The state of play: AI capability in 2026

The capability leap is real and measurable. On SWE-bench Verified, where models resolve real GitHub issues, scores that sat in the single digits in 2023 reached roughly 72% and now approach 100% on frontier systems (Stanford HAI, 2026 AI Index). The frontier is also no longer a single-country story: the top US–China model gap on the Arena leaderboard narrowed to about 2.7% as of March 2026 (Stanford HAI).

But capability now carries a compute bill. The shift to test-time — or reasoning — compute means models “think” before they answer: median inference energy rises from about 0.31 Wh per query to roughly 3.91 Wh under a long 5,000-token reasoning workload, a ~13x jump (Stanford HAI, 2026 AI Index). At the fleet level, the IEA projects global data-center electricity to roughly double to about 945 TWh by 2030, with AI the primary driver.

That compute is highly concentrated. Epoch AI estimates the US hosts roughly three-quarters of the world’s AI-supercomputer performance, with the US and China together holding close to 90%. Capability may be converging; the hardware behind it is not.

• Benchmarks are nearing saturation: SWE-bench Verified went from single digits to ~72% and climbing (Stanford HAI, 2026 AI Index).
• Reasoning is expensive: per-query inference energy can rise ~13x, and data-center demand heads toward ~945 TWh by 2030 (IEA).
• Compute is concentrated: ~75% of AI-supercomputer performance sits in the US, and the US plus China hold about 90% (Epoch AI).