Amazon AI chief says AWS will catch OpenAI and Anthropic on frontier models within a year

Peter DeSantis, Amazon's senior vice president overseeing semiconductors, AI, and quantum, told CNBC he hopes Amazon will compete with OpenAI and Anthropic on frontier models within the "coming year" after acknowledging the company's models have not been at the frontier for the most demanding workloads. For AI builders and enterprises relying on AWS AI infrastructure, Bedrock, and Nova2, this signals that Amazon is betting on a long-term infrastructure advantage rather than a quick model catch-up.

What happened

DeSantis said Amazon has been taking a "deliberate approach" to get its foundations right in data, architecture, and infrastructure. Amazon's AI strategy is two-pronged: Bedrock, a marketplace for third-party models accessible to AWS customers, and Nova2, its in-house model released in December. DeSantis said Nova2 has about 50,000 customers but admitted, "I'm not sure we're there yet" with frontier capability. The company also focuses on custom semiconductors under the Trainium and Graviton brands. DeSantis drew parallels to Nvidia in chip development, noting that Amazon can design, produce, and deploy its own chips.

"We're one of a very few ... players who have the ability to design a chip, design the physical attributes of that chip, and then do the production of that chip. And so I think when you're thinking about us, you should be thinking about us relative to them [Nvidia]," DeSantis said.

Currently, Amazon effectively rents out its compute capacity via its cloud division Amazon Web Services, with Anthropic among its biggest customers. But CEO Andy Jassy said in April that the company could consider selling racks of its Trainium chips to third parties. DeSantis said there is no timeline for this yet, but explained the thinking behind it.

"I think that we're going to see an explosion in innovations for how people want to deploy AI infrastructure, and we want to be a part of that," DeSantis said. He left the door open to Amazon also selling its Graviton chips to third parties. "Graviton is at the centre of our strategy with chips, and it will continue to be. So, today we're not thinking about deploying that outside of AWS, but who knows," DeSantis said.

Caveats and practical notes

DeSantis's comments are forward-looking statements, not a product roadmap. No specific model names, benchmark scores, or release dates were given. Nova2's current 50,000 customer count is impressive for a young model but does not indicate frontier quality. The timeline of "coming year" remains vague. The potential to sell Trainium racks has no timeline and may not materialize. Graviton chips are described as remaining inside AWS for now.

Implications for AI builders

If you are building on AWS, your model selection and infrastructure costs depend on whether Amazon delivers competitive frontier models. Bedrock provides access to third-party models, including Anthropic, which is one of Amazon's biggest compute customers, but Nova2 is the tightly integrated Amazon model. A stronger Nova2 could reduce latency and simplify procurement. Additionally, if Amazon starts selling Trainium racks to third parties, it could create an alternative to Nvidia hardware for AI inference and training, potentially lowering costs.

Practical implications for operators

Builders should treat Nova2 as a solid mid-tier option rather than a frontier model for now. Benchmarks are not provided in this announcement, so monitoring Nova2's progression is prudent. A significant capability jump within a year could make it worth evaluating for latency-sensitive workloads on AWS. The potential external sale of Trainium racks means you should watch for pricing and availability announcements that could reshape AI infrastructure procurement. On Bedrock, the marketplace approach means you are not locked into Amazon's models; you can still use Anthropic, Cohere, or others.

Caveats

These are forward-looking statements and not a roadmapped product plan. No specific model names, benchmarks, or dates were provided. Nova2's 50,000 customers are impressive for a young model but do not prove frontier capability. The "coming year" timeline is uncertain. The Trainium external sale possibility has no timeline and may not materialize. Graviton chips are described as remaining inside AWS for now. Evaluate based on actual capabilities when they arrive, not aspirations.

FAQs

What is Amazon Bedrock and how does it fit into AWS AI offerings?

Bedrock is a managed service that provides access to a variety of foundation models from companies including Anthropic, Cohere, and Amazon's own Nova2. It is a key part of Amazon's two-pronged AI strategy, allowing AWS customers to choose third-party models without managing infrastructure.

What is the Nova2 model and how does it compare to OpenAI and Anthropic models?

Nova2 is Amazon's in-house large language model released in December. DeSantis stated it has about 50,000 customers but acknowledged it is not yet at the frontier for the largest workloads. No direct benchmarks were provided. Nova2 is considered behind the latest releases from OpenAI and Anthropic.

What are Trainium and Graviton, and how do they support AI workloads on AWS?

Trainium is Amazon's custom chip designed for AI training and inference. Graviton is its general-purpose Arm-based processor used in EC2 instances. Both are central to Amazon's strategy to optimize performance and cost for AWS AI workloads. DeSantis compared Amazon's chip design capabilities to Nvidia's.

How does AWS plan to catch up to OpenAI and Anthropic in the coming year?

DeSantis cited a deliberate focus on data, architecture, and infrastructure, along with leveraging custom Trainium and Graviton chips. The goal is to have a model that is considered among the most capable, but no specific milestones or roadmaps were shared.