Amazon spent years building its chip business away from the spotlight. Now that business has crossed a $25 billion annual revenue run rate, placing the company much deeper inside the AI infrastructure race than many people may realise.
The story did not begin with ChatGPT, generative AI or the current rush to build giant data centres. It started in 2015, when Amazon acquired Israeli chip design company Annapurna Labs.
At the time, the logic was fairly practical. AWS wanted processors built specifically for cloud computing instead of depending entirely on general-purpose hardware from outside suppliers.
A decade later, that decision looks much bigger.
Amazon’s Chip Strategy Started Before the AI Boom
Amazon acquired Annapurna Labs years before artificial intelligence became the centre of almost every major technology investment.
The company believed that designing its own processors could help AWS deliver stronger performance, lower operating costs and better energy efficiency. It also gave Amazon more control over the hardware running inside its cloud infrastructure.
That early work eventually produced several chip families with very different jobs.
Trainium handles AI training and inference. Graviton powers general cloud computing and an increasing number of agent-based AI workloads. Nitro manages much of the networking, storage and security work operating behind AWS services.
Together, those chips have turned into one of Amazon’s fastest-growing businesses. Amazon says its custom silicon operation now exceeds a $25 billion annual revenue run rate and continues to grow at triple-digit percentages year over year.
Trainium Is Amazon’s Answer to Expensive AI Computing
Training a large AI model requires an extraordinary amount of processing power. Frontier developers may run hundreds of thousands of chips continuously for weeks or months.
Then comes inference—the equally expensive process of running the finished model whenever someone asks a question, generates an image or uses an AI agent.
Amazon designed Trainium specifically for those workloads.
Unlike general-purpose processors, Trainium chips focus on the calculations used to train and operate AI models. Amazon argues that this specialised design gives customers better price-performance than relying entirely on conventional GPU infrastructure.
Trainium3, the latest generation discussed by the company, delivers up to 40% better price-performance than Trainium2. Amazon has also developed Trn3 UltraServers that can connect as many as 144 Trainium3 chips inside one integrated system. According to Amazon, those systems deliver up to 4.4 times more computing performance than Trainium2 UltraServers.
Those numbers sound technical. The business argument is simpler: models can train faster, consume less energy and cost less to operate.
That matters when the bill involves entire data centres rather than a few servers.
Anthropic and OpenAI Are Making Large Trainium Commitments
Amazon’s AI chip ambitions are no longer based only on internal workloads.
Anthropic has committed to using up to five gigawatts of current and future Trainium capacity for training and running its Claude models. Amazon says Claude workloads already operate across more than one million Trainium2 chips.
OpenAI has also committed to consuming two gigawatts of Trainium capacity through AWS infrastructure, with the arrangement expected to begin in 2027.
These agreements give Amazon something its custom chip programme badly needs: serious external validation.
Convincing companies to leave familiar GPU software ecosystems is difficult. Developers have spent years building tools and workflows around established hardware platforms. A cheaper chip is not automatically useful when moving an entire model introduces new engineering work.
Large commitments from Anthropic and OpenAI suggest Amazon is beginning to overcome that problem, at least among customers willing to operate at enormous scale.
Graviton Could Matter Just as Much as Trainium
Trainium attracts attention because it directly competes for AI workloads. Graviton may be the quieter part of the strategy.
Graviton processors handle the general computing work behind websites, databases, applications and cloud services. That role is expanding as AI systems become more agentic.
An AI agent does not only generate an answer. It may search databases, call software tools, process payments, write code or complete several tasks in sequence. Those activities require far more than specialised AI accelerators. They also generate heavy demand for conventional CPU resources.
Amazon says Graviton is already used by 98% of its top 1,000 EC2 customers. More than 130,000 customers currently use Graviton-based servers, while over half of the new processing capacity added to AWS runs on Amazon-designed Graviton chips.
Meta has agreed to deploy tens of millions of Graviton cores for CPU-heavy agentic AI workloads. Uber uses Graviton for its rider-and-driver matching systems and is also piloting Trainium3 for model training.
Amazon is covering both sides of the workload: specialised AI processing through Trainium and the supporting cloud operations through Graviton.
Project Rainier Shows the Scale Amazon Is Chasing
Amazon is not treating custom chips as isolated components. It is connecting them into enormous computing systems.
Project Rainier is one of the clearest examples.
The AWS cluster was built to support Anthropic and contains a vast network of Trainium chips operating across multiple data centres. Earlier deployments used nearly half a million Trainium2 chips, while Amazon now says Claude models run on more than one million of them.
This is where Amazon’s cloud advantage becomes difficult to ignore.
Designing an AI chip is only one part of the job. Those chips must also be manufactured, installed, connected, cooled and supported by software that lets developers use them without fighting the hardware.
AWS already operates the data centres, networking systems and cloud services surrounding those processors. Amazon can design the chip and build the environment where customers rent it.
That vertical control could become one of its strongest advantages.
Amazon Wants More Control Over the Cost of AI
The AI infrastructure market has been shaped by high hardware prices, limited chip availability and heavy dependence on a small number of suppliers.
Amazon’s custom silicon programme gives the company another route.
Building its own processors may lower the cost of running AWS services. It can also reduce Amazon’s exposure to supply shortages while giving customers another option for large-scale AI training and inference.
There is a competitive motive too.
Microsoft, Google and other major cloud companies are also developing custom processors. None of them wants its AI future controlled entirely by an outside chipmaker.
Amazon’s response has been unusually broad. It has processors for model training, inference, general computing, networking, storage and security. Trainium4 is already under development, while Graviton continues to evolve around the growing demand created by AI agents.
Amazon Is Becoming an AI Infrastructure Company
Amazon will not stop being a retailer or cloud provider. Its chip business does show how much the company’s identity has changed.
AWS once made computing infrastructure available through the internet. Amazon is now designing the processors inside that infrastructure, connecting them into huge AI clusters and selling access to some of the largest model developers in the world.
The $25 billion revenue run rate is significant, but the larger story sits underneath it.
Amazon wants to control more of the AI stack—from silicon and servers to foundation models and managed services such as Amazon Bedrock. Bedrock already runs most of its inference workloads on Trainium, according to the company.
The chip programme that began as a cloud-efficiency bet has become a central part of Amazon’s AI strategy.
It took ten years to reach this point. The next phase may move much faster.
