In short
Rubbish Talk rates CNBC's op-ed claim that the "U.S. lead over China in AI is all but gone" a 5/10 because benchmark convergence data backs the thesis, but the headline omits that America still holds a massive, largely undiminished lead in compute capacity and AI investment.
The Verdict
Selective. The benchmark-convergence argument is genuinely strong and well-supported, but the headline's blanket claim that the lead is "all but gone" glosses over the dimensions where the U.S. advantage remains overwhelming: raw compute, chip manufacturing, and capital.
What actually happened
This is a labelled opinion piece, not a news report, arguing that China's AI ecosystem (DeepSeek, Kimi K3, Qwen, Hunyuan) has matured to the point that Washington is now reacting to Chinese breakthroughs rather than shaping the field. The author frames this as a shift to "ecosystem statecraft," where China competes on cost, deployment and standards rather than just benchmark scores.
Key facts
- Stanford's 2026 AI Index found the benchmark performance gap between the best US and Chinese models had narrowed to roughly 2.7%, down from 17.5 to 31.6 percentage points in May 2023.
- Epoch AI's tracking tells a more cautious story: since 2023, every frontier-capability model has originated in the US, with Chinese models trailing by an average of seven months (range four to fourteen), and no Chinese model has yet surpassed OpenAI's o3.
- The US spent roughly 23 times more than China on private AI investment in 2025 ($285.9 billion versus $12.4 billion), and separate reporting puts the US compute-installation lead near 10x and chip-manufacturing capacity near 35x.
What to watch for
Watch whether the next Stanford or Epoch update shows the benchmark gap widening again as new frontier models (Claude, GPT, Gemini) ship, since leaderboard convergence has been volatile month to month. Also watch whether future coverage separates "model capability" from "compute/investment," since conflating the two is where headlines like this one drift from the data.
