Amazon Web Services just confirmed what a lot of people in the industry have been feeling: the AI boom has completely outrun the cloud’s ability to keep up.
On a podcast with Andreessen Horowitz, Amazon CEO Andy Jassy said AWS capacity is largely sold out through 2027, with significant commitments already stretching into 2028. He was pretty blunt about it. Capacity for 2026 is already gone. The big blocks for 2027 are mostly taken. And a good chunk of 2028 is spoken for too.
Amazon still keeps some capacity aside for on-demand use and services like EC2 and Bedrock, but the large, multi-year deals that AI labs and big enterprises want? Those are mostly booked.
Why This Matters
This isn’t a normal “we’re growing fast” update. It’s a signal that the physical limits of the cloud are starting to show.
Building data centers takes years. Securing enough power is getting harder. High-bandwidth memory for AI chips is constrained. Even regular DRAM is under pressure. Jassy pointed out that the shortage of data center space, power, and memory isn’t something that magically ends in 2027.
Amazon is spending heavily to try to keep up. The company has guided to around $220 billion in capital expenditures for 2026. Jassy even said they could spend more and it still wouldn’t be a bad use of capital, because demand is that strong. Despite all that spending, AWS is still leaving some customer demand on the table.
What’s Driving the Crunch
The biggest factor is AI. Training and running large models requires enormous amounts of specialized compute. AI labs and big companies are locking in capacity years in advance so they don’t get left behind. At the same time, traditional cloud workloads are still growing, so AWS has to balance both.
This creates a difficult trade-off. Amazon has to decide how much capacity to allocate to AI customers versus the rest of its business, and how to manage limited supplies of power, memory, and data center space.
What It Means Going Forward
For customers, this means a few things:
- If you need large blocks of capacity for AI training or inference, waiting until next year to shop around is risky.
- Multi-year deals are becoming the norm for anyone who wants guaranteed access.
- On-demand capacity will still exist, but the big, flexible pools that companies used to rely on are getting tighter.
For the industry, it’s another clear sign that infrastructure — not just models or algorithms — is the real bottleneck right now. Power, chips, memory, and data center construction are all lagging behind the demand created by generative AI.
Amazon reports its third-quarter earnings on October 29, so there will likely be more details then. But Jassy’s comments already paint a pretty clear picture: the cloud is no longer in a position where it can simply absorb whatever demand comes its way.
For the next couple of years at least, capacity is going to remain a scarce resource.