AI, Honestly

This Episode

The Trump-Xi summit just wrapped in Beijing. Both leaders agreed on one thing: Iran cannot have a nuclear weapon. They agreed on almost nothing about AI. The framework they produced — a mutual incident notification system — is real, and thin. This was round one. But the US-China AI relationship isn't only playing out in summit rooms. It's in county permit hearings in Virginia and Ohio, in UN standards bodies, in fusion labs, and in 162 declassified files describing physics no one has explained yet.

Cold Open

Two Headlines From One Summit

Iran. Boeing. A Board of Trade. China buying oil from Texas and Louisiana — unconfirmed by Beijing. And AI: a notification framework that analysts called "significant for its existence, thin on its substance." Time called AI "the elephant in the room." This was round one. Trump doesn't show all his cards in round one.

The Summit

What Was Agreed, What Wasn't, and What It Means

Iran nuclear alignment is the most consequential result — and the most surprising, given China's economic relationship with Iran. The H200 chip story is more interesting than coverage suggests: the US cleared the sales before the summit. China told its firms not to buy. Substantive AI governance — what each nation's systems can do, and to whom — remains entirely unresolved. That is the largest open item between the two nations.

The Stakes

What Winning and Losing the AI Race Actually Means

Coverage says "the US has to win the AI race." Nobody explains what winning and losing mean. Kyle walks five stakes: military decision cycles, industrial productivity compounding, who writes global standards, who builds infrastructure for the Global South, and whose values are inside the model that answers the question. The summit addressed exactly none of it substantively.

The Home Front

The Race Is Happening at Your County Permit Hearing

PJM projects a 6 GW grid shortfall by 2027. Half of announced data center capacity isn't being built — five-year transformer lead times. And the American Edge Project documented Chinese and Russian state media messaging showing up near-verbatim at US local permit hearings in Maine, Ohio, and Oklahoma. The permit process is a leverage point. Foreign state media knows it. The REA didn't ask county boards whether they wanted electricity.

The Feel-Goods

Fusion, an Open Question, and the Career Story Nobody's Telling

AI found Rentosertib — Phase IIa trial, Nature Medicine — the first AI-designed molecule to show safety and efficacy in real patients. The same tool is now pointed at fusion energy. DuctGPT at Ames Lab. Stellar-AI at Princeton. DeepMind + Commonwealth Fusion targeting early 2030s. And 162 Pentagon files describing physical observables that current physics cannot explain. Nobody has asked the obvious question yet. Also: 340,000 unfilled data center jobs — and the people being recruited are nuclear engineers, aerospace cooling specialists, and military technicians.

Takeaway: The Five Stakes

"The US has to win the AI race" is the sentence everyone says. Nobody explains what winning and losing actually mean. Here's what's on the line — five stakes, in order of how often they get named in policy coverage.

Stake 1 · Military

Faster decision cycles — and a board that can say no

AI-assisted targeting, logistics, and intelligence surveillance compound over a long conflict. But the asymmetry isn't just capability — it's governance. Anthropic said no to Pentagon targeting work. OpenAI said yes. China's state-directed AI doesn't have an Anthropic situation. The question of what the AI is allowed to do gets answered by the state. One side in any conflict has a board that can say no. That's the asymmetry.

Stake 2 · Economic

A productivity multiplier that compounds over decades

China doesn't need better models. It needs good-enough models deployed faster, at scale, into real industrial workflows. DeepSeek R1 is free. State subsidies cover up to 50% of data center energy costs. On deployment speed and cost structure, the US lead in frontier model capability may matter less than assumed. This isn't a single-year GDP gap — it's a directional shift that determines which economy is larger in 2040.

Stake 3 · Global Standards

Standards encode values permanently

In July 2024, China secured adoption of its AI governance resolution by the UN General Assembly — 143 co-sponsors. In July 2025, Beijing unveiled a Global AI Governance Action Plan, positioning China not as a participant but as the world's convenor. The US went to a bilateral summit and got a notification framework. China got the UN. Everything built on a technical standard inherits its assumptions.

Stake 4 · The Global South

Five billion people — and who builds their infrastructure default

China is not trying to out-compete OpenAI in San Francisco. It's building the default AI infrastructure for the people who aren't in the US or the EU. State-backed financing, no human rights conditions, no governance strings. Chinese surveillance AI already runs in 140+ cities worldwide — cameras, facial recognition, social scoring — embedded in courts, hospitals, financial systems. Once a country builds on that architecture, the switching cost is enormous. You don't win infrastructure competition by having better values. You win it by showing up.

Stake 5 · Values in the Model

Every AI model is a values document

A Chinese model trained under state direction will not surface content challenging the Communist Party. It will not give a balanced answer on Taiwan. It will not acknowledge Tiananmen Square. Those are features, not bugs. When that model becomes the default health advisor in an Indonesian hospital — the educational tool in a Nigerian school — the legal research assistant in a Kenyan court — the values of Beijing's government are embedded in those institutions' daily operations. Not through coercion. Through convenience.

Three things to watch

The September 24th Washington meeting is round two. Watch what the US walks in with — on standards, on the Global South, on deployment. A strategy shows up differently than a photo op.
The infrastructure problem has an answer — treat it as a national coordination problem, not a county permit problem. The REA didn't ask county boards whether they wanted electricity. The answer to foreign state media in your county hearing is not more county hearings.
Watch what happens when someone points a serious research team at 162 files. Kyle's not saying it's aliens. He's saying the tool that found physics in plasma is available — and the question hasn't been asked.

Sources

The Cast

Kyle

Host. Opinionated. Expect a history drop.

Kate

The correspondent. Tight, sourced, no spin.

Morgan

The heartbeat. "Well, why though?"

Full Transcript

Loads on open

▶

Loading transcript...