Amazon's AI Resurgence: AWS and Anthropic's Multi-Gigawatt Trainium Expansion (semianalysis.com)

Amazon is staging a major AI comeback through its partnership with Anthropic, fueled by a multi-gigawatt expansion of AWS datacenters powered primarily by Amazon’s custom Trainium2 chips. Despite lagging Nvidia and Google in raw chip performance, Trainium2 offers a competitive total cost of ownership (TCO) advantage—particularly in memory bandwidth efficiency per dollar—which aligns well with Anthropic’s aggressive reinforcement learning training approach. This bet on custom silicon combined with Anthropic’s scaling strategy could help AWS regain lost ground in the GPU-dominated cloud AI market, where it has struggled due to networking inefficiencies and slow time-to-market compared to rivals like Microsoft Azure and Google Cloud. Anthropic’s rapid revenue growth—from $1B to $5B annualized in 2025—and its expanding $13B funding round underscore its rising influence as AWS’s anchor AI customer. While a significant portion of Anthropic’s current cloud spending still flows to Google Cloud for TPU-powered inference, new AWS campuses with over 1.3 gigawatts of IT capacity are nearing completion, poised to significantly increase AWS’s share of Anthropic’s AI training workload by the end of 2025. This infrastructure buildout, paired with Anthropic’s hands-on role in the Trainium roadmap, signals a deep hardware-software co-design trend, potentially positioning Anthropic and AWS alongside Google DeepMind as leaders in tightly integrated AI system development. Amazon’s AI resurgence is therefore not just about scale but strategic alignment—leveraging Trainium’s cost efficiencies and Anthropic’s training demands to carve out a unique niche in the fiercely competitive AI cloud ecosystem. While challenges remain, especially in networking technology and ecosystem software, this partnership illustrates an alternative path in AI cloud dominance beyond Nvidia GPUs and signals a renewed AWS growth trajectory surpassing 20% year-over-year by 2025.