Everyone is watching OpenAI $OPENAI ( 0.0% ) and Google fight over the best AI model. But while that battle gets all the attention, Amazon AI chips are quietly running the infrastructure underneath all of it.

The real money in AI does not go to whoever builds the smartest model. It goes to whoever owns the pipes, the power, and the compute that every model depends on. And right now, that is Amazon.

This article breaks down how Amazon $AMZN ( ▲ 3.81% ) AI built a chip empire most people have never heard of, why it already generates over $20 billion a year, and what it means for anyone building or using AI tools today.

Most people think of AI as chatbots and image generators. But all of that needs computing power to run, and computing power needs chips.

Right now, NVIDIA dominates the AI chip market. Their GPUs are what most companies use to train and run AI models. The problem is NVIDIA chips are expensive, hard to get, and create a single point of dependency for the entire industry.

Amazon saw this problem early. So they built their own.

Amazon has three custom chips you should know about.

Amazon's own CPU. It handles everyday computing tasks on AWS and gives about 40% better value for money compared to standard chips. Right now, 98% of AWS's biggest customers are using Graviton.

Amazon's answer to NVIDIA for training and running AI models. It costs about 30% less than similar NVIDIA options. The next version, Trainium2, is almost sold out. Trainium3 is almost fully booked after its first shipments in 2026.

The invisible layer of the whole system. It takes care of security and networking across thousands of AWS servers running at the same time, all behind the scenes.

Put together, these three chips are bringing in over $20 billion in revenue every year, growing at triple digits year over year. This is not a small test. This is a real business.

NVIDIA builds chips and sells them to whoever will buy. Amazon builds chips and uses them inside AWS, then passes the cost savings to customers.

This creates a different kind of advantage.

When you use Amazon Bedrock to access Claude or other AI models, the inference is often running on Trainium chips. When you spin up an EC2 instance on AWS, it likely runs on Graviton. You never see the chip, but you feel it in the price.

Here is a simple way to think about it. NVIDIA is like a company that sells engines to everyone.

Amazon is like a taxi company that built its own engine, cheaper and optimized for their own roads. Now they are starting to sell rides to everyone else, and the rides cost less because they made the engine themselves.

Amazon CEO Andy Jassy wrote in his 2025 shareholder letter that demand for Trainium chips is so high that it is quite possible they will start selling racks of chips directly to third parties soon.

If that happens, Amazon enters the chip market as a standalone business. At current production rates, Jassy estimates this could generate around $50 billion in annual revenue, larger than AMD's recent run rate, at semiconductor margins of 40 to 50%.

AWS AI revenue hit a $15 billion annual run rate in early 2026. That is the first time Amazon publicly shared this number.

To put it in context, three years after AWS launched as a product, its total revenue run rate was $58 million. 3 years into the AI wave, it is $15 billion. Jassy called it the fastest commercial adoption of any technology Amazon has ever seen.

Amazon plans to spend approximately $200 billion in capital expenditures in 2026. Most of it goes toward data centers, power, servers, and networking.

The OpenAI contract alone is worth over $100 billion, meaning the capacity being built is largely pre-sold before it even exists.

One project worth knowing about is Project Houdini. This is Amazon's internal initiative to speed up data center construction by shifting from manual on-site building to factory-assembled modular units.

The goal is to cut tens of thousands of labor hours and bring new computing capacity online much faster.

Beyond chips, Amazon is also dealing with a constraint most people overlook, which is electricity. AWS added 3.9 gigawatts of new power capacity in 2025 and plans to double total power capacity by end of 2027.

Nearly half of all planned U.S. data centers in 2026 are delayed, not because of money, but because the power grid cannot keep up. Amazon locked in power agreements early, which gives them a structural advantage competitors cannot easily copy.

85% of global IT spending still runs on company-owned servers, not the cloud. AWS, at a $142 billion annual run rate, represents only about 15% of total global IT spend.

That means the cloud migration is far from finished. Even before adding a single AI workload, just the base shift from on-premise servers to cloud is one of the longest growth stories in enterprise technology right now.

Add AI on top of that and the numbers get hard to wrap your head around. Jassy has internally projected that AWS alone could reach $600 billion in annual revenue by 2036. Amazon's entire company did $717 billion last year.

He is saying one division, AWS, could approach that number in a decade, at operating margins three times higher than the rest of the business.

At 35% operating margins, $600 billion in AWS revenue generates $210 billion in annual operating income from one division alone.

This is the infrastructure play in its clearest form. The Medici family in 15th century Florence did not paint, sculpt, or build cathedrals. They financed everyone who did and took a cut of every transaction.

OpenAI is building the models. Anthropic is building the models. Everyone is building on top of AWS. Amazon collects the infrastructure fee from all of it.

If you are a founder, operator, or creator working with AI tools, a few things are worth knowing.

The cheapest way to run AI models at scale is increasingly through Amazon Bedrock, because Trainium chips make inference cheaper than standard GPU options.

You can access frontier models including Claude, Llama, and others through a single API without managing your own servers.

The companies building on top of the cheapest infrastructure win on margin. Right now, that infrastructure is powered by Amazon AI chips.

The AI race looks like a competition between models. But underneath, it is a competition between infrastructure providers. And in that race, Amazon is not just competing. It is the ground everyone else is running on.

⚠️ Disclaimer: This newsletter is for informational purposes only, just for fun and knowledge. This is not investment advice. Your money, your responsibility!