34 Amazon Research Awards Build on Trainium recipients announced

Amazon is handing out chip time instead of cash, and it's making a specific bet with this round: that responsible AI research is compute-starved, not idea-starved. The company just named 34 recipients of its Build on Trainium awards, a $110 million program that gives university researchers credits to run experiments on AWS's custom AI silicon rather than waiting in line for scarce Nvidia GPUs.

This particular cycle, announced under the Fall 2025 "Responsible AI" call for proposals, narrowed the scope on purpose. Instead of funding anything vaguely AI-related, Amazon asked for work in five buckets: AI safety and alignment, multilingual language models, representation engineering, sustainability paired with small language models, and synthetic data generation. The winning list reads like a cross-section of what's keeping AI researchers up at night in 2025 — hallucination detection at Northwestern, differential privacy for synthetic data at Georgia Tech and UC Santa Barbara, youth-safety guardrails for multimodal generation at University of Illinois Urbana-Champaign, and a project at MIT benchmarking how AI agents misuse tools through the Model Context Protocol.

The scale of some of this work is the real story. Yida Wang, an AWS principal applied scientist, pointed to UIUC researchers testing topology-aware parallelization on mixture-of-experts models with up to a trillion parameters, spread across 1,024 Trainium chips. That's not a laptop experiment — that's the kind of infrastructure normally reserved for well-funded labs at OpenAI or Google DeepMind, now available to a university team without them needing to buy a single GPU. Separately, University of Washington researchers are building an inference framework meant to squeeze more useful tokens out of every dollar spent on Trainium hardware, which, if it works, benefits everyone else building on the same chips.

Winners also get more than raw compute. Amazon throws in access to over 700 internal datasets, a dedicated research contact for consultation, and hands-on tutorials for actually wringing performance out of Trainium's architecture, which is notoriously different to program against compared to CUDA-based GPUs. That support layer matters because Trainium adoption has lagged Nvidia's ecosystem largely due to tooling friction, not raw chip capability.

What's notable is how global the recipient list is — Sydney, Imperial College London, National University of Singapore, Chinese University of Hong Kong, Dalhousie in Canada. Amazon isn't just subsidizing American computer science departments; it's quietly building a worldwide base of researchers fluent in Trainium, which happens to double as a long game against Nvidia's grip on AI research infrastructure.