Amazon has been making its own chips for years. Until recently, however, few people talked about it as a major business. That has just changed. At the company's first-quarter earnings conference in April, CEO Andy Jassy announced that its custom silicon division had surpassed an annual revenue run rate of $20 billion. And he says that still does not come close to showing how large it really is.
Amazon sells its own chips primarily to itself, specifically to its AWS cloud division. But if it offered them freely on the market like Nvidia or AMD, the annual revenue run rate would climb to $50 billion, according to Jassy. "If our chip business operated independently and sold chips manufactured this year both to AWS and other customers, as other major chip companies do, our annual revenue run rate would be $50 billion," Jassy said. "If we can estimate it this way, our custom silicon division ranks among the world's three largest manufacturers of data center chips."
One of the three largest in the world. And yet it practically did not exist five years ago. Growth exceeds 100 percent year over year, and in the first quarter alone, the division grew by almost 40 percent compared with the previous quarter. The division sells three types of chips. Graviton processors are designed for general-purpose computing. Trainium powers artificial intelligence training. And Nitro handles security and networking.
Trainium is the flagship chip
The clear star is Trainium. This chip is Amazon's direct response to Nvidia's dominance in artificial intelligence. And demand for it is enormous. How enormous? Amazon has amassed orders and agreements worth more than $225 billion. For this single product line alone. To put that figure into perspective, it is roughly one and a half times what the entire AWS cloud business currently earns in a year. Amazon has essentially presold a future version of AWS that is larger than today's.
Who is ordering Trainium? The world's two leading artificial intelligence labs, Anthropic and OpenAI. Both have signed multiyear contracts for enormous amounts of computing power. OpenAI has reserved roughly two gigawatts of capacity, while Anthropic has reserved as much as five gigawatts. Other companies, such as Uber, are joining them.
Anyone wanting to buy Trainium now is out of luck. They have to wait. Trainium2 offers about 30 percent better price-performance than comparable graphics chips. And it is largely sold out. Amazon began shipping Trainium3 in early 2026; it delivers 30 to 40 percent more performance for the same price than the previous generation and is almost entirely sold out. And Trainium4? It will not become widely available for another year and a half, yet customers have already reserved a large share of its production.
That demand is not limited to Trainium. A similar situation has arisen with Graviton processors. Two major AWS customers asked whether they could buy all available Graviton capacity for 2026. Amazon declined. Today, Graviton is used by 98 percent of EC2's thousand largest customers and, according to Jassy, offers more than 40 percent better price-performance than comparable x86 processors, a market dominated by Intel and AMD.
Trainium also powers Amazon Bedrock, which handles most of its artificial intelligence workloads. Bedrock is used by more than 125,000 customers and by nearly 80 percent of Fortune 100 companies.
Will Amazon sell chips to others as well?
Will Amazon start selling its chips directly to customers, as Nvidia does? The entire industry is waiting for the answer. Jassy has so far been cautious. He says the company is balancing its own demand against potential external sales. "I expect that we will most likely start selling them within the next few years," he said. In any case, there would not be many potential buyers. A few major artificial intelligence labs, yes, but competing cloud companies probably not.
When a company designs its own chips, it does not have to pay inflated prices for hardware from other manufacturers. Amazon's management estimates that, at scale, Trainium will save it tens of billions of dollars in capital expenditure and deliver operating margins several hundred basis points higher than if it relied on third-party chips. For a business the size of AWS, a few hundred basis points means billions of additional dollars.
That enormous volume of orders serves another purpose for Amazon. It justifies the company's astronomical spending on data center construction. Management repeatedly says it is spending with confidence because a large portion of the capacity has already been contracted in advance. The servers are only now being installed, but customers are already waiting for them.
Sources: theregister.com and faq.com
