Amazon's Jassy Says AI Inference Is Still 10x Too Expensive and Points to Trainium
Amazon CEO Andy Jassy says AI inference, the computing used to run models, remains roughly 10x too expensive for widespread business-software adoption, with custom Trainium chips key to lowering costs, as the company plans about $220 billion in 2026 capital spending and says its buildout will pressure free cash flow for a couple of years.
Amazon (AMZN) CEO Andy Jassy says AI inference, the computing used to run models, remains roughly 10 times too expensive for widespread adoption in business software, and that custom chips are key to lowering the cost, according to the summary supplied. Jassy said AI labs consume enormous amounts of computing, while most businesses use AI only for specific tasks, and that broader adoption across everyday business software could eventually become the largest market.
What Jassy said
| Topic | Comment |
|---|---|
| Inference cost | Roughly 10× too expensive for widespread business-software adoption; costs must justify rebuilding existing applications |
| Capital spending | About $220 billion in 2026 |
| Capacity | Major commitments sold out for 2026, largely booked for 2027 and already covering "a good chunk" of 2028 |
| Chips | Amazon's AI chips offer 30–40% better price-performance than comparable processors; the dominant supplier, not named, remains an important partner |
| Trainium commitments | More than $225 billion, including agreements with Anthropic and OpenAI |
| Cash flow | Free cash flow pushed negative as hardware arrives before power connections; the strain should last "a couple of years" |
| Software and staffing | Amazon expects to rewrite nearly all its code and operate with smaller teams |
| Pricing | Amazon will not raise prices on core customers to chase AI-lab demand |
The summary did not name the supplier Jassy called important, though the context points to Nvidia (NVDA), the dominant AI chip maker.
The chip and capacity picture
On Amazon's April 29 earnings call, Jassy said the company's chip business had passed a $20 billion annual run rate, according to The Register, and that Amazon had more than $225 billion in revenue commitments for Trainium. The same report said OpenAI committed to roughly two gigawatts of Trainium capacity through AWS, ramping in 2027, and Anthropic to up to five gigawatts across current and future generations.
The price-performance claim has a more specific form on that call. Per The Register, Jassy said Trainium2 offers about 30% better price-performance than comparable GPUs, and Trainium3, which began shipping in early 2026, is 30–40% more price-performant than Trainium2.
On the July 31 earnings call, Jassy said Amazon expected to spend about $220 billion in cash capital expenditure in 2026, roughly $20 billion more than first planned, with higher memory prices driving the increase, Fierce Network reported. AWS revenue rose 36.7% to $42.2 billion in the second quarter, and AWS's order backlog reached $496 billion. Jassy said Amazon aims to have double its 2025 power capacity by the end of 2027, and that it expected free cash flow headwinds until the new infrastructure comes online.
Why it matters
Jassy's argument is that today's AI economics fit labs and a handful of specific business tasks, but not the rewrite of everyday applications. If inference gets much cheaper, the demand could widen well beyond the labs that dominate usage now. Amazon is betting that its own chips are the route to that cost curve, while accepting several years of heavy spending and weaker free cash flow to build the capacity.
Sources: The summary supplied (Oct. 23); The Register (Apr. 30); Fierce Network (July 31 earnings call). The remarks on inference cost, sold-out capacity, free cash flow, rewriting code, smaller teams and pricing are as given in the summary supplied. This is informational, not a recommendation to buy or sell.
Readers also read
MarketCatalyst LLC is not a registered investment advisor and does not manage client assets. Content on this platform is provided for informational and educational purposes only. It is not investment advice, and MarketCatalyst is not a stock-picking or trade-alert service. Trading stocks and options involves risk, including the possible loss of principal. Consider your own goals, time horizon, and risk tolerance, and consult a qualified financial advisor before making any investment decision.