A tenth of the cost, 2026 mostly booked

The stamp on this IT Home page is 2026/10/24 9:51:20. It reports an interview with Amazon chief executive Andy Jassy that a16z published on the 23rd, local time. He says that for AI to be widely used in enterprise software, inference cost still has to fall to about a tenth of where it is. His line is that inference cost is still an order of magnitude too high if AI is to be used as widely as possible. About a tenth, and one order of magnitude, are two wordings of the same judgment. This is how far he thinks the cost still has to fall. It is not a bill that already shows the cut.

Capital spending and the booked years stay apart. Amazon plans to spend 220 billion dollars on capital expenditure this year, and the page says that still will not build enough infrastructure for AI demand. Jassy says the main supply of compute for 2026 is mostly booked, most of 2027 has been reserved by customers, and a sizable share of 2028 is already sold out. He thinks demand is strong enough that spending still more would not be waste. Mostly booked, most, and a sizable share are three different wordings. They do not mean all three years are sold out.

The shape of demand is his comparison. Jassy compares AI demand to a barbell. One end is AI laboratories, using an amount of compute he calls unprecedented. The other end is enterprise customers, using AI mainly for concrete tasks such as customer service and fraud detection. What he says is not yet well developed is the middle: the everyday software companies already run. That software has not widely taken up AI. He thinks the largest future AI demand may come from that stretch. May come from is his judgment. The page names inference cost as the main constraint on adoption: if the gain does not cover the higher operating expense, a company has little reason to rebuild the software it has.

The chip figures are an account, not a new test run on this page. The page says Trainium's price-performance is about 30 to 40 percent higher than comparable processors, so the same money buys more compute, and that the chip business plays a role in bringing inference cost down. About 30 to 40 percent is introduced as an account. Amazon has previously disclosed that Trainium has more than 225 billion dollars of future revenue commitments, including agreements with Anthropic and OpenAI. A future revenue commitment is not sales already booked as revenue. In the interview Jassy does not name NVIDIA directly. He stresses that the dominant supplier in the market remains an important partner for Amazon. He uses the older Graviton decision as his own comparison, and says the same story is playing out in AI. The comparison is his. Currency conversion stays off the page.