이미지: 아마존 AGI
$110 Million Program to Support University AI Research
Amazon has unveiled the 34 recipients selected for the Fall 2025 cycle of its "Build on Trainium" program, a $110 million computing credit initiative designed to support AI research and university education using AWS Trainium infrastructure. This round's recipients are teams that submitted research proposals in five priority areas of Responsible AI: AI safety and alignment, multilingual language models, representation engineering, sustainability and small language models, and synthetic data generation.
Amazon said proposals were evaluated based on the quality of their scientific content and their potential impact on the research community and society.
Resources Available to Recipient Researchers
Recipient researchers can access more than 700 Amazon public datasets and utilize AWS AI/ML services and tools through AWS promotional credits. Each team is also assigned an Amazon research staff member for advisory support, and the program provides tutorials and hands-on sessions related to AWS Trainium.
Yida Wang, Senior Principal Applied Scientist at AWS AI, explained that the program gives researchers robust and scalable access to Amazon's proprietary AI chips. He noted that a UIUC research team is studying topology-aware parallelization strategies for mixture-of-experts models with up to 1 trillion parameters across as many as 1,024 Trainium chips, while a University of Washington team is developing an inference optimization framework to improve LLM inference performance on Trainium.
List of Research Projects Selected This Cycle
The 34 research teams selected this cycle belong to a range of universities including UIUC, UCLA, CMU, MIT, Johns Hopkins, UC San Diego, Northwestern, and Georgia Tech. Research topics span federated adversarial co-training for LLM security and robustness, hybrid public-private approaches to differentially private synthetic data generation, multimodal scam content detection and mitigation, algorithm-system co-design for sparse and quantized LLMs, multilingual speech-to-text large language models, safe pretraining for LLM agents, memorization-aware preference optimization for machine unlearning, multi-stage and multi-objective alignment for multilingual language models, synthetic data scaling for low-resource language model pretraining, certified robustness for quantization- and pruning-aware training of vision-language models, LLM hallucination detection and mitigation, and safety benchmarking of tool use by LLM agents.
The full list of recipients is available on Amazon Science's official website.
How to Participate and Program Constraints
Build on Trainium operates as a call-for-proposals program for university research teams. This round's recipients were selected through the Fall 2025 cycle's call for proposals. The source material does not disclose specific details about the schedule or application process for the next cycle. Program details should be checked separately on Amazon's official Research Awards page.
