My thought is that any competition to Anthropic is going to need to actually have the compute to serve all of Anthropics customers. Chinese labs certainly do not. It seems to be the crux of all these compute agreements.

Even then, Anthropic has the luxury of knowing exactly what hardware (even what datacenter!) will serve their models. This gives them the ability to co-design their model for the exact chips it will run on, something that Open Weights models are completely unable to do - even to the point that Deepseek fully open-sourced their inference engine, revealing that they're still running on only H100s and Huawei Ascend chips (https://github.com/deepseek-ai/DeepInfer/blob/main/README.md).

Anthropic will be running their 2027 Opus and Fable models on Vera Rubin, a chip which produces tokens for about 1/100th the cost of the H200. Seriously. 100x more efficient on GLM 5.3: https://inferencex.semianalysis.com/inference/glm-5-3 compared to what DeepSeek is able to use for inference.

Combined with Kimi K3.1's underperformance I'm honestly not too bullish on Chinese AI being able to catch back up. As an example, according to https://artificialanalysis.ai/, Kimi K 3.1 would need to be run 10x cheaper than its current inference cost in order to hit the Paraeto Frontier (matching 6.1 Sol Medium)