Amazon appoints new leaders for AI, infrastructure teams

Amazon CEO Andy Jassy announced a leadership shuffle affecting its AI and infrastructure teams on December 17.

Peter DeSantis, a longtime Amazon executive, will lead a new group overseeing advanced AI models, custom silicon development, and quantum computing.

Pieter Abbeel, co-founder of robotics firm Covariant and an AI researcher, will head Amazon’s frontier model research within its AGI team while continuing work with the robotics division.

AWS Utility Computing leaders will remain in charge of their current areas, with some added responsibilities, and more details on the new AWS structure are expected soon.

Rohit Prasad, who has overseen Alexa and Amazon’s AGI efforts, will leave the company at the end of the year after more than a decade.

Amazon said these changes aim to better align leadership with its expanding focus on AI and infrastructure technology.

🔗 Source: Amazon

🧠 Food for thought

Implications, context, and why it matters.

Amazon plans $125 billion in 2025 capex 1. Most targets AI infrastructure and custom silicon (AWS-designed chips) 2. That ranks among its largest spending cycles 3.
Peter DeSantis now leads custom silicon (Trainium for training and Inferentia for inference) 2. AWS demand for AI services already exceeds available chips 2.
Migration speed is the bottleneck 4. Models must compile with the AWS Neuron Software Development Kit (SDK) 4, which needs engineering work to tune workloads 4.
Oppenheimer pegs each extra gigawatt at $3 billion in AWS revenue 5. That payoff relies on customers choosing Amazon’s chips rather than NVIDIA across clouds.
Inferentia can cut inference costs by up to 70% 4. Trn1 training can be up to 50% cheaper than comparable Elastic Compute Cloud (EC2) virtual machine instances 6. Models need recompilation 4. Many workloads run best under roughly 10 billion parameters 4.
System integrators (specialist consulting firms that design, implement complex IT systems) plus consultancies can build migration practices 7. Work spans moving GPU workflows to Trainium or Inferentia 7. It covers fixing compile errors and removing runtime bottlenecks 7.
AWS added more than 3.8 gigawatts of data center power in the past 12 months and plans to double that through 2027 5. Early Neuron SDK experts can stake out the market before it matures.