Ben Thompson on Big Tech, China, and the AI Boom Running Out of Money - [Invest Like the Best, EP.487]

Summary

Ben Thompson and Patrick O’Shaughnessy dig into AI through the lens of geopolitics, capital markets, and industrial bottlenecks. Thompson argues that a U.S. “win” in AI could be strategically dangerous, that China’s dependence on TSMC and other chokepoints makes the situation unstable, and that the current frontier may be sustainable longer than people think. He also says the binding constraint may be capital rather than compute: AI buildouts have long lead times, inference is expensive, and the industry could run out of money before revenues fully mature. From there the discussion turns to where value capture is likely to land—Google and Meta’s ad businesses, Amazon’s internal-first hardware flywheel, Apple’s customer access, Microsoft’s enterprise middleware strategy, and Nvidia’s hidden financing exposure. The throughline is that AI may be transformative, but its near-term winners will be determined by economics, infrastructure, and commodity-like supply dynamics rather than pure model performance.

Notes

• Thompson says a U.S. AI victory could be strategically dangerous if it created overwhelming military and national-security superiority.
• He thinks U.S. dependence on China for fabs, actuators, and precursor manufacturing is underappreciated.
• He frames decoupling as economically unrealistic unless forced by conflict or other hard constraints.
• He compares the AI geopolitical equilibrium to Taiwan: tense, but potentially sustainable longer than expected.
• He places OpenAI and Anthropic at the frontier, with Google, Grok, and Meta chasing.
• He says Chinese teams are likely 6 to 9 months behind via model distillation.
• He distinguishes free R&D from paid inference: open-source models still incur real marginal serving costs.
• He warns that inference can be expensive enough to matter materially in product economics.
• He worries the industry may hit a timing gap where capex outruns revenue and capital gets scarce.
• He says the AI financing stack moved from free cash flow to debt markets in about a year.
• He analogizes AI infrastructure to the railroads because buildouts take years while financing comes due immediately.
• He says the nearest-term AI question may be whether there is enough money, not just enough compute or electricity.
• He interprets Google’s equity issuance as symbolically important for how the AI capex game is being funded.
• He compares Google Search to See’s Candies: high-margin, but with limited reinvestment runway.
• He argues Google’s move toward AI may prioritize larger absolute profits over prior ultra-high margins.
• He is bullish on AI overall but skeptical that general-purpose reasoning is proven in messy, unverifiable domains.
• He says coding and math are clear AI strengths because they are verifiable domains.
• He sees medicine as a huge AI opportunity, but one constrained by regulation and access to records.
• He argues consumer AI should be monetized with advertising because consumers resist paying.
• He says consumers do not want to pay for software or be productive, reinforcing ads as the likely consumer model.
• He says Microsoft’s usage-based enterprise pricing is awkward because it breaks the familiar headcount-linked software budget model.
• He notes Microsoft’s E7 plan as roughly $100 per user per month plus usage charges.
• He says AI compute shortages are partly the result of underinvestment in 2023 and 2024.
• He says new AI capacity financed today mostly shows up in 2028 and 2029.
• He says AI infrastructure should be viewed like a commodity market with sunk costs and marginal pricing.
• He warns that the AI boom can be transformative and still hit a financing gap before revenues arrive.
• He says memory is a classic boom-bust commodity market with huge upfront capex and delayed capacity.
• He says leading-edge memory fabs are now in the billions of dollars because they rely on EUV machines.
• He cites Samsung’s strategy of investing through downturns to beat Japanese memory rivals.
• He says the memory market has consolidated to a disciplined oligopoly of three players.
• He argues TSMC’s conservatism effectively pushed risk onto big tech customers.
• He says the relevant risk often shows up as foregone revenue, not just losses.
• He sees Intel’s foundry push as a rational response to TSMC dependence and scarcity.
• He says compute scarcity ultimately helped Intel by making a foundry partner more valuable.
• He says Amazon is compelling because it is its own first-best customer.
• He says Amazon can iterate silicon inside managed services like Redshift without customers even seeing the chips.
• He thinks that internal scale gives Amazon a real hardware development advantage.
• He says Apple’s ecosystem means suppliers come to Apple because it owns customer access.
• He says Apple’s physical-goods distribution and retail footprint make it more insulated than software-only AI firms.
• He contrasts Apple’s deterministic product model with AI’s probabilistic nature.
• He argues OpenAI and Anthropic have the biggest upside and riskiest footing because of their mission intensity.
• He says it is often more reckless for digital firms not to be on the frontier.
• He characterizes Microsoft’s strategy as rational but non-frontier, which leaves more cash flow available.
• He says AI threatens Microsoft’s systems-of-record moat because AI is good at repetitive migration work.
• He says the biggest AI monetization may be incremental gains inside Google and Meta rather than the frontier labs.
• He says a few percentage points of ad improvement can create billions in value.
• He points to Facebook advertising as the clearest commercial use case.
• He says bandwidth became world-changing because it became a commodity with free marginal distribution.
• He says Nvidia’s margins are not truly intact once financing backstops are included.
• He says Nvidia’s Neo cloud backstops and equity stakes lower customers’ cost of capital by shifting risk.
• He says the expected value of those financing exposures is economically equivalent to a price cut.
• He says Google and Amazon may become Nvidia’s biggest long-term chip competitors.
• He says hyperscalers sell chips as commodities, not differentiated products.
• He says the hyperscalers’ scale and lower cost of capital are their key advantage over Neo clouds.
• He says CUDA’s moat is weakened because models care less about hardware choice.
• He argues power scarcity would favor Nvidia if token efficiency becomes decisive.
• He says the U.S. has already brought more power online than expected through behind-the-meter generation, West Texas gas, and nuclear restarts.
• He argues the most durable bubble-era asset may be power infrastructure, not GPUs.
• He points to dark fiber and railroad infrastructure as examples of durable value left behind after bubbles.