AWS is seeing a surge in interest for its custom chips, Trainium and Inferentia. As companies aim to cut costs and lessen reliance on Nvidia GPUs, these chips are proving transformative. The evidence suggests this is not merely a side venture but a vital business component for Amazon.
Many organizations are shifting inference tasks from Nvidia to Inferentia, achieving cost reductions of 80-90%. Such savings are substantial, indicating a clear path from concept to market viability for AI products.
When it comes to training, Trainium chips are carving out their place in the technology landscape. While they don't yet have the extensive ecosystem of Nvidia’s CUDA platform, organizations that adapt their workloads to Amazon's hardware find the price-performance ratio attractive, making the switch worthwhile.
For example, Anthropic, an AI safety lab working on the Claude model family, uses a cluster of around 500,000 Trainium chips for its Rainier project. This deployment is reportedly yielding a fivefold uptick in computing power compared to prior AI models. Amazon has invested $8 billion in Anthropic and has designated AWS as its principal cloud and training partner.
Additionally, OpenAI, known for its ChatGPT, plans to deploy about 2 gigawatts of Trainium capacity, with expectations for ramp-up starting in 2027. By early 2026, AWS's custom silicon efforts, which include Trainium, Inferentia, and the Graviton processors, are projected to exceed a $20 billion annual run rate.
While Nvidia remains the leading force in AI training, holding a substantial advantage thanks to CUDA, its proprietary software ecosystem, the landscape is competitive. Nvidia’s latest GPU architectures continue to raise the bar on performance metrics, critical for training advanced models.
Looking ahead, all three major cloud providers—Amazon, Google, and Microsoft—are committing significant resources to custom silicon development. Google has introduced its sixth-generation TPU line, while Microsoft is advancing its Maia chips.
The upcoming 18 months will be crucial. Major capacity increases in 2027 may confirm whether Amazon's custom silicon ambitions can result in lasting market share growth or expose challenges in manufacturing and software development.