NVIDIA: A Deep Value-Investing Study

Industry stage and competitive landscape. AI infrastructure is clearly still in a high-growth phase. By IDC's measure, the global server market reached $444.1 billion in revenue in 2025, up 80.4% year over year; in the fourth quarter of 2025, revenue from servers with embedded GPUs grew 59.1% year over year and made up more than half of total server-market revenue. In other words, the main profit pool NVIDIA occupies is still expanding rapidly, not a mature, shrinking industry.

But that does not mean low risk. NVIDIA itself lists its competitors in the 10-K: AMD, Intel, Huawei, as well as large cloud players building their own chips such as Alphabet, Amazon, Microsoft, Alibaba, and Baidu; on the networking side there are Broadcom, Arista, Cisco, and Marvell. Reuters reporting also flags that AMD is increasingly seen by the market as the primary merchant-GPU challenger, that Broadcom is sharing in the AI-infrastructure pie from another direction via custom ASICs, and that large customers' in-house chips take aim at NVIDIA's biggest soft spot: whether "high-priced standardized GPUs" are optimal in every inference scenario.

Is this a business I can understand? If you take only a single line, "NVIDIA makes money selling high-performance GPUs to data centers," that is an oversimplification. A more accurate statement is: NVIDIA makes money by bundling compute hardware, interconnect networking, system architecture, the software toolchain, and the developer ecosystem into a hard-to-replace modern AI-infrastructure platform. For a long-term investor this business is understandable, but not exactly simple; it is far more complex than Coca-Cola, yet far more real than many pure-concept AI companies. Business understandability score: 4/5. Would I hold this business if the stock market closed for 5 years? Yes, provided the purchase price is reasonable.

The Moat

NVIDIA's moat is not from a single source; it is a stack of multiple layers.

Start with the most critical layer: switching costs and the ecosystem moat. The company discloses that more than half of its engineers work on software, supporting 6,000 applications, and that it owns large software assets including CUDA, CUDA-X, SDKs, APIs, industry frameworks, and AI Enterprise; meanwhile, all major CSPs, AI model makers, and enterprises use its platform. For a customer, buying the GPU is not the hard part; migrating models, training pipelines, inference frameworks, cluster management, developer habits, and operations systems is the hard part. For a large AI customer, switching suppliers is not as simple as pulling a card out of a socket; it means rewriting the toolchain, retraining engineering workflows, and re-architecting the system.

The second layer is scale advantage and system integration. NVIDIA does not just sell chips; it co-designs GPUs, CPUs, NVLink, InfiniBand/Ethernet, system racks, power, cooling, and software. In FY2026 the company spelled it out in bold in its proxy materials: demand for Blackwell systems drove simultaneous growth in Data Center compute revenue and networking revenue, and the NVLink fabric is becoming a system-level selling point. Competition at this layer is not a "single-chip benchmark" contest but a contest over "the throughput, power consumption, deployment speed, and software maturity of an entire AI factory."

The third layer is brand and developer mindshare. In AI training and high-end inference infrastructure, NVIDIA has reached a position close to the industry standard. The brand here is not a consumer brand but a "default-choice brand": when CIOs, model labs, and cloud providers discuss large-scale deployment, NVIDIA is often the first candidate. What the brand means in the enterprise-infrastructure market is lower trial-and-error cost, stronger purchasing confidence, and easier access to external talent support.

The fourth layer is the combination of capital efficiency and supply-chain control. It builds no fabs itself, so its capex intensity is very low; yet it locks up critical capacity through long-term supply, prepayments, and capacity agreements. FY2026 capex was about $6.042 billion, less than 3% of revenue, an extremely rare asset-light trait for a company selling advanced compute platforms; at the same time, the company acknowledges that long manufacturing lead times, non-cancellable orders, and prepaid capacity bring inventory and commitment risk. In other words, it enjoys asset-light returns while also bearing the complexity of supply-chain maneuvering.

Is the moat widening? My judgment is: on the whole it is still widening, but certain segments are facing their first real test. The widening part is that NVIDIA has expanded from a "single-point chip advantage" to a "whole AI-data-center platform advantage." The narrowing risk is that the industry's center of gravity is gradually expanding from training toward large-scale inference, agentic AI, and cost-sensitive scenarios, where custom ASICs, in-house chips, open networking, and purpose-built inference architectures all have a greater chance to break in. Several Reuters reports emphasize this point: AMD, Broadcom, and hyperscaler in-house silicon are not zero-noise.

How long, and how much capital, would a competitor need to replicate this? Replicating a single GPU might take a few years; replicating CUDA + libraries + frameworks + developer habits + system integration + networking + the supply chain + the brand is a far longer, far more capital-intensive task. AMD and the large cloud players have already shown that money is not the only barrier; ecosystem time matters just as much. Huawei's progress in China, meanwhile, shows that if political forces keep pushing domestic substitution, parts of the market really can have replication accelerated by "non-market factors."

Does it have pricing power? Yes, but not unlimited. FY2025 gross margin was 75.0% and FY2026 still 71.1%, nearly a luxury in the hardware industry; it shows NVIDIA still has significant pricing power and value-capture ability. At the same time, the decline in FY2026 gross margin versus FY2025 is a reminder that some of the past year's ultra-high profitability also included cyclical supply-demand tightness and a favorable product mix. My judgment is that the high margin reflects both a structural advantage and the cyclical bonus of the AI-investment wave.

Moat-strength score: 4.5/5. Not a 5, because regulation and customer in-house chips have already begun to pry at its most profitable parts; and not below a 4, because in the real world very few companies possess such ecosystem, system, and cash-flow advantages all at once.

Management and Capital Allocation

Start with the most important fact: Jensen Huang remains founder and CEO, beneficially owning about 870.6 million shares as of March 23, 2026, roughly 3.58%; directors and officers together hold about 3.94%. This is not "symbolic ownership" but genuine economic alignment with long-term shareholders. The company also discloses that Mr. Huang meets or exceeds the share-ownership requirement, and that its insider-trading policy prohibits executives from hedging, pledging, or using company stock on margin. For a long-term investor, this alignment of interests is a clear plus.

Honesty and long-term orientation. Management does not dodge hard topics in its risk disclosures. The 2026 10-K very directly acknowledges that the company is "effectively foreclosed" from the China data-center compute market, and spells out that H20 led to $4.5 billion in charges related to inventory and purchase commitments because of export restrictions; it further admits that these restrictions have helped competitors build a larger developer and customer ecosystem in China. Such statements do not paper over the trouble, indicating that management is at least willing, in writing, to lay out the bad news clearly.

Compensation structure and rationality. In its proxy materials the company emphasizes that NEO compensation is centered on at-risk, performance-based pay, using annual and multi-year PSUs, revenue and Non-GAAP Operating Income targets, relative-TSR targets, a 4-year service period, clawbacks, and prohibitions on hedging and pledging. In FY2026 Jensen Huang's total compensation was about $36.34 million, much of it equity- and performance-related; at the same time, "other compensation" for the CEO included nearly $4.046 million in security, travel, and related costs. My stance is that the overall pay architecture is on the rational side, but the CEO's security/personal-related costs are on the high side and still warrant shareholder attention.

Capital allocation. NVIDIA's order of cash use is very clear: reinvest first, then maintain high liquidity, then buy back stock heavily, with dividends a very small part. FY2025 buybacks were $34 billion and FY2026 about $40.39 billion; dividends over the same periods were only $834 million and $974 million. From the results, the large-scale buybacks of the past two years, carried out when the stock was already richly valued, look more like actively using ample cash flow to offset equity-based compensation and defend per-share value than the textbook "buying the dip heavily when undervalued." This is not bad, but it is not textbook-perfect either.

M&A and capital-allocation risk. The materials reviewed show the company had acquisition activity in FY2026, which notably raised goodwill and intangible assets on the balance sheet. Because this report did not further reconcile the price and return of each transaction in full, I do not want to draw a sweeping conclusion. What can be confirmed is that, so far, the main axis of NVIDIA's capital allocation is still not large-scale "empire-building M&A" but organic R&D + locking in supply-chain capacity + buybacks. That is more restrained than many tech leaders.

Management and capital-allocation score: 4/5. Excellent, but not flawless. The biggest plus is founder leadership, long-term vision, and high-intensity execution; the biggest minus is that heavy buybacks during a high-valuation phase may not create equal value for new shareholders, and a few governance details still call for vigilance.

Financial Quality and Owner Earnings

Among large technology companies, NVIDIA's financial quality is among the very few samples that simultaneously combine growth, margins, cash flow, and a sound balance sheet. Over FY2024, FY2025, and FY2026, revenue rose from $60.922 billion to $130.497 billion and then to $215.9 billion; gross margin went from 72.7% to 75.0% and back down to 71.1%; operating profit climbed from $32.972 billion to $81.453 billion and then to roughly $130.4 billion; and operating cash flow rose from $28.090 billion to $64.089 billion and then to $102.718 billion. This shows it is not an AI company that "only tells stories" but one that has already converted its story into audited, large-scale cash flow.

That said, profit and cash flow are not perfectly synchronized every year. FY2026 net income of $120.067 billion exceeded operating cash flow of $102.718 billion. The key cause of the gap was not bad debt or runaway receivables but two factors: first, the company recognized sizable non-operating fair-value gains — FY2026 "Gains on non-marketable equity securities and publicly-held equity securities, net" were $8.918 billion; second, working capital clearly tied up cash, with FY2026 accounts receivable up $15.399 billion and inventory up $11.324 billion. So if you are a long-term owner, you must give more weight to cash flow than to a single year's accounting net income.

The table below is enough to illustrate its financial leap over the past three years:

Data note: FY2024/FY2025 are taken from the FY2025 10-K; FY2026 is taken from the FY2026 proxy materials and the FY2026 10-K cash-flow and EPS disclosures. The FY2026 operating margin and net margin are approximations computed from disclosed revenue, operating profit, and net income.

Balance sheet. As of the end of FY2026, the company held $10.605 billion in cash and $51.951 billion in marketable securities, totaling $62.556 billion; short-term debt was $999 million and long-term debt $7.469 billion, leaving a very ample net-cash position. Total assets were $206.803 billion, total liabilities $49.510 billion, and shareholders' equity $157.293 billion. This means the main risk NVIDIA faces today is not financial-leverage risk but business, regulatory, and valuation risk.

Working capital and capex intensity. In FY2026, accounts receivable rose from $23.065 billion to $38.466 billion, and inventory from $10.080 billion to $21.403 billion, both growing fast; but revenue grew faster over the same period, and the company still generated more than $100 billion in operating cash flow, so this looks more like a "supply-demand expansion side effect" of rapid growth than a sign of cash strain. On capex, FY2026 "Purchases related to property and equipment and intangible assets" were about $6.042 billion, and relative to $215.9 billion in revenue, capex intensity is below 3% — very light. The conclusion is that this company is not one that "needs more cash the faster it grows" but one that "throws off more cash the faster it grows," though it does tie up more working capital in stages.

Changes in share count. From FY2024 to FY2026, the diluted weighted share count fell from 24.940 billion shares to 24.804 billion shares and then to 24.514 billion shares, showing that the large buybacks are not just offsetting SBC but genuinely increasing per-share value. A value investor must still press one question further: at what price the buybacks happen. If large amounts are repurchased at clearly overvalued levels, the long-term rate of return is not necessarily optimal.

Signs of financial fraud or aggressive accounting. Based on the materials reviewed, I have seen no clear evidence of financial fraud. But I will flag two items that make the valuation "look cheaper": first, FY2026 net income contains large investment fair-value gains; second, although SBC is non-cash, it is not "zero-cost" to shareholders. So a truly conservative analysis should center on operating cash flow, free cash flow, and the change in per-share value rather than on net income and PE alone.

A conservative estimate of Owner Earnings. Following the "long-term owner earnings" approach, I prefer to start from operating cash flow.

• Fact: FY2026 operating cash flow was $102.718 billion.

• Fact: FY2026 asset-related purchase outlays were about $6.042 billion.

• Inference: Approximating free cash flow to equity on this basis, FY2026 "low-subjectivity Owner Earnings" was about $96.68 billion.

• A more conservative inference: If FY2026 SBC of $6.386 billion is also treated as a real economic cost to shareholders, then "stricter-basis" Owner Earnings was about $90.29 billion. At the current price of about $225.32, the valuation the market assigns NVIDIA equals roughly 57x FY2026 FCF or about 61x Owner Earnings on the stricter basis. This is not a cheap price.

Valuation and Margin of Safety

This section needs a special emphasis on distinguishing fact, assumption, and inference.

Facts. As of May 15, 2026, NVDA trades at about $225.32. Against FY2026 diluted EPS of $4.90, the static GAAP PE is about 46x; against approximate FY2026 FCF of $96.68 billion and the equity value implied by the current price, the FCF yield is about 1.8%. Over the same period the U.S. Treasury-published 10-year Treasury yield was about 4.59%. This means what you buy in NVIDIA today is not a "current yield" but "the return after many years of high growth have been delivered."

Method One: discounted Owner Earnings. Here I do not use "ultra-optimistic" parameters but assumptions more consistent with a value-investing framework.

• Conservative scenario: Start from $90.3 billion of stricter-basis Owner Earnings, grow 8% annually for the first 5 years and 4% for the next 5, discount at 10%, with terminal growth of 3%.

• Neutral scenario: Start from $96.7 billion of approximate FCF/Owner Earnings, grow 12% for the first 5 years and 6% for the next 5, discount at 9%, with terminal growth of 3.5%.

• Optimistic scenario: Start from $96.7 billion, grow 16% for the first 5 years and 8% for the next 5, discount at 8.5%, with terminal growth of 4%.

Under these assumptions, my inference is:

• Conservative intrinsic-value range: about $90 to $130 per share

• Fair intrinsic-value range: about $130 to $180 per share

• Optimistic intrinsic-value range: about $180 to $230 per share

These are not "revealed truth"; they simply tell you that the current price sits roughly near the upper edge of my optimistic scenario, not within the conservative range. So buying today is not asking that "the company stay excellent" but asking that "the company stay extraordinarily excellent." The underlying cash-flow, net-income, and share-count data supporting these inferences all come from the FY2026 audited statements.

Method Two: relative valuation. Here I try hard to avoid the trap of "peers are all expensive, so it isn't expensive." NVIDIA's current static profile is extremely high margins, extremely high growth, extremely high capital efficiency, an extremely wide moat, and, along with all of that, extremely high valuation multiples. Compared with AMD, Broadcom, and the large customers' in-house routes, NVIDIA's quality is usually higher and its ecosystem stronger, but "stronger" does not equal "worth buying at any price." If a stock needs more than 15% high growth sustained for many years merely to prove "it isn't expensive now," then it offers no typical margin of safety to a new investor. Reuters reporting on AMD, Broadcom, hyperscaler in-house chips, and intensifying competition in the inference phase also shows the market will not hand the profit pool to NVIDIA alone.

Method Three: asset or liquidation value. This company is not suited to being priced on a liquidation-value basis. Although its on-balance-sheet net cash is strong, even after subtracting debt from cash and marketable securities the net cash is only about $54.09 billion, just a few dollars per share; the real value comes almost entirely from future cash flows and the moat, not from realizable book assets. This implies two conclusions: first, it is a genuine franchise asset; second, the asset approach offers you almost no valuation-floor protection.

Below is a price framework that is more practical to execute:

By this framework, the current price of about $225.32 is closer to the "good company but bad price" side than to "a cheap good company."

Is the margin of safety adequate? My answer is: not adequate. The most fragile valuation assumption is not "will the company collapse?" but "can the company keep sustaining ultra-high growth and ultra-high margins for years at an enormous scale?" If growth comes in below expectations, if gross margin keeps sliding from the 71%–75% range, or if the valuation multiple contracts from 46x GAAP EPS toward a level closer to a mature leader's, new investors can easily run into years of low returns even if the company itself stays excellent.

Risk Comparison and Final Judgment

The most important risk. NVIDIA's risk is not a short-term drawdown in the share price but the permanent loss of capital triggered when, after buying at the current high valuation, the fundamentals turn out "slightly worse than the market's overly high expectations."

First is competitive risk. AMD is increasingly seen by the market as the leading merchant-accelerator challenger; Broadcom, via custom ASICs, takes a different slice of AI-chip demand; and large cloud players and platform customers keep developing in-house chips to optimize cost, power, and specific workloads. NVIDIA still leads, but leading does not mean market share never declines.

Second is regulatory and geopolitical risk. The company discloses plainly that, as of the end of FY2026, it has been "effectively foreclosed" from the China data-center compute market; H20 has already led to $4.5 billion in charges; and this kind of restriction affects not only China but also encourages overseas customers to "design-out" U.S. chips and turn to alternatives. In other words, regulatory risk is not a margin note; it has already entered the income statement for real.

Third is supply-chain risk. NVIDIA depends on TSMC and HBM suppliers as well as on advanced-packaging capacity such as CoWoS; the 10-K repeatedly stresses long lead times, non-cancellable orders, prepaid capacity, supplier concentration, and Asian geographic concentration. As long as demand keeps rising, this risk looks like "part of the moat"; once demand turns, it will show up in inventory, gross margin, and purchase-commitment losses.

Fourth is customer-concentration and cyclical risk. In FY2026 two direct customers accounted for 36% of total revenue; receivables are also highly concentrated. NVIDIA's business is not naturally smooth like consumer staples; it is in substance highly exposed to the spending cadence of very large AI-infrastructure budgets.

Fifth is overvaluation risk. The current static earnings yield is only about 2.2%, below the 10-year Treasury at 4.59%; the FCF yield is about 1.8%. So a buyer is not buying "current return" but "growth that remains strong many years from now." Once growth is merely "very good" rather than "better-than-expected good," the valuation may revert.

The strongest bear case. If I had to write the strongest argument for a short seller, I would put it this way: This company has gone from an "exceptional company" to an "exceptional consensus," and the capital market's assumptions about it are too demanding — the market does not ask that it stay excellent but asks that, on a far larger revenue base, it keep sustaining years of high growth with a near-monopoly share, near-peak margins, an almost error-free product cadence, and a broadly stable regulatory environment. Break any one link in that chain and the returns for new buyers are very likely to disappoint. This bear case does not require the company to collapse; it requires only that the company "fail to keep miraculously beating expectations." That is precisely where a high-valuation leader is most dangerous.

What facts would make me admit I was wrong? If the following facts appear over the next few years, I would conclude that the "Watch/cautious" verdict needs to be revised upward, perhaps to a buy: First, Data Center revenue keeps growing rapidly on a larger base while gross margin holds near about 70%; second, customer in-house chips and ASICs do not significantly erode the position of the CUDA/full-system/networking platform; third, free cash flow keeps growing strongly while the share count keeps falling steadily; fourth, the market price falls back into a more reasonable range. Conversely, if the following facts appear, I would admit the bull thesis is broken: a sharp stall in growth for several consecutive quarters, a marked step-down in gross margin, falling order concentration among major customers who shift to in-house or competing chips, export controls continuing to spill over into more markets, and working capital and inventory deteriorating persistently while cash flow falls behind. These are the signals that would truly trigger a re-rating.

Compared with other opportunities. Against the 10-year Treasury, NVIDIA clearly does not offer a higher "current yield"; what it offers is higher uncertain growth. Against a broad index, NVIDIA's advantage is business quality and profitability; its disadvantage is single-company risk and almost no margin of safety. If you believe the company can keep dominating AI infrastructure over the next decade, NVIDIA may outperform the index; but if you require "not losing much even if the judgment is slightly off," the index is usually friendlier. My subjective conclusion is: if I could hold only 5 assets, at the current price I would not put NVIDIA in the top five for fresh capital; if the price returns to a more conservative range, it would again qualify to compete for that spot.

Investment checklist.

The judgments above draw collectively on the company's disclosed business, customer concentration, cash flow, returns on capital, and risk factors.

Final judgment.

【Final Rating】 Watch

【One-Sentence Investment Thesis】 NVIDIA is one of the strongest AI-infrastructure platforms of our era, but the current share price looks more like buying a future of "sustained flawless delivery" than buying "value with a margin of safety."

【Core Bull Case】 First, the data-center AI platform has become one of the de facto industry standards, with extremely strong ecosystem and switching costs. Second, revenue, margins, and cash flow have all been verified by audited results, not merely told as a narrative. Third, the balance sheet is extremely strong and almost free of leverage risk. Fourth, the founder still holds a deep stake, giving strong strategic continuity. Fifth, the total addressable market for AI infrastructure is still expanding rapidly.

【Core Bear Case】 First, the current valuation demands too much of future growth. Second, the lockout from the Chinese market and export controls have already genuinely hurt profits and competitive position. Third, customer in-house chips, AMD, Broadcom/ASICs, and cost competition in the inference phase are all intensifying. Fourth, customer and order concentration is high, and the business is not naturally smooth. Fifth, the larger the industry grows, the harder it is to sustain ultra-high growth.

【Key Assumptions】 For the investment to hold, the following must be true: the CUDA and full-stack-system moat is not rapidly weakened; aggregate customer capex stays on a high-growth path for the next several years; gross margin broadly holds at a high level; export controls do not spill over more broadly; and management keeps converting cash flow into per-share value.

【Fair Buy Price】 The range I lean toward is $120 to $160. This is because the fair value, discounted under conservative-to-neutral assumptions, clusters more in this band, leaving some room for error. $160 to $220 is better suited to "existing holders assessing whether to keep holding"; above $220 looks more like paying up for the optimistic scenario.

【Target Holding Period】 If bought at the right price, I believe the holding horizon should be 5 to 10 years or more; if there is no margin of safety at purchase, then even a longer horizon will not automatically repair the return.

【Expected Annualized Return】 This part is an inference, not a fact. Buying at the current price, my rough ranges are: conservative scenario 0% to 4%; neutral scenario 4% to 8%; optimistic scenario 9% to 14%. The reason even the neutral return is not very high is that the current price has already discounted a great deal of success in advance.

【Maximum Loss Risk】 If AI capex cools, the valuation multiple contracts, and gross margin steps down from its high while the company is merely "excellent but no longer magical," new investors could entirely plausibly face a 40% to 60% long-term drawdown; layer on worsening regulation, lost share, or a cyclical downturn, and an even larger decline is not impossible in an extreme scenario. The true permanent-capital-loss scenario is not bankruptcy but the return being erased over the long run after buying at too high a price.

【Tracking Metrics】 Going forward I will keep tracking: Data Center revenue growth; whether gross margin holds at a high level; whether operating cash flow and free cash flow keep pace with net income; whether receivables and inventory keep growing faster than revenue; whether major-customer concentration keeps rising; whether the share count keeps falling; export controls and the status of the China market; the penetration of AMD/ASIC/in-house chips on the inference side; the networking attach rate; and changes in software- and support-related deferred revenue.

【Signals That Trigger a Reassessment】 Including but not limited to: a marked stall in Data Center growth for several consecutive quarters; gross margin falling persistently below its past high range; major customers clearly shifting to in-house or competing chips; regulatory restrictions expanding to more products/regions; free cash flow notably weaker than net income while working capital keeps deteriorating; and the most central part of the moat, the CUDA/system ecosystem, being weakened.

【Final Recommendation】 Soberly put, NVIDIA deserves respect, but not necessarily a purchase at any price. If you already own it and your tax burden and opportunity cost are low, I lean toward "hold it while continuously verifying the moat"; if you do not yet own it, by Buffett-style value-investing standards I would advise you to put this company on a high-priority watch list rather than chase the price immediately. What truly makes long-term investors a lot of money is often not the half about "finding the greatest company" but the half about "acting only when the price gives you a sufficient margin of safety."

Open questions and limitations. This report's core judgments rest on the FY2026 annual report, the FY2026 proxy materials, official interest-rate data, and credible news; but for a same-day, like-for-like cross-comparison of peers' EV/EBITDA, P/FCF, and ROIC, I did not establish a fully unified, like-for-like peer dataset within the materials reconciled this time, so I deliberately stay restrained on the related peer-valuation comparison to avoid drawing seemingly precise but actually misleading conclusions from mixed measures. What the current conclusion most depends on is not peer relative valuation but the relationship among NVIDIA's own cash flow, its moat, and the current price.

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