Global AI compute to hit 200M H100-equivalents by 2028, fueling agentic loop toward ASI

新智元 · wechat · 2026-08-03

According to EpochAI, global AI chip compute now equals 20M H100 GPUs, doubling every 9 months, projected to reach 200M by end of 2028. This compute will power an accelerating loop: more chips → stronger models → agents take on more work → faster next-gen model development.

Data center scale: EpochAI tracks 74 large AI data centers with 12.5M equivalent chips; largest is xAI's Colossus2 (1.112M), followed by Microsoft, Meta, Amazon, and OpenAI clusters in the hundreds of thousands.

Capital expenditure: 2026 capex for Amazon, Alphabet, Meta, Microsoft, Oracle is projected at $750B combined; AI infrastructure spending expected to grow from $318B in 2025 to over $1T by 2029.

Agent capability leap: Remote Labor Index shows leading models completing tasks rose from 2.5% (Oct 2025) to 16.1% (Jul 2026). Inference share of AI cloud spending expected to reach 55% in 2026 and >65% by 2029.

AI self-optimization: GPT-5.6 Sol optimized production systems, cutting service costs by 20% and boosting token generation efficiency by 15%, prompting OpenAI to cut prices by 80%. DeepMind CEO Hassabis says AI change could be 10x the Industrial Revolution.

Original post →

More from AGI Musings

AGI Musings channel →