Matt Garman on the $580B Backlog Behind AWS's Growth
AWS's CEO argues its 40%-growth forecast is less optimism than bookkeeping — a $580B committed backlog, demand spread across every industry, and a buildout now capped by power and chips rather than by customers.
Accelerating, Not Decelerating
<strong>A $184B business growing 39% would normally be slowing down; AWS says it has accelerated for six straight quarters instead.</strong>
we're a $184,000,000,000 run rate business that's growing 39%, and we've accelerated every single quarter for the last six quarters.
The $580B Already on the Books
<strong>The confidence to forecast 40% growth two years out rests on a $580B backlog of committed spend — much of it already contracted.</strong>
if you see in our backlog, it's grown to be $580,000,000,000. So we have that much committed to spend on AWS, and it's growing really rapidly every single quarter.
One to Five Percent In
<strong>Enterprises have transformed only a few percent of their work with AI, and Garman sees that heading toward half.</strong>
we see customers on that journey where 3% is gonna grow to 10% is gonna grow to 50% of their business that's transformed.
Spread, Not Concentrated
<strong>AWS's growth comes from every industry at once, which Garman contrasts with rivals leaning on a couple of AI labs.</strong>
you look at some of the others that are growing, they have really high concentration with just a couple of these labs. For AWS, that's not the case. We're spread across almost every single industry that you can imagine.
The Anthropic Number Is a Customer Number
<strong>What looks like Anthropic buying compute is really tens to hundreds of thousands of AWS customers running Anthropic models through Bedrock.</strong>
most of that spend is actually spread across customers too because a lot of that spend is in Bedrock. And so it's actually our customers that are spending on Anthropic in Bedrock.
A Chip Every Fourteen Seconds, Still Short
<strong>The thing capping AWS's growth is not demand but a supply bottleneck that moves day to day across power, memory, chips, and network.</strong>
It's DRAM some days. It's chips other days. It's sometimes it's network components other days.
Trainium Leaves the Building
<strong>Next year AWS will sell its own Trainium chip for use outside its cloud, because some customers want to run it on their own premises too.</strong>
But certain sets of customers also tell us that they wanna run these things on premise.
Spend Now, Monetize for Thirty Years
<strong>AWS pays for capacity up front — backed by $80B of supplier financing — then earns it back over asset lives of five to thirty years.</strong>
there's a very large amount of CapEx that goes up front, and then we monetize that over periods of time, whether it's five or six years for servers, whether it's twenty or thirty years for data centers and power.