Build on Trainium call for proposals: Fall 2026

Build on Trainium: Agentic, recursive improvement of kernels and frameworks is a Fall 2026 call from the Amazon Research Awards. Amazon Trainium is an AI chip for training and inference of machine learning models. The page says researchers can use the Neuron software development kit and the Neuron Kernel Interface, and that Amazon has a Trainium research cluster of up to 40,000 Trainium chips on Amazon EC2 Trn2 instances.

The call asks for work that turns open source foundation and code models into specialised optimisation agents. It names supervised fine tuning and reinforcement learning, including long horizon training where an agent generates, compiles, runs, profiles, and repairs a kernel before a reward arrives. It also asks for reward designs that fit the task, and for learning that improves the agent's tools and knowledge, not only its weights. The page encourages teams to release models and training data openly.

Applicants are encouraged to request AWS Promotional Credits in one of two ranges: up to $50,000, or up to $250,000 and beyond. Awards also include AWS Trainium training resources, tutorials, and hands on sessions with Amazon scientists and engineers. Awards are one time unrestricted gifts. The budget should list expected costs in USD and should not include administrative overhead. The awards panel sets the final amount. Use of the credits follows the AWS Promotional Credit terms. Decision letters go out in February 2027.

The principal investigator must be a full time faculty member at an accredited academic institution, or a permanent researcher at a non governmental organisation with recognised legal status equivalent to 501(c)(3) status in the United States. They must be at or above the age of majority. Each principal investigator may submit only one proposal to the Amazon Research Awards in this call period. Amazon employees cannot receive an award. The rules also bar recent Amazon staff, their immediate family or household, paid government employees other than at an accredited academic institution, and people under US export controls or sanctions. The program is void in Cuba, Iran, Syria, North Korea, and the Crimea, Luhansk and Donetsk regions of Ukraine.

The submission period is 1 October to 4 November 2026 at 11:59 p.m. Pacific Time. Proposals should follow the proposal template and are encouraged to stay within 4 pages, not counting appendices. The call asks principal investigators to say whether they plan open source contributions of data or code. Submit through the call page.