Trainium/Inferentia traction grows
Amazon’s Trainium and Inferentia chips are picking up adoption among major model players — coverage cites Anthropic, OpenAI and Apple as moving inference workloads to Amazon silicon for cost/efficiency. That adoption is shifting inference economics and forcing hybrid infra conversations. (x.com)(x.com)(techcrunch.com)
AWS agreed to supply OpenAI with 2 gigawatts of Trainium compute capacity as part of a broader strategic partnership that included a reported $50 billion investment from Amazon in OpenAI. (aboutamazon.com) TechCrunch reported that Amazon has deployed roughly 1.4 million Trainium chips across its fleet and that Anthropic’s Claude runs on more than 1 million Trainium2 chips. (techcrunch.com) AWS says Trainium2 powers the majority of inference traffic on its Bedrock managed model service and made Trn2 instances and Trn2 UltraServers generally available to customers in late 2024 and 2025. (techcrunch.com) Apple executives have confirmed ongoing use of AWS’s Inferentia and Graviton silicon for services like search and said the company is evaluating Trainium2 for pretraining components of Apple Intelligence. (cnbc.com) AWS’s Inferentia-powered Inf1 instances claim up to 2.3x higher throughput and up to 70% lower cost per inference versus comparable EC2 options, figures AWS markets to justify migrations away from general-purpose GPU instances. (aws.amazon.com) On March 13, 2026 AWS announced a multi‑year collaboration with Cerebras to pair Trainium servers for “prefill” with Cerebras CS‑3 wafer‑scale engines for “decode,” a disaggregated inference design AWS says will deliver multiple‑times higher token throughput for Bedrock customers. (press.aboutamazon.com) Consulting and industry analyses note inference now drives the bulk of AI operational costs and is pushing organizations toward three‑tier hybrid models (public cloud for training, private infra for high‑volume inference, edge for latency‑sensitive workloads). (deloitte.com)
