NVIDIA Corporation

On 30 June we published NVIDIA at $200.09 and concluded "Too hard": a magnificent business we could not value with the confidence we demand. For those who bought anyway, we suggested a margin of safety nearer the low $150s, with an alert at $155. The shares are $225.07 today. Here is the account.

Four right, three partly right, one wrong — and the wrong one is embarrassing rather than expensive. The calendar check that now opens every report on this site exists because of mistakes like it: a dividend changed in May, our data showed April's rate, and nobody looked.

One grade needs more than a row. Two days ago we wrote a new rule into our playbook after Microsoft: we will not demand a bargain for a wonderful business trading inside our own estimate of its value. Does June's NVIDIA verdict break it? We think not, and the reason matters. "Too hard" is not "too expensive". It says we cannot estimate the value with confidence — so there is no range to be inside. The rule governs verdicts about price; this one was about knowledge. The $155 alert, though, sat below the $247 our own DCF showed in June, and we should not have offered a number we did not believe in. We will not do that again.

In June we described NVIDIA as the seller of shovels in a gold rush: it did not need to know which miner would strike gold, because every miner needed shovels. That was the heart of its appeal. It is no longer quite true.

Over the past year NVIDIA has become one of the largest investors in artificial intelligence. Including public holdings, its equity investments stood at about $99bn on 26 July, with roughly $25bn more committed. The largest single cheque, about $30bn, went to OpenAI; it has also taken part in funding rounds for Anthropic and xAI, among many others, and on 26 September it was reported to be weighing a $10bn stake in Anthropic's planned IPO. Many of these companies are among the largest buyers of NVIDIA's chips.

Now look at what that does to the income statement. When a company NVIDIA owns raises money at a higher valuation, NVIDIA marks its stake up and books the gain. In the second quarter those gains on equity securities were $7.8bn before tax. For the first half of the fiscal year, NVIDIA's GAAP profit was $118.0bn; its own adjusted figure, which strips out those gains, was $99.5bn. About $18.5bn — roughly 16% of reported profit — came from the rising value of stakes, not from selling chips.

★ This is not an accounting trick; it is disclosed, and NVIDIA's own adjusted numbers exclude it. It is something subtler — a loop. NVIDIA invests in AI companies. Those companies use the money, and money raised from others on the strength of NVIDIA's endorsement, to buy NVIDIA chips. Their growing chip purchases make them look more valuable at their next funding round. That higher valuation becomes a gain on NVIDIA's books. Each step is rational. Together they mean that part of NVIDIA's demand and part of its profit now depend on the same thing: the continued willingness of investors to fund AI laboratories at rising prices.

Buffett has a view on vendor financing, which is what this resembles at one remove: the seller's profits look best exactly when the buyers are most dependent on the seller's own money. A business that finances its own demand has made its earnings more cyclical, not less. NVIDIA's core economics are no less extraordinary than they were in June. But the question "who is paying for all these chips, and with whose money?" now has NVIDIA's name in the answer.

The second quarter of fiscal 2027 (to 26 July) produced revenue of $96.2bn, up 18% on the previous quarter and 106% on the year; gross margin was 75.0%; GAAP earnings were $2.46 a share and adjusted $2.22. NVIDIA guided the third quarter to $108bn ±2% at a 74% gross margin — and, notably, assumes no data-centre compute revenue from China at all. Jensen Huang's summary on the day: AI "has reached its inflection point… compute is revenue."

Since the start of this fiscal year NVIDIA has stopped reporting gaming, workstations and cars separately. It now splits the data centre by who buys, and that split is more revealing than the old one:

Read the second row twice. The fastest-growing customers are not Microsoft or Google; they are the newer "AI clouds" — CoreWeave, Nscale, Nebius and their peers — plus enterprises and governments ("sovereign AI" more than tripled year on year). Several of those AI clouds are companies NVIDIA has invested in, guarantees or rents capacity back from (Parts IV and VI). The hyperscalers, who are building their own chips (Part V), are growing their NVIDIA purchases more slowly — still doubling, but more slowly.

What management expects next. Vera Rubin should be about a fifth of data-centre revenue in the third quarter, in what Colette Kress called "the fastest product ramp in NVIDIA's history". Gross margin is guided to 74% in the third quarter and to "bottom in Q4 in the 71% to 72% range" as the new platform ramps. For fiscal 2028 NVIDIA expects revenue to grow about 70% — "a supply-constrained outlook", meaning it believes it could sell more if it could make more.

Capital returned: about $26bn in the quarter ($19.7bn of buybacks and $6.0bn of dividends); about $99bn of buyback authorisation remaining after May's additional $80bn; and the quarterly dividend, now $0.25, paid again on 1 October.

In June we called NVIDIA's moat "wide today, of genuinely debated durability" and scored it 8 out of 10. Readers arriving here for the first time deserve the full argument, and readers who were with us in June deserve to know whether three months of evidence moved it. Start with what the moat is made of, because it is not a chip.

Layer one: CUDA, the language AI is written in. NVIDIA released CUDA in 2006 and lost money on it for a decade — "it cost the company billions; we didn't make money for 10 years," as Ian Buck, who led its creation, put it this spring. Its annual report now counts more than 7.5 million developers using CUDA and its other tools, and "hundreds of domain-specific software libraries". The switching cost is not the chip; it is the code, the tuned libraries and the engineers who know them. When SemiAnalysis benchmarked AMD's MI300X against NVIDIA's H100 in December 2024, it found AMD's out-of-the-box software "very difficult to work with" — a gap AMD has worked hard to close since.

Layers two and three: the network and the rack. Tens of thousands of chips must behave like one machine. The wiring that does it — NVLink, InfiniBand and Spectrum-X, from the 2020 Mellanox purchase — produced $31.4bn of revenue in fiscal 2026, up 142%. And NVIDIA now sells whole racks, the NVL72, seventy-two GPUs wired as a single computer, rather than loose chips. The proof that it works is the price: by Kress's figures NVIDIA earns about $18bn per gigawatt of capacity built on Hopper, $25bn on Grace Blackwell and $40bn on Vera Rubin. Customers pay more per unit of power each generation — the opposite of a commodity.

Layers four and five: the cadence and the supply. A new platform every year means a rival aiming at this year's product meets next year's. And NVIDIA has committed $279bn to suppliers — up from $119bn a quarter earlier, mostly memory — securing components competitors must then fight over. That is a moat money can buy, and NVIDIA has the money.

★ Our moat verdict: still intact — wider in the stack, narrower at the very top, and now partly paid for. The layers NVIDIA controls (network, systems, cadence, supply) all got stronger. The one it cannot control — whether its biggest customers want an alternative — got weaker. We keep the dial at 8.

A moat keeps competitors out. It does not stop the people inside the castle from building a second castle. NVIDIA's largest customers are its most capable competitors: Google designs the TPU, Amazon the Trainium, Microsoft the Maia, Meta the MTIA, and OpenAI is designing its own chip with Broadcom. They do not need to beat NVIDIA everywhere; they need only to run their own most common workloads — mostly inference, the cheaper and more predictable job of running a trained model — on chips that cost them less. Every workload they move is a workload NVIDIA cannot price at 75% gross margin. The data shows the shift at the edge rather than at the core: IDC counted spending on servers using non-GPU accelerators up 238% year on year in the second calendar quarter, against 28% for GPU servers, making the alternatives roughly a quarter of accelerated-server spending.

Part II described NVIDIA's $99bn of equity stakes. The 10-Q filed in August shows the stakes are only one of several ways NVIDIA now uses its balance sheet to help customers buy its products:

There are two honest readings of that table. The first is that it strengthens the moat. NVIDIA has the strongest balance sheet in the industry and is using it as a competitive weapon: an AI laboratory or a new cloud that receives NVIDIA's capital, credit or guarantee builds on NVIDIA's platform rather than a rival's, and the Ohio site will host nothing else. On the August call Kress said NVIDIA had "invested nearly $50 billion in the frontier AI labs" and expects laboratories supported by its balance sheet to be "roughly a quarter of our business next year". She anticipated the criticism herself: "We know some will call this circular financing. We see it differently."

The second reading is that it manufactures part of the demand it then reports. Revenue that depends on NVIDIA's own guarantee or rent-back promise is not the same quality as revenue from a customer paying with its own money. If AI funding tightens, the same customers who stop buying are the ones whose shares NVIDIA owns, whose leases it guarantees and whose unsold capacity it has agreed to take. Buffett has a phrase for profits that depend on the seller financing the buyer: they look best exactly when they are least reliable.

We think both readings are true at once. A moat that has to be subsidised is still a moat; it is just no longer free — and what NVIDIA spends to keep customers, from $2.2bn of stakes two years ago to $99bn today, is rising faster than its revenue.

Sources: NVIDIA 10-K for fiscal 2026 (developer count, networking revenue); 10-Q and CFO commentary of 26 Aug 2026 (supply commitments, guarantees, AI cloud agreements, receivables); Q2 FY27 earnings call transcript, 26 Aug 2026 (revenue per gigawatt, Rubin ramp, Kress quotations); SemiAnalysis, 22 Dec 2024; Epoch AI on Google's TPUs; IDC Q2 2026 server data via The Next Platform; CNBC, 4 Sep 2026 (stakes two years ago).

In June we wrote that "the one to watch is not AMD — it is NVIDIA's own customers," and cited estimates that custom chips might take 15–25% of the market by 2030. That looks too slow. Three months of filings and earnings calls show the alternatives arriving faster than we expected — and NVIDIA growing anyway.

The pattern across the rows. Nobody is beating NVIDIA at what it does best. What is happening is subtler: every large buyer is building or buying a second source, and the second sources are now big enough to matter. OpenAI's announced commitments split roughly 12 gigawatts to NVIDIA, 10 to Broadcom and 6 to AMD. Anthropic, a company NVIDIA has invested in, has announced up to 5 GW on Amazon's Trainium and a reported $125bn five-year TPU lease, against up to 1 GW of NVIDIA systems bought directly — although its IPO filing is also reported to show up to $84.5bn through 2029 for NVIDIA-based capacity rented from SpaceXAI, cancellable at 90 days' notice. NVIDIA is in Anthropic's data centres; it is no longer the default. Epoch AI's estimate of AI computing power shipped in the last quarter of 2025 gave NVIDIA 67% — a commanding share, but not the 80–90% of training we quoted in June.

And yet NVIDIA's data-centre revenue grew 117% and management says it is supply-constrained into 2028. Both things are true because the market is growing faster than any one supplier can serve. The question is not whether NVIDIA loses share — it is losing some at the margin — but what happens to its prices and margins when the market stops growing faster than supply. That day has not come. When it does, the second sources its customers are building today will be the reason NVIDIA cannot hold 75%.

China was once one of NVIDIA's largest markets. In April 2025 the United States required licences for the H20, NVIDIA's China-specific chip, and NVIDIA took a $4.5bn charge for inventory it could not sell. In January 2026 Washington moved to case-by-case licences for the more powerful H200, capped at half of US volumes and subject to a 25% duty; Beijing promptly told companies not to buy it "unless necessary". In the latest quarter H200 shipments to China were under 1% of data-centre revenue, NVIDIA booked another $0.4bn charge, and its outlook assumes none. Meanwhile Huawei is ramping its Ascend chips for a domestic market NVIDIA no longer serves. For valuation this is now upside, not risk: any reopening would be a bonus the forecasts do not contain.

Sources: AMD Q2 2026 results, 4 Aug 2026; AMD–Meta, 24 Feb 2026; AMD–Anthropic, 22 Jul 2026; Broadcom FY26 Q3 results, 2 Sep 2026; Epoch AI, AI chip market share by designer; Anthropic–Amazon, 20 Apr 2026; Reuters on Anthropic's TPU lease, 1 Oct 2026; Reuters on China's H200 curbs, 14 Jan 2026; Amazon Q2 2026 call (30 Jul 2026); NVIDIA 10-Q and call (26 Aug 2026). Gigawatt figures are announced commitments, not deliveries; Anthropic's TPU figure (~3.5 GW) is as reported from a Broadcom filing.

Buybacks versus dilution. NVIDIA spent $40.4bn buying back 282 million shares in fiscal 2026 and another $39.8bn for 203 million shares in the first half of this year. Against that it paid stock-based compensation of $6.4bn last year — about 3% of revenue — and has $19.4bn more already granted and waiting to vest. The net result: the diluted share count fell from 24,804 million in fiscal 2025 to 24,285 million in the latest quarter, about 2%. That is honest arithmetic, not window dressing — the buybacks genuinely shrink the company, slowly, after absorbing employee awards. But note what $80bn bought: two per cent of a company priced at around twenty-five times next year's earnings.

The investments, and what comes back. This is where the score is decided. In the first half of fiscal 2027 NVIDIA put $31.0bn into private companies — almost as much as it spent on its own shares. Here are the largest positions we can verify, and what NVIDIA receives besides the equity:

Some of these look superb as investments: the Intel stake has multiplied about six-fold, and the SpaceX stake turned a bet on xAI into a $21bn holding. Huang's own view, on the August call: "the only regret that I have is that I didn't invest more and sooner." But returns on stakes are not the test we apply to a chipmaker's management. The test is whether capital goes where it compounds the core business at the highest return — and an equity book that went from about $2.2bn two years ago to $99bn, invested largely in the company's own customers, is a different business from the one we analysed in June. We keep capital allocation at 8, down from 9 in June: the buybacks and the dividend are exemplary; the venture book is large, growing and harder to judge every quarter.

Jensen Huang is 63 and has run NVIDIA since he co-founded it in 1993. Asked about succession on Lex Fridman's podcast in March 2026, he answered: "I'm famous in saying that I don't believe in succession planning" — his approach is to pass on what he knows "as often and continuously" as he can, and he has said he hopes to die on the job. The 2026 proxy statement does not discuss succession at all; the annual report mentions it only in a generic risk factor. Two of his children, Madison and Spencer Huang, hold senior product roles, which has prompted press speculation; NVIDIA has made no statement on a transition.

We do not think NVIDIA's culture depends on one man in the way Berkshire's once seemed to depend on one. But the company's strategy — CUDA a decade early, Mellanox, the rack, and now the investment book — has been Huang's in every important respect, and a board that has not described a plan for replacing him is a key-person risk an owner should price, especially in a business whose customers buy a roadmap as much as a product.

Two adjustments and one new fact. First, about $18.5bn of first-half profit came from gains on stakes (Part II), which is why adjusted earnings ($4.09 a share) sit below GAAP ($4.85). Second, working capital: receivables rose from $38.5bn to $63.1bn in six months on "extended payment terms" for large customers (Part IV), which is why second-quarter operating cash flow ($24.1bn) was well below profit ($59.7bn). The new fact: in June NVIDIA issued $25bn of bonds "for general corporate purposes", taking debt from about $8.5bn to $33.5bn while holding $56.6bn of cash and debt securities. A company producing over $100bn a year of free cash did not need to borrow; it chose to, in the quarter it committed tens of billions to customers.

On 18 May 2026 NVIDIA raised its quarterly dividend from $0.01 to $0.25, an increase of 2,400%, and approved another $80bn of buybacks. We missed it in June — Part I. At $225 the yield is only 0.44%, which is why the change is easy to overlook. But look at it in dollars rather than as a yield: NVIDIA paid $2$243m of dividends in the first quarter of the fiscal year and $6.0bn in the second. At $0.25 a share on about 24.1 billion shares, it has committed to roughly $24bn a year — more than the annual profit of most companies in the S&P 500. The yield is symbolic; the commitment is not.

The verdict on the dividend: safe because it is small relative to today's cash, and symbolic because it is small relative to the price. Nobody should own NVIDIA for income. The signal is that the board now believes the cash generation is durable enough to promise a fixed share of it every quarter — a promise worth remembering because test five shows the one scenario, a capex bust like fiscal 2023, in which it would have to be paid from savings.

On forecasts NVIDIA is cheaper than it looks: 24 times this fiscal year, 14 times next, 10 times the year after. On the question of what those forecasts rest on, it is harder to own than in June. They now assume not only that AI spending keeps compounding and that customers keep choosing NVIDIA over their own chips, but also that the laboratories NVIDIA has invested in keep raising money at rising valuations — because that is where part of the demand, and part of the profit, now comes from.