Ben Thompson argues a decisive American AI and military lead would be dangerous because China's best response could be attempting to blow up TSMC.
Thompson warns AI return on investment may not generate enough revenue to fund spending, comparing it to the 1870s railroads that ran out of money despite lasting value.
Thompson argues Nvidia's guarantees and backstops amount to a hidden price cut, with hyperscalers like Google and Amazon posing its biggest long-term threat.
Thompson Warns the AI Boom May Run Out of Money
Ben Thompson argues the AI buildout faces a funding problem. Spending is racing ahead of revenue, compute behaves like a boom and bust commodity, and inference carries real marginal costs that search and ads never did. Each big company is positioned differently, and TSMC sits at the center of the geopolitical risk.
Takeaways
AI spending may outrun the revenue funding it. Patrick cites about $800bn of AI capital expenditure this year and $1.3tn next year. Tech has already moved through debt markets and Google is issuing equity. Thompson compares it to 1870s railroads, which ran out of money yet kept operating.
Compute behaves like a boom and bust commodity. Container ships take two years to build, and prices rose from about $3,000 to $17,000 during COVID. Samsung wiped out Japanese memory rivals by investing through downturns. TSMC has offloaded its overcapacity risk onto big tech, which bears the cost of foregone revenue.
Inference costs break the aggregator playbook. Search and ads had near-zero marginal costs, while AI inference has real ones, and open-source models still cost money to run. Microsoft's $100 per user E7 plan with usage charges breaks the link to headcount, and consumers resist paying for software.
A decisive US AI lead would be dangerous. China's best response to a decisive American lead would be trying to blow up TSMC. Today's equilibrium favors the US, with Chinese labs distilling models to stay six to nine months behind. Manufacturing dependence on China will not be fixed outside a conflict.
Google may need Berkshire's profit-dollars mindset. Berkshire moved from maximizing margins to maximizing absolute profit dollars, as with BNSF free cash flow. Google's very high margin search business may need the same thinking for AI, which is why its equity issuance matters.
Nvidia faces pressure from cheaper-capital hyperscalers. Nvidia's guarantees and backstops work as a hidden price cut. Google and Amazon have a lower cost of capital and are its biggest long-term threat. Thompson says CUDA's moat has shrunk, and power may become the real constraint.
Thompson's read on each major player
Amazon has the most interesting setup because it builds for itself first, as AWS did. Its Trainium chips now run Anthropic workloads.
Apple benefits from owning customer access. Thompson says it is fine for Apple not to lead in AI because it excels at deterministic products.
Microsoft is not on the frontier and is running an IBM-style middleware playbook, backed by $40bn of free cash flow last quarter.
Meta is the most interesting frontier lab because of Mark Zuckerberg's founder energy.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
OpenAI and Anthropic are the riskiest bets with the biggest upside, and SpaceX AI is the weakest.