OpenAI’s Altman says AI computing costs should fall to $20B

OpenAI CEO Sam Altman said the five-year cost of building and running one gigawatt of AI computing capacity, which he put at about $70 billion, should come down to around $20 billion. He made the comments on Stripe co-founder John Collison's Cheeky Pint podcast, which Apple Podcasts lists as published on Thursday, October 29. The figure was a goal. Altman did not announce a new budget, construction plan or timeline for getting there, TokenPost reported.tokenpost.com+1

A target, not a plan

Altman said estimates of what a gigawatt costs vary. He used $70 billion as the cost to build and run one for five years and said $20 billion would make him "much happier," according to The Next Web, which said it had not verified the numbers. Collison clarified that the $70 billion covered both installation and operation, and Altman agreed. "I think if that number were 20 instead of 70, we should be much happier," Altman said, according to TokenPost.thenextweb.com+1

To close that gap, Altman pointed to cheaper energy, more supply and thinner gross margins across the supply chain, and he said the industry should keep pushing for 10-fold yearly performance gains from smarter models, TokenPost reported. He called himself pro-deflation and said chips also need to get cheaper, TMTB noted. Crypto Briefing reported that Altman doubted cutting margins alone would be enough while energy supplies stay tight. It also cited semiconductor analyst Dylan Patel's estimate that a gigawatt now costs more than $60 billion to build, and his view that costs could fall to around $20 billion without the margins collected by chipmakers such as Nvidia ↘ 5.28% and Broadcom ↗ 8.84%. Altman had talked about a similar $20 billion goal in October 2025, TokenPost noted.tokenpost.com+2

Robots and in-house data centers

In the same interview, Altman said OpenAI is "learning how to build our own data centers" and has "a robotics program that we're going to point at making lots of data centers," FourWeekMBA reported from the episode's transcript. He said self-replicating data centers built by AI-controlled robots are "within grasp," The Next Web reported. When Collison said data centers face permitting and component bottlenecks, Altman answered, "We could make more components with robots."fourweekmba.com+1

Altman also said memory chipmakers, which once spent $20 billion in capital expenditure to earn $19 billion, are now each on track for about $100 billion in yearly free cash flow, FourWeekMBA reported.fourweekmba.com

The stakes for OpenAI

Cost per gigawatt matters to OpenAI because it has made more than $1.4 trillion in infrastructure commitments tied to revenue growth forecasts, Crypto Briefing reported. One of these is Project Camellia in Georgia, where Georgia Power The Southern Company ↗ 1.58% plans to deliver 3.2 gigawatts in phases between 2028 and 2032, according to FourWeekMBA. Asked how the market was handling high memory and GPU prices, Altman gave a one-word answer: "Bad."cryptobriefing.com+1