Amazon AI Chief: Nova 2 Aims to Catch Up with OpenAI and Anthropic 'Next Year'
Miles Bennett
Amazon AI head Peter DeSantis admitted Nova 2 has not yet reached frontier level, but said the company expects to compete head-on with OpenAI and Anthropic 'in the coming year' — putting a verifiable timeline on Amazon's self-built model ambitions.
What did Amazon actually admit?
DeSantis said it plainly: "I'm not sure Nova 2 is at the very frontier yet, but that is our goal."
He acknowledged Amazon's AI models are not yet frontier-grade on the largest, most demanding workloads.
This means → Amazon has publicly drawn a "not-yet-there" line under its own models, then staked a timeline on closing the gap — giving the market a window it can hold them to.
If they're behind, why call it deliberate?
DeSantis's explanation: Amazon prioritized data, architecture, and infrastructure over chasing benchmark rankings early.
In plain terms = Amazon chose to build the road before racing on it — betting that foundational capability matters more than near-term scores.
Nova 2 has already attracted roughly 50,000 enterprise customers — a sign Amazon is running an "enterprise-first, model-ceiling-later" playbook, the opposite direction from OpenAI's consumer-first approach.
What role do Amazon's custom chips play?
Amazon runs two chip lines: Trainium (AI training) and Graviton (general compute), both offered mainly as rental capacity through AWS.
DeSantis drew a direct comparison to Nvidia, calling Amazon one of a handful of companies that can design, physically engineer, and produce chips end-to-end.
This means → Amazon is not just chasing the model layer — it wants supply-side autonomy in compute. Higher chip self-sufficiency lowers training costs and Nvidia dependence at the same time.
Will these chips be sold externally?
CEO Andy Jassy said in April that Amazon is considering selling Trainium chip racks directly to third parties — but DeSantis gave no timeline.
On Graviton, the tone was more cautious: "We're not thinking about deploying it outside AWS today, but who knows."
In plain terms = Trainium external sales are "under consideration"; Graviton external sales are "not on the table but the door isn't locked" — two different speeds of openness.
How will the 'coming year' promise be tested?
Amazon also operates Bedrock — a marketplace letting enterprise customers call models from multiple providers — running in parallel with the self-built Nova line.
This reflects Amazon's real logic: even if its own models lag, Bedrock can still lock in enterprise customers. Whether Nova catches up affects margin structure, not client relationships.
Whether the Nova family reaches the frontier tier within a year will be the key checkpoint for Amazon's "infrastructure-first, model-second" strategy.
Content is for reference only, not financial advice.