Andy Jassy says AWS is sold out for 2026 and largely for 2027 despite $220 billion capex

Jassy says AWS has sold its capacity years ahead

Amazon chief executive Andy Jassy says Amazon Web Services has already sold almost all of the computing capacity it can bring online for the next two years. Speaking on an episode of the a16z Show published on 23 October 2026 with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg, Jassy said 2026 capacity is sold out, 2027 is largely sold out, and a good chunk of 2028 is spoken for too.

The remarks were quickly picked up by Reuters, which summarised them as AWS capacity being largely sold out through 2027 with significant commitments extending into 2028, and by WebProNews. They are the clearest public statement yet from a hyperscaler chief executive that the shortage of AI computing is measured in years, not quarters.

The comments matter because AWS is the largest cloud provider and the main infrastructure partner for Anthropic, and it now also sells OpenAI models through its Bedrock service. When the biggest supplier says its pipeline is booked, everyone who rents compute, from start-ups to large companies planning AI projects, has to plan around scarcity.

What he actually said

According to Reuters, Jassy said the company could invest a lot more and it would not be a bad use of capital, given that 2026 capacity is already sold out. Reuters quoted him directly: 2027 is largely sold out, AWS always leaves a certain amount for on-demand use and for its inference service Bedrock, or EC2, but has largely sold out all the big chunks, and a good chunk of 2028 is sold out.

That carve-out is important. EC2 is the AWS service that rents out virtual servers, and Bedrock gives customers access to AI models from Anthropic, OpenAI, Amazon and others through an API. Jassy's point is that the large, multi-year blocks of capacity are committed, while a reserve is kept for customers who buy by the hour. In other words, a developer can still spin up a server tomorrow, but a company that wants a large new cluster in 2027 is joining a queue.

Jassy did not give numbers for how much capacity is committed, to whom, or at what price, and Amazon did not publish any new figures alongside the podcast. The statements are his characterisation of the order book, not audited disclosures.

Why $220 billion is still not enough

Amazon said earlier this year that it expects to spend about $220 billion on capital expenditure in 2026, a figure Horowitz translated on the show into roughly $600 million a day. Jassy said that even at that level Amazon is leaving a meaningful amount of demand on the table. AI Lookout has tracked how that money is landing, including Amazon's plan for a 16 billion euro German investment with 4.8 billion euros for AWS.

He also said he does not expect the squeeze to be a 2027 problem only. Supply constraints will show up in different components for several years, he said, and he named memory as one of them: with a finite supply, Amazon has to decide how much goes to high-bandwidth memory for AI accelerators and how much to ordinary DRAM for its other cloud services. That echoes what chip suppliers are saying, as when Nvidia added Samsung Electro-Mechanics to expand AI chip substrate capacity.

Jassy defended the spending as disciplined. Data centre buildings take about two years before they earn money but last for decades, while servers and networking gear are ordered only months ahead and pay back in about three years, so Amazon can see demand before it commits the larger share of the money. He conceded that free cash flow suffers in years when many buildings are under construction at once. According to the episode description, he also believes the build-out could ultimately help make AWS a $1 trillion annual run-rate business.

The barbell of AI demand

Jassy described AI demand as a barbell. At one end are the AI labs, which he said consume enormous amounts of compute, unlike anything he has seen, along with a handful of runaway applications such as ChatGPT and Claude Code. At the other end are companies using AI to cut costs in areas like customer service and fraud detection.

The middle, which he expects to become the biggest segment, is the ordinary production software that enterprises already run. Most of it does not yet use AI inference in any pervasive way. Jassy predicted that within five to seven years quite a few, if not most, enterprise applications will, but said it never happens in one fell swoop.

The gating factor, in his view, is price. WebProNews quotes him saying inference is still an order of magnitude too expensive for enterprises to use it everywhere. If that is right, today's sold-out capacity is mostly serving the two ends of the barbell, and the larger wave of demand has not arrived yet.

Chips are Amazon's answer to the cost problem

Amazon's plan for bringing inference costs down runs through its own chips. Jassy called the 2015 purchase of Annapurna Labs one of the best acquisitions in Amazon's history. The team first built networking cards, then the Graviton CPU, which according to summaries of the episode he said gives about 40 percent better price-performance than comparable x86 processors, and then Trainium, the AI accelerator he said is about 30 to 40 percent better on price-performance than comparable chips. Amazon has also been making Trainium easier to program, for example by opening TorchNeuron so PyTorch code runs on Trainium without a rewrite.

Jassy said Anthropic has made a multi-year, multi-gigawatt commitment to Trainium. He described Amazon's investments in both Anthropic and OpenAI as partnership decisions first, saying many AWS customers had wanted OpenAI models but could not get them before the companies reached their recent deal to offer them in Bedrock.

These price-performance figures are Amazon's own and were not independently tested. They also do not solve the shortage directly: custom chips still need fabs, memory, power and buildings, the same inputs that are scarce for everyone else.

Smaller teams and a security clock

Beyond capacity, Jassy spent part of the conversation on how AI is changing Amazon itself. According to published summaries of the episode, he said the Bedrock team rebuilt Amazon's inference engine with six engineers in 76 days, work he estimated would once have taken 40 people a year. He predicted fewer people doing today's tasks in coding, customer service and analysis, smaller cross-functional teams, and new jobs in AI, robotics and space.

He also argued that the nearest real risk from AI is cybersecurity rather than science fiction. Models now find far more vulnerabilities than humans did, he said, so applications will need to be patched within minutes, and companies whose code cannot be safely updated by agents will struggle. That is in line with what security researchers have been reporting all month as AI-found bugs pile up.

Jassy also pushed back on criticism of data centres over water, emissions permits and power prices, citing Amazon's own figures. Those arguments are contested, and local concerns are concrete, as in research suggesting Texas data centres could push smog-forming emissions back to 2004 levels.

What it means for customers and rivals

For AWS customers, the practical message is to plan capacity early. Large reservations for 2027 are already scarce, and the on-demand pool Jassy described is a buffer, not a guarantee of large GPU or Trainium clusters. Horowitz asked whether AWS would push long-standing customers aside for AI labs willing to pay more. We are not going to squeeze our customers, Jassy said, comparing it with the pandemic, when AWS helped customers cut their bills.

For rivals, the comments cut both ways. A sold-out AWS leaves room for Microsoft, Google, Oracle and specialist GPU clouds to win overflow business. But WebProNews notes the episode came at the end of a bad week for AI infrastructure stocks, with CoreWeave, Nebius and IREN each down well over 30 percent in October according to CNBC, a sign that investors are questioning how long the build-out can continue at this pace even as demand stays high.

For the AI labs, a booked-up supplier strengthens the case for the long-term compute deals they have been signing, and for building or financing their own data centres.

What to watch on 29 October

Amazon reports third-quarter results on 29 October. That call will show whether AWS revenue growth matches the picture of overwhelming demand, whether the company raises its spending plans beyond about $220 billion, and whether it puts numbers on its order backlog.

Three questions are worth tracking after that. The first is whether 2028 fills up the way 2026 and 2027 have. The second is whether inference prices fall fast enough to unlock the enterprise middle of the barbell Jassy described. The third is whether the shortage moves from chips to memory, power and buildings, which he suggested will be constrained for several years. Until then, Jassy's remarks are the best public signal that the AI compute crunch is not easing.