"In the three months ended September 30, 2026, our underlying compute efficiency improvements were partially offset by the delivery of additional compute, including $5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments." aka as we buy this extremely expensive compute and can't sell it for more money (e.g. $) our efficiency goes down.
It's a wonderful technology, world-changing. It's been a great business.
It's not obvious any use case will take over the 2026 token-maxxing S Curve.
Let's hope they make it out alive!
"As of and subsequent to September 30, 2026, we have entered into purchase commitments and lease agreements, including short-term commitments, as well as leases not yet commenced, with an aggregate committed spend of approximately $895.4 billion, substantially all of which is related to compute and infrastructure. These commitments include approximately $447.5 billion expected to be incurred before December 31, 2030, and substantially all of the remainder is expected to be incurred between 2031 and 2045. "
In Q3, I think they struggled with the government ban on Mythos/Fable, Opus 5 not being so good, OpenAI models being more token efficient, and they decided to dedicate a higher percentage of compute to training.