Nvidia Corp

Strong returns and little debt, but the price fails the margin of safety test, so the case rests on price.

NVIDIA designs graphics processors, data centre networking, systems and software, and sells them mostly to cloud providers and AI model makers. Revenue for the fiscal year ending 25 January 2026 was $215.9B, and the company made no loss in any of the fifteen years of filings we reviewed. Returns on invested capital averaged 94%, and total debt of $8.5B sits against net cash of $2.1B, so the balance sheet passes the little-debt test. Operating margin was 60% in the latest year against an average of 38%, with a volatility of 16%, which is why the consistent earning power test fails. On valuation, the base value is $103.79 per share, the bear value is $59.10 and the bull value is $167.03, against a price of $214.13, giving a base margin of safety of -106%.

NVIDIA designs graphics processors (GPUs), data centre processors, networking equipment, systems and software, and sells them to cloud providers, AI model makers, enterprises, gamers and car makers. Its two reported segments are Compute & Networking and Graphics. Revenue for the latest year was $215.9B, and most of it came from data centre customers buying AI computing systems. The company does not own factories. It relies on outside foundries such as TSMC to make its chips, which keeps capital needs low but leaves supply in other hands.

• Data centre AI computing systems, bought by cloud providers and AI model makers
• Networking products that connect large clusters of chips
• Annual product launches that replace earlier chip generations
• Gaming and professional graphics cards, which are more mature markets
• Export licences that decide whether some products can be sold in China

Inside the circle? The largest profit pool depends on one fast-moving AI buildout, customers who buy in lumps, and export rules set by governments. Operating margin has swung widely over the period (16% volatility), and the company says demand is hard to forecast with precision. A careful owner can describe the moat, but cannot reasonably estimate the economics ten years out from these filings.

Three things appear to protect the profits: the CUDA software platform, the developer base that uses it, and a large patent and intellectual property portfolio. The filing says the developer and installed base strengthen the platform, which makes it harder for a rival chip to replace. Returns on capital back this up, with return on invested capital averaging 94% and reaching 71% in the latest year. The moat is narrow, though: operating margin averaged 38% with a volatility of 16%, export rules have already closed the China data centre market, and the filings do not show how long the developer advantage will hold. At the current price the margin of safety test fails, so what the moat is worth is a separate question from what you would pay for it.

NVIDIA returns cash mainly through share buybacks, with a small dividend. The average payout is 7.5% of profit, so most surplus cash either stays in the business or goes to buybacks. Buffett's rule is that buybacks make sense only below intrinsic value (the discounted value of cash the business can pay out over its life). At $214.13, the stock is above the base value of $103.79, so the buybacks may be made at a price above our estimate. The filing also shows large sums going into private companies and guarantees for partners, and it says there is no assurance of a return on those investments. On candour, the filing reports a large charge tied to export rules and admits that demand is hard to forecast, which is a good sign. Its own language is still promotional in places, and the filing does not give a clear view of how much of the AI spending flows back to customers who are also investees.

• Buybacks: the company repurchased shares in the latest year, and the share count has changed by -0.6% a year over the period. Buffett would ask whether those shares were bought below intrinsic value, and the filings alone cannot settle that.
• Dividends are small. The payout ratio (the share of profit paid out as dividends) averaged 7.5%, so the company keeps most of its profit in the business.
• Ecosystem investments: the company invested heavily in private companies and gave guarantees to partners. Some of those partners buy its products, so the returns depend partly on its own customers succeeding.

• FailsDemonstrated, consistent earning powerFails

  No loss years across 15 years of filings, with operating margin varying by 16% around its average.

  Why this test
• FailsMargin of safetyFails

  At $214.13, the price is above the base-case value of $103.79 (the bear case is $59.10), so there is no margin of safety (-106%). The standard lens asks for 25%.

The engine places NVIDIA outside your circle of competence, and the filings explain why. Buffett asks whether you can estimate a business's future, and these filings do not allow a ten-year estimate: the company says its 'data center scale AI infrastructure company reshaping all industries' is still early, and it describes demand as hard to estimate with any reasonable degree of precision. The pricing power test, which Buffett calls the single most important decision, is mixed. The filing says the company has 'had to reduce average selling prices, including due to our channel pricing programs', which weakens the case that customers pay whatever the price is. On owner earnings, free cash flow of $96.7B and normalised owner earnings of $90.1B pass the test, and capital needs are modest, with capital spending at 1.3× depreciation. The return on equity figure of 56% is not reliable, because a small equity base inflates it. Buffett's retained earnings test passes, and the share count has fallen, but buybacks only create value below intrinsic value, and at $214.13 the shares are above the base value of $103.79. Interest rates act as gravity: an owner earnings yield of 1.7% against a ten-year Treasury yield of 5.1% is a large gap for a business whose future cash flows are hard to see.

• Operating margin compared with its long-run average

  The latest margin is far above the average, and a fall back toward the average would reduce the value of the base case.

• Gross margin and inventory provisions in each new annual report

  The filing ties provisions and purchase commitments to demand forecasts, and large provisions would show that pricing power or demand is weaker than it looks.

• Revenue share from the largest direct and indirect customers

  The filing says revenue is concentrated among a limited number of customers, and a shift in one of them would change the business quickly.

• Size and return of investments in private companies and partner guarantees

  The filing says there is no assurance of a return on these investments, and they are growing alongside the core business.

• Export licence decisions and any China-related charges

  The filing says the China data centre market is effectively closed, and new rules could reduce revenue or add charges.

Questions the filings couldn't answer

• Can the developer base and CUDA platform keep customers when open-source models run on rival hardware? The filing says that could reduce demand, and it does not show the scale of the risk.
• Will the sums invested in private companies and the partner guarantees return value to owners? The filing gives no assurance either way.
• How much of the AI spending flows back to customers who are also investees? The excerpts do not give a clear answer.