In the first half of 2026 the AI infrastructure buildout has been firmly top of mind.1 The November 2022 launch of ChatGPT brought large-language models into the spotlight, and with each passing year you could feel the diffusion into nearly every conversation about the future of work, education, and culture. To feed the insatiable appetite for generative AI, the industry is making capital investments without recent precedent. From a recent column in The Economist, emphasis mine:
This year the five firms [Amazon, Google, Meta, Microsoft and Oracle] will spend $800bn filling warehouses with computers to run artificial-intelligence models... at around 40% of their revenues this year, the cloud giants’ capital expenditures will surpass those of the oil industry during the shale boom in the 2010s and the telecoms industry during the dotcom bubble in the 1990s.
Many are asking if this a speculative bubble that will collapse like the end of the dotcom boom, but this time with a technology sector that is a larger share of the economy and a much larger share of the S&P 500. Leaving aside the argument that bubbles are actually good by compressing technological progress to build otherwise impossible things, this investment reflects the sincere belief of these firms that generative AI will bring about dramatic economic changes. From Microsoft CEO Satya Nadella's November 2025 interview with Dwarkesh Patel:
In some sense this goes back again to, essentially, what’s the economic growth picture going to really look like? What’s the firm going to look like? What’s productivity going to look like? ...what took 70 years, maybe 150 years for the Industrial Revolution, may happen in 20 years, 25 years. I would love to compress what happened in 200 years of the Industrial Revolution into a 20-year period, if we’re lucky.
If this narrative is correct, this implies some disturbing vulnerabilities. If the current era is a second industrial revolution, the GPUs used to run AI models are as important as basic electrical transmission. And nearly all of those GPUs are produced in semiconductor fabrication plants (fabs) run by the Taiwan Semiconductor Manufacturing Company (TSMC). The only problem with this arrangement is that Taiwan is about 110 miles off the coast of a China that claims the de-facto independent island as its territory. America's CPU resilience is in a mediocre but acceptable place; unlike the rest of their rivals who contract out all manufacturing to TSMC, America's Intel owns many advanced semiconductor fabs in America. But it has struggled to compete in the latest generation semiconductor manufacturing technology and their GPU production is a rounding error as compared to Nvidia. And while TSMC now has a fab in Arizona, as of 2026 the facility is still dependent on packaging and final assembly based in Taiwan.
The Polymarket odds for a Chinese invasion of Taiwan by the end of 2027 stand at 17% as of 16 May 2026. The fragile fab equipment wouldn't make it through a war in one piece, and if Taiwan were to fall, America would likely sabotage TSMC lest it fall into Chinese hands. It's worth noting that the AI investment era did not create this vulnerability; if the island was captured in 2021 this would have meant a complete stop to iPhone production given that Apple contracts all of their CPU production to TSMC. But when taking future economic growth into account, the stakes have been raised: that $800bn in 2026 capex only pencils out if you can actually acquire GPUs for these new data-centres.
But there are many outcomes short of a Chinese invasion that the industry needs to be prepared for. AI model use is priced by the token; it varies between tokenisers but a text token is roughly a syllable. For risks both existential and less-than, having option contracts on AI token prices could provide some means of hedging and a stronger price signal than today's prediction markets. But everyone I've talked to on Wall Street who's in a position to know says that synthetically creating options against a basket of token prices can't be done without baking in some other asset prices. And writing literal options doesn't seem to have the blessing of regulators yet. While the marginal cost of a lot of CPU compute reduces with incremental users, the costs to use AI models scales linearly with tokens. This means companies built around roughly the existing token economics will be more sensitive to price changes than they are for CPU compute. Some firms that rely on LLMs to serve customers will be able to tolerate a 10% increase in frontier model pricing, but not all will. While technological improvements would suggest that token prices will go down over time, demand has not looked very price sensitive in the last two years. And if you are choosing between more model calls or hiring an employee, why would it be?
Putting this all together the technology industry has placed an incredibly large and difficult to hedge bet that A) the political status quo will be maintained in Taiwan and B) generative AI models will be a load-bearing part of the economy in the near future. While ordinary software engineers have little say in semiconductor manufacturing or geopolitics, there are things the profession can do to be less exposed to the great token wager.
Small editorial note: this post assumes less prior knowledge and explains more details than most others; I wanted this to be understandable to people who do not read Stratechery everday↩︎