AWS Has Already Sold a Good Chunk of Its 2028 Capacity

Amazon Web Services has sold out its capacity for this year and largely sold out 2027, Amazon CEO Andy Jassy said on an episode of the a16z Show released Friday. “A good chunk of 28 is sold out,” he told Andreessen Horowitz co-founder Ben Horowitz and host Erik Torenberg. AWS always leaves a certain amount for on-demand customers, he said.

The episode came out at the end of a bad week for AI infrastructure stocks. CoreWeave, Nebius and IREN have each lost well over 30% in October, CNBC reported, and SK Hynix dropped plans on Thursday to take its Solidigm storage unit public, citing market volatility. CoreWeave had already slid on debt worries earlier this month before Goldman Sachs upgraded it.

Even $220 billion won’t cover it

Amazon expects about $220 billion in capital spending in 2026, which works out to roughly $600 million a day. Jassy said that still leaves a meaningful amount of demand on the table, and he doesn’t expect the shortage to end in 2027. Supply constraints will show up in different components for several years, he said. Memory is one of them. With only so much to go around, Amazon has to decide how much goes to HBM for AI accelerators and how much to ordinary DRAM for the rest of its cloud business.

AWS’s AI business has passed a $25 billion run rate in the first three years of the generative AI wave, according to Jassy. AWS itself was at about $60 million a year at the same point.

Longtime customers won’t get bumped

Horowitz asked whether AWS felt pressure to shift capacity from longtime customers to AI labs willing to pay almost anything. “We’re not going to squeeze our customers,” Jassy said, adding that AWS would make sure they can keep scaling. He compared it to the pandemic, when AWS sent its field teams out to help customers cut their bills even though the company knew its revenue would take a hit.

Customers can reserve capacity, too. On Wednesday AWS began letting Aurora and RDS users reserve database capacity, billed at on-demand rates whether or not the instances run.

Jassy also said inference is still “an order of magnitude too expensive” for enterprises to use it everywhere. Bringing that cost down is where Amazon’s Trainium chips come in. “Our chips business is a big piece of that,” he said. Amazon holds its third-quarter earnings call on Oct. 29 at 5 p.m. ET.

Gartner predicts 80% of neocloud companies will shut down or be absorbed by 2030, CNBC reported. Nvidia CEO Jensen Huang, speaking at a company conference in Berlin this week, was more upbeat: “demand for the neoclouds is off the charts.”