Amazon Closes San Francisco AGI Lab, Leaving Its Research Strategy Less Clear

GeekWire, citing The Information, put the team at about 80 people at its peak. More than a dozen of the Adept hires had already left, including Luan, before the closure. The report does not establish how many of the peak team remained or are affected now.

Continuity claims meet a leadership reset

Abbeel joined Amazon when it licensed Covariant’s technology and hired its team. Amazon said in the company-confirmed report that he will continue to lead the frontier-model research lab. His background in robotics and the Covariant deal explain why he is central to Amazon’s remaining agent and model research, but Amazon has not said how that lab relates organizationally or geographically to the closing site.

Nova Act also remains available through AWS and is being used by customers, Amazon said. That is material evidence against treating the site closure alone as the end of its agent effort: the product remains on offer. It is not evidence, however, that the whole research program, staffing base or original technical agenda will be preserved.

The organizational sequence reinforces the uncertainty. Rohit Prasad, the Amazon executive who oversaw AGI, left at the end of 2025; Luan left in February. In December, Amazon placed AGI under Senior Vice President Peter DeSantis, a longtime Amazon executive who also leads a group covering silicon development and quantum computing, as the report detailed.

Amazon’s spokesman said in that report that the company has been building large AI models for several years and is sharpening its focus on initiatives that matter most to customers. The spokesman described the eliminated roles as difficult decisions within parts of AGI, while saying Amazon would continue investing in areas important to customers’ future. That is Amazon’s explanation, not a disclosure of why particular roles or projects were selected.

The companywide backdrop is also broader than this lab: Amazon cut 16,000 jobs in January, the report noted. Public posts cited by another report said the new AGI cuts reached model-customization and post-training groups; Amazon has not confirmed the affected teams or the full scope.

The money points to reallocation, not a settled verdict

The cuts sit beside continuing spending on the infrastructure and deployment side of Amazon’s AI strategy. One account puts Amazon’s projected 2026 capital expenditure at $200 billion while describing investment in AI infrastructure, custom silicon and foundation models. The account does not describe that figure as an AGI Lab budget, so it cannot show how much research capacity the closure removes.

Separately, AWS is pursuing a $1 billion effort to embed engineers with businesses building AI agents, according to the report. AWS also continues to offer agent-oriented tools including Bedrock AgentCore. These are deployment investments, not a like-for-like substitute for a research lab working on long-horizon agent capabilities.

The tension has produced different interpretations. One analysis frames the closure and leadership exits as Amazon stepping back from a frontier-model race in favor of selling AI infrastructure and business-facing agents. The company’s retained research lab, live Nova Act service and stated spending make that conclusion premature; at the same time, they do not resolve what research capacity has been lost.

What would resolve the question

Amazon needs to disclose the affected headcount, the destination of the closing lab’s projects and whether employees can transfer into Abbeel’s group or other AGI teams. It also needs to show whether the research agenda—reasoning, planning, world models and agents for physical settings—will survive in a form that can feed AWS products.

Until then, the best-supported reading is not that Amazon has abandoned AI research, nor that nothing material has changed. It has closed a specialized lab that had about 80 people at its peak while retaining selected research and commercial programs. The missing staffing and transfer information is the evidence needed to tell a focused reorganization from a deeper reduction in the company’s agent-research capacity.