Andy Jassy says AI inference still far too costly for broad business use, points to Amazon’s chips

Amazon CEO Andy Jassy said AI is still not cheap enough at the “inference” stage, the moment when a system generates answers for each user or application request. In remarks reported by Yahoo Finance, Jassy characterized inference as still about 10 times too expensive for broader business use and suggested that hardware choices, including Amazon-designed chips, could be part of the fix.

The distinction matters because training costs, while significant, are typically paid during model development and refinement, whereas inference is incurred repeatedly at scale for every prompt, search, recommendation, or automated decision. That recurring cost structure is one reason many companies have been cautious about putting generative AI into everyday workflows, especially when budgets are tight.

Jassy’s comment also implicitly contrasts the economics of experimentation with the economics of production. If each call to an AI system remains costly, deployment tends to be limited to high-value use cases or tightly constrained volumes. Lowering inference costs is therefore central to the argument that AI can spread from pilot programs to broad, user-facing software and enterprise automation.

In the same reporting, Jassy pointed to Amazon chips as a practical lever to reduce those inference costs. The CEO did not lay out specific pricing targets, performance benchmarks, or a timeline in the excerpted report, but the core message was that Amazon believes its chip approach can materially change the cost curve for running AI workloads.

Amazon’s broader positioning has long tied AI execution to its cloud platform. Through its corporate newsroom, the company often frames technology development, including cloud infrastructure and workplace or customer-facing innovations, as aimed at improving performance and efficiency for customers using AWS services.

Even with Jassy’s confidence in a hardware-driven path, important details remain unclear from the reported comments alone. The post did not specify which Amazon chip products were being referenced, how much cost reduction the company expects, or whether the improvement would apply equally across model types, languages, and workload sizes.

What to watch next is whether Amazon follows through with more concrete disclosures, such as operational metrics tied to inference performance, cost per request, or evidence from customer workloads. Additional specificity could also emerge in future AWS updates, earnings discussions, or technical publications that translate the “10 times too expensive” critique into measured improvements.

For now, Jassy’s remarks add to the industry narrative that the gating factor for mass-market enterprise adoption is not just model capability, but the unit economics of running those models continuously and reliably at scale.