A few headlines from the past weeks, read together, tell a strange story.
xAI, now merged into SpaceX, is renting out its Colossus data-centre. Anthropic reportedly agreed to something like $1.25 billion a month for access to roughly 220,000 GPUs, mostly for running Claude (ActuIA). Google contracted capacity worth around $920 million a month as a bridge while its own data centres scale (CNBC). Meta is exploring a business called Meta Compute to lease its idle hardware, even as it keeps spending billions to build more (24/7 Wall St.).
Look at who is renting to whom. These are not cloud providers. They are AI labs and platform companies renting compute to each other, including to direct competitors. A year ago the story was that nobody could get enough GPUs. Now some of the biggest buyers have enough to sublet.
That shift is small today, but it points at something large, and it lands directly on the one strategy that made Nvidia the most valuable company in the world.
The strategy nobody names correctly
Nvidia had its best year ever. Revenue hit $215.9 billion in fiscal 2026, up 65% year on year, at a 71% gross margin. Almost every frontier AI model in the West trains and runs on its hardware.
Most people describe Nvidia as the company that sells the best chips. True, but that is not the strategy. The strategy is what it does with the profit.
In fiscal 2026, Nvidia put $17.5 billion into private companies and infrastructure funds, most of them early-stage AI labs that turn around and spend that money on Nvidia compute. Add another $3.5 billion in land, power, and shell guarantees for data-centre partners. There is a finalising deal with OpenAI, on top of existing commitments to Anthropic, CoreWeave, and xAI.
Read that again. Nvidia invests in its own customers so those customers can afford to buy from Nvidia. The investee builds on CUDA, consumes Nvidia hardware, and its growth then validates the next round of funding. Jensen Huang has more or less admitted the logic: the labs could not raise enough through normal venture channels, so Nvidia stepped in to make sure the demand existed.
This is not diversification. It is demand manufacturing. Capital goes out one door and comes back as revenue through another, and every dollar deepens a customer's dependence on the Nvidia roadmap.
Why it made sense
It is worth being fair to the play, because in context it was rational.
Through 2023 to 2025, the binding constraint on the whole industry was supply. Nvidia could not make chips fast enough. Allocation queues ran into quarters. In that world, the risk to Nvidia was never "will anyone buy." It was "will my customers survive long enough, and stay funded enough, to keep buying at the scale my roadmap assumes." A young lab with a brilliant model but no capital is a customer that evaporates before it places its next order.
So Nvidia used its balance sheet to remove that risk. Fund the labs, guarantee they can pay, keep the flywheel spinning. When demand outstrips supply, manufacturing more demand looks almost free. You are not creating buyers out of nothing; you are underwriting buyers who genuinely want the product and simply cannot afford it yet. The scarcity does the hard work of making the investment safe.
That is the crucial assumption hiding inside the strategy. It only works while compute is scarce.
The assumption is starting to break
Which is what makes the rental headlines matter. When xAI rents out Colossus, when Meta stands up a business to lease idle GPUs, the signal is that the crunch is loosening at the top. Buyers who a year ago were desperate for allocation now have enough slack to become sellers.
Some of this is timing rather than glut. Google is renting as a bridge while its own capacity comes online. But that is precisely the point: the buyers Nvidia funded to guarantee demand are now building their own supply and monetising the overflow. The supply problem Nvidia spent billions to solve is quietly turning into a supply surplus, at least in pockets, and a demand-manufacturing machine is only safe while the opposite is true.
Once compute stops being scarce, the economics invert. Manufactured demand is demand you paid to create. If the underlying scarcity that made it a bargain disappears, you are left having capitalised your own customers into a market that no longer clears at the prices your valuation assumes. Circular financing, where a company funds the buyers of its own product, looks like genius on the way up and like the telecom and dot-com booms on the way down. The difference between the two is almost entirely whether real, unsubsidised demand keeps pace.
The buyer list is the whole risk
If your growth depends on a handful of buyers you also fund, the shape of that buyer list matters more than any roadmap. Here is what it looks like.
Who buys Nvidia's chips
Disclosed customer share of total revenue · Nvidia 10-K (FY2026)
In FY2025 three customers each sat at 11–12% (34% combined). By FY2026, two customers alone made up 36% — and the 10-K notes a further AI lab buying indirectly through those clouds.
In fiscal 2025, the top three customers each sat around 11 to 12% of revenue. Uncomfortable, but spread out. One year later, a single customer is 22% and a second is 14%. Two buyers now account for 36% of the company. And the 10-K quietly adds that one more AI lab contributes "a meaningful amount" of revenue by renting Nvidia-powered cloud from those same customers, so the real concentration is worse than the bars show.
The 10-K does not name any of them. Analysts generally assume the direct buyers are the large hyperscalers and system integrators, and that the lab buying indirectly is OpenAI, but Nvidia only ever discloses them as "Customer A" and "Customer B." The anonymity is itself telling: the company's fortunes now turn on a group so small it can be counted on one hand and labelled with single letters.
Concentration like this is fine while everyone is aligned and buying. It becomes the central risk the moment those same buyers decide they would rather build than buy, or discover they have more compute than they need.
The buyers are building, too
They are doing both. Every one of these customers has the capital and the engineers to make their own chips, and they are all doing it: Google's TPU, AWS Trainium, Meta's MTIA, Microsoft's Maia, and the OpenAI-Broadcom partnership.
The reason is not disloyalty. When your capacity plan depends on an allocation queue measured in quarters, you build a second pipeline just to hit your own roadmap. Sole-sourcing tens of billions in annual capex from one vendor is a risk any serious infrastructure team has to hedge. Custom silicon does not need to beat Nvidia everywhere. It only has to absorb enough of the predictable inference load to make the buyer a genuinely diversified consumer.
The software lock-in is loosening too. CUDA was the moat: choosing another chip meant abandoning the ecosystem. But Triton, a hardware-agnostic kernel language written inside OpenAI, one of the very labs Nvidia funds, is now the default backend for PyTorch's compile path. The same code can target an Nvidia GPU, an AMD GPU, or in principle a custom ASIC. The abstraction that erodes CUDA is being authored by the customer Nvidia's strategy was built to bind.
Put the two trends side by side. The buyers have enough compute to rent it out, and they are building their own chips to need Nvidia less. Both point the same way: away from the scarcity that made manufacturing demand a safe bet.
The actual mis-step
It is tempting to say Nvidia bet wrong on GPUs versus custom silicon. It did not. General-purpose hardware really is more adaptable to fast-moving AI research, and that is a defensible engineering position.
The mis-step is commercial, not technical. Nvidia let its biggest customers walk into custom silicon without offering them a credible semi-custom option of its own. It ceded that ground to Broadcom, Marvell, and the hyperscalers' internal design teams, when it could have kept the interconnect and runtime while letting buyers tune the accelerator. NVLink Fusion, announced in 2025, is the right idea. It is just too small a response to how fast the buyers are moving.
What this means
Nvidia is at once the biggest winner of the AI build-out and the company most exposed to its fragility. The demand-manufacturing engine is brilliant, and while compute was scarce it was close to riskless. But it concentrates revenue in exactly the buyers most capable of replacing it, it depends on a scarcity that is beginning to loosen, and it has handed those buyers both the reasons and the tools to need Nvidia less.
None of this means Nvidia is in trouble now. It is dominant, and dominance buys time. The question is what it does with that time: build a real semi-custom offering, push the moat up to the system and networking layer where the software abstraction cannot reach, and spread demand beyond five hyperscalers into sovereign, enterprise, and edge markets that do not sublet their spare capacity back into the market.
Manufacturing your own demand is one of the most powerful moves in business when your product is scarce and your customers are starved. It is one of the most dangerous when that stops being true, because you no longer own a moat. You own a bill. The open question for the next two or three years is which of those Nvidia is holding.
This piece builds on a strategy audit I wrote for my MSc in Digital Business Management. Financials are from Nvidia's FY2026 10-K; the compute-rental figures are from press reporting linked inline above.