In just one weekend, Claude managed to run its cutting-edge model on a brand new AMD MI355X rack, without a single line of code modified by a human engineer.
It seems the CUDA moat that NVIDIA (NVDA) spent 20 years building has been crossed. The story begins with AMD sending a rack equipped with the MI355X to Anthropic. The team braced for a difficult integration, expecting to adapt to a new platform and software stack.
However, Anthropic had a secret weapon: Claude. The team asked one engineer to connect Claude to the machine and simply instructed it to "get this machine running."
When they returned to their desks after the weekend, a continuously rising performance curve was already on the screen. Not only was it running, but performance was steadily improving with each iteration. Anthropic co-founder and Chief Computing Officer Tom Brown shared this story, and AMD CEO Lisa Su then revealed that when she heard only one Anthropic engineer was handling the MI355X, her first instinct was to send an AMD team to assist. The team returned and told her the help wasn't needed—everything was already done.
This weekend experiment quickly led to a real deployment plan. Anthropic recently announced it will deploy up to 2GW of Instinct GPUs in the AMD Helios system, with the first 1GW scheduled to launch in the first half of 2027. Furthermore, both companies will directly use Claude to optimize AMD workloads and accelerate ROCm software development.
A New Reader for Chip Manuals
Claude's ability to achieve a "dimensionality reduction" attack on the CUDA ecosystem stems from AMD providing an AI-driven toolkit for GPU tuning. The ROCm.AI platform, unveiled at AMD's Advancing AI 2026 event, is a comprehensive GPU toolbox designed for Claude, Codex, and Cursor.
The entry point is called AMD Skills, which contains a repository of verified ROCm knowledge. When an AI Agent receives this knowledge, it can independently set up environments, deploy models, read logs, and diagnose faults, all while using the ROCm CLI to interact with real hardware. Developers just state their goal, and the Agent finds its own path.
Importantly, AMD has even changed how it writes chip manuals. AMD's Corporate Vice President of AI Software and Solutions, Anush Elangovan, revealed that each generation of AMD GPUs will publish an instruction set and provide an AI-readable ISA. After the Agent understands the manual, it can directly start tuning performance. AMD has delegated this task to its Hyperloom system, which launches an inference service, runs a baseline, identifies CPU and GPU bottlenecks, tests different configurations, generates custom kernels, and then re-runs the optimized solution. In a live demo, Hyperloom increased the output speed of the MiniMax M3 model by 38%. AMD also ran 14,000 models through it in one go, consolidating the test results into reusable experience for future tasks.
AMD is also collaborating with frontier model companies to train this capability. Elangovan's point was that they are teaching frontier models to "speak AMD programming natively." Going forward, when evaluating a chip, peak compute and memory bandwidth will be important, but whether an AI can read, utilize, and tune its performance will be a key metric on the spec sheet. The new developer chip companies must now win over is the Agent.
A Dimensionality Reduction Strike from ASI
CUDA's journey from 2006 to today has been built by countless engineers writing code line by line. It has a complete ecosystem: compilers, math libraries, debuggers, performance analysis tools, and development documentation. What is even harder to replicate is the accumulated "know-how" of generations of engineers. This knowledge—how to tune operators, speed up communications, allocate memory, and identify common issues in distributed deployments—is locked in the minds of a few experts. Any newcomer, even with a chip of similar performance, must walk this long software road from the beginning. Chip fabrication can progress quarterly, but ecosystem accumulation can only be measured in years.
This is precisely the gap that Agents are filling. An Agent can read platform documentation, invoke performance analysis tools, find bottlenecks, then modify code, compile, and test. If one approach fails, it tries another; if a benchmark improves, it continues in that direction. While training a top-tier human GPU engineer takes years, launching a new Agent only requires starting another task. A single engineer can deploy a fleet of Agents to debug errors and locate performance bottlenecks in parallel. The configurations, kernels, and lessons learned from one successful run can immediately become the starting point for the next batch of Agents.
Compressing a multi-year catch-up process into a task-based calculation is where AI's greatest value lies. This is the dimensionality reduction strike against the CUDA ecosystem by ASI-level (Agentic Systems and Intelligence) capabilities. Chinese engineers are now at the forefront of this AI-driven transformation of the GPU software stack, and the foundational experience they are accumulating is being organized into capabilities that more Agents can call upon.
A new starting point for catching up with CUDA has appeared: hand over the hardware interface and software tools to AI, and let the Agents get to work. For the first time, a 20-year ecosystem gap can be closed by a fleet of Agents working tirelessly day and night.
Comments