Amazon Web Services has released TorchNeuron, a tool that lets AI startups move their applications onto Trainium, Amazon’s AI chip, in a direct challenge to the software advantage that has helped Nvidia become the default hardware for AI.
TorchNeuron works with PyTorch, the free, open source toolkit used to build, train and run AI models, and Amazon says it allows PyTorch apps to run on Trainium by changing a single line of configuration, with no code rewrite. Amazon estimates that about 85% of AI apps are written in PyTorch, where Nvidia’s CUDA software has long been deeply embedded. While AMD and Apple chips already offer similar porting options, AWS argues that Trainium also gives startups access to its entire cloud, positioning it as a stronger choice than Nvidia-aligned neoclouds such as CoreWeave, Nebius and Firmus.
Colin Brace, VP of Technology Software at AWS Annapurna Labs, said startups building on AWS have the infrastructure to move from prototype to production already in place. He added that Trainium’s value comes from its price-performance and from handling both training and inference on one platform, citing AI voice provider Deepgram, which he said now processes 70% more audio per dollar after moving to Trainium.
Amazon is also contributing code upstream to PyTorch to remove ties to specific chips, aiming to make the framework more hardware-agnostic for all developers. Brace said the bigger story is removing the lock-in that had denied developers real choice, a shift that could have significant consequences for Nvidia if successful.
Nvidia, meanwhile, has released new open-weight models built for always-on agents and edge computing. The Nemotron 3.5 Super VL is a 120-billion-parameter model for enterprise agents that Nvidia says matches the larger Nemotron 3 Ultra in instruction-following and tool use. The company claims Super VL completes tasks faster than competing models from Anthropic and OpenAI when post-trained on agentic tasks such as document intelligence, financial accounting and medical insurance claims, and says it is used by companies including Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce.
Kari Briski, Nvidia’s vice president of Open Models and Software, said organizations running agents on Nvidia infrastructure now have an open, optimized model for every tier, and that unlike closed frontier models, Super VL can be post-trained on a company’s own data and deployed on its own infrastructure. The move comes as fast-growing AI labs increasingly turn to alternatives such as Trainium, Cerebras and their own custom chips.
Nvidia also introduced two lightweight Bolt models designed to run locally on AI PCs powered by its RTX Spark chip and on DGX Spark workstations, which partners including Microsoft and Dell are now bringing to market. Briski said users can delegate everyday tasks without sending workloads to the cloud, addressing privacy concerns around consumer AI assistants. Nemotron 3.5 Bolt 74B-A9B targets local agentic tasks such as multistep reasoning, tool calling and web browsing, while the smaller Bolt E4B is aimed at conversational agents.
Super VL and Bolt 74B-A9B are available now, with Bolt E4B due later this month.
