AWS capacity is already committed years ahead, Amazon chief executive Andy Jassy said, even as inference remains an order of magnitude too expensive for everyday enterprise software. On an a16z podcast with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg, released on October 23, he said 2026 is sold out, 2027 is largely sold out, and a good chunk of 2028 is committed, India Today reported. Large organizations planning production AI cannot assume spare capacity or cheap inference.
Key Takeaways
- AWS capacity is sold out for 2026 and largely sold out for 2027.
- A good chunk of 2028 capacity is already sold out, Jassy said.
- Inference remains about ten times too costly for pervasive enterprise use.
- Trainium offers 30% to 40% better price-performance than comparable chips.
What Sold-Out AWS Capacity Means Through 2028
Jassy put dates on the calendar.
“’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, as India Today and The Hans India reported.
Mint, noting Reuters inputs, and BW Businessworld reported the same picture: most of 2027 is sold out, a substantial share of 2028 is committed, and a slice is held for on-demand Bedrock and EC2.
The limit is physical, not only a building schedule. “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, India Today reported.
Accelerators need high-bandwidth memory; conventional cloud still needs DRAM. Analysis: AI workloads and the existing estate are now bidding for the same power, space, and memory.
| Year | AWS capacity status |
|---|---|
| 2026 | Sold out |
| 2027 | Largely sold out; some held for on-demand Bedrock and EC2 |
| 2028 | A good chunk sold out |
A Barbell of AI Demand
Jassy described demand as a “barbell,” Mint and BW Businessworld reported. Laboratories consume computing resources at unprecedented scale. Businesses use AI for specific tasks such as customer service and fraud detection.
Everyday business software, the middle of that market, has yet to adopt AI widely and remains largely untapped. He expects that segment could eventually become the largest source of demand.
Analysis: customer service, fraud controls, and line-of-business tools are where most enterprises will meet both the queue and the price, not in laboratory-scale training alone.
Inference Costs Still Block Everyday Software
Booked capacity does not make AI cheap to run. “It’s still an order of magnitude too expensive in my opinion for it to be used as pervasively as possible,” Jassy said, as Mint reported. Mint framed that remark as a need for costs to fall roughly tenfold before AI is used as widely as possible.
Inference is the computing required when a model responds to a prompt or performs a task—the meter on each interaction, not the training run. Bedrock is the inference service held partly for on-demand use; EC2 provides virtual servers. Analysis: CX and product leaders should treat pervasive in-application AI as a unit-cost problem until that gap narrows.
Trainium, Capex, and the Remaining Gap
Amazon’s cost path is its own silicon, the approach it took with Graviton when it wanted better performance per dollar. “We saw that same movie in the AI space,” Jassy said, as Mint reported. Trainium, he said, offers 30% to 40% better price-performance than comparable processors. “Our chips business is a big piece of that,” he said.
Mint reported more than $225 billion in future Trainium revenue commitments, including agreements involving Anthropic and OpenAI, and said Amazon’s investment in Anthropic has supported the chip effort.
On supply, Amazon plans about $220 billion of capital expenditure in 2026, with much of that year’s cloud capacity already booked. He said the company could justify investing more; India Today reported he suggested spending could exceed that plan. The next checkpoint is third-quarter results on October 29, when investors will look for cloud growth and infrastructure spending, Mint reported.
Analysis: what buyers should weigh:
- Large blocks for 2026 through 2028 are largely committed; new multi-year programs should assume a queue.
- On-demand holdbacks for Bedrock and EC2 fit bursty use better than always-on production.
- Trainium’s 30% to 40% edge is not the tenfold inference drop Jassy says pervasive use requires.
- Power, data center space, and memory stay tight, and he does not expect that shortage to end in 2027.
