The story begins with AMD delivering Anthropic a rack outfitted with MI355X GPUs.
By convention, a new platform means months of manual software-stack adaptation. The Anthropic team braced for a long fight.
But they had Claude on hand. One engineer connected the machine and gave it a single instruction: “Go. Get this machine running.”
When the weekend ended, the engineer returned to a screen showing a performance curve that kept climbing. Claude had not just gotten the hardware running — it was tuning performance up, round after round, on its own. The human never wrote a line of code.
That is the story Anthropic co-founder and chief compute officer Tom Brown told on stage at AMD Advancing AI 2026. Lisa Su, sitting beside him, picked it up: when she heard only one Anthropic engineer was handling the MI355X, her first instinct was to send the AMD team to help. Her team came back reporting the offer was declined — everything was already handled.
The CUDA moat NVIDIA spent 20 years building was crossed by an AI model in a single weekend.
Not Luck — the Architecture of Chip Adaptation Changed
Claude’s feat was not Anthropic black magic. It followed from AMD rebuilding how chips get adopted. At the same event, AMD launched ROCm.AI — a GPU toolbox designed for AI agents. The main entry point, AMD Skills, packages verified ROCm knowledge: an agent equipped with it can install environments, deploy models, read logs, and diagnose failures on its own, calling real hardware through the ROCm CLI. The developer states the goal; the agent finds the road.
The detail that matters most: AMD changed how it writes the manual. Company VP of AI software and solutions Anush Elangovan disclosed on stage that every AMD GPU generation publishes its instruction set along with an AI-readable ISA. Once an agent reads the documentation, it can start tuning performance directly.
For the first time, chips have a non-human readership.
Automated tuning goes to Hyperloom: it spins up inference services, establishes a baseline, hunts CPU and GPU bottlenecks, tests configurations, generates custom kernels, and re-runs the whole thing. In the live demo it lifted MiniMax M3’s output speed by 38%. AMD has also run it across 14,000 models in one pass, banking the results as reusable experience for the next round. And AMD is training this capability with the frontier labs themselves — in Elangovan’s words, making frontier models “natively fluent in AMD programming.”
The Weekend Experiment Became a $5 Billion Order
The experiment connected immediately to real deployment. On July 22, AMD and Anthropic announced a strategic partnership: AMD will invest up to $5 billion in Anthropic and deploy up to 2GW of Instinct MI450-series GPUs to run Claude, with the first gigawatt targeted for the first half of 2027. Beyond the money, the two will use Claude directly to optimize AMD workloads and accelerate ROCm development, and AMD’s engineering and product teams will adopt Claude broadly.
The deal’s significance exceeds its size: Claude’s compute stack now spans four vendors — NVIDIA, Google TPU, AWS Trainium, and now AMD Instinct. Anthropic is no longer locked to any single silicon supplier.
How AI Itself Dismantled the CUDA Moat
CUDA’s moat, built since 2006, is an ecosystem coded line by line by countless engineers: compilers, math libraries, debuggers, profilers, documentation — everything, present and accounted for. Harder to copy is the accumulated feel of generations of engineers: how to tune operators, speed up communication, allocate memory, where distributed deployments break. That knowledge lives locked in a few thousand expert heads. A newcomer can build a chip with comparable silicon in quarters, but the software trail takes years to walk.
Agents plug exactly that gap. They read platform documentation, call profiling tools, find where the speed is stuck, then edit code, compile, and test. A failed approach gets swapped for the next; a better benchmark gets pursued further. Training a top GPU engineer takes years; spinning up another agent takes one more task. One engineer can loose a swarm of agents to debug errors and locate bottlenecks in parallel — and every configuration that works, every kernel written, every pitfall hit becomes the starting point for the next batch.
Compressing a years-long chase into a per-task cost — that is AI’s structural strike on traditional chip ecosystems.
The Developer Roster Just Added a New Class: Agents
The implications run far past one partnership. When you evaluate a chip, peak compute and memory bandwidth still matter — but “can AI read it, drive it, and tune its performance” now belongs on the same spec sheet.
For AMD, this is a textbook corner-exit overtake. CUDA’s moat was typed out by human engineers, and AI does not need to re-walk those 20 years. If agents can read the manual, tune directly, and produce reusable configurations, AMD’s deficit was never time — it was making its silicon legible to AI.
For the industry, the competition has gained a dimension: the ecosystem war shifts from “who has more developers” to “who lets AI develop better.” Chip vendors now court a second class of developer — the agent.
Chinese engineers are already at the front line of this AI-driven rebuild of the GPU software stack, and their accumulated low-level experience is being packaged into capabilities more agents can invoke. The new starting line in the chase after CUDA is here: hand the hardware interfaces and software tools to AI, and let the agents get to work.
For the first time, a 20-year ecosystem gap can be handed to a swarm of agents working around the clock to close it.
References:
-
Synced/36Kr: “CUDA’s 20-Year Moat Collapsed in a Weekend — Claude Ported AMD’s New GPU Alone”
-
AMD press release: “AMD and Anthropic Announce Strategic Partnership”
-
The Next Web coverage
-
Digital Applied analysis
