Amazon Web Services has effectively sold out its cloud capacity through 2027, CEO Andy Jassy said in a video podcast with Andreessen Horowitz co-founder Ben Horowitz published on October 23. According to Reuters, Jassy said the 2027 capacity is largely spoken for, with significant commitments already extending into 2028, and that Amazon could spend far more on infrastructure without it being a bad use of capital. Inference costs remain the weak point: Jassy called inference “roughly ten times too costly” and said AWS must bring them down. With third-quarter results due October 29, the picture is of a cloud business constrained not by demand but by its ability to build fast enough.
What Jassy Told The Podcast
Jassy made the remarks on a video podcast hosted by Horowitz and Andreessen Horowitz general partner Erik Torenberg, Reuters reported on Friday. He drew a clear line through the next two years: 2026 capacity is already gone, 2027 is largely sold out, and a good chunk of 2028 is spoken for. Amazon always reserves some headroom for on-demand use, plus its inference service Bedrock and EC2, the cloud servers customers rent to run applications. Our explainer on Amazon’s AI stack maps how Bedrock sits between Alexa+ at the consumer edge and Nova, Amazon’s own model family.
The pressure comes from what Jassy described as insatiable demand from AI labs and enterprises, and he said Amazon is struggling to keep pace with it. His answer to the spending question was blunt: Amazon “could invest a lot more and it wouldn’t be a bad use of capital,” a signal that the constraint is construction speed rather than boardroom appetite.
The $220 Billion Buildout
Earlier this year Amazon said it would spend about $220 billion in capital expenditure for 2026, a figure that already ranked among the largest corporate investment plans on record. Jassy’s remarks suggest that number is a floor rather than a ceiling: if 2026 capacity is sold out before the year has even begun, every additional data center Amazon can bring online has a customer waiting.
That spending is already landing in concrete. Amazon’s €16 billion German investment for 2027, with €4.8 billion for the AWS Frankfurt region, shows where the money goes: cloud regions where customers need local capacity for AI workloads. Expect more such regional packages as Amazon races to convert its sold-out order book into live servers.
Inference Must Get Cheaper
Jassy’s second message was about price. Inference — running trained models for real users — is “roughly ten times too costly,” he said, and driving that cost down is now an AWS priority. When capacity is scarce and every inference run is expensive, the economics of AI applications tighten from both ends at once.
Amazon’s cost lever is its own silicon. The TorchNeuron software Amazon launched for its Trainium chips lets PyTorch code run without modification, a direct attempt to cut the cost of training and inference on Amazon-built hardware. If those savings materialize, they feed straight back into the capacity story: cheaper inference means each sold-out server serves more customers.
Why It Matters
Amazon’s sold-out cloud reframes the AI race as a construction race. The binding constraint on AI growth is no longer model quality or even chip supply alone — it is finished data centers with power and cooling, and AWS, the world’s largest cloud provider, is telling customers the queue now runs into 2028.
For enterprises, that means planning AI deployments around availability, not just price. For rivals, it is an opening: Microsoft and Google will absorb the demand AWS cannot serve, and every quarter of delay hands them a chance to convert impatient customers. For chipmakers, the sold-out sign is the demand guarantee behind every fabrication expansion. And Jassy’s inference-cost warning points at the next battleground — whoever makes each model run cheapest wins the margin when capacity finally catches up.
What to watch: Amazon reports third-quarter results on October 29, when investors will press for updated capital expenditure guidance and any sign that the 2028 order book is still growing.
Background
Amazon launched AWS in 2006 and built it into the world’s largest cloud provider and the profit engine of the company. Capital spending across the hyperscalers surged after generative AI took off in 2023, with Amazon, Microsoft and Google all racing to add data center capacity; Amazon answered the cost side with its own Trainium and Inferentia chips from its Annapurna Labs unit. The sold-out remarks mark the first time Amazon has publicly said demand has outrun even its record buildout.
Source: Reuters
0 Comments