The AI boom is getting so big that even the companies providing the computing power are struggling to keep up. Amazon CEO Andy Jassy says Amazon Web Services (AWS) has already sold out its cloud computing capacity for 2026, with most of its capacity for 2027 also booked.

And if you think 2028 is still far away, a good chunk of that year's capacity is gone, too.Speaking on an a16z video podcast released on October 23, Jassy said demand for AI infrastructure is growing rapidly, leaving Amazon with limited room to meet new customer requirements. He also suggested that the company could spend more than its planned $220 billion capital expenditure for 2026 to expand its infrastructure.AWS capacity is booked, and AI demand keeps growingJassy explained that Amazon has to balance demand for AI computing with the needs of its traditional cloud business.

Both are growing quickly, but resources such as memory, electricity and data centre space are limited.“If you think about it, there’s only so much data center and power today. There’s only so much memory, and I don’t think that that shortage ends in 2027,” he said.

The shortage is not just about building more data centres. AI systems need specialised memory called high-bandwidth memory (HBM) to power AI accelerators, while Amazon's regular cloud services also depend on DRAM.

With supplies limited, the company has to decide how to distribute its resources.Jassy said the situation is one reason Amazon could justify investing more money in infrastructure. “'26 is obviously at this point sold out. '27 is largely sold out. We always leave a certain amount for on-demand in our inference service, Bedrock or EC2, but largely sold out, all the big chunks.

A good chunk of '28 is sold out,” he said.AWS offers cloud computing services such as EC2, which provides virtual servers, and Bedrock, which lets customers build AI applications using different AI models.Amazon wants to make AI cheaper to runThere is another challenge beyond finding enough computing capacity: the cost of running AI models. Even as businesses experiment with AI tools, the expense of using them at scale can make widespread adoption difficult.Jassy believes inference, the process through which an AI model responds to a prompt or performs a task, needs to become much cheaper.“It's still an order of magnitude too expensive in my opinion for it to be used as pervasively as possible,” he said.

In other words, costs may need to fall by around ten times for AI to become more practical across everyday business operations.Amazon is working on its own AI chips, called Trainium, to improve performance and reduce costs. Jassy said these chips offer 30 per cent to 40 per cent better price-performance than comparable alternatives.- Ends