Stratechery's Ben Thompson: Nvidia Is Already Cutting Prices, It's Just Hidden in Circular Deals
Stratechery founder Ben Thompson on the Invest Like the Best podcast lays out where the AI capex cycle is fragile, who is best positioned, and why the "US must beat China" narrative is being overplayed, August 21, 2026
Ben Thompson, the writer behind Stratechery and the aggregation theory framework that has shaped a decade of tech investing, used a wide-ranging conversation on the Invest Like the Best podcast to challenge several of the AI trade's most comfortable assumptions. The most striking argument was not about who wins the AI race, but about whether the industry can fund itself long enough to find out.
Nvidia's Margins Are Already Being Cut, Just Not Where Anyone Can See It
Thompson's sharpest new framework concerns Nvidia's recent wave of circular financing arrangements, including the $500 billion structure tapping pension funds and insurance floats, and Nvidia's practice of backstopping Neocloud purchases through 2030. He argues these deals are functionally price cuts that simply don't appear in the income statement. "Nvidia is providing a 25% backstop, and if you actually ascribe a value to that, to Nvidia taking equity in the Neoclouds or guaranteeing they're going to buy all their compute to 2030, why do they do that? So the entity in question can get a lower cost of capital," he said. "Risk never disappears. It just sort of appears somewhere else. Taking on risk has a price." In his view, Nvidia is holding steady on headline pricing while absorbing real downside risk off balance sheet, a dynamic he says the market isn't pricing correctly through simple discounted cash flow models.
The bigger structural threat to Nvidia, in his telling, isn't AMD or custom silicon start-ups but the hyperscalers themselves. Google has already agreed to sell roughly 20% of its TPU capacity to Anthropic, and Amazon's Andy Jassy has signaled that Trainium chips will eventually be sold externally too. Because hyperscalers have a structurally lower cost of capital than Neoclouds, and because they are increasingly willing to sell chips as commodities rather than differentiated products, Thompson sees this as Nvidia's central long-term vulnerability. He also floated a counterintuitive read on power: prolonged power abundance in the US, which he says has "surprised" him and likely surprised Jensen Huang as well, has removed a scarcity dynamic that would have otherwise reinforced Nvidia's moat around token efficiency.
TSMC's Conservatism Handed the Risk to Big Tech, and Handed Intel a Lifeline
Thompson's account of the current compute shortage places the blame squarely on TSMC's investment pace in 2023, 2024 and 2025, when growth rates declined even as AI demand accelerated. Because fab lead times exceed data center lead times, he argues the shortage will worsen before it improves, with capacity from today's capital commitments not manifesting until 2028 or 2029. "TSMC has offloaded risk onto the big tech companies," he said, describing how the foundry's structural bias toward avoiding overcapacity, since a fab is expected to run for 30 years, has left hyperscalers "foregoing so much revenue and so many profits because we don't have enough compute."
That dynamic, he argues, is what finally makes Intel investable as a foundry partner after more than a decade of failed attempts. Thompson has been writing since 2013 that Intel needed to build a genuine foundry business, but says no rational customer would previously choose Intel over a smoothly functioning TSMC. Acute scarcity changes the math. "The scarcity is what ultimately saved Intel," he said, predicting a major new foundry partnership announcement in the near term. He drew a pointed parallel to memory makers and to Iran's leverage over the Strait of Hormuz: both are levers that lose their power the moment they are actually used, because the world routes around them permanently afterward. "No one's going to let themselves get in this situation again," he said of chip customers' likely response to today's shortages.
Amazon's Internal Flywheel Versus Apple's Deliberate Pass
Asked which large technology company has the most compelling setup, Thompson pointed immediately to Amazon, citing its pattern of building infrastructure for its own use before commercializing it externally, from AWS itself to Graviton and Trainium chips. "The early versions were terrible," he said of Amazon's custom silicon, "but if you're on Amazon and you're using some of their managed services... they can put all their crappy processors underneath the services they're selling, and that gives them the volume and the capacity to iterate them and get better." He sees this self-referential scale advantage as durable and largely insulated from AI disruption at the core retail and logistics business.
Apple's decision to sit out the frontier model race, Thompson argues, is defensible rather than a missed opportunity, given the mismatch between Apple's deterministic manufacturing culture and AI's probabilistic nature. "Apple is the king of deterministic products," he said. "You ship that iPhone, you ship it once, and it's got to be good... I'm generally fine with Apple not doing AI. I want them to keep making great devices." The open question he raised is whether Apple repeats Microsoft's historic mistake with mobile, assuming the phone will always remain the center of computing even as ambient AI potentially shifts the interface elsewhere.
Microsoft's IBM Playbook and Meta's Necessary Recklessness
Thompson drew a direct historical line from IBM in the 1990s to Microsoft today, arguing that Microsoft's decision to stay off the frontier and instead build middleware for enterprises mirrors Lou Gerstner's turnaround strategy of embedding IBM as the trusted layer between legacy systems and the internet. "Microsoft is: we will help you figure out AI in a way where you're not giving away the crown jewels to these companies," he said, noting this is why Microsoft generated $20 billion of free cash flow and paid a $10 billion dividend last quarter despite not chasing frontier models. He called the strategy "rational" but also "desperate in an existential way," since Microsoft's entire enterprise software business, particularly the user-interface layer that products like Codex or Claude Code increasingly threaten, sits directly in AI's crosshairs.
By contrast, he argued it is Meta that faces the most interesting bind: "there's a very good case to make that it is actually more reckless to not be on the frontier if you're a digital company." He credited Mark Zuckerberg's willingness to blow up an existing research team and rebuild from scratch as a rare display of founder conviction, while flagging that Meta's core advertising business could actually benefit disproportionately from large language models, since Meta's ad marketplace functions as a live, global verification engine for AI-generated creative and predictive ad matching. He was critical, however, of Meta's public communication around advertising's societal value, a gap he says has persisted since Sheryl Sandberg's departure and one he believes has made it harder for the company to win investor patience for continued AI spending, particularly after roughly $100 billion in cumulative Reality Labs losses.
OpenAI's Late Pivot to Advertising
Thompson was direct in criticizing OpenAI's initial refusal to build an advertising business, arguing the company repeated Dropbox's early mistake at far larger scale by betting on consumer subscriptions instead. "They literally had OpenAI replaying the Dropbox story, but at a hundred-x size, being like, no, we're going to sell subscriptions to consumers. And they did. They sold a lot, but they didn't sell enough." He argues that had OpenAI leaned into advertising as soon as ChatGPT became a hit, both Google and Meta would be facing far more competitive pressure today, since advertiser-funded pricing has no consumer elasticity ceiling the way subscription pricing does. OpenAI has since begun rolling out ad-adjacent features, including retailer connections for purchase tracking, which Thompson says he is watching closely.
The Financing Gap and the Railroad Comparison
The macro warning threading through the conversation is about capital, not compute or power. Thompson noted the industry has moved rapidly down the funding stack, from free cash flow, to corporate debt, to Google issuing equity, calling the latter "shocking" given Google's balance sheet strength. "What's after that? Where's the money come after that?" he asked, warning of "a big blow up" if revenue generation doesn't catch up to capital deployment in time. He compared the moment to the 1870s railroad buildout, noting that railroads also required financing far ahead of returns and that "the world just ran out of money," yet the infrastructure ultimately proved transformative to GDP regardless. He pointed out the irony that Berkshire Hathaway's railroad profits are now funding its stake in Google's AI buildout, calling it "quite literal."
AI's Capability Limits and the China Framing
On the technology itself, Thompson said he remains unconvinced that AI's dominance in verifiable domains like coding and math generalizes cleanly to the much larger set of unverifiable, real-world economic tasks, though he stressed the economic opportunity in already-solved domains is still "massive" even without further model improvement. He was openly skeptical of Silicon Valley's geopolitical framing around China, calling the idea of durable US AI-driven military superiority "very problematic" and arguing that dependency on Chinese manufacturing is underappreciated and unlikely to be resolved outside of an actual conflict. "Everyone can use a good bogeyman," he said. "I think from the AI trade perspective nothing works better than we have to beat China." He described the current competitive gap, with the US frontier roughly six to nine months ahead of Chinese open-source models, as a reasonably stable equilibrium, though one whose durability he considers genuinely uncertain.