Jassy's thesis: inference costs are too high
According to a report by IT之家, Amazon CEO Andy Jassy believes that for artificial intelligence to spread through enterprise software, inference costs must fall to about one-tenth of their current level. The statement appears in an interview published on October 23 by venture capital firm a16z. Jassy explained that, in his view, for adoption to be as broad as possible, inference costs remain an order of magnitude higher than what would be needed.
Jassy frames the issue primarily in the context of applications requiring repeated responses and tasks: every response or task completed by a model consumes compute capacity. If the benefits of introducing AI do not offset the increase in operating costs, businesses, according to the CEO, have little incentive to rebuild their existing software. According to the source's account, this is the main reason adoption remains limited.
Excess demand and the bet on proprietary chips
On the capacity front, Amazon plans $220 billion in investments this year but, according to Jassy, is unable to build enough infrastructure to meet AI demand. The company has reportedly already exhausted most of its compute availability for 2026, with much of 2027 booked by customers and a significant share of 2028 sold. Jassy said that with such strong demand, further large investments would not be a waste.
Jassy describes demand as a "barbell": on one side, AI labs, which absorb unprecedented capacity; on the other, enterprises, which use AI mainly for specific tasks such as customer service and fraud detection. The still underdeveloped segment would be the middle one—software used by companies every day—which in his view could become the main source of AI demand in the future.
Amazon's technological answer is its proprietary Trainium chip, presented as a cheaper alternative to market-dominant processors. Jassy did not explicitly name competitors, limiting himself to describing the dominant supplier as an important partner. The source reports that Trainium's price-performance is 30–40% better than comparable processors and that the company has already secured future revenue commitments of more than $225 billion tied to Trainium, including agreements with Anthropic and OpenAI. The CEO compared the current situation to that of Graviton processors for cloud servers, developed by Amazon after suppliers failed to meet its requirements for performance at the same cost.
What is confirmed and what remains to be verified
The material distinguishes between data and forecasts: the $220 billion in investments and the revenue commitments for Trainium are presented as company information; the need to cut inference costs tenfold and the reading of demand as a "barbell" are Jassy's personal assessments. It is not specified how such a reduction would be measured, over what time horizon it should occur, or what share of the savings would actually depend on proprietary chips. It therefore remains to be verified whether Trainium will deliver the stated savings and whether the middle segment of enterprise software will expand as expected.
