Anthropic told IPO investors that its Claude Mythos Preview model completed the low-level software work needed to run one of its models on AMD MI355X GPUs at useful performance in about three days, with basically no human intervention, versus six to 12 months and dozens of engineers previously. Mythos ported or wrote almost 100% of the kernels, a result that directly challenges Nvidia's CUDA moat.
CUDA
Nvidia's parallel computing platform and programming model, the key software barrier tying AI models to its GPUs.
Curated Intelligence
Nvidia's board announced that Jensen Huang has resigned as CEO 33 years after co-founding the company, opening a succession process. The piece traces his three reinventions, CUDA as a general-purpose computing platform, the pivot to deep learning after ImageNet, and the Mellanox acquisition that moved Nvidia from GPU boards to integrated AI datacenter racks, and warns that visionary-led firms often coast on momentum after their founder departs. The board must decide whether the next CEO is hired to execute today's business, reinvent it, or both.
Anthropic co-founder Tom Brown says the company has escaped single-vendor dependence on Nvidia via a multi-chip strategy, training and serving Claude across all four major chip families as one interchangeable compute pool. Its frontier model Mythos reportedly wrote or ported nearly 100 percent of the kernels for an AMD GPU migration itself in just three days, versus the usual six to twelve months. The claim directly challenges Nvidia's CUDA moat, and Anthropic plans to deploy up to two gigawatts of AMD Instinct MI455X GPUs.
A developer has open-sourced samir-vllm, a teaching project that builds a small LLM inference engine from scratch in C++ and CUDA, covering everything from loading SafeTensor weights to executing each transformer stage on GPU. It ships with 85 pages of handwritten notes and over six hours of video, explicitly written without AI assistance, making it a solid structured resource for engineers who want to understand inference internals.
The US DOJ is scrutinizing Nvidia's $12.93 billion acquisition of Hugging Face, fearing control over a key AI distribution hub could harm competition. Regulators are also investigating Nvidia's sales practices, cloud investments, and ecosystem lock-in risks associated with its CUDA software.
Amazon has released TorchNeuron, an open-source tool that enables developers to run PyTorch code natively on its custom Trainium chips with only a configuration change and no code rewriting. Designed to challenge NVIDIA's CUDA ecosystem dominance, the tool aims to reduce migration costs and enhance hardware agnosticity, debuting at the North American PyTorch Conference.
AWS released TorchNeuron, a new tool enabling PyTorch-based AI apps to run on Amazon's Trainium chips with just one configuration change and no code rewrite. This move directly targets Nvidia's software moat (CUDA) by decoupling PyTorch from Nvidia hardware, offering developers genuine choice and challenging Nvidia's dominance in AI infrastructure.
Lenovo has opened blind pre-orders for the YOGA Pro 15 RTX Spark laptop, featuring the first-in-class NVIDIA RTX Spark N1X superchip. The device delivers up to 1 Petaflop FP4 local AI performance via integrated Blackwell GPU and Grace CPU cores with native CUDA support, aiming to shift personal productivity toward autonomous local agents.
rGPU is an open-source tool that offloads computation to remote NVIDIA GPUs while keeping applications local. It offers two integration paths: a native PyTorch device backend supporting `rgpu`, and a CUDA shim intercepting low-level APIs for existing binaries. This solution addresses remote training and inference needs in scenarios with limited local compute or resource isolation requirements.
HKUST researchers introduced AccelEval at NeurIPS 2026, the first benchmark evaluating LLMs' ability to convert correct CPU code into efficient GPU implementations. Covering 6 domains, 42 tasks, and 3 input scales, it integrates 43 CUDA optimization strategies to analyze acceleration principles. The study aims to verify if AI can achieve significant end-to-end performance gains while maintaining correctness.