Amazon CEO Andy Jassy said on October 23, 2026, that Amazon Web Services' cloud computing capacity is largely sold out through 2027, with significant commitments extending into 2028. Speaking on a video podcast with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg, Jassy said Amazon is struggling to keep pace with insatiable demand from AI labs and enterprises. Capacity for 2026 is already sold out, he said, adding: "2027 is largely sold out. We always leave a certain amount for on demand and our inference service bedrock, or EC2, but largely sold out all the big chunks. A good chunk of 2028 is sold out."
Amazon had earlier said it would spend about $220 billion in capital expenditure for 2026, and reports third-quarter results on October 29. For startups, the headline is a new operating reality: the scarcest resource in AI is now compute itself, booked years ahead.
What "sold out" does and does not mean
Jassy's wording matters. AWS still reserves capacity for on-demand use, for its Bedrock inference service, and for EC2, its service providing cloud servers for running applications. A startup that needs a modest amount of standard compute tomorrow is not locked out. What is largely gone is the big, guaranteed, long-horizon capacity — the large training reservations and bulk commitments that AI labs and large enterprises buy in advance.
Analysis: this creates a two-tier market. Large buyers lock in capacity years ahead and pay for certainty. Smaller buyers live in the on-demand tier, where availability and price can move. That is workable for many products, but it is a poor foundation for a business plan that assumes unlimited, cheap, on-tap training compute at a moment's notice.
The playbook: five moves
First, separate training from inference in your planning. Training is lumpy, bursty, and capacity-hungry; inference is continuous and cost-sensitive. Many startups conflate the two and then discover that their inference bill, not their training run, determines whether the unit economics work. Design for inference from day one: smaller models, distillation, caching, batching, and routing simple queries to cheaper models.
Second, treat efficiency as a moat, not a chore. When compute is scarce, the team that serves ten times as many users per GPU has a structural advantage over the team that simply buys more GPUs. That means investing in model selection, quantization, prompt engineering that reduces token use, and observability that shows cost per request, per customer, and per feature. Investors will increasingly ask for those numbers; founders should have them before the meeting.
Third, use reservations and commitments deliberately. Reserved and committed-use pricing rewards predictable demand with lower rates, but it punishes over-commitment. The sensible approach is to reserve a conservative baseline that you are confident you will use, keep growth capacity flexible, and revisit the mix quarterly. Locking in a huge commitment to impress a roadmap is how a startup turns a capacity problem into a cash problem.
Fourth, build genuine multi-cloud optionality, not just a slide that says "multi-cloud." That does not mean running everywhere at once, which doubles operational complexity. It means keeping your workloads portable: containerised services, infrastructure defined in code, storage abstractions that do not trap you, and at least one tested path to a second provider for your most critical workload. When one provider's region or instance family is constrained, portability is the difference between a delay and a stop.
Fifth, price scarcity into your product. If your costs are dominated by inference, your pricing must reflect usage, not just seats. Usage-based tiers, rate limits on expensive features, and premium tiers for heavy users are not just monetisation choices; they are capacity management. A product that lets a handful of power users consume unlimited frontier inference at a flat price will discover its capacity problem in its cloud bill.
What to tell investors
Capacity scarcity changes the diligence conversation. Founders should be ready to answer: how much compute do you need for the next 18 months, where is it contracted, what happens if your primary region is constrained, and what is your cost per successful customer outcome? A credible answer includes contracted capacity, a fallback, and a cost curve that improves with scale — not just a hope that more capacity will appear.
There is also an opportunity here. Scarcity rewards companies that sell efficiency: inference optimisation, workload orchestration, spot-market management, and tools that help others do more with less compute. If the hyperscalers are sold out, the next layer of value may be in making the capacity that exists go further.
The era when a startup could assume infinite elastic cloud is ending, at least for AI-heavy workloads. The winners will be the teams that plan for scarcity early, measure cost per outcome obsessively, and keep a second path open. That is less glamorous than announcing a bigger model, but in a sold-out market, it is what keeps a company running.
Angeled.com connects founders with investors.
Sources: Reuters via https://www.reuters.com/business/retail-consumer/amazon-cloud-capacity-sold-out-through-2027-ceo-jassy-says-2026-10-23/ ; https://wncy.com/2026/10/23/amazon-cloud-capacity-sold-out-through-2027-ceo-jassy-says/ ; https://coincentral.com/amazon-amzn-stock-jassy-says-aws-capacity-sold-out-through-2027/
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



