Nvidia’s Risky Business (Stratechery Article 8-11-2026)

---

1. Key Themes

Theme 1: The AI Infrastructure Buildout Is Increasingly Debt- and Equity-Funded — and the Risk Pool Is Widening

The hyperscalers have moved well beyond using free cash flow to fund AI infrastructure, and the capital sources are becoming progressively riskier and less conventional.

"After raising a combined $108 billion in all of 2025, these four companies [Oracle, Meta, Alphabet, Amazon] have, as of July 7, already raised $194 billion this year. Unsurprisingly, spreads are rising, and 86% of the bonds issued this year are already trading at higher yields than at issuance. Cover for recent issuance has fallen to less than 2x, from 5x in February."

The parallel to 1870s railroad financing is explicit and alarming:

"The $500 million that went into U.S. railway bonds annually during the boom years of the early 1870s would today be the equivalent of $600 billion, roughly what is projected to be invested by major tech companies in 2026."

---

Theme 2: Nvidia Is Creating a New Asset Class to Sustain GPU Demand — But at the Cost of Absorbing Real Risk

Nvidia is partnering with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to unlock $500B+ in third-party capital for AI infrastructure, reframing GPUs as productive infrastructure assets.

"We have moved from an era in which companies bought chips and built data centers project by project to one in which AI factories can be financed as productive infrastructure — with repeatable platforms, long-term institutional capital and a diverse customer base that uses compute to create revenue."

But Thompson notes Nvidia isn't entirely off the hook:

"The company is backstopping opportunities with up to 25% residual-value based financing, suggesting that Huang believes his 'investable asset class' pitch much more than the market does. That is, in a certain sense, a price cut."

---

Theme 3: Google Is Pivoting from Frontier AI Lab to AI Infrastructure Monopolist

DeepMind's leadership exodus (Hassabis, Dean) and Google Cloud's explosive growth (82% YoY) signal that Google is deprioritizing frontier model competition in favor of dominating the compute layer.

"Whether or not Google is competing for the frontier, they are absolutely competing to dominate AI infrastructure. And, in a world where intelligence is a commodity, TPUs in particular are a big deal."

SemiAnalysis crystallizes the internal power shift:

"More than 20% of total TPU shipments from 3Q26 to 4Q27 are being sold directly to Anthropic... Whereas Gemini and GCP used to desperately fight for compute allocation, it's now clear that Thomas Kurian won."

---

Theme 4: TPUs (and Custom Silicon) May Be a Durable Cost Advantage Over Nvidia GPUs

Anthropic's decision to convert compute costs from marginal to capital (buying TPUs outright for its own data centers) suggests TPUs are competitively cheap enough to be strategic infrastructure.

"It seems likely that TPUs are cheaper than Nvidia GPUs; Anthropic may have built for TPUs (and Amazon's Trainium chips) because only Google and Amazon had the wherewithal to fund them, but at this point that ability may very well be a significant advantage."

"The fact that Anthropic is straight up buying TPUs for its own data centers (converting compute costs from marginal costs to capital costs) suggests that is the case."

---

Theme 5: CUDA's Moat Is Eroding as the AI Stack Moves Up the Abstraction Layer

The frontier labs are the most important Nvidia customers, and they're increasingly departing from CUDA dependency — threatening Nvidia's software lock-in precisely when it matters most.

"Anthropic has not been dependent on CUDA for years, and OpenAI is moving in that direction, at least for inference. If those companies win then Nvidia's profits will be squeezed."

Thompson had flagged this structural vulnerability as early as 2024:

"The use cases for those GPUs is happening at a much higher level than CUDA frameworks (i.e. on top of models); that, combined with the massive incentives towards finding cheaper alternatives to Nvidia, means both the pressure to and the possibility of escaping CUDA is higher than it has ever been."

---

2. Contrarian Perspectives

Perspective 1: Google Losing the Frontier Model Race May Actually Be Good for Google's Business

The consensus view is that DeepMind's leadership collapse is a disaster for Google. Thompson argues the opposite — surrendering the frontier model competition may free Google to dominate the more profitable infrastructure layer.

"What is fascinating about Google's position is that these machinations do not necessarily mean the Berkshire Hathaway bet was a bad one; indeed, it's arguably good news."

SemiAnalysis supports this: Google Cloud revenue growth actually accelerated after the internal power shift to Kurian, from 32% YoY a year ago to 82% YoY last quarter, with margins expanding from 21% to 36%.

---

Perspective 2: Nvidia's New Financing Partnerships Are a Stealth Price Cut and Admission of Demand Risk

The market will likely read the Apollo/BlackRock/KKR announcement as a bullish expansion of Nvidia's ecosystem. Thompson reframes it as a sign that organic demand has hit a ceiling and Nvidia is absorbing risk to stimulate it.

"Whereas equity dilutes the upside for investors without adding risk to the company, this structure preserves Nvidia's margins by finding new pools of capital willing to bear risk... The implication of that backstop is [Nvidia's profits] already are [being squeezed]."

The signal: Nvidia wouldn't need to backstop 25% of residual value if buyer confidence in GPU-as-asset were truly strong.

---

Perspective 3: The Most Dangerous Capital in This Cycle Isn't the Hyperscaler Bonds — It's the Insurance Floats and Pension Money Coming In Last

The 1873 analogy points to the Jay Cooke dynamic: the most catastrophic damage came when risk was distributed to retail investors and safety-seeking capital. Thompson sees the asset manager partnerships as a direct analog.

"To the extent Nvidia competes through novel funding mechanisms that, at the end of the day, draw on things like insurance floats and pension funds and other long-run liabilities that are the bread and butter of the asset managers the company is partnering with, the risk — unmarked, unlike equity — is considerably higher."

---

3. Companies Identified

---

4. People Identified

---

5. Operating Insights

Insight 1: Frontier AI Companies Should Prioritize Custom Silicon Dependency Over CUDA Convenience

The companies pulling away from Nvidia (Anthropic on TPUs, OpenAI moving for inference) are gaining a structural cost advantage. For operators building AI products at scale, evaluating non-CUDA inference paths isn't just a cost exercise — it's a strategic hedge against being squeezed when Nvidia's pricing power reasserts itself.

"Anthropic has not been dependent on CUDA for years, and OpenAI is moving in that direction, at least for inference. If those companies win then Nvidia's profits will be squeezed."

Insight 2: Converting Compute from Marginal Cost to Capital Cost Is a Signal of Conviction — and Competitive Advantage

Anthropic's move to buy TPUs outright rather than rent them signals that when you have high conviction in sustained compute demand, owning infrastructure collapses your long-run cost structure. Operators with predictable inference workloads should model this lease-vs-own calculus carefully.

"The fact that Anthropic is straight up buying TPUs for its own data centers (converting compute costs from marginal costs to capital costs) suggests that [the cost advantage of TPUs] is the case."

Insight 3: Platform Positioning Beats Vertical Integration When Demand Is Uncertain

Google Cloud's willingness to sell TPUs to Anthropic — a direct Gemini competitor — is generating explosive revenue growth (82% YoY) and margin expansion (21% → 36%). Operators building platforms should consider whether "selling to competitors" is actually the highest-return posture when a market is uncertain.

"As a platform player, we have to allow our technology to be monetized in as many ways as possible and we don't see it as a zero sum."

---

6. Overlooked Insights

Insight 1: Bond Market Demand for AI Infrastructure Debt Is Already Deteriorating Badly

The article notes that bond market oversubscription ("cover") has collapsed from 5x in February to under 2x today, and 86% of bonds issued this year are already trading at higher yields than issuance — a sign of forced selling or mispriced risk. This is a leading indicator of credit stress that is buried beneath the headline narrative about Nvidia's financing innovation.

"Cover for recent issuance has fallen to less than 2x, from 5x in February."

Insight 2: DeepMind Had a Working AI Chatbot a Full Year Before ChatGPT but Was Blocked from Releasing It

Briefly mentioned as evidence of Google's bureaucratic culture rather than Hassabis's failure, this fact reveals the magnitude of opportunity cost from organizational risk-aversion in AI — and implies that incumbent culture, not capital or talent, may be the binding constraint on AI competitiveness.

"Remember that DeepMind had an AI chatbot 1 year before ChatGPT but was not allowed to release it due to fears of disrupting their core business."