Uber adopts Amazon Trainium and Inferentia AI chips in AWS, breaking from NVIDIA GPU reliance

Uber (UBER) is shifting a meaningful share of its AI workloads onto Amazon’s (AMZN) in-house Trainium and Inferentia chips, a move that matters far beyond one customer’s cloud bill.

A global platform is putting real production AI work onto Amazon-designed silicon inside AWS, showing that high-stakes workloads can run somewhere other than NVIDIA’s default stack. The broader backdrop is that Amazon’s AI ambitions are expanding on multiple fronts, including reported talks around a $50 billion OpenAI investment.

Amazon is offering large customers a cheaper path for training and inference. That is true, but it undersells what changed here. When AWS supplies the cloud, the software layer, and the chips, Amazon gains more control over the economics and roadmap of AI infrastructure. That kind of vertical integration can be powerful, but it also raises the stakes on execution and reliability, especially as Amazon faces scrutiny over issues such as cloud outages linked to internal AI coding tools.

Three forces make this shift important

First, cost. Amazon has been pushing Trainium and Inferentia as lower-cost alternatives for AI training and inference, and Uber’s decision gives that pitch real weight. At Uber’s scale, small improvements in price-performance can move margins. If AWS can deliver acceptable performance at lower cost on its own chips, it keeps more of the value inside its own stack instead of passing a large share to NVIDIA.

Second, bargaining power. For years, cloud providers largely served as channels for NVIDIA’s hardware. AWS is changing that equation by becoming both the cloud platform and the chip supplier. That gives Amazon more room to shape pricing, capacity, and integration around its own infrastructure. It also gives customers another option when GPU supply is tight or NVIDIA pricing is hard to swallow. That does not erase NVIDIA’s advantage, but it does chip away at the dependency that helped preserve its pricing power.

Third, market psychology. Enterprise technology markets run on defaults, and in AI the default has been simple: buy NVIDIA. Uber’s move creates a concrete counterexample. This was a production decision by a company with large-scale, live AI needs. Uber’s relationship with Amazon is also becoming relevant in other areas of the business, including plans to use the Uber platform to launch Zoox robotaxis across the U.S.. That alone will influence procurement discussions, RFPs, and architecture reviews across large companies already looking for lower-cost compute or more predictable supply.

This does not mean Amazon’s chips now match NVIDIA across the board. NVIDIA’s H100 and newer systems remain the standard for many cutting-edge and general-purpose AI workloads, especially where peak performance, software maturity, and broad model support matter most. Uber’s shift is better read as proof that the market is opening, not as proof that NVIDIA has lost the crown.

If AWS can persuade more major customers to run meaningful workloads on Trainium and Inferentia, Amazon captures more of the AI value chain itself: the infrastructure, the integration, the customer relationship, and more of the margin. Uber is a real-world proof point that this model can move beyond theory.

What to watch next

Three signals will show whether this is an isolated win for AWS or the start of a broader shift.

First, watch for follow-on customers. If other large-scale companies move production AI workloads onto Trainium or Inferentia, that would show Uber is not a one-off. Named deployments matter more than pilot programs.

Second, watch for independent performance evidence. AWS has made the economic case for its chips, but third-party benchmarks, customer case studies, and audited enterprise results will matter more than vendor claims. The key question is simple: do these chips deliver strong enough price-performance on real workloads to justify switching?

Third, watch NVIDIA’s behavior. If NVIDIA starts adjusting pricing, bundling more aggressively with cloud partners, or moving faster to defend hyperscaler relationships, that would suggest the pressure is real. Supplier strategy often reveals competitive stress before market share data does.

Uber’s decision does not break NVIDIA’s hold on AI infrastructure. It does show that the hold is finally being tested in production by a customer that matters. If more enterprises follow, the AI stack becomes less dependent on a single chip supplier and more controlled by the hyperscalers that run the cloud. That would reshape where power and profit sits in AI.

Amazon (AMZN) and Uber (UBER) have a Disruption Score of 2.

Click here to learn how we calculate the Disruption Score.