As projected AI capital spending by the world’s largest technology companies keeps climbing, questions around when an AI bubble might break — and whether Nvidia can keep its outsized margins — are getting louder. In a recent interview, Stratechery founder and tech analyst Ben Thompson said the AI boom is reaching a key turning point.
His argument has two parts. One is that Nvidia’s seemingly untouchable net profit profile is being propped up by what he described as a form of circular financing that reduces the company’s real economics. The other is that the easing of U.S. power constraints is undercutting the electricity-based advantage that had helped lock in Nvidia’s lead, giving hyperscalers more room to catch up.
Circular financing may preserve shipments, but it also shifts risk back to Nvidia
Thompson addressed a long-running criticism of Nvidia’s GPU pricing: that such elevated pricing looks commercially unnatural and hard to sustain over time. In his view, Nvidia’s response has not simply been product strength. It has also involved financial engineering designed to keep demand moving.
He said Nvidia has taken stakes in newer cloud providers, or Neoclouds, and provided backstops of about 25% in exchange for commitments to keep purchasing Nvidia compute through 2030.
The structure lowers funding costs for those Neoclouds and makes it easier for them to buy more GPUs. But Thompson’s point was that the risk does not vanish. It gets reassigned. As he put it, 「risk doesn’t disappear, it only moves」.
If the compute market swings into oversupply, or if the Neoclouds can no longer keep buying, Nvidia could end up exposed to capacity that no one wants. Looking at free cash flow and expected value across the whole arrangement, Thompson said this kind of high-risk investment is effectively a reduction in profit. In practice, he framed it as an indirect price cut to customers.
More available power gives Amazon and Google extra time to improve their own chips
In Thompson’s telling, power scarcity should have been one of Nvidia’s strongest cards in AI infrastructure. He said that in a world short on electricity, energy efficiency becomes the deciding metric, and Nvidia’s chips still rank near the top in token efficiency. Under those conditions, customers facing power limits would be pushed toward Nvidia.
That setup has changed over the past two years, he said, because the U.S. has shown far more energy elasticity than many expected. Thompson pointed to natural gas generation in West Texas, restarted nuclear plants, and Elon Musk’s deployments on the grid as examples of supply coming in above expectations.
That is positive for the broader U.S. economy, but not necessarily for Nvidia. With more electricity available, hyperscalers with deep resources — including Amazon with Trainium and Google with TPU — have more time to improve the efficiency of their in-house chips. That extends the runway for competitors to chip away at Nvidia’s technical barriers.
If the AI boom breaks, the infrastructure may still outlast the bubble
Thompson also addressed fears that the AI buildout could end in a major blowup. His answer was historical. Healthy bubbles, he said, tend to leave behind physical infrastructure that changes how the world works.
He pointed to the railway bubble of the 1870s. Even though duration mismatch drained capital and ruined many investors, the rail network remained in place and kept operating, contributing enormously to GDP over the decades that followed. He drew a similar parallel to the internet bubble, when large amounts of dark fiber were laid down and later became part of the foundation for Google’s rise and its search business.
By the same logic, Thompson said that even if the current AI investment cycle ends in excess compute capacity and a market correction, the massive buildout of power infrastructure created to support that demand could become the most valuable long-term legacy of the boom. He tied that outcome to a world of energy abundance, where plentiful and cheap power supports economic activity and broader human progress.