AWS is adding one GPU every 14 seconds, chief executive Matt Garman told Alex Heath in an interview for the Sources podcast, according to Heath’s written summary published on 10 November. By our arithmetic, that pace works out to about 6,170 GPUs a day and about 2.25 million a year. Heath writes that Garman still “expects capacity shortages for another couple of years”.
On putting data centers in space, Heath writes that Garman “thinks the economics and practical challenges will keep that prospect out of reach for another decade or two”. SpaceX’s prospectus says its first orbital AI compute satellites could deploy “as early as 2028”, as we reported when SpaceX sought 1 million orbital data center satellites.
Business Pill · GPUs Explained
An evergreen Business Engineer explainer on why AI runs on GPUs. It teaches the general idea only and does not discuss this interview.
The key insight:AWS chief executive Matt Garman told Alex Heath that AWS is adding one GPU every 14 seconds, about 2.25 million a year by our arithmetic, and Heath writes that Garman thinks data centers in space are out of reach for another decade or two. At that pace, FundaAI’s reported 432,000-GPU SpaceX orbital order for 2027 is about 70 days of AWS additions, by our arithmetic.
Heath recorded the interview in Seattle on the Friday after the midterm elections. Running models already accounts for the majority of AWS’s AI workloads, Garman says, according to the summary. Asked how far along the inference market is, he answered: “I don’t think it’s even 1%”. He says each 50 percent drop in inference prices has brought an order of magnitude more usage, with revenue continuing to rise.
Heath writes that Amazon thinks AWS could become “a trillion-dollar revenuebusiness within the next five to ten years”. The interview also covers how Garman decides who gets compute, why he reserves capacity for startups that have not been founded yet, and why he sees AWS’s risk differently from providers dependent on debt and a handful of large customers. Our earlier piece found that the AWS backlog is now 3.6 times a year of AWS sales.
Millions of GPUs: AWS at its stated pace of one GPU every 14 seconds, per Alex Heath’s Sources summary of his interview with Matt Garman, for one quarter and one year (our arithmetic), beside FundaAI’s reported estimates for SpaceX’s 2027 orbital rack order and lifetime orbital program, which SpaceX had not confirmed when we reported them.
The Pace in Numbers
A day has 86,400 seconds. At one GPU every 14 seconds, that is about 6,170 GPUs a day, about 563,000 a quarter and about 2.25 million a year, by our arithmetic. The summary gives the pace as a rate; it does not say which chips count as GPUs, whether Amazon’s own Trainium chips are included, or over what period the rate was measured.
Set Against the Orbital Plans
FundaAI, a research firm, wrote that SpaceX has ordered 6,000 space-specific racks for 2027, equal to 432,000 GPUs, and put the program’s lifetime volume at about 2.16 million GPUs; SpaceX had not confirmed the order when we covered the reported 6,000-rack order. At AWS’s stated pace, 432,000 GPUs is about 70 days of additions and 2.16 million is about 350 days, by our arithmetic.
SpaceX measures its launch throughput in mass. Its 10-Q reports 1,517 metric tons of mass to orbit in the nine months to 30 September, against 1,627 a year earlier; 1,266 tons of the 2026 figure were its own payloads. Our piece on how SpaceX makes money breaks down the segments behind that launch business.
Heath asked Garman about orbit after a conversation with Google chief executive Sundar Pichai about putting data centers in space. On Earth, Heath writes: “Of the alternatives we discuss, he’s much more interested in deserts”.
The Structural Read
As we read the interview summary, the two companies are describing different constraints. Garman, per Heath, still expects capacity shortages for another couple of years at a build rate of about 6,170 GPUs a day, and points to deserts rather than orbit as the alternative site. SpaceX’s pitch, in its prospectus and FCC application, is that orbit itself becomes a site, with the first satellites as early as 2028.
Matt Garman, AWS chief executive, to Alex Heath (Sources), on how far along the inference market is
FOR SCALE One GPU every 14 seconds is about 563,000 a quarter and about 2.25 million a year, by our arithmetic.
FOR THE ORBITAL COMPARISON FundaAI’s reported SpaceX figures, 432,000 GPUs for 2027 and about 2.16 million over the program, equal about 70 and 350 days of AWS additions at the stated pace, by our arithmetic.
FOR THE TIMELINE SpaceX’s prospectus says as early as 2028; Garman, per Heath, says a decade or two, which by our arithmetic is 2036 to 2046.
The framework says: “The structural bottleneck isn’t computing, it’s electrons.”
As we read it, both answers in this interview are answers to that bottleneck: a ground build rate that still leaves AWS short of capacity, and a preference for deserts over orbit as the next site. The framework predates the interview and does not discuss it.
Years from 2026 until data centers in orbit: SpaceX’s prospectus (as early as 2028) and AWS chief executive Matt Garman, per Alex Heath’s Sources summary (a decade or two). Year counts are our arithmetic.
What Is Not Established
The 14-second figure and the decade-or-two horizon are as Heath’s written summary gives them; we have not heard the podcast audio. The FundaAI figures are a research firm’s estimate as reported. Setting a stated build rate beside an order estimate shows scale only: it does not show that AWS could supply orbital compute, or that the two compete for the same customers.
AWS and SpaceX were not contacted for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
AWS’s stated pace of one GPU every 14 seconds is about 2.25 million GPUs a year, by our arithmetic, enough to match FundaAI’s reported 2027 SpaceX orbital order in about 70 days. Garman, per Heath, puts data centers in space a decade or two away; SpaceX’s prospectus says as early as 2028.
A note on sourcing. Garman’s remarks come from Alex Heath’s written summary of his Sources interview, published on 10 November 2026; we have not heard the podcast audio. The SpaceX mass-to-orbit figures are from its 10-Q for the quarter to 30 September 2026. The SpaceX prospectus date and the FundaAI rack and GPU estimates are as we reported them in earlier pieces; FundaAI is a research firm, and SpaceX had not confirmed its figures when we reported them. Per-day, per-quarter, per-year and day-count figures are our arithmetic. AWS and SpaceX were not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
