Amazon Shuts Its AGI Lab and Cuts Jobs to Chase Enterprise AI Instead

Amazon has closed its San Francisco AGI Lab and cut jobs in its artificial general intelligence group, shifting focus to enterprise AI via AWS. The lab, opened December 9, 2024, with David Luan and Pieter Abbeel as leaders, lost Luan in February and AGI head Rohit Prasad by end of 2025. Amazon is winding down most Nova models, including Nova Premier, Nova Omni, Nova Reel, and Nova Canvas, while raising its 2026 capital spending plan from $200 billion to $220 billion, with AWS revenue growing 37% year over year.

Amazon has closed its San Francisco AGI Lab and cut jobs inside the group, but the company isn't quitting AI. It's choosing the AWS customer over the leaderboard. Amazon just gave its AI researchers a blunt message: building the smartest model in the world isn't the main plan. Reuters reported on July 22, 2026, that Amazon cut jobs in its artificial general intelligence group, and The Information reported that the company closed the San Francisco AGI Lab it created in late 2024 around talent from Adept. This is not a small personnel shuffle. The lab was barely 18 months old, and it was supposed to give Amazon a sharper answer to OpenAI, Google DeepMind and Anthropic. Amazon opened the AGI SF Lab on December 9, 2024, with David Luan and Pieter Abbeel named on Amazon Science as the leaders attached to the project. Luan had co-founded Adept, worked at OpenAI and led large language model work at Google. He left Amazon in February, according to GeekWire. Then Rohit Prasad left too. GeekWire reported in December 2025 that Prasad, the executive Amazon had put in charge of AGI in 2023, was leaving at the end of that year. Andy Jassy's memo put Peter DeSantis, a 27-year Amazon veteran known for AWS infrastructure, custom chips and quantum computing, over the newly combined AI organization. You can call that a reorg. Don't stop there. A lab built around agent research loses its founding leader, then the broader AGI boss leaves, then the lab shuts. That is a strategy changing shape in public. The Models Left Behind Business Insider reported on July 28 that Amazon has started winding down most of its flagship Nova models, including Nova Premier, Nova Omni, Nova Reel and Nova Canvas. Some employees described those models as being in KTLO status, short for keep the lights on. They can still support existing customers, but they are no longer where the company is putting its best engineering attention. That detail matters because Nova was Amazon's own answer to the foundation-model race. Nova Premier was the high-end model. Reel handled video generation. Canvas handled images. Omni was meant to cover multimodal work. Those are not side projects. They are exactly the categories where OpenAI, Google and Anthropic have spent the past two years trying to prove whose model can see, talk, reason and generate better than the rest. Amazon's public statement tried to make the move sound orderly. The company told Reuters it had been building large AI models for several years and was sharpening its focus on initiatives that matter most for customers, including by eliminating some roles in parts of the AGI organization. Fair enough. But corporate language often sands down the hard edge of a decision. The hard edge here is simple. Amazon is narrowing the work. What remains is more pointed. Business Insider said resources are shifting toward frontier model research led by Pieter Abbeel, with a new flagship model expected at AWS re:Invent later this year. Nova 2 Sonic, Nova 2 Lite, Nova Forge and Nova Act still sit in the picture. That is not the same as trying to maintain a full spread of text, image, video and multimodal flagships at once. AWS Wants The Customer Work Here's the thing about Amazon and frontier models: it never fully acted like OpenAI. It put up to $8 billion into Anthropic, offers Claude through Bedrock, and sells customers access to a menu of models rather than insisting everything has to be Amazon-built. That is a very AWS way to behave. If your customer wants Claude, Llama, Nova or something else, you sell the rails. The money points in the same direction. AP reported on July 31 that Amazon raised its 2026 capital spending plan from $200 billion to $220 billion after strong second-quarter results, with AI and other technology driving the increase. Amazon also said AWS revenue grew 37% year over year, according to the same AP report. That is the business Amazon understands best: buy the infrastructure, rent it out. The new AWS Forward Deployed Engineering group fits that perfectly. Amazon announced on June 30 that AWS would invest $1 billion in a team embedding thousands of engineers with customers to build agentic AI systems. Amazon named the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh and Southwest Airlines as early customers. This is not a lab demo. It is Amazon sending people into customer environments to get production systems built. For you, if you run on AWS, the practical effect may be more important than another benchmark score. Bedrock still gives you model choice. Nova Act is still available as a browser-agent service. The company is also putting engineers beside customers who have moved past pilots and now need AI systems tied to their own data and workflows - governance included. That is where budgets get approved. Frankly, Amazon looks tired of pretending it has to win every layer of AI by itself. Nvidia made a fortune selling chips to every side of the model war. AWS wants its own version of that trade: provide the compute and the deployment help while other labs absorb the brutal cost of chasing the next breakthrough. That does not mean Amazon is done with frontier research. Abbeel's team appears to be getting more focus, not less. But the San Francisco AGI Lab closure tells you which version of ambition lost. The broad, Adept-flavored push toward agentic AGI has been cut down. The AWS customer machine is still very much alive. Also read: NXP is chasing Ambarella for $3.3 billion because it needs eyes, not just brains https://startupfortune.com/nxp-is-chasing-ambarella-for-33-billion-because-it-needs-eyes-not-just-brains/ • California's AI Transparency Act Takes Effect With Fines That Compound Daily https://startupfortune.com/californias-ai-transparency-act-takes-effect-with-fines-that-compound-daily/ • NXP Is Circling Ambarella in a $3.3 Billion Bet on AI Vision Chips https://startupfortune.com/nxp-is-circling-ambarella-in-a-33-billion-bet-on-ai-vision-chips/