The AI Race Is Becoming One for Compute
With every quarter comes another jump in AI benchmark scores, parameter counts, context windows, or reasoning performance. Indeed, companies capable of training the strongest systems appear to hold the most valuable advantage.
However, that advantage is becoming harder to defend.
Model capability is compressing across labs and countries. Stanford’s 2026 AI Index found several leading developers clustered closely together in human preference rankings, while the performance difference between the strongest US and Chinese models had fallen to 2.7% by March. Chinese and American models have traded places near the top of rankings since early 2025.
Kimi K2 was one of the important early demonstrations of this trend. Moonshot AI trained a one-trillion-parameter Mixture-of-Experts model that activates 32 billion parameters per token and released its weights publicly. Its reported performance was competitive across coding, mathematics, and agentic tasks, including a 65.8 score on SWE-Bench Verified.
Since then, the Kimi family and...
Copyright of this story solely belongs to hackernoon.com. To see the full text click HERE
