Amazon founder Jeff Bezos has identified the company’s expanding semiconductor operations as a potential new pillar of its business, as demand for artificial intelligence computing drives investment in specialised processors.Bezos said Amazon’s portfolio of custom chips, including its Trainium artificial intelligence accelerators and Graviton central processing units, could develop into one of the company’s most durable businesses. The assessment places semiconductor design alongside the online marketplace, Prime subscription service and Amazon Web Services as a possible long-term engine of growth.
The strategy reflects Amazon’s attempt to control more of the technology underpinning its cloud platform while reducing the cost of running data centres. Custom processors allow AWS to design hardware for particular workloads rather than depending entirely on chips supplied by companies such as Nvidia, Intel and Advanced Micro Devices.
Amazon’s chip operations, which also include Nitro infrastructure processors, have reached an annualised revenue run rate of more than $20 billion. The business is growing at more than 100 per cent a year, supported by the rapid expansion of artificial intelligence services and rising demand for lower-cost cloud computing.
Trainium is designed to train and run large artificial intelligence models, while Graviton processors handle general cloud workloads using technology based on Arm architecture. AWS says its latest Trainium systems offer better price performance than comparable graphics-processing infrastructure, although performance can vary depending on software, model architecture and workload configuration.
Demand has been strengthened by large commitments from artificial intelligence developers. Amazon has secured more than $225 billion in prospective revenue commitments linked to Trainium infrastructure, including multi-year agreements involving Anthropic, OpenAI and other major customers.
Anthropic is working closely with Amazon’s Annapurna Labs chip-design division to optimise future Trainium processors for Claude models. The collaboration gives Amazon feedback from one of the industry’s largest artificial intelligence workloads while helping Anthropic secure access to computing capacity at a time when advanced processors remain costly and heavily contested.
Trainium2 capacity has been largely sold out, while the next generations are intended to improve processing speed, memory bandwidth and energy efficiency. Trainium3 was introduced as a more powerful platform for large-scale model development, and Trainium4 is expected to arrive in 2027 with greater compatibility across artificial intelligence computing systems.
Graviton is also becoming increasingly important as artificial intelligence applications create demand for conventional computing alongside specialised accelerators. Artificial intelligence agents require processors to manage databases, applications, security controls and communications, even when the most computationally intensive model operations are handled by accelerators.
Meta has agreed to use millions of Graviton processors for artificial intelligence-related workloads, demonstrating that Amazon’s custom chips are gaining traction beyond companies traditionally associated with AWS infrastructure. Uber and other large enterprises are also adopting Trainium for selected machine-learning operations.
Amazon is examining whether it could eventually sell Trainium processors directly to customers for installation in their own data centres. Such a move would represent a major departure from the current model, under which the chips are primarily accessed through AWS cloud services.
Direct sales could place Amazon in closer competition with Nvidia, which dominates the market for artificial intelligence accelerators. Amazon, however, remains one of Nvidia’s largest customers and continues to offer Nvidia-powered cloud instances. Its strategy is therefore based on expanding customer choice rather than replacing external suppliers entirely.
The semiconductor push is tied to Amazon’s planned capital expenditure of about $200 billion in 2026, with most of the investment directed towards artificial intelligence infrastructure, data centres, networking equipment and power capacity. Management has argued that much of the spending is supported by contracted customer demand rather than speculative construction.
The scale of the investment has nevertheless raised questions about whether technology companies are building artificial intelligence infrastructure faster than revenue can develop. Amazon faces similar scrutiny to Microsoft, Alphabet and Meta, all of which are directing unprecedented amounts of capital towards computing capacity.
Custom processors could help offset those pressures by lowering the cost of each unit of computing sold through AWS. They also give Amazon greater control over supply, product development and integration between hardware and cloud software.
Topics
Technology