On May 15, 2026, as President Donald Trump touched down in Beijing for a high-stakes summit with Xi Jinping, Anthropic —the company behind Claude— released a document that reads less like a research paper and more like a foreign policy brief on corporate letterhead.
Titled "2028: Two Scenarios for Global AI Leadership," the paper paints two possible futures for the US-China AI race. In one, democracies hold the lead. In the other, the Chinese Communist Party (CCP) surges ahead and shapes global AI rules in its own image. The message is unequivocal: the window to act is measured in months, not years.
The Timing Is No Coincidence
The paper dropped —deliberately or not— during Trump's Beijing trip, the first by a US president in years. Along for the ride was Nvidia CEO Jensen Huang, eyeing a thirty-billion-dollar deal to sell H200 chips to Chinese firms.
The Trump administration had already authorized those sales back in December 2025. This very week, reports confirmed that ten Chinese companies —including Alibaba, Tencent, and ByteDance— received the green light to buy the H200, Nvidia's second-most powerful AI chip. Chinese tech stocks surged.
For Anthropic, that timing is a five-alarm fire.
The Compute Argument
The core of Anthropic's case is simple: the most critical ingredient for developing frontier AI is access to advanced chips —"compute"— and America controls it. Export restrictions, they argue, have been "incredibly successful."
Citing CFR data, Anthropic projects Huawei will produce just 4% of Nvidia's total processing performance in 2026, and only 2% in 2027. That gap, they say, is what's holding China back.
But here's the problem: Chinese labs are only a few months behind on model intelligence. How, if the hardware is locked down?
Anthropic has two answers: loopholes and distillation.
Loopholes: Chips in the Cloud, Not at the Border
US law regulates chip sales, not remote access to them. The Financial Times reported that Alibaba and ByteDance are training frontier models on export-controlled American silicon housed in Southeast Asian data centers.
A bill to close that loophole passed the House 369 to 22 in January 2026. It's been sitting in the Senate ever since.
Distillation: Industrial Espionage or Business as Usual?
This second point is more controversial. Anthropic accuses Chinese labs of running "large-scale distillation attacks": thousands of fake accounts systematically extracting outputs from American frontier models to clone their capabilities.
Back in February 2026, Anthropic had already named DeepSeek and other Chinese rivals as practitioners of this technique. Now they want Congress to explicitly legislate it as illegal.
The irony was not lost on critics: an AI company that trains on content created by millions of humans —and paid $1.5 billion to settle copyright lawsuits— is accusing competitors of "copying." The Register called it a "staggering lack of self-awareness."
Moreover, "scraping an API to train a smaller model" describes a fair chunk of what the AI industry has done to itself for years. Anthropic's counter: the difference is scale, state involvement, and strategic stakes.
The Two Scenarios for 2028
The paper lays out a binary contrast:
Scenario 1: Democracies Lead
The US defends its compute advantage. Policymakers close export loopholes, disrupt distillation attacks, and accelerate AI adoption across allied nations. Democracies set global AI rules and norms. Under this scenario, Anthropic estimates a 12-to-24-month lead in frontier capabilities by 2028.
Scenario 2: China Takes the Lead
Washington fails to act. Chinese labs exploit the gaps, catch up to the frontier, and eventually overtake it. AI norms are shaped by authoritarian regimes. The most capable systems enable what Anthropic calls "automated repression at scale." And all of it, they note, "on the back of American compute."
The paper pulls no punches about how the CCP already uses AI: speech censorship, mass surveillance in Xinjiang with facial recognition and biometric data collection, offensive cyber operations, and the export of repression technology to other autocracies.
If the CCP leads frontier AI, Anthropic warns, the world's most capable systems will enable automated repression at a scale human enforcers could never achieve alone.
The "Gatling Gun" Moment
Perhaps the paper's most effective anecdote involves Mozilla Firefox.
In April 2026, Anthropic shared a model called Mythos Preview with select partners through Project Glasswing. Firefox's security team used it to find and fix more security bugs in a single month than in all of 2025 combined —roughly twenty times their monthly average.
The paper quotes a Chinese cybersecurity analyst reacting to the release: China is "still sharpening our swords while the other side has suddenly mounted a fully automatic Gatling gun."
The implicit argument: if a single model release can produce that kind of step change, a three-month gap doesn't stay a three-month gap for long.
What Anthropic Wants (And What It Doesn't Say)
The recommendations boil down to three items:
- Tighter chip export controls, including foreign data center access
- Legal designation of distillation attacks as illegal
- Promotion of the American AI stack abroad so allies adopt US technology, not Chinese
A study by the Institute for Progress cited in the paper estimates that with stricter controls, the US sector would have roughly eleven times more compute than China.
Now, the part Anthropic says more quietly: tighter export controls directly benefit American frontier labs. Anthropic is one of them. A restricted Chinese market means one fewer competitor.
But dismissing the paper as mere corporate protectionism would ignore that the CCP is already using AI for purposes democracies find unacceptable, and that a neck-and-neck race between American and Chinese labs could lead both sides to cut safety corners for fear of falling behind.
The Criticisms: It's Not Black and White
Anthropic's paper has drawn fire from multiple angles:
The DeepSeek factor. The release of DeepSeek's R1 model in early 2025 shook markets precisely because it proved China can innovate without access to the most advanced chips. Numerous reports show Chinese firms making significant strides with domestically developed silicon. Beijing has even tried to discourage its tech companies from buying Nvidia chips.
The copyright hypocrisy. The Register was particularly sharp: accusing China of copying while training on content created by others shows a staggering lack of self-awareness. Anthropic paid $1.5 billion in 2025 to settle lawsuits from authors whose work was used without permission to train its models.
Europe refuses to pick a side. Many European countries view both American and Chinese AI supremacy as threats to democracy. There's a concerted push for "digital sovereignty" to reduce reliance on US technology.
Trump's strategy is erratic. While Anthropic calls for tighter controls, the Trump administration just approved H200 sales to ten Chinese companies. Reuters reported that export controls were "not a major topic" during the Beijing trip. The pendulum of US policy swings between containment and commerce.
What's Next?
The House bill to close the remote compute loophole still awaits a Senate vote. NIST's safety evaluations of frontier models continue. And the next Mythos-tier release —whenever it lands— will test whether Anthropic's twelve-to-twenty-four-month lead projection holds up or starts looking optimistic.
What's clear is that the next two years will define who sets the rules of the game. And Anthropic, by publishing what amounts to a legislative wishlist under the banner of research, has made it clear it's ready to be a first-order political player in that contest.
The question isn't whether AI will reshape the global balance of power. It already is. The question is who gets to write the rules when the transformation accelerates.
Sources: Anthropic (2028: Two Scenarios for Global AI Leadership), The Register, Financial Times, Reuters, CNBC, Council on Foreign Relations, Institute for Progress, NIST/CAISI, eWeek, The Outpost, TechTimes.
Top comments (0)