全球 AI 算力与 Token 消费:推理时代的算力经济学

Executive Summary · English

(Data updated through 2026-08-01) Inference now accounts for roughly two-thirds of global AI compute, and Goldman Sachs projects monthly token consumption to grow 24x to 120 quadrillion by 2030. Hard events in late July reinforced the "inference era" thesis: Moonshot AI open-sourced the 2.8-trillion-parameter Kimi K3 weights on July 26 as promised, then closed a $3.5B round at a $35B post-money valuation; Anthropic released Opus 5 on July 24 ahead of its October listing push; OpenAI cut GPT-5.6 Luna prices by 80% on July 30 and DeepSeek shipped V4-Flash on July 31, igniting a full-scale token price war; the four hyperscalers now plan roughly $725B of combined 2026 capex, and Alphabet's stock fell despite raising guidance to up to $205B; DeepSeek's second funding round (~$71B valuation) was reportedly paused on July 25. This report is based on public information and does not constitute investment advice.

• 01Inference now takes roughly two-thirds of global AI compute (Deloitte TMT Predictions 2026; ~1/3 in 2023, ~1/2 in 2025); Goldman Sachs' May report "Decoding the Agentic Economy" projects monthly token consumption growing 24x to 120 quadrillion by 2030; agentic token flow on OpenRouter surpassed direct human usage around early February 2026, with agent requests consuming ~15x the tokens of human chat.
• 02Open source and share: Moonshot AI released Kimi K3 weights on July 26 (2.8T parameters, 1M context — the largest open-weight release in history, one day ahead of schedule); Chinese models hit a record 58% of OpenRouter token share in July (briefly 63% in the first week), holding all top-five spots; Bloomberg reported July 29 that Moonshot closed a $3.5B round at a $35B post-money valuation and plans a pre-IPO round in August at up to $50B.
• 03A token price war erupted in the last week of July: Anthropic released Opus 5 on July 24 ($5/$25 per 1M tokens with an adjustable effort dial); OpenAI cut GPT-5.6 Luna by 80% to $0.20/$1.20 and Terra by 20% to $2/$12 on July 30; DeepSeek shipped V4-Flash 0731 ($0.14/$0.28) on July 31; Alibaba's Qwen3.7 Flash (July 27) and Google's Gemini 3.6 Flash (July 21) landed in the same window.
• 04Dual constraints of power and capex: xAI's $1B+ acquisition of APR Energy (1GW+ of mobile generation) won final clearance, but the NAACP and environmental groups filed a Clean Air Act suit over up to 35 allegedly unpermitted gas turbines (allegations pending), and TechCrunch reported July 31 that SpaceX will not remove all related turbines for another year; Alphabet's Q2 capex hit $44.9B with full-year guidance raised to $195-205B and free cash flow turning negative at -$5.9B; Meta guided $130-145B; the four hyperscalers plan ~$725B combined 2026 capex (~+77% YoY).
• 05Endpoints and agent commercialization: OpenAI's first hardware, the $230 Codex Micro, sold out in ~12 hours on July 15, first deliveries July 24, with eBay resales around $800 and listings up to $1,850; Apple sued OpenAI and io for trade-secret theft on July 10 (allegations denied by OpenAI); SoftBank became Sierra's exclusive Japan partner on July 14, with pilot resolution rates rising from 83% to 97% as enterprise token spend shifts to outcome-based pricing.
• 06Capital-market repricing: Anthropic confidentially filed its S-1 on June 1 targeting an October Nasdaq listing led by Goldman Sachs, JPMorgan and Morgan Stanley, expected to raise over $60B (May Series H valuation ~$965B); Zhipu placed new H-shares at HK$1,588 on July 9, raising ~HK$31.4B, and launched its "Touch High" plan on July 11; DeepSeek, after a record $7B first round in June, reportedly paused its second round (~$71B valuation) on July 25 per Bloomberg while preparing a STAR Market listing.