put it on the watch list than squeeze it into the portfolio right away.

Investment Checklist

The judgments above synthesize the company's latest 10-K, 10-Q, proxy, official developer materials and recent industry/competitive news.

Open questions / limitations. What this report most needs to keep tracking, but which is still not fully settled, is the true growth quality across full-year FY2027 under the combined effect of "an extra fiscal week, the Blackwell/Rubin delivery cadence, structural changes in inference demand, and the evolution of export policy"; in addition, the impact of Broadcom and cloud providers' in-house chips on NVIDIA is more a 2-to-5-year evolution curve than a proposition that can be falsified in a single quarter.

Final Investment Verdict

【Final Rating】 Watch

【One-Sentence Investment Thesis】 NVIDIA is an exceptionally fine long-term platform business, but at the current price investors are paying more for the high expectation of "continued perfection ahead" than for the margin of safety of "cheap enough today."

【Core Bull Case】

• The data center business has become the overwhelming core, with platform depth extending from GPU to CPU, networking, systems and software.

• CUDA and the developer ecosystem create high switching costs and strong network effects.

• Cash flow and the balance sheet are extremely strong, with almost no financial-leverage risk.

• Founder governance, incentive alignment and the long-term R&D cadence are all relatively excellent.

• Even through the 2023 headwind, the company sustained profitability and positive free cash flow, showing this is not "pure boom-bubble profit."

【Core Bear Case】

• The current valuation demands too much of long-term growth and high margins, with an insufficient margin of safety.

• Competition on the inference side is fiercer, and custom ASICs and cloud providers' in-house chips may erode the profit pool.

• The China market is constrained long term by export controls, and the company itself admits it is effectively excluded from the local data center compute market.

• Both customer and supply-chain concentration are rising.

• GAAP net income is significantly distorted by investment gains, and the headline PE understates the true operating valuation.

【Key Assumptions】

• Total AI demand will not collapse obviously over the next 2 to 3 years, but will only spread from training to inference and broader industries.

• CUDA and the complete platform capability will remain enough to fend off some in-house and custom-chip substitution.

• Export controls will not escalate further on a broad scale to more core products.

• Gross margin can stay high over the long run, and working-capital consumption will not run out of control.

【Fair Buy Price】 If you are a balanced, conservative-leaning long-term investor, the ideal buy range I give is 80 to 110 dollars; if you accept a smaller margin of safety for a high-quality company, you can treat 110 to 140 dollars as a "discussable range," but not cheap. The basis is the fair value range derived above from discounting conservative/neutral Owner Earnings.

【Target Holding Period】 10 years or more. For this kind of company the real source of return is the compounding of the ecosystem and platform, not the beat/miss of any one quarter. The premise is that the buy price must not over-borrow the good news of the next decade in advance.

【Expected Annualized Return】 Near the current price of about 220 dollars, my subjective estimates are:

• Conservative scenario: -5% to 0%

• Neutral scenario: 2% to 6%

• Optimistic scenario: 8% to 12%

This is not because the business is poor but because the buy price is already very high. If it pulls back further to a more reasonable buy point in the future, the expected annualized return would improve markedly. The ranges above derive from the valuation scenarios in this report and are inferences and estimates.

【Maximum Loss Risk】 In the worst case, if AI spending growth falls markedly, custom-chip substitution accelerates, gross margin declines and the valuation multiple compresses from its current high, a permanent capital loss of 50% to 70% in the share price is not unimaginable. Note that this does not require the company to "deteriorate into failure," only to step down from a "perfect platform" to a state that is "still excellent, but no longer scarce."

【Tracking Metrics】

• Data center revenue year-over-year and quarter-over-quarter growth

• The attach rate and growth of networking revenue within the data center

• GAAP/Non-GAAP gross margin trend

• Operating cash flow and working-capital occupancy

• The scale of inventory and purchase commitments

• Changes in the revenue share of the top two customers

• Whether the dominance of CUDA and the developer ecosystem shows any loosening

• The penetration of Broadcom and cloud providers' in-house chips on the inference side

• Export controls and changes in the China market

• The net effect of shareholder returns and SBC (whether the share count keeps falling)

【Signals That Trigger a Re-evaluation】

• A clear stall in data center revenue over several consecutive quarters

• Gross margin persistently breaking below about 65%

• Customers shifting markedly to in-house/custom chips

• Export restrictions escalating and spilling over into more businesses

• The company beginning to sustain growth through large, low-return acquisitions

• FCF persistently and markedly below net income, with insufficient explanation

• The share count no longer falling, with buybacks mainly becoming a "treading water" offset to SBC

【Final Recommendation】 Put coolly, NVIDIA deserves a permanent place on the core watch list and deserves continued holding by low-cost long-term holders; but for a "balanced, conservative-leaning" long-term value investor, what matters more today is not to prove it is great, but to admit that the market has very likely already recognized its greatness. If you buy now, you are relying on continued upside surprises for many years to come; if you can wait for a more reasonable price, what you rely on looks more like "the business itself." My advice is: study it, track it, respect it, but do not abandon discipline on the buy price just because it is a first-rate business.

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