In the last newsletter we looked at what everyone else is building. France, India, Korea, Japan, the Emirates, each one picking the two or three pieces of the AI stack it could actually win and building around those. The piece ended on a claim. The US and China aren't ahead because their engineers are smarter. They're ahead because they control most of the supply chain. In this newsletter edition we explore this claim.

The reality is the US and China don't control the whole thing either. They control the ends. The middle belongs to three other countries, and both superpowers depend on all three.

The two halves

Let's start with what each side owns.

The US owns the high-value front end. Nvidia designs the chips almost everyone trains on, and it controls CUDA, the software layer that makes those chips usable. Add the cloud platforms from Amazon, Microsoft, and Google, plus capital markets deep enough to finance data centers the size of small towns. That's a commanding position in design, software, and money.

China owns the back end. Not the fanciest chips, but scale, electricity, and the ability to point state banks, provincial governments, utilities, and state-owned companies at the same goal at the same time. Where the US builds by convincing private companies to spend, China builds by deciding to build.

So we've got two very different superpowers, with very different approaches, and the interesting part is what neither of them makes.

The three chokepoints

An advanced AI chip passes through three countries that are neither American nor Chinese, and there's no quick way around any of them.

Taiwan does the manufacturing. TSMC manufactures most of the world's leading-edge chips, including nearly all of Nvidia's. Its share at the cutting edge sits north of 85 percent. If you want the best chips made, they're made in Taiwan, whether you're in California or Shenzhen (or Alabama).

South Korea makes the memory. Advanced accelerators need high-bandwidth memory sitting right next to the processor, and that market is essentially SK Hynix, Samsung, and America's Micron. SK Hynix alone runs somewhere around half of it. It's one of the hardest parts of the whole chain to produce, and China barely factors in.

Then there's the Netherlands, which nobody thinks about in an AI conversation. The most advanced chips require a process called extreme-ultraviolet lithography, and exactly one company on earth makes the machines that do it. ASML, in the Netherlands. It's not the leading supplier, it's the only supplier. Each machine depends on thousands of specialized parts from Germany, Japan, and the US, and China has never been allowed to buy one. In the first quarter of 2026, ASML shipped 42 of these systems. Twenty-two went to Taiwan, eleven to South Korea, six to the US. Zero to China.

So the "US versus China" framing is a little off. Both countries design chips. Neither can make an advanced one without a Taiwanese fab, Korean memory, and a Dutch machine that China can't even purchase. The race isn't two runners pulling away from the pack. It's two assemblers of the same global supply chain, each trying to build domestic replacements for the parts they don't control. Neither has managed it yet.

The money nobody counts

You've probably seen the number that says America is running away with this. Stanford's 2026 AI Index put US private AI investment at $285.9 billion in 2025 against $12.4 billion in China. Twenty-three to one.

Stanford also warned, in the same report, that the number is misleading. It measures private investment, venture rounds and corporate deals, the stuff that shows up cleanly in a database. That's most of the story in America, where four companies alone are on track to spend somewhere north of $700 billion on AI infrastructure in 2026, nearly all of it private money.

In China, private investment isn't where the money mostly comes from. China funds its buildout through channels a venture chart never sees. State banks lending below market rates. Provincial guidance funds. Electricity sold to data centers at a discount, in some energy-rich provinces cut by as much as half. One government semiconductor fund, the third in a series, raised about $47.5 billion by itself. None of that lands in the private-investment column, which is exactly why the 23-to-1 figure undercounts by design. Nobody can produce a clean total for China's state spending, that's part of the point, but it clearly runs into the tens of billions a year on top of what the charts capture.

There's a tradeoff underneath it all. American money is fast to form and prone to speculative excess. Chinese money moves on political command and builds redundant capacity, hundreds of local computing centers, some running at under 30 percent utilization, that Beijing started reining in this summer. They're different failure modes without a clean winner.

The Workarounds

Cut off from the best chips, China has leaned on efficiency. Smaller models trained more cleverly, which is where next week's piece picks up. Some of that efficiency also comes, according to US officials, from copying.

In early 2026 US labs and then the White House accused Chinese firms of "adversarial distillation," systematically extracting the behavior of American frontier models through fake accounts and industrial-scale querying. Anthropic alleged that three Chinese labs used roughly 24,000 fraudulent accounts against Claude. It's worth being plain about what these are. Allegations, with documented suspicious activity behind them, but no independent finding that any specific Chinese model was built this way. Ordinary distillation is a normal technique. Proving a rival copied a model rather than trained a good one independently is genuinely hard, and the public evidence hasn't closed that gap.

Where it actually stands

On model quality, the gap has nearly closed. Stanford's early-2026 index put the best US model only about 2.7 percent ahead of the best Chinese one, and by this summer the top few systems worldwide sit within a rounding error of each other. American models still trade the overall lead, Chinese models top the open-weight rankings, and the edge now shows up in cost and reliability more than raw score (more on this in the next newsletter). It's no longer accurate to call Chinese systems a generation behind.

On compute and hardware, the US still leads by a lot. More frontier chips, bigger clusters, more reliable ones.

On supply-chain independence, neither side is close. The US leans on Taiwan, Korea, and the Netherlands. China leans on foreign tools and memory plus whatever it can build or acquire. China's remaining gap is physical, not intellectual. It's chips, yields, memory, and machines, not ideas.

That's why these two lead and everyone else watches. It isn't that either one built the whole machine, it's that each controls enough of it, capital and demand on one side, power and coordination on the other, to organize an entire ecosystem around the parts they buy from someone else.

In Part 3, we'll look at the small models, and why the most interesting corner of this race is the one that needs the least of everything above.

Know someone who’d enjoy this?

Don’t miss the next issue.