Nvidia: The Widest Moat, the Narrowest Customer List

What this desk is watching

Nvidia sells the compute that trains and serves frontier models: GPUs, the networking that lashes them into clusters, and the software that makes the whole assembly programmable. In the quarter ended 26 April 2026 that business produced $81.6bn of revenue, up 85 per cent on the year, of which $75.2bn came from the Data Center segment — a 92 per cent increase — at a 74.9 per cent GAAP gross margin.¹ A hardware company earning gross margins in the mid-seventies is doing something a hardware company should not be able to do, and this dossier is about the mechanism that lets it, and the two facts that could take it away.

The margin and the developer base are the moat; the customer concentration is the risk. Read the first two against the fourth, and against a China market the company now says it is foreclosed from, and you have the whole position in one line.

The scale, and what it is compounding against

Nvidia's fiscal year ends in late January, which makes the trailing arithmetic worth stating plainly. Full-year fiscal 2025 revenue was $130.5bn, up 114 per cent, with Data Center at $115.2bn.⁴ Fiscal 2026 then rose to $215.9bn in total, up 65 per cent, with Data Center reaching $193.7bn and a full-year gross margin of 71.1 per cent.³ The single quarter just reported, at $81.6bn, is larger than the entire company was two fiscal years earlier, when fiscal 2024 revenue totalled $60.9bn.⁴

The sequential path inside the Data Center line is the part that matters for anyone trying to judge whether the curve is bending. Data Center revenue was $51.2bn in the quarter ended October 2025, $62.3bn in the quarter ended January 2026, and $75.2bn in the quarter ended April 2026 — sequential gains of roughly 22 and 21 per cent, not a plateau.⁶

Within that Data Center figure, the compute component — the GPUs themselves, as distinct from networking — set its own record at $60.4bn, up 77 per cent, which tells you the accelerators, not just the switches around them, are still the demand engine.² Management guided the current quarter to $91.0bn of revenue, plus or minus 2 per cent.²

The moat is software wearing a hardware price tag

A moat has to name a mechanism, and Nvidia's is not simply "best GPU." It is that the industry's inference and training code is written against CUDA, Nvidia's proprietary programming layer, and the tuned libraries that sit on top of it. The company reports over 7.5 million developers worldwide building on CUDA and its adjacent tools, and it says plainly that "the large and growing number of developers and installed base across our platforms strengthens our ecosystem and increases the value of our platform for our customers."⁵ That is a network effect stated in a filing, not a slogan: each developer-hour spent optimising against CUDA raises the cost of moving the workload to anyone else's silicon.

The second layer is that Nvidia increasingly sells the rack, not the chip. It packages accelerators with its own networking and with software — the quarter shipped Dynamo 1.0, which the company says lifts inference throughput on its Blackwell GPUs by up to seven times² — so that a buyer comparing a rival accelerator is not comparing chips but comparing whole systems, including the software that extracts the performance. The financial signature of that lock-in is the gross margin: 71.1 per cent for fiscal 2026 as a whole, and 74.9 per cent in the latest quarter.³ Commodity hardware does not clear those numbers; a franchise with pricing power does.

The narrow base beneath the wide moat

The risk sits in the same filings as the strength. In the quarter ended April 2026, three direct customers accounted for 21, 17 and 16 per cent of total revenue — 54 per cent of the company between them — and Nvidia separately notes that a single "AI research and deployment company" contributes a meaningful share indirectly, through cloud providers that buy on its behalf.¹ These are the hyperscalers and the largest labs, and their combined capital-expenditure budgets are, functionally, Nvidia's revenue line.

That concentration cuts both ways. It has been a tailwind while a handful of extraordinarily well-capitalised buyers race to build capacity; it becomes the central fragility the moment any one of them slows its orders, builds its own accelerator, or simply decides it has bought ahead of demand. The customers most able to fund Nvidia's growth are also the ones most able, and most motivated, to design around it. A moat measured in millions of developers does not protect against a demand base measured in single-digit numbers of buyers.

The capital returns tell you how management reads its own durability. In the latest quarter the company repurchased $20.2bn of stock and paid $243m in dividends,¹ then authorised a further $80bn of buybacks and raised the quarterly dividend from one cent to twenty-five cents a share.⁷ A board returning capital at that rate is signalling confidence that the cash flows are not a one-cycle event — a claim the concentration data invites you to test rather than accept.

The market it has already lost

The clearest thing that has changed is geographic. Nvidia states in its fiscal 2026 annual report that, as of the end of that year, it was "effectively foreclosed from competing in China's data center compute market."⁵ This is not a forward risk the company is hedging; it is a market it is describing as already closed to its accelerators under US export rules.

The numbers around that foreclosure are concrete. A year earlier Nvidia took a $4.5bn charge against excess H20 inventory and purchase obligations when the rules tightened; in the latest quarter it released $103m of that reserve as it sold or settled some of the stranded stock.¹ Beginning in February 2026 the US government granted licences allowing small volumes of the H200 part to specific China-based customers — but any H200 shipped under that programme carries a 25 per cent tariff, a structure that reopens the market only at the margin and on Washington's terms.¹ For a company growing Data Center revenue 92 per cent a year, the striking part is that it is doing so with one of the world's two largest compute markets treated, in its own words, as closed.

The tape

Walled off from the business and secondary to it: Nvidia's shares trade on the Nasdaq Global Select Market under the ticker NVDA, and the current quote and its range are worth reading against the fundamentals above rather than in isolation.⁸ With the current quarter guided to $91.0bn of revenue² and roughly 24.2 billion shares outstanding,¹ the market is not being asked to price whether Nvidia dominates this cycle — the filings settle that — but whether mid-seventies gross margins and 54 per cent customer concentration can both persist as the buyers build their own silicon and the China market stays shut. The business is the story; the tape is the footnote.

Go deeper

Start with the primary documents, all free on SEC EDGAR. The Q1 FY2027 Form 10-Q carries the customer-concentration note, the H20/H200 language, and the cash-flow statement;¹ read the concentration disclosure across the last several 10-Qs to see whether the top-three share is rising or falling. The fiscal 2026 Form 10-K is where the CUDA ecosystem and the China-foreclosure language live, alongside the full risk-factor set.⁵ For the trend, line up Data Center revenue quarter by quarter from the earnings releases and compute the sequential growth rate yourself; the question worth answering is whether the compute component alone² is still accelerating or merely large. Then hold all of it against the next print: management has guided to $91.0bn, and the gap between that number and the eventual result is the cleanest read on whether the AI factory is still being built at the pace the CEO describes.