Anthropic Wants America to Win the AI Race. Its Strategy Makes That Impossible.
China Spends 23x Less on AI. Its Models Are Almost as Good. And They’re Free. That’s the Problem Anthropic Won’t Solve.
On May 14, 2026, while Donald Trump was shaking hands with Xi Jinping in Beijing with the CEOs of Nvidia, Apple, Tesla, and BlackRock standing behind him, Anthropic published the most aggressive policy essay the company has ever written. The timing was not an accident. The title gives it away: “2028: Two Scenarios for Global AI Leadership.”
Their argument, stripped of diplomatic padding: the United States and its allies must maintain their lead in artificial intelligence over China, because the alternative is an authoritarian regime writing the rules of the most transformative technology in human history. They say the Chinese Communist Party is already using AI for censorship, surveillance, and repression, and that if China reaches the technological frontier, it will be worse.
Their diagnosis is excellent. Their proposed solution has a hole in it large enough to drive the entire open-source ecosystem through.
Why 2028 is the Line?
The essay, when read between the lines, makes Anthropic’s logic evident, even if not clarified. 2028 is their internal estimate for when AI reaches the point of recursive self-improvement: systems capable of accelerating their own development without human intervention. Once that threshold is crossed, the first to arrive takes a potentially irreversible lead. Better AI produces better AI faster, and the gap compounds with every cycle.
That is the finish line.
Not a benchmark score, not a revenue number.
The moment when the technology improves itself.
On that specific point, the essay’s argument is structurally sound. Anthropic published the paper calling on the US to lock in a 12-to-24-month compute lead by closing chip-smuggling loopholes, restricting remote data center access for Chinese firms, and containing what they call distillation attacks — where foreign actors extract the knowledge embedded in American models by training smaller models on their outputs.
The computational advantage, they argue, is decisive. The United States controls the hardware pipeline through Nvidia, TSMC in Taiwan, and ASML in the Netherlands, the only company that makes the extreme ultraviolet lithography machines required for cutting-edge chip fabrication. China cannot manufacture these chips domestically and does not currently have access to the most advanced fabrication technology. Export controls are designed to keep it that way.
On paper, that is a crushing structural advantage. The story is already more complicated.
The Numbers That Undercut Anthropic’s Own Argument
While Anthropic was publishing its essay, the US government was authorizing the sale of 75,000 Nvidia H200 chips per company to roughly ten Chinese firms, including Alibaba, Tencent, and ByteDance. Beijing’s response was not to buy. A quiet directive from Chinese authorities encourages domestic firms to put local chip suppliers ahead of American ones.
That response is exactly the scenario Anthropic’s essay describes as catastrophic: on the day the Chinese ecosystem no longer needs Nvidia, the United States loses its primary leverage. And the data suggests the gap is closing faster than the essay acknowledges.
The Stanford AI Index 2026, a 423-page annual report from Stanford’s Institute for Human-Centered AI and the most credible independent assessment of the global AI landscape, quantified the situation in terms that are hard to dismiss. The performance gap between the best American AI model and the best Chinese AI model has collapsed to 2.7%. In May 2023, that gap was between 17.5 and 31.6 percentage points. As of March 2026, Anthropic’s Claude Opus 4.6 leads with an Arena score of 1,503. ByteDance’s Dola-Seed-2.0-Preview sits at 1,464. Thirty-nine points apart.
Half the Planet Uses AI. Stanford Published the 400-Page Report That Explains What Happens Next.
Here’s a piece of data I want to share, which I suspect most individuals haven’t internalized.
The spending disparity makes those numbers genuinely striking. US private AI investment reached $285.9 billion in 2025. China’s was $12.4 billion.
Twenty-three times fewer. And the performance gap is 2.7%.
The Stanford report adds context that goes further. China holds 69.7% of global AI patents. It dominates AI publication volume. It installs nine times more industrial robots than the United States. And the migration of AI researchers to the United States has dropped 89% since 2017. America holds two assets: money and chips. If it loses the battle for adoption, those assets alone may not be enough.
Anthropic’s own essay concedes this point, almost in passing. According to their writing, China can make up for its limitations in raw intelligence by more quickly incorporating less advanced models into its economy, doing so more affordably, and exporting them assertively. They acknowledge that adoption can matter as much as performance.
That admission undermines their central thesis on their own terms.
The Open Source Contradiction
This is where Anthropic’s argument collapses structurally.
Their third proposed solution, after strengthening chip export controls and protecting American models against distillation, is to promote the global export of American AI. On its face, this is a coherent strategy. If you want American AI to become the world’s default infrastructure, you need the world to adopt it.
But Anthropic is against open source.
Their essay explicitly redefines open-source AI as a national security risk and frames fine-tuning on frontier model outputs as a form of industrial espionage. The business model is built on closed, proprietary access. Their most capable model, Mythos, is restricted to a private consortium of twelve named partners. They will not release it publicly, and they have not given the European Union meaningful access to it, despite the European Commission confirming four or five meetings with Anthropic that produced nothing concrete.
Meanwhile, OpenAI has offered the EU access to GPT-5.5 CyberNet, their defensive cybersecurity variant.
The structural problem is straightforward. A developer in Europe will pay $30 per million output tokens to run Claude when a comparable open-source Chinese model is free. Not because they have a political preference, but because it is economically the only viable option for most workloads. When a Chinese open-source model does 95% of the job at one-tenth the cost, the conversation about geopolitical alignment becomes academic soon.
This is not a theoretical risk. It is already happening. Given that Chinese AI models are free and perform nearly as well, European companies and developers in the Global South find them an obvious choice for implementation. This is exactly how you lose a race for influence: not by having the less capable model, but by having the less available one.
Anthropic wants America to win. But it refuses to deploy the one distribution strategy that actually works at a global scale. Keeping models locked behind proprietary access and expensive API pricing while the competition gives theirs away is a strategy that protects Anthropic as a company. It does not protect the American position in this race.
The Pentagon Problem
The credibility issue does not stop at open source.
On May 1, 2026, the Department of Defense signed agreements with eight companies to deploy AI on classified networks: Google, OpenAI, Nvidia, Microsoft, Amazon, SpaceX, Oracle, and Reflection. Anthropic was not among them. The company had been classified as a supply chain risk for the defense industrial base, a designation normally reserved for adversarial foreign entities.
The reason: Anthropic refused to give the Pentagon unrestricted access to Claude without safety guardrails. They would not provide a version of the model without the limitations that constrain the consumer product. The Pentagon wanted full authorization for any lawful use. Anthropic said no.
On the level of principle, that position is defensible and arguably admirable. It presents a substantial problem regarding strategic credibility. When the same company that refused to work with its own country’s defense establishment without conditions publishes a 20-page essay arguing that American AI must dominate for reasons of global security, the tension is visible.
You can respect the ethical stance. But it makes it harder to position yourself as a strategist for national defense when you are currently on a quasi-adversary list maintained by the institution you are trying to advise.
What Europe Sees From the Stands
Here is the part that matters most if you are reading this from Belgium, Germany, or anywhere in the EU.
Europe does not have a frontier model. It does not have an Nvidia, nor does it have a domestic chip fabrication ecosystem. It depends on both Washington and Beijing for the infrastructure that will define the next decade of economic and security capabilities. In this contest between two superpowers, Europe is watching from the stands.
Anthropic writes an essay saying American AI must protect democracies. The largest democracy allied with the United States comes asking for help with cybersecurity. The answer, so far, is “We’ll talk later.” That gap between rhetoric and action is what determines where adoption goes next, and adoption is the game.
It’s not guaranteed that the nation with the top AI model will emerge victorious in the race for AI leadership. It will be won by the country whose model is the most used. And right now, the United States is spending 23 times more than China and holding a 2.7% performance lead, while China is open-sourcing its technology for anyone to use tomorrow morning.
The next two years will settle this.
Not in ten years, not in five.
The window Anthropic describes in their essay is real. Whether their strategy for keeping it open actually works is the question they have not answered yet.
Thanks for reading. Share your thoughts in the comments so we can discuss this sensitive topic where AI is a bridge between tech and geopolitics.