Amazon quietly built a $20 billion chip empire right under NVIDIA's feet
While the entire technology community focuses its attention on NVIDIA, OpenAI or SpaceX's famous IPOs, there is a number that few people pay attention to but is worth more than many noisy deals combined: Amazon's own chip segment has just surpassed the internal revenue mark of 20 billion dollars a year, and if Amazon sells chips out like regular chip companies do, that number will jump to about 50 billion dollars. This is what Andy Jassy, CEO of Amazon, said himself during the first quarter of 2026 business results report, and it is enough to put Amazon in the group of three largest data center chip companies in the world, although few people consider Amazon a semiconductor company.
What do the numbers say?
Amazon's home-made chips Amazon's custom silicon segment includes three main product lines: Graviton general-purpose processors, Trainium artificial intelligence training chips, and Nitro security chips that handle background tasks for cloud infrastructure. According to Jassy, this segment is growing more than 100% each year, and Amazon currently holds more than $225 billion in revenue commitments for the Trainium line alone. To easily visualize the level of this mutation, Jassy directly compares it with the history of AWS itself: in the first three years after AWS launched, this segment only achieved revenue of $58 million a year, while in the first three years of the current AI wave, AWS's AI revenue alone exceeded $15 billion, nearly 260 times more.
Who's lining up to buy Trainium
What makes this story interesting is not just the scale, but who is placing the orders. OpenAI commits to approximately 2 gigawatts of Trainium capacity through AWS to run its most advanced models, a contract that will ramp up from 2027. Anthropic goes further, committing to securing up to 5 gigawatts of Trainium capacity now and in future generations to train and run models. These are two of the world's leading AI labs, and both rely significantly on the infrastructure of a direct competitor in the cloud segment.
OpenAI is one of the companies committed to using Trainium through AWS In addition, Meta signed an agreement to deploy tens of millions of Graviton cores for AI tasks capable of self-action (agentic AI), while Uber uses both Graviton4 and Trainium3 for its ride-hailing and delivery platform. According to Jassy, the reason Meta chose Graviton is that the chip offers up to 40% better performance per price than any other x86 processor, and is currently used by 98% of the 1,000 largest customers of EC2, AWS's virtual server service.
Why is Amazon's model different from Google or Meta?
Google builds TPU, Meta builds MTIA, OpenAI recently made its own Jalapeño chip with Broadcom, but Amazon's approach has an important difference: they do not start from the ambition to compete with NVIDIA, but from AWS's own huge internal operating needs, then turn that capacity into a separate revenue stream. Trainium2, the current generation, has about 30% better performance-for-cost than comparable GPUs and is almost sold out. Trainium3 has just started shipping in early 2026, the performance on price is 30 to 40% better than Trainium2, and is also almost fully booked. Even Trainium4, which is still about 18 months away from widespread availability, has most of its capacity reserved. In other words, Amazon is selling something that hasn't even been produced yet, and there are still buyers.
The bigger picture behind the numbers
The chip story was just one part of a more impressive quarter for Amazon. Total revenue in the first quarter of 2026 reached $181.5 billion, up 17% over the same period, while AWS alone reached $37.6 billion, up 28%, the fastest growth in 15 consecutive quarters. The Bedrock platform, which accesses AWS's platform models, processed more tokens in the first quarter than in all previous years combined, and customer spending on the platform increased 170% in just one quarter. Amazon's net profit reached $30.3 billion, a sharp increase compared to $17.1 billion in the same period last year, although a large part of it came from the $16.8 billion profit recorded thanks to its investment in Anthropic itself, meaning Amazon is benefiting double, both selling infrastructure to Anthropic and profiting from shares in that company.
The broader context is also worth mentioning: TSMC, the world's largest chip processing factory, has a full production schedule until 2027, forcing a series of large companies from NVIDIA to Tesla to find alternative sources of supply, of which Samsung is a widely mentioned option. In that context, the fact that Amazon already has a smoothly operating chip supply chain, serving its huge demand, becomes a significant advantage. NVIDIA probably doesn't need to worry right away, because demand for general-purpose GPUs is still far exceeding industry supply. But as more and more major technology companies build their own computing power, the question is no longer whether NVIDIA will lose customers, but rather how much of the future AI chip pie will be divided, and whether the winner will always be the loudest name in the news.