Global inference share · 28d tokens

US models · OpenRouter 28d

CN models · MiMo-V2.5 #1 (10.5T/wk)

Two of every three LLM calls worldwide run Chinese models — and the global top 5 by volume is 100% China: MiMo · DeepSeek · Tencent HY3 · GLM · Qwen.→ CHINA

Frontier training run · public cost

US runs · Grok 4 ≈ $388M (Epoch)

MiMo-V2.6 · RL phase (live log)

Xiaomi's entire livestreamed RL bill (Pro + Flash) is ~1/30 of one US frontier run — scopes differ, the efficiency gap doesn't.→ CHINA

Flagship API cost · $/1M input tokens

GPT-6 Astra

DeepSeek V4 Pro

7.6x cheaper at the flagship tier — China keeps the floor under frontier pricing.→ CHINA

Value-tier API cost · $/1M input tokens

Gemini 3.8 Flash

MiMo-V2.6-Flash (Xiaomi)

5.4x under Google's fast tier — Xiaomi's 309B/15B MoE (MIT, 1M ctx) prices for volume, not margin.→ CHINA

Open-weight flagships · tracked

GPT-6 · Claude · Gemini closed

DeepSeek · Qwen · GLM open

Every tracked Chinese flagship ships open weights (MIT/Apache); all three US flagships stay closed.→ CHINA

Max flagship context window

GPT-6 Astra

DeepSeek V4 Pro

The context war ended in a ~1M tie — Astra edges it. RAG builders: both fit a mid-size codebase.→ USA

Consumer AI adoption · national stats

adults 18–64 · US RPS survey (Feb 2026)

share of population · CNNIC (Dec 2025)

Denominators shown on purpose (they differ by design): the US surveys working-age adults, China counts the whole country — 602M users, +141.7% in 12 months. Comparable evidence: US leads intensity, China leads velocity.→ USA