AWS Capacity Sold Out Through 2027: What the Cloud Crunch Means for Your Business

Amazon's cloud is running out of room. Speaking on a video podcast with Andreessen Horowitz on 23 October, Amazon CEO Andy Jassy said that AWS capacity for 2026 is sold out, that 2027 is "largely sold out", and that "a good chunk of '28" is already committed as well.

Think about that for a moment. The largest cloud provider in the world, in the middle of a construction programme worth roughly $220 billion this year alone, is telling the market that most of its compute for the next two years is already spoken for.

If you run a business, this can sound like someone else's problem, something for AI labs and Fortune 500 infrastructure teams to argue about. It is not. The capacity crunch will shape what hosting costs, how cloud contracts get written and how realistic your own AI plans are over the next couple of years. Let's unpack it.

What Jassy actually said

The remarks came in a conversation with a16z co-founder Ben Horowitz and general partner Erik Torenberg, and were reported by Reuters on 23 October. Jassy described demand from AI labs and enterprises as "insatiable" and admitted that Amazon is struggling to keep pace with it. He also said the company "could invest a lot more and it wouldn't be a bad use of capital", which is a striking thing to hear from a CEO already spending at historic levels.

Two details matter before anyone panics. First, Jassy said "largely" sold out, not entirely. Amazon keeps a pool of capacity in reserve for on-demand use and for its EC2 and Bedrock services, so a company spinning up ordinary servers tomorrow will still get them. Second, what is sold out are the large, long-dated reservations, the multi-year blocks that AI labs and big enterprises book years ahead.

Still, the direction is clear. Jassy does not expect the shortage to end in 2027, and Amazon reports its third-quarter results on 29 October, where analysts will push for specifics. Google's Sundar Pichai has made similar comments about compute staying tight into 2027, so this is not one vendor's bottleneck. It is the state of the industry.

Why AWS capacity is sold out even at $220 billion

The obvious question is how a company can spend about $220 billion in a single year and still turn customers away. Spread out, that is roughly $600 million of investment every day, and it still leaves what Jassy called "a meaningful amount of demand on the table".

The answer is that money does not convert into capacity instantly. Jassy named three constraints: data centre space, electrical power and memory. A data centre building is typically committed about two years before it earns its first dollar of revenue, even though it will then operate for decades. Power is even slower, because grid connections and new generation take years to approve and build. You cannot shorten any of that by writing a bigger cheque.

There is also a financial rhythm to it. Servers inside those buildings pay for themselves in about three years and last five or six, while the buildings themselves are thirty-year assets. That mismatch is why Amazon's free cash flow looks strained during the buildout even though the underlying business is growing. The spending is enormous, the lead times are long, and demand keeps arriving faster than the concrete can be poured.

Who is buying all the capacity

Jassy described the demand as a barbell. On one end sit the AI labs, which reserve vast amounts of compute for years ahead to train frontier models. On the other end sit ordinary enterprises, which increasingly run AI inference in production, answering customer questions, processing documents, generating content. Amazon's generative AI business alone reached a $25 billion annual run rate within about three years of existing.

The labs are the reason the long-dated blocks are gone. When a single training customer books capacity through 2028, that supply simply leaves the market. Enterprises then compete for what remains, and their side of the barbell is growing too, even though Jassy admits inference is still roughly an order of magnitude too expensive for the broadest enterprise use cases.

For everyone else, the practical meaning is simple. The cloud has stopped behaving like an infinite tap. It now behaves like a market with limited stock, where the biggest buyers reserve early and the rest adjust around them.

Will your cloud bill go up?

Amazon has not announced any price increases tied to the shortage, and nothing in Jassy's remarks suggests one is imminent. The reserve pool means everyday workloads, websites, databases, business applications, keep running as before. If your company's footprint is a few dozen instances and some managed services, you will probably not notice anything on Monday morning.

The pressure shows up in subtler ways. When a provider is selling everything it can build, its motivation to discount weakens. Negotiated savings on committed spend may get less generous. Quota increases for GPU instances or the newest machine families may take longer to approve, and some instance types may simply be unavailable in your preferred region for a while. Spot capacity, which lives off spare inventory, gets thinner when there is less spare inventory to begin with.

In other words, the crunch does not arrive as a price rise on your invoice. It arrives as friction: slower approvals, fewer options, less negotiating room. Businesses that plan around that friction will barely feel it. Businesses that discover it mid-project will feel it a lot.

What businesses building digital products should do

Start with an audit of what you actually use and what limits you are near. Check your service quotas, note which instance families your workloads depend on, and request increases before you need them rather than during a launch week. Quota requests are free; delayed launches are not.

Next, talk to your provider like a buyer in a tight market. Account teams know which regions and machine families have open inventory, and they will tell you if you ask. If your baseline usage is stable, committing to one or three years of it through savings plans or reserved capacity locks in both availability and price while both are still reasonable.

Then look at your architecture. Applications packaged in containers, defined in infrastructure-as-code and free of hard regional assumptions can follow capacity to wherever it exists, including a second provider if it ever comes to that. Portability used to be an insurance policy that mostly paid for peace of mind. In a constrained market it becomes a practical bargaining chip.

Finally, be deliberate about AI features. A smaller model with caching behind it often serves a product feature just as well as a frontier model called on every request, at a fraction of the compute. Efficiency work that used to be hard to justify now has a clear business case, because every unit of compute you do not need is a unit you do not have to fight for.

Compute is now a planning constraint

The deeper shift is in how businesses should think about infrastructure. For fifteen years the cloud trained everyone to treat compute like tap water, always there, metered, never scarce. That assumption quietly broke in 2026. Compute now belongs on the same planning list as hiring, office space and inventory: a resource with lead times, worth securing ahead of need.

This matters most for anyone budgeting a product build for 2027. If the plan involves meaningful AI workloads, the sensible moment to think about where that compute comes from is during planning, not after the first failed quota request. Teams that bring infrastructure questions into the roadmap early will ship on schedule. Teams that leave it to the end inherit whatever is left on the shelf.

It is also worth remembering that constraints pass. Amazon, Google and Microsoft are building at a pace with no precedent in the history of infrastructure, and the power projects signed today come online over the next few years. The squeeze is real, but it is a phase, and businesses that stay flexible through it will be well placed when supply finally catches up.

Final notes

Jassy's comments are a rare moment of a cloud provider openly saying demand has beaten supply. For AWS, it is a quality problem. For the rest of us, it is a planning signal: capacity is no longer guaranteed by default, pricing leverage has moved towards providers, and efficiency has turned into a competitive advantage rather than a nice-to-have.

None of this calls for alarm. It calls for a few unglamorous habits: know your quotas, commit where your usage is predictable, keep your architecture portable and make your AI features earn their compute. The companies that do this quietly over the next six months will sail through the crunch, and most of their competitors will wonder why everything suddenly takes longer.