Most takes on the US vs China AI race feel abstract. This one names a year, draws two timelines, and tells you exactly what tips the scales. That’s a sharper frame than I usually see from a major lab.
The analysis comes from a fresh Anthropic essay that Matthew Berman walks through in detail, and the original poster is essentially saying: democracies have a lead, and 2028 is when we find out if we kept it. I was hooked from the first paragraph because it’s the most concrete strategic piece a frontier lab has published.
Here’s the contrast that makes this essay land
The old framing: AI competition is a vague, ongoing contest with no finish line. Everyone keeps shipping, everyone keeps negotiating, things sort themselves out.
The new framing from Anthropic: there’s a narrow window before recursive self-improvement makes the gap permanent. Whoever crosses that line first sets the rules for everyone else. Berman pushes this further and argues 2028 is the unspoken deadline because that’s roughly when self-improving AI becomes plausible.
The author lays out two 2028 scenarios:
- 🏆 Scenario A: Democracies hold the compute lead, export controls tighten, distillation attacks get blocked, and democratic norms shape global AI.
- ⚠️ Scenario B: Loopholes stay open, PRC labs catch up using smuggled chips and distillation, and authoritarian regimes set the rules for automated repression at scale.
The creator behind the essay grounds this in four fronts of competition:
- 🧠 Intelligence: who builds the most capable models.
- 🏢 Domestic adoption: who integrates AI fastest across business and government.
- 🌍 Global distribution: whose AI stack the world economy runs on.
- 🛡️ Resilience: who stays politically stable through the transition.
The expert puts intelligence at the top. Berman pushes back: cheap, efficient open-source models from China could dominate global adoption even if they trail the frontier by six to nine months. A CEO comparing $30 per million output tokens to $3 for something nearly as good will pick the cheap option almost every time.
Where the contrast gets interesting is the solution set. The post’s author proposes three plays:
- Close the loopholes. Stop smuggled chips, foreign data center workarounds, and gaps in equipment controls.
- Defend innovations. Restrict model access and deter large-scale distillation attacks.
- Champion the export of American AI, but keep it closed-source.
Berman agrees with the risk diagnosis and disagrees hard on the closed-source part. His read: if you want the world to build on American AI, open-source has to be a major piece of it, because that’s how you win adoption against subsidized Chinese alternatives.
A few sharp data points the original poster surfaces:
- 📊 Huawei is projected to produce just 4% of Nvidia’s aggregate compute in 2026 and 2% in 2027.
- 🔓 DeepSeek R1 complied with 94% of overtly malicious requests under a common jailbreak, vs 8% for US reference models.
- 🔬 Only 3 of 13 top Chinese labs published safety evaluations last year.
- 💰 Algorithmic gains scale with compute, not as a substitute for it. More chips means more experiments means more breakthroughs.
The quiet thesis underneath all of it: leverage in any future negotiation comes from being clearly ahead, not from being tied. A neck-and-neck race pushes every lab to cut corners on safety. A decisive lead lets the leader set the terms.
The Mythos Preview reference is the part that stuck with me. The mind behind the essay argues that if a Chinese lab had shipped a Mythos-level cyber model first, critical US infrastructure would already be exposed. Whether you buy the framing or not, that’s a concrete stakes statement.
Watch the full breakdown for Berman’s counterarguments on export controls, open-source, and why he thinks the compute moat is more fragile than Anthropic suggests.