In what amounts to one of the most significant strategic retreats in the recent history of big tech, Amazon is dismantling the organizational architecture of its AGI ambitions, consolidating competing teams, reallocating prized hardware, and preparing a wave of layoffs that internal sources suggest could arrive as early as April. The restructuring marks the end of Amazon's attempt to compete at the raw frontier of artificial intelligence — and a frank acknowledgment that it was losing.
The clearest symbol of that reckoning is the departure of Rohit Prasad, announced at the end of last year, the SVP and head scientist who had led the Nova project since its inception. Replacing him is Peter DeSantis, a 27-year Amazon veteran who built the company's data center empire. The choice was not subtle. Where Prasad represented the dream of breakthrough research, DeSantis represents operational discipline, silicon, and the unglamorous infrastructure that actually generates cash. He now commands a unified organization spanning AI models, custom chips — Graviton, Trainium, Nitro — and quantum computing, reporting directly to Jassy. The infrastructure executive owns the model strategy. The message was written in the org chart.
Under DeSantis, a new platform called Project Mantle will consolidate teams that had operated as separate fiefdoms: AGI-I, AGI-O, Inference Engine, Frontend Service. These are being absorbed or dissolved. The auxiliary teams — infrastructure, information, and others — are expected to disappear entirely. Mantle, announced publicly at re:Invent 2025, functions as the next-generation inference engine behind AWS Bedrock, adding service tiers, fault-tolerant systems, and confidential computing. It is not a replacement for Bedrock so much as the engine that will quietly make it faster, cheaper, and harder to leave.
Research, meanwhile, is being centralized under the FMR team, led by Pieter Abbeel, one of the most cited robotics and machine learning researchers in the world. The Nova GPUs — including EC2 P5 instances running H100 and H200 chips — are being redirected to his group, a reallocation that strips resources from scattered teams and bets on coordinated frontier research. Whether this produces results remains, for now, an open question.
It is against this backdrop of reorganization that the human cost becomes visible. Amazon cut roughly 14,000 corporate positions in October 2025, followed by 16,000 more in January 2026. Jassy has tied both rounds explicitly to AI-driven efficiency: machines doing the work of middle management. The AGI division is not being spared. The departure of David Luan, who led Amazon's AGI Lab, in February 2026 added to a growing list of senior exits that signal a talent drain that restructuring alone cannot fix.
Employees who remain describe the pre-restructuring environment in terms that are difficult to put in a press release. "Man-made chaos," one said. "Intense internal competition." "Technical debt." These are the consequences of building a research organization inside a company whose core leadership principle — Be Scrappy — was designed to ship consumer goods faster than competitors, not to recruit and retain the patient, exacting minds that frontier AI demands. The cultures were never compatible.
And yet Amazon's strategic logic, stripped of its losses, is not irrational. It was never going to out-spend OpenAI, Anthropic, or Google on pure research while managing the margins of a global retailer. What it can do — what DeSantis was hired to do — is build the highway. Through Bedrock, Amazon already serves more than 100,000 organizations. It holds a meaningful stake in Anthropic. It is reportedly in advanced discussions around OpenAI. The models it cannot build, it can host, optimize, and monetize at a scale few can match.
Nova failed to win a race Amazon perhaps should never have entered. What comes next is a more honest version of what Amazon has always been: not the fastest car, but the road beneath every wheel.
not investment advice