Nvidia's grip on AI hardware has a new pressure point, and it is price. Amazon's AI chips are winning business by costing less, with one client finding an Amazon combination 80% cheaper than the Nvidia equivalent.
That figure, reported by The Information's Katherine Perloff, is the kind of number that makes buyers look past brand loyalty.
Why customers are switching
The pull is not only price. Compute is scarce, and Nvidia chips are hard to get. Companies short of supply will try the alternative in front of them.
Amazon is filling that gap with two chips. Trainium handles training, the work of tuning models for specific business needs. Inferentia 2 handles inference, the running of those models, though both can do inference in practice.
What the price claim leaves out
The 80% saving deserves scrutiny. It is not clear whether the advantage comes from a lower cost per token or from chips that use tokens more efficiently.
Amazon's own training ran about an hour faster in one test. Nvidia counters that its chips are more efficient, which could make them cheaper over a full workload even at a higher sticker price. The headline saving and the real saving may differ.
The business behind the chips
The scale is already meaningful. Amazon's chip business runs at a $20 billion rate, covering its Graviton CPUs and AI chips, though that counts only customers renting the silicon.
Andy Jassy has suggested that selling chips outside Amazon's own data centres, the way Nvidia does, could push the business toward a $50 billion run rate. That estimate is his, and it is unproven.
The structural move
The clearest signal is organisational. Amazon has pulled its chip business into a separate unit outside AWS, alongside its own AI models and quantum work.
A standalone unit can take outside investment and chase partners. Regulated industries are already asking about running Amazon chips in their own data centres through hybrid setups such as AWS Outposts.
The pattern across hyperscalers
Amazon is not alone in this. Google has formed a joint venture with Blackstone to build rentable compute, and was in talks with Meta about using its TPUs before that deal.
The hyperscalers are turning their in-house chips into franchises that compete with the company they all still depend on. Each saving they prove against Nvidia weakens the assumption that Nvidia sets the price.
Amazon has the volume and now the structure. What it has not produced is a clean, like-for-like number that settles whether its chips are cheaper, or just available.
