Nvidia Makes Chips. Its Business Is the System Around Them.

In the quarter ended July 26, 2026, Nvidia booked $96.2 billion of revenue, roughly $89 billion of it from data center, at a 75% gross margin. How does a company that owns almost none of the factories behind its chips end up organizing an highly profitable industrial system around its architecture?

In short, Nvidia stopped being a chip company some time ago. It is closer to the operating layer of the AI economy. Whether that layer holds depends on several numbers most coverage skips, and they point in uncomfortable directions.

The Moat is Made of Software and Habit

Nvidia's history includes a near-death lesson in standards. Its first major chip, the NV1, bet on quadratic surfaces while the industry, steered by Microsoft's DirectX, moved to triangle-based 3D graphics. The company pivoted to the Riva 128 in 1997, reportedly selling over a million units in four months. The lesson stuck: you can build something brilliant and lose if the ecosystem moves somewhere else.

In 2006, Nvidia applied it with CUDA, a programming model that let developers use GPUs for general computing. The spend looked irrational for years. Then AlexNet won ImageNet in 2012 using GPU training, and a decade of accumulated tooling suddenly sat under the entire deep learning boom. Nvidia now says more than 7.5 million developers use its software.

When switching costs are high, buyers are not simply comparing one accelerator against another on raw performance. They are weighing years of production code, trained engineering teams, software libraries, workflow integration, and accumulated debugging expertise.

That means a competitor cannot win simply by building a faster chip. It has to overcome the value of an entire software and developer ecosystem. Hardware advantages can narrow or disappear over time, but software ecosystems become harder to displace as more developers use them, more tools are built around them, and more institutional knowledge accumulates.

That moat is psychological as much as technical. An executive betting hundreds of millions on infrastructure picks the option with the largest trained workforce and the clearest evidence that other serious companies already use it. IBM once looked unavoidable. So did Intel. Default status erodes the moment something becomes cheaper or easier, which is why Nvidia spent about $7.1 billion on R&D in a single quarter of fiscal 2027.

From Chip to Rack to Data Center

The $7 billion Mellanox acquisition in 2020 moved the unit of competition from the processor to the cluster. Training large models means thousands of processors exchanging data at extreme speed; idle compute is wasted margin. Nvidia reported about $31.4 billion of data center networking revenue in fiscal 2026, and by the second quarter of fiscal 2027, data center was over 92% of quarterly revenue. The company sells system economics: deployment speed, power efficiency, reliability. Its price hikes tell the same story; Nvidia has told its largest customers to expect AI server price increases above 15% for 2026, driven by memory costs, and expects them to pay.

The fabless model explains the margins. Nvidia outsources fabrication to TSMC and Samsung, memory to SK Hynix, Micron, and Samsung, and keeps the architecture, software, and customer relationships. Per $100 of revenue in the latest quarter, about $25 covered all costs of revenue, leaving roughly $66 of operating income.

The Numbers that Should Make You Pause

Fabless removes factory ownership; it does not remove supply risk. At the end of fiscal 2026, Nvidia disclosed about $95 billion of purchase and supply obligations. Six months later, that figure reached $279 billion of future supply and capacity commitments, largely memory. A company famous for not owning fabs has effectively committed to buying inputs on a scale larger than most national industrial programs, and International Business Times reported that Nvidia now has $99 billion riding on the AI industry through investments and commitments.

Then the bottleneck moved past the chip entirely, and Nvidia followed. Filings show $36 billion of AI cloud commitments, $20 billion of certain data center lease commitments, and, in August 2026, guarantees capped at $105 billion tied to SB Energy affiliates leasing roughly 4.25 gigawatts of data center capacity for an OpenAI affiliate in Ohio. The guarantee is contingent, not cash owed today. But a chip supplier putting its balance sheet behind customers' real estate raises a fair question: when a supplier helps finance the buyers of its own product, is reported demand a market signal or a closed loop?

The concentration data sharpens the point. In the first half of fiscal 2027, three direct customers represented 16%, 15%, and 13% of revenue, and one unnamed AI research company indirectly drove a meaningful share through cloud intermediaries. Meanwhile, those same customers, Amazon, Google, Microsoft, Meta, are building their own accelerators, as we tracked in our articles covering Nvidia's GPU era and in AMD's Instinct MI350 push.

The Bigger Nvidia Gets, the Bigger the Risk

Export controls already showed how fast commitments turn into losses. After the April 2025 US licensing requirement on H20 chips shipped to China, Nvidia took a $4.5 billion charge and described itself as effectively foreclosed from China's data center compute market; H200 shipments under later licenses ran under 1% of data center revenue.

The best argument against the bearish view is that Nvidia’s advantages reinforce one another. CUDA keeps developers tied to the platform, its networking technology improves system performance, its scale helps secure supply, and widespread adoption brings even more developers into the ecosystem. A competitor therefore has to challenge more than the chip itself; it has to overcome the entire stack.

The bearish argument starts from the same strength but reaches a different conclusion. The more Nvidia’s ecosystem expands, the more dependent the company becomes on continued growth in AI infrastructure spending. In that sense, its moat and its exposure are increasingly connected.

Both arguments can be valid. Nvidia’s $96.2 billion in quarterly revenue, more than three and a half times its roughly $27 billion in total fiscal 2023 revenue, shows how dramatically the market itself has expanded. The key question now is whether that level of demand can remain durable without Nvidia continuing to help stimulate and support the spending behind it.