The BBC has published a rare inside look at the laboratory where Amazon designs its artificial intelligence chips, offering a glimpse into one of the more closely held operations within the company's hardware division.

Access to the facility is unusual. Amazon, like other hyperscalers developing custom silicon, restricts press visits to its chip design operations. The BBC report marks one of the few occasions journalists have been allowed to observe the working environment directly.

Amazon has invested heavily in custom accelerators to reduce its dependence on third-party suppliers such as Nvidia for the compute capacity that powers its AI services. The company's silicon portfolio includes the Trainium chips used for training large AI models and the Inferentia line used to run them in production. Both families serve Amazon Web Services, the cloud unit that represents the company's largest source of operating profit.

The design lab sits at the centre of that effort. Teams there handle the engineering work that determines how Amazon's data centres deploy compute for machine learning workloads at scale. Custom silicon gives AWS a lever to tune price and performance for its own workloads rather than relying solely on merchant chip vendors.

The BBC's account focuses on the people and the workspace behind that programme. Chip design is a discipline with long lead times and high engineering headcount requirements, and the report describes the environment in which those engineers operate day to day.

Amazon is not alone in this strategy. Google builds its Tensor Processing Units, Microsoft has its Maia series, and Meta has developed its own inference accelerators. Each company seeks to control the silicon layer of its AI stack as model training and inference costs rise.

For AWS customers, the outcome of the lab's work appears in the instance types offered on the cloud platform. Instances based on Trainium and Inferentia provide alternatives to GPU-based instances, and AWS has positioned them as cost-effective options for specific workload classes.

The full BBC report, including photographs and details gathered during the visit, is available on the broadcaster's website. Die Signal readers tracking the custom-silicon strategies of the major cloud providers will find it a useful data point on how Amazon organises its chip engineering effort.