AWS Capacity Is Nearly Sold Out: What Amazon's $220 Billion Build-Out Means for Your Business

Ask most business owners what the cloud actually is and you will hear some version of the same answer: computing you rent by the hour, as much as you want, whenever you want it. That answer quietly stopped being true this week. Speaking on a video podcast hosted by Andreessen Horowitz, with the remarks widely reported on 23 October, Amazon CEO Andy Jassy said that AWS capacity is effectively sold out years in advance: "26 is obviously sold out. 27 is largely sold out. A good chunk of 28 is sold out."

Keep in mind the scale here. Amazon is spending about $220 billion on capital projects this year, the bulk of it on data centres, chips and power, and it still cannot build fast enough. Jassy put it plainly: "Even spending $220 billion of capex this year, we're leaving a meaningful amount of demand on the table."

For twenty years, businesses have planned their websites, apps and internal systems around the assumption that cloud capacity is infinite and instantly available. That assumption now needs a second look, and not only for companies training AI models. Let's sink in!

What Andy Jassy Actually Said About AWS Capacity

The exact wording matters, because "sold out" sounds more dramatic than the reality. What Jassy described is a situation where the large, reservable blocks of AWS compute, the kind AI labs and big enterprises book on multi-year contracts, are already committed through 2026, mostly committed through 2027, and partly committed into 2028. He was careful to add a caveat: "We always leave a certain amount for on-demand and our inference service Bedrock, or EC2."

So no, the AWS console is not about to show an out-of-stock page when you try to launch a server for your website. Ordinary workloads still get served. What has changed is the headroom. The spare capacity that used to sit idle, ready to absorb whoever showed up, is thinner than it has been at any point in the cloud era.

It is also not just an Amazon story. Google CEO Sundar Pichai has said compute will stay tight into 2027, and every major provider has been telling investors some version of the same thing for months. Jassy himself does not expect the shortage to ease by 2027. When the people selling the product tell you openly they cannot meet demand for years, it is worth listening.

Why the Cloud Is Running Short

Jassy described the demand as a barbell. At one end sit the AI labs, which reserve enormous clusters of accelerated compute for years at a time to train frontier models. At the other end sit ordinary enterprises, which are now moving AI from pilots into production, which mostly means inference, the day-to-day running of models inside real products and workflows. Both ends keep growing, and they compete for the same buildings, chips and electricity.

The constraints are physical, not organisational. "There's only so much data center and power today. There's only so much memory," Jassy said. Power is the most stubborn one. Loudoun County in Virginia, home to roughly 230 data centres and often called the heart of the internet, paused approvals of new data centre projects in September until its electricity supply picture becomes clearer. Amazon, for its part, is backing twelve nuclear reactors in Indiana with around 960 megawatts of capacity, a project measured in years, not quarters.

That is the core of the problem. Demand for compute moves at software speed, and supply moves at construction and power-grid speed. No amount of money closes that gap quickly, which is exactly why $220 billion in one year still leaves demand on the table.

The Numbers Behind the Squeeze

A few figures from the same conversation help explain why Amazon keeps building anyway. Jassy said the AI business inside AWS has passed a $25 billion annual revenue run rate within the first three years of the generative AI wave, growing several hundred times faster than AWS itself did in its early years. Nobody walks away from a curve like that.

At the same time, he was surprisingly frank about costs. Inference, in his words, is "still an order of magnitude too expensive in my opinion for it to be used as pervasively as possible." That is the head of the largest cloud provider saying that running AI in production costs roughly ten times more than it should. Amazon's partial answer is its own Trainium chips, which Jassy claims offer 30 to 40 percent better price-performance than comparable hardware.

Put those two statements together and you get a useful picture for any business adding AI features to its products. The capacity to run them is scarce, the cost to run them is still high, and the provider itself expects both problems to persist for a while. Plan with that in mind, not with the marketing version.

What This Means for Cloud Prices and Availability

Here is the part that matters for your budget. No price increases have been announced, and the competitive pressure between AWS, Microsoft and Google still works in the customer's favour. Scarcity, though, tends to show up in quieter ways before it shows up on the invoice. Quota increase requests take longer to approve. Certain instance families go missing in certain regions. Spot capacity, the discounted spare compute many companies use for batch jobs, gets thinner and interrupts more often. The best discounts start to require longer commitments.

It helps to be honest about which side of the barbell your business sits on. A company website, an online store, a booking system or a typical SaaS backend uses a rounding error of the capacity an AI lab consumes, and that kind of workload is not about to be squeezed out. Where you will feel it is anywhere near GPUs: AI chatbots, recommendation engines, image or document processing, anything calling a model endpoint at scale. Those workloads now live in the contested part of the cloud.

There is a second-order effect too. When reserved capacity gets scarce and expensive at the big providers, demand spills over to smaller GPU clouds, and that market has its own problems.

Be Careful With the Smaller Providers

A whole generation of so-called neoclouds has grown up renting out Nvidia-based compute, often financed with debt. Analyst Dan Niles expects around 75 percent of neocloud operators to not exist in their current form within three years, pointing at former crypto miners who pivoted to AI as the weakest group, and at unsecured debt from some operators trading at yields above 15 percent as a warning sign. Investor Chamath Palihapitiya adds another uncomfortable statistic: roughly 1 percent of companies account for 90 percent of AI spending, and about 80 percent of that comes from unprofitable, venture-backed firms.

None of this means smaller providers are bad choices. Many offer excellent prices precisely because the big clouds are full. But if a critical part of your product runs on a provider that may be gone or restructured in two years, that is a business risk, not just a technical detail. Treat cheap GPU capacity the way you would treat any supplier with shaky finances: use it, benefit from it, and have an exit ready.

Practical Takeaways for Businesses Building Digital Products

First, find out what you are actually using. Most companies we audit are surprised by their own cloud bill, and you cannot plan capacity you have never measured. Review your spend, your quotas and your reservations before you need more of any of them, and request quota increases early rather than during a launch week.

Second, if AI features are on your roadmap for 2027, talk to your provider's account team now about which instance types and regions have real availability, and consider locking in reserved capacity or savings plans while pricing is still calm. A conversation that costs nothing today may save you an unpleasant one later.

Third, design for portability. Containerised workloads, infrastructure as code, and an abstraction layer between your product and any specific AI model make it far easier to shift between providers when one of them has no room for you. For inference specifically, keep a fallback endpoint with a second provider, even a modest one. And resist the urge to panic-buy capacity you do not need; over-provisioning in response to shortage headlines is how cloud bills double without any new revenue to show for it.

If you are not sure where to start, this is a very normal thing to outsource. A short architecture review by an experienced development partner will usually pay for itself in the first quarter.

Final Notes

The cloud is not ending, and for everyday workloads it will keep feeling infinite for a long time yet. What has changed is that the margin of spare capacity behind that feeling has shrunk, and the people who run the largest cloud on earth are saying so in public. Capacity planning, a discipline many businesses happily forgot when they left their server rooms behind, is quietly back.

The businesses that will ride this period comfortably are the ones that know their own usage, commit deliberately rather than by default, and keep their systems portable enough to go where the capacity is. None of that requires a huge budget. It requires attention, and it is a lot cheaper to pay that attention now than in 2027.