Amazon Deepens Custom Silicon Investment in Texas
Amazon is accelerating its investment in custom silicon with a strategic focus on building out its AI infrastructure in Texas. The company is increasing its deployment of proprietary Trainium, Inferentia, and Graviton chips to optimize performance and cost for both internal and AWS workloads. This Texas-focused strategy is also aimed at improving supply chain control and reducing geopolitical risk.
Amazon's custom silicon journey began long before the current AI boom, rooted in its 2015 acquisition of Austin-based Annapurna Labs for approximately $350 million. This subsidiary became the core of its chip design efforts, which started as early as 2012 with the AWS Nitro System, now the foundation for all modern EC2 instances. The strategic pivot to in-house chips is a direct challenge to the high-margin dominance of NVIDIA, whose data center GPUs command gross margins around 75%. Developing custom silicon allows hyperscalers to break from third-party release cycles, optimize hardware for specific cloud workloads, and mitigate supply chain risks tied to geopolitical tensions. For AI training, Amazon's latest Trainium3 chip delivers 2.52 petaflops of FP8 compute and is over four times more energy-efficient than its predecessor. AI leader Anthropic is a key partner, utilizing a massive "Ultracluster" with hundreds of thousands of Trainium chips to train its frontier models, validating the architecture at scale. On the inference side, where low latency and cost-per-query are critical, the Inferentia2 chip provides 4x the throughput and 10x lower latency than the first generation. AWS claims customers using the full Trainium-Inferentia stack can achieve up to 50% lower total cost of ownership compared to traditional GPU-based instances. [27
