Amazon Earnings: The $578 Billion AWS Backlog Behind a Bigger 2027 Capex Bill
Amazon's earnings missed by seven cents and the stock fell, while the AWS backlog nearly tripled to $578 billion and the capex guide leaves a fourth quarter that looks too small for a bigger 2027.
Amazon ($AMZN) reported third-quarter earnings Thursday night, and the stock dropped more than 6% after hours even as AWS posted its fastest growth in nearly five years. Andy Jassy told analysts Amazon expects to spend "meaningfully more in 2027" than this year. The AWS backlog is why he can say it so casually. It stood at $578 billion at the end of September, up from $200 billion a year earlier.
Why Amazon's earnings missed while AWS beat
AWS revenue came in at $46.0 billion, up 39.3%, ahead of the $45.3 billion StreetAccount had. Operating income for the whole company hit $28.1 billion, which cleared the top of the $22.5 to $26.5 billion range Amazon gave in July. Below that line, interest expense tripled to $1.7 billion and other income swung to a $703 million loss. Earnings per share landed at $1.89 against the $1.96 analysts expected, per LSEG. (Last year's third quarter also carried a $9.5 billion pre-tax gain on Amazon's Anthropic stake, which is why earnings fell from a year ago at all.)
You can sell a stock on that. But the 10-Q says a lot more than the income statement does.
The AWS backlog names its two biggest new tenants
Amazon's 10-Q names customers in its backlog footnote. OpenAI expanded its existing $38 billion AWS commitment by $100 billion over eight years in the first quarter. Anthropic expanded its own commitment by more than $100 billion over ten years in the second. Those two expansions are worth more than $200 billion together, which is over half of what the backlog gained in a year (Amazon doesn't say how much of each is booked yet). I'd guess they're why the weighted-average contract life stretched to 6.1 years from 3.8 years a year ago (we tallied Anthropic's side of this in our September piece on its compute commitments).
Both labs are buying Amazon's own chips. In its first-quarter release, Amazon said Anthropic will secure up to 5 GW of Trainium capacity, and that OpenAI committed to about 2 GW of Trainium that starts ramping in 2027. Trainium is Amazon's in-house AI accelerator, the alternative to renting Nvidia GPUs. Jassy said on the call that revenue commitments for Trainium and Graviton (Amazon's in-house server CPU) have each grown more than 20 times year over year.
Amazon capex is running ahead of its own guide
Amazon spent $75.9 billion in cash capex in the third quarter, the figure finance chief Brian Olsavsky gave on the call. That brings the nine-month total to $172.2 billion. Amazon still says the full year will be "approximately $220 billion," which leaves about $48 billion for the fourth quarter. That would be a sharp step down from the quarter it just reported, right before a year it says will be bigger.
So either the third quarter pulled spending forward, or $220 billion is a floor. I lean toward the floor. Jassy also said Amazon is installing an AI accelerator "about every 14 seconds," and a company that says that doesn't usually cut its spending by more than a third the next quarter.
Which suppliers the AWS order book lands on
TSMC ($TSM) is first, because it makes the chips. Trainium2 is built from two 5nm compute dies joined with TSMC's CoWoS packaging, so every Trainium shipped is a wafer order and a packaging order at TSMC. Jassy didn't name the foundry, but he said "everybody is producing their advanced chips with the same provider" and "everybody gets an allocation." Amazon also plans to start selling its chips to outside customers in late 2027, which means asking TSMC for more.
Marvell ($MRVL) signed a five-year, multi-generational agreement with AWS in December 2024 that includes custom AI products. Alchip (a Taiwanese chip-design house) is reportedly designing Trainium3.
Memory is the cost line Amazon keeps flagging. Management called higher memory and component prices a headwind on the call, and memory makers like Micron ($MU) and SK Hynix are the ones charging those higher prices. Nvidia ($NVDA) still has a seat too, since Amazon listed a new AWS agreement with Nvidia among this quarter's deals.
The view
The market sold a seven-cent miss on earnings per share. I'm watching the fourth-quarter capex print instead. If it comes in anywhere near the third quarter's pace, the $220 billion guide was low and 2027 starts from a higher base, with TSMC's allocation to Amazon growing right along with it.
Amazon has $578 billion of signed work and 6.1 years to deliver it. A lot of that starts as a wafer order at TSMC.