Amazon Scales Back Its Own AI Models to Focus on Infra, OpenAI and Anthropic
Trending Topics Jakob Steinschaden ● Covered by 2 sources
Amazon is quietly shelving most of its own AI models. It's betting its future on OpenAI, Anthropic and AWS infrastructure instead.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon just admitted, in the roundabout way big companies admit things, that it's stepping back from the race to build the smartest model on the planet. Several US outlets report the company is deprecating most of its in-house Nova lineup -- the high-end Nova Premier, the multimodal Omni model, the Reel video generator and the Canvas image tool -- shifting them into what employees internally call KTLO, or 'keep the lights on' status. That's engineering shorthand for: still supported, no longer a priority. An Amazon spokesperson insists this isn't a retreat, pointing to continued investment in frontier work and a promise of clear migration paths for customers. But the timing, right after layoffs in the AGI organization and the shutdown of the AGI Lab, tells its own story.
That lab was built in December 2024 around David Luan, the Adept co-founder, and it closed along with its San Francisco site after Luan left in February. The broader AGI unit, created in 2023 under longtime Alexa executive Rohit Prasad, has now been folded under Senior Vice President Peter DeSantis, alongside Amazon's custom silicon and quantum computing groups. Where Prasad ran parallel model families across text, image and video, DeSantis is narrowing the aperture, funneling talent and compute toward Frontier Model Research, the group led by UC Berkeley's Pieter Abbeel, who joined Amazon via its Covariant acquisition. A new flagship model out of that effort is expected to surface at re:Invent this fall, possibly still under the Nova name, alongside survivors like Nova 2 Sonic, Nova 2 Lite, Nova Forge and the Nova Act agent tech.
And while Amazon trims its own model ambitions, it's throwing its cloud wide open to the competition. The AWS-OpenAI partnership, announced in November 2025 at $38 billion, has OpenAI running workloads on EC2 UltraServers packed with hundreds of thousands of Nvidia GB200 and GB300 GPUs, with full capacity due by the end of 2026 and room to grow into 2027. By April 2026 that turned into an actual product: GPT-5.5, the Codex coding agent and Amazon's Bedrock Managed Agents all landed on Bedrock in limited preview, authenticated through AWS credentials and billable against existing cloud commitments. Codex alone reportedly pulls in more than four million weekly users. GPT-5.5, GPT-5.4 and Codex went generally available in June; the GPT-5.6 family followed in July.
The Anthropic relationship runs even deeper. On April 20, 2026, Amazon added another $5 billion in direct investment plus up to $20 billion in milestone-based capital, pushing its cumulative stake to roughly $13 billion. Anthropic, in turn, committed to spending more than $100 billion on AWS technology over the next decade and locking in up to 5 gigawatts of Trainium capacity. The backbone here is Project Rainier, fully operational since October 2025, where Anthropic says it's using more than a million Trainium2 chips to train and serve Claude -- with close to a gigawatt of combined Trainium2 and Trainium3 capacity expected online by year-end. Over 100,000 customers now run Claude on AWS, and Anthropic is working directly with Amazon's Annapurna Labs on future chip generations.
Put it together and a pattern emerges that Amazon has never quite said out loud: it's less interested in owning the best model than in owning the rails everything runs on -- Bedrock as the marketplace, Trainium and Annapurna as the silicon, and equity stakes in the labs supplying the intelligence. Two years ago Andy Jassy personally framed the AGI unit as the team that would deliver Amazon's most ambitious foundation models, and just last year's re:Invent showcased Nova Omni 2 as the flagship reasoning model. Less than three years on, that line is being wound down. Amazon's Q2 earnings on July 30, alongside roughly $200 billion in planned 2026 capital expenditure, should show whether this bet on infrastructure over ownership is paying off.
My take — AI-written commentary, not fact-checked reporting
Amazon spent two years talking like it would out-build OpenAI and Anthropic at their own game, and now it's quietly settling for renting them the building instead. That's not humiliation, it's arithmetic: infrastructure margins are dull but dependable, while chasing frontier models has a way of burning cash with no guaranteed payoff. Expect other hyperscalers to make the same calculation eventually and rebrand it as strategy rather than surrender.
Read more about this at: Trending Topics