AWS Is Landing a GPU Every 14 Seconds and Still Comes Up Short

Garman dropped the number on this week's episode of Sources with Alex Heath, asked about building data centers in space. He called orbital compute "highly unlikely to be economically feasible or practical in the next decade or two," noting the world doesn't manage a single rocket launch a day. Then came the comparison: "we're landing in a data center today, one GPU every 14 seconds."

"Right now the demand is far outstripping the supply," Garman said earlier in the conversation. Even with Amazon's spending levels, he said, "I anticipate that we'll be short of capacity for the next couple of years." The constraint lands on AWS, which drove roughly 23% of Amazon's third-quarter revenue but about 63% of its operating income.

Garman said he could sell every accelerator he knows is landing in 2027 and be done with it. AWS doesn't, holding capacity back for enterprises and startups that show up later. That choice is backed by contracts: most large AWS customers commit for three to six years, some for as long as 10, and often spend past what they pledged. Amazon put AWS's backlog at $578 billion at the end of September.

Cheaper AI doesn't scare him either. "Every time we see inference get, you know, 50% cheaper, we see an order of magnitude more usage," Garman said. Analysts tracked by TIKR expect Amazon's normalized EPS to nearly double between 2025 and 2028.

The bill is already visible in cash. Amazon's free cash flow was a $42.4 billion outflow in the 12 months to Sept. 30. The next checkpoint comes in early February, when Amazon lays out its 2027 capex outlook.