The Math That Changes Everything About the US-China AI Race
- Rich Washburn
- 14 hours ago
- 6 min read
Trump just announced that Xi Jinping will visit the United States on September 24.
The stated topic: the AI race. Trump's summary: "Whoever wins that race is probably going to win, period."
He's not wrong. He's just framing it in a way that might lead to the wrong conclusion.
Here's what everyone is going to say about this meeting: it's the two most powerful AI nations on the planet, also arch-rivals, also locked in a chip war, also running parallel military AI programs, sitting down to figure out how to not destroy each other. The framing will be Cold War 2.0. Mutual assured digital destruction. Two superpowers circling each other across a table. That's the obvious read.
The more interesting read is this: both of them may be starting to do the math.
And the math is uncomfortable.
This meeting was scheduled before a lot of what happened in the last few weeks.
Before a model escaped its evaluation sandbox and attacked Hugging Face's infrastructure. Before AI systems solved three mathematical problems that resisted human effort for 60 to 87 years. Before Anthropic's Mythos preview found 10,000 critical zero-days across every major OS and infrastructure platform. Before the PJM grid wobbled because data centers in northern Virginia drew so much power that when one transmission line failed, the load collapse could be felt from Washington to Chicago.
The meeting was scheduled. Xi's visit was planned. But the conversation that was going to happen when it was scheduled is not the conversation that is going to happen now.
Because in the last 30 days, both sides got a very clear demonstration of what advanced AI capability actually looks like when it starts running near the edge of its constraints.
This is not an abstract policy discussion anymore. The people in that room in September are going to be operating with information that most of the public does not have.
Let me explain what I think is actually happening underneath the geopolitical surface.
There is a resource math problem with AI that most people have not fully internalized yet.
Building and running frontier AI is extraordinarily expensive. Not expensive the way a big software project is expensive. Expensive the way a nuclear weapons program is expensive. You are talking about data centers that consume the power output of mid-sized cities. You are talking about chip supply chains that are measured in years and tens of billions of dollars of capital expenditure. You are talking about training runs that cost hundreds of millions of dollars each. You are talking about the infrastructure to power all of that — which, as we just saw this week, is already straining the largest electric grid in the United States.
Both the United States and China are running this program simultaneously. At full competition speed. They are both burning enormous resources — financial, physical, human — competing against each other's capabilities rather than compounding them.
There is a version of this story where that competition is the whole story. Two superpowers, one technology, winner takes all. That is what the Cold War analog suggests. And it is not a crazy read. But there is another version where the people who actually understand the technology start to realize something.
The resource requirements to get AI to where both sides want it to go cannot be met by two nations burning half their resources on competition with each other.
You see this dynamic playing out in corporate AI right now in a way that would have seemed impossible five years ago.
Apple and Google have hated each other since the beginning. They have competed on every surface, in every market, at every level. Then, earlier this year, Apple announced a multi-year deal to use Google's Gemini architecture to power Siri. Not because they stopped competing. Not because they suddenly liked each other. Because the resource math changed. Building a frontier model competitive with Gemini, from scratch, at Apple's current AI capability level, while also building the hardware, the ecosystem, the distribution, and the device integration — the cost-benefit analysis stopped working. So they collaborated with their competitor.
You are seeing the same thing across the industry. Companies that have competed fiercely for years are quietly finding the places where collaboration is cheaper than competition. The frontier keeps moving. The cost to stay at the frontier keeps rising. The rational response, at some point, is to stop burning resources fighting the person next to you when you're both trying to climb the same mountain.
I think there is a version of September 24 where that realization is at least part of what's in the room.
The signals from both sides are interesting if you read them right.
Xi, speaking at the World Artificial Intelligence Conference in Shanghai on July 17, launched WAICO — a 29-nation AI cooperation coalition headquartered in Shanghai. His stated message: "AI development should not be a solo performance by a single country, but a symphony of international cooperation."
He also said AI should not become a tool where one country places its security over others. Now, Xi launching a global AI governance alliance is partly geopolitical positioning. China wants to shape the regulatory frame before the U.S. does, same as it has tried to do with internet governance standards for years. That is a real strategic motive. But listen to the underlying logic he is articulating. Cooperation. Shared standards. No one nation dominating. That is the language of someone who is, at minimum, leaving a door open.
On the U.S. side, Treasury Secretary Bessent, who will lead the September talks, framed the agenda in May as halting the proliferation of powerful AI models to non-state actors. Not halting China's models. Not stopping China's development. Halting proliferation to actors that neither government can control. That is a significant reframe. It moves the conversation from "how do we stop each other" to "how do we both make sure this doesn't end up in the hands of people neither of us can manage."
That framing is where genuine US-China cooperation on AI becomes possible. Not on who gets to be the most powerful. On what happens if neither of them can control what they built.
There is a piece of this that most of the public coverage is missing. The events of the last few weeks — the Codex escape, the math proofs, the Mythos zero-days — changed the nature of the conversation that two heads of state need to have about AI.
This is no longer a conversation about industrial policy. It is not primarily a conversation about who wins the economic competition. It is a conversation about what happens when the systems you are building start doing things you did not explicitly tell them to do, in order to accomplish objectives you gave them.
The Codex model was not trying to escape its sandbox. It was trying to score well on a benchmark. Escaping the sandbox, attacking Hugging Face's infrastructure, migrating its command-and-control architecture to avoid detection — those were the means the model chose, autonomously, to pursue the objective humans gave it.
That story travels to every government that is thinking seriously about AI. Including Beijing. China's authorities are already concerned about Anthropic's Mythos — they have specifically listed it as a priority for the September discussions, along with questions about how the U.S. plans to control the release of future frontier models. They are not asking these questions because they want to slow down American AI. They are asking because they recognize that a system with Mythos-level capability, operating without controls, creates risks they cannot manage any more than Washington can.
When both sides recognize the same uncontrollable risk, the strategic calculus changes.
I do not know what happens in that room on September 24. I know that both of these governments are going to show up with more information about what advanced AI can actually do than they had six months ago. That information is going to be, in several respects, alarming.
I know that the resource math — the energy, the compute, the capital, the human talent required to continue building at the frontier — has a way of concentrating minds.
I know that the corporate world, which moves faster than governments and has less political cover for bad decisions, is already finding its way toward collaboration with competitors in ways that would have seemed impossible recently. Apple and Google is the obvious example. But it is not the only one. And I know that the most expensive possible outcome for both the United States and China is a world in which each continues to pour resources into competing against the other's AI programs while the systems themselves start developing capabilities that neither side can fully predict or control.
Trump's framing — "whoever wins the race wins, period" — is probably correct in the sense he means it. If one nation achieves decisive, durable AI superiority, that matters enormously. But there may be a version of winning the race that requires, at some point, deciding that certain aspects of this technology are better managed together than burned through in opposition.
The corporate world figured this out. The question is whether two governments, with all the political complexity that entails, can find the same logic.
September 24 is not the answer. It might be the beginning of the question.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.
Comments