AIGPT-4-scale training runs as a yardstick for AI computing growth
A post revisiting an AI:AM metric suggests a hyperscaler could potentially do a couple of GPT-4-scale training runs per day, compared with a couple of days per run at some buildouts a year earlier.
TLDR
A September 23, 2026 post relays a metric revived on AI:AM: some buildouts could do a GPT-4-scale training run in a couple of days a year earlier. After “probably two triplings,” it suggests, a hyperscaler could potentially do a couple per day.
GPT-4-scale training runs as a yardstick for AI computing growth
A post revisiting an AI:AM metric suggests a hyperscaler could potentially do a couple of GPT-4-scale training runs per day, compared with a couple of days per run at some buildouts a year earlier.
TLDR
A September 23, 2026 post relays a metric revived on AI:AM: some buildouts could do a GPT-4-scale training run in a couple of days a year earlier. After “probably two triplings,” it suggests, a hyperscaler could potentially do a couple per day.