Buried inside Nvidia's blockbuster Q1 results last week was a number that does not behave like a semiconductor number is supposed to. According to CFO Colette Kress, rental prices for the company's H100 GPUs, the chips that powered the first wave of generative AI training in 2022, have risen 20 percent so far in 2026. Even the older A100, launched in 2020, is up 15 percent. Chips this far into their lifecycle are meant to depreciate, not appreciate. They are doing the opposite.
The headline figures from Nvidia's report were vast. Revenue of $81.6 billion, up 85 percent year on year. Data center revenue of $75.2 billion, up 92 percent. Networking revenue of $14.8 billion, almost triple last year. Guidance for next quarter at $91 billion, several billion above what Wall Street had penciled in, and crucially without any contribution from China. The board authorised another $80 billion in buybacks and raised the dividend twenty-fivefold, from one cent to twenty-five. Sequential revenue growth, an analyst at Motley Fool pointed out, has now run for 14 consecutive quarters.
But the more interesting signal arrived from Nvidia's competitor. Reuters reported this week that AMD has asked its manufacturing partners to accelerate production timelines because AI-driven demand is pushing the global CPU market back into shortage territory. Every AI server needs a host CPU alongside its GPUs. Every rack needs networking silicon, HBM memory, power distribution and cooling. The shortage that for two years was framed as "not enough H100s" has migrated to every other component in the bill of materials.
This is why CEO Jensen Huang spent last week talking about CPUs rather than GPUs. Nvidia's new "Vera" central processor, unveiled alongside the Rubin GPU architecture, gives the company access to what Huang now describes as a $200 billion market. Speaking to reporters in Taipei on Saturday, he confirmed that figure includes China, despite no shipments of even the export-tier H200 having actually been delivered there yet. The Vera platform is the most explicit signal so far that Nvidia is no longer just selling accelerators; it is selling the whole rack.
For customers, the consequences are uncomfortable. Enterprises that finally secure GPU allocations now discover that the surrounding infrastructure is the new bottleneck. Lead times for electrical equipment have crept past lead times for the servers themselves. North American data center vacancy has fallen to a record 1.6 percent. BlackRock's Larry Fink, speaking at the Milken Institute conference earlier this month, called compute an emerging asset class and said simply that the world does not have enough of it. CME has started listing GPU futures.
The bear case is the obvious one: every semiconductor cycle eventually ends, supply catches up, prices collapse. The counter-argument, which Nvidia spent the whole earnings call making, is that AI workloads compound. Every model trained creates inference demand. Every deployed agent generates data that feeds the next training run. The demand curve does not roll over; it stair-steps upward. So far the numbers support that view: data center networking revenue rose 199 percent. That is not a single product cycle.
There are tensions in this picture, though, worth naming clearly. Nvidia's 75 percent gross margin assumes its pricing power is intact, but the same shortages that prop up H100 rental prices also drive customers toward Google's TPUs, AMD's MI series, and the custom ASICs the hyperscalers are quietly building. The fact that Nvidia is itself moving into CPUs is a defensive move as much as an offensive one. Vera is a hedge against the possibility that the bottleneck the company cannot solve, the one Larry Fink calls a shortage of electricity, eventually breaks the model.
The takeaway for anyone planning AI deployments is unglamorous. The era of treating "buy GPUs" as a procurement strategy is over. The shortage is now a full-stack problem, and the companies that locked in CPUs, networking, racks and power eighteen months ago are the ones turning models into products today.