Amazon is taking a big swing at NVIDA with plans for a new "Ultracluster" supercomputer, powered by its own Trainium chips. The AI supercomputer will support Anthropic, an AI startup Amazon has heavily invested in, and aims to rival NVIDA's dominance in AI model training.
Key Points:
- Amazon Web Services (AWS) wants to reduce AI training costs and provide an alternative to NVIDA's expensive GPUs.
- Anthropic, backed by Amazon's $8 billion investment, will use this supercomputer for its AI projects.
- The Ultracluster could be one of the largest AI supercomputers in the world by 2025.
Driving the news: AWS announced plans for the Ultracluster at its annual re:Invent conference, highlighting a push to make AI model training cheaper and more efficient. The new supercomputer, called Project Rainier, is set to include hundreds of thousands of Trainium chips and promises to offer a cost-effective option for companies seeking AI infrastructure. According to AWS VP Dave Brown, it will deliver five times the computing power Anthropic currently uses.
Between the lines: AWS also introduced the Ultraserver, a new server model with 64 Trainium chips, designed to lower AI model training costs by up to 40%. This move comes as Amazon tries to shake up NVIDIA's 95% hold on the AI chip market.
Ultraserver Benefits: AWS claims that its Ultraserver can drastically cut AI model training costs, with Apple’s early testing showing potential savings of 50% for AI applications.
The big picture: The AI chip market is booming, with projections of growth from $117.5 billion in 2024 to $193.3 billion by 2027, per IDC. Amazon's entry with Trainium aims to diversify options for companies wanting alternatives to NVIDIA's high prices and supply constraints.
Strategic Diversification: Amazon is not severing ties with NVIDIA. AWS is also working on Project Ceiba with NVIDIA, indicating a balanced strategy—offering its own chips while still collaborating with the GPU leader.
Zoom out: Amazon’s AI ambitions don’t stop at Anthropic. It’s been expanding its in-house hardware lineup for years, starting with Graviton chips. Now, Trainium chips are the next step in reducing reliance on NVIDIA and expanding AI capabilities.
What's next: AWS is already working on Trainium3 chips, promising 4x more performance than current versions when they launch in late 2025.