Nvidia Stock at a Record High: Why the AI Build-Out Splits Its Customers in Two

Nvidia set a record high on October 2, its first since May. Four days earlier its board had added $150 billion to the buyback, leaving $235 billion authorised through fiscal 2028, and on October 2 Morgan Stanley's Joseph Moore made it the bank's top chip pick again, on a valuation he called "very undemanding". The record is easiest to understand by starting with his argument, stated at its strongest, because the rest of this piece disagrees with one part of it.

Put together, Moore's case is that when the hard part of AI becomes finding sites and money rather than chips, the winner is the supplier that every kind of builder can buy from. Nvidia sells through more clouds in more countries than anyone, and the half of its revenue that comes from buyers other than the largest hyperscalers is, on his account, probably growing faster than the half that does. Our extension of his case: those buyers will never design a chip, so every gigawatt they finance is a gigawatt for Nvidia.

There is a strong number behind it. Over the year to Q2 FY2027 Nvidia added about $48 billion of quarterly Data Center revenue, roughly four times Broadcom's AI gain over the same span, from a base nearly eight times larger (year-ago bases derived from the growth rates each company reported). A reader who thinks custom silicon is a footnote can point out that after a year of the fastest growth in the industry, Broadcom's AI line is still about a sixth of the two companies' combined total.

And there is Microsoft. It has its own inference chip, Maia 200, which it says delivers "30% better performance per dollar than the latest generation hardware in our fleet", and yet Nvidia's Q2 FY2027 release names Microsoft Azure, alongside Google Cloud, among the first clouds racking Vera Rubin. Owning a chip team has so far meant buying Nvidia and building in-house, not choosing between them.

We agree with most of this, and disagree with one step: that a constraint on building capacity lifts all of Nvidia's customers equally. The same constraint makes each gigawatt scarcer, and the buyers with workloads big enough to fill gigawatts are the ones with the scale, and now the evidence, to put the scarce ones into chips they design themselves. Moore's half is real. The other half is moving faster.

The rate is the leading indicator, and the rates are not close. Broadcom's AI line grew at nearly twice Nvidia's Data Center pace, is guided faster again for the quarter ahead, and on the September 2 call its chief executive said Broadcom has "secured the supply to again double AI revenue to approximately $115 billion" in fiscal 2027. Part of that line is networking, and some Broadcom Ethernet ships into Nvidia clusters, so the ratio slightly flatters custom silicon; the custom-accelerator core alone is still growing from a far smaller base into far larger named commitments.

Amazon reports the same direction from inside a hyperscaler. Its Q2 2026 release says the chips business passed a $25 billion run rate "growing triple-digit percentages year-over-year", with Anthropic and OpenAI both making multi-year, multi-gigawatt Trainium commitments. That figure includes Graviton CPUs and Nitro, so it is not all accelerators. Google's TPUs, both for its own use and for Anthropic, are built with Broadcom and so already sit inside Broadcom's line.

Start with Nvidia and AMD. The framework rates Nvidia's learned interfaces strong, which is CUDA, and AMD's weakened, along with AMD's network effects. That is the moat working with most buyers: a rival selling a general-purpose accelerator has to persuade them to rewrite software they already run. AMD's stock page records where it has succeeded, with OpenAI, Meta and Anthropic all committing gigawatts to its MI450. Those are the same gigawatt-scale buyers moving to custom silicon, which supports the split: buyers that write their own stacks can leave CUDA for AMD as readily as for a TPU. With everyone smaller, AMD has not yet broken it. AMD's data-centre revenue sits outside the Broadcom and Nvidia ratio, so that ratio understates how much of the marginal gigawatt is leaving CUDA. Nvidia's bundling, the full rack with networking, is rated intact rather than strong, so the moat the framework leans on is the software, not the box.

Now read the custom-silicon rows, which do not need to win that fight. Alphabet and Amazon are rated strong on scale economics, and the notes behind those ratings name the TPU and the silicon AWS buys at scale. Broadcom's strength sits in transaction embedding, the co-design work built into its customers' chip roadmaps. A TPU or a Trainium chip is not sold to a developer who has to leave CUDA; it is built for a buyer that already writes the software for its own models, at a scale large enough to pay for a chip team. CUDA is a switching cost, and a buyer designing its own chip is not switching. Meta sits in the same group on the strength of its named Broadcom deployments rather than a pillar that measures chips, since its scale economics are rated intact on the ad business.

CoreWeave's row is the other side of the split. It has no chip and no model of its own; its business is renting Nvidia capacity, often to the same labs that are building custom chips elsewhere. Its own weakened embedding and scale economics say nothing directly about Nvidia's moat, but on our reading a buyer in its position has no route to AI compute except the general-purpose rack and the CUDA stack, and that is the half of the market Moore is counting on.

• Capacity caps the number of chips, not the budget. A buyer that can only build so many gigawatts cannot spend its way past the limit with more accelerators. It can only get more work out of each one, and a chip built for one family of models wastes less of it than a chip built for every workload.
• The supplier's margin becomes the prize. Nvidia's GAAP gross margin was 75.0% in Q2 FY2027, and Morgan Stanley expects its revenue per gigawatt to rise. A buyer filling gigawatts has a direct incentive to keep part of that for itself. On the call Broadcom's chief executive relayed OpenAI's claim that Jalapeño "outperforms the Grace Blackwell Ultra in performance per watt, latency, throughput, and power" on OpenAI's workloads, and added in Broadcom's own voice that a co-developed chip can do this "at half the cost of a GPU". Neither claim has been independently tested.
• The software cost is already paid. CUDA's moat is the rewrite it forces. The labs and hyperscalers on Broadcom's list write their own training and serving stacks, so supporting a new chip is part of their normal engineering rather than a cost they avoid.
• Only a few buyers can make the trade. Designing a chip takes years, a scarce team and a workload large enough to fill gigawatts. Broadcom's chief executive acknowledged that the "vast majority of compute demand" comes from a concentrated group of frontier-model developers. That concentration is why custom silicon grows fast, and why it stops where that group stops.

Microsoft is the honest counter-example, and its answer is "both": on our reading it keeps buying Nvidia because its cloud customers ask for CUDA and because a custom chip covers the workloads its owner understands best, not every workload it hosts. If that pattern holds across Google, Amazon and Meta too, custom silicon adds capacity alongside Nvidia rather than taking it, and our split describes a mix rather than a migration.

That is why the claim below is about share of the marginal gigawatt and not about anyone abandoning Nvidia. A hyperscaler can buy more Nvidia every year and still put a rising share of each new gigawatt into its own chip. Nvidia's revenue can keep growing while its share of the largest buyers' spending falls; the ratio in the first evidence table is the number that separates the two.

Custom silicon also carries its own fragility. Broadcom's growth depends on six customers, and two of the three it named on the call, Anthropic and OpenAI, fund their spending largely from outside capital rather than their own cash flow. A pause in frontier-lab financing would hit Broadcom's line harder and faster than Nvidia's broader book, and it would move the ratio toward Nvidia without telling us anything about CUDA. Moore's half, the sovereigns and enterprises, does not share that dependence.

None of this makes Nvidia a weak company. The matrix rates its learned interfaces, network effects, embedding and scale strong, and the merchant challenger has dented none of them. What the build-out's constraint changes is the size of the market that moat governs. CUDA protects Nvidia with every buyer too small to design a chip, and it protects it less with the few buyers large enough to do so, who are among its largest end customers. Morgan Stanley's case needs the first group to outgrow the second; the latest reported quarter shows the second's custom spending growing faster.

That is a narrower claim than "custom silicon will beat Nvidia", and it is meant to be. For a reader holding Nvidia, the number to watch each quarter is not Data Center growth on its own but the ratio in the first evidence table. For a reader looking for the build-out's other winners, the matrix points to the companies with workloads big enough to design around, and to the builder they share.

The lesson generalises past chips. When a scarce input caps a market, the moat that matters is not only the one that keeps rivals out but the one that keeps the largest customers from building it themselves. Nvidia's moat keeps rivals out, which is what it was built to do. Whether it also keeps its largest customers buying the marginal gigawatt is what the next two quarters start to answer.