Claude’s 3-day AMD port signals CUDA lock-in collapse
The gist
Anthropic’s lightning-fast three-day port of Claude to AMD GPUs marks the most convincing collapse yet of Nvidia’s CUDA software lock-in.
What to know
- Anthropic moved Claude Mythos Preview onto AMD’s MI355 GPUs in just three days by keeping most model logic architecture-independent and only rewriting deep matrix-multiplication kernels.
- AI developers like Alibaba and leading labs are proving true hardware flexibility, with models like Qwen 3.8 27B running natively at speed on RISC-V and multi-accelerator setups thriving.
- The old norm of months-long, CUDA-centric onboarding for new chips has shrunk to days, signaling a dramatic erosion of Nvidia’s competitive software moat.
Portability Becomes Real
Anthropic’s rapid Claude port to AMD proves that unified model logic and open hardware specs have finally made true hardware agnosticism practical, not just aspirational.
Anthropic’s November 2026 Claude port to AMD GPUs landed as more than a one-off engineering feat because it showed that the software layer around a frontier model had become portable enough to move quickly onto a different accelerator stack. That speed was plausible because, as AMD said in the September 2026 Explainer, ROCm functions as “this unified tool chain and layer that allows for customers to program on our hardware,” and because the company publishes its ISA specifications so “anyone in the world can pick up the compilers, pick up the tools program to our GPUs.”
Anthropic described the same breakthrough from the model-builder side in its November 13 interview: the company now “unif[ies] the part that is common,” keeping batching, speculative decoding, and KV-cache logic “completely architecture independent,” while confining chip dependence to “the innermost part of the matrix multiplication.” That division of labor is what makes hardware agnosticism real rather than rhetorical: if the serving system and model logic stay stable while only the deepest kernels change, then moving Claude onto AMD hardware in days becomes evidence that the old software-hardware coupling is breaking down.
Custom Kernels Rewrite the Stack
As leading labs target multiple backends with custom kernels and orchestration layers, CUDA’s dominance is undermined by a new era of cross-platform AI deployment and interoperability.
What is changing is not just where models run, but how they are built to run: Sequoia Capital argued the old assumption that every AI company needed deep CUDA dependence is fading because “it's not like there's 10,000 model companies that are each… each need CUDA compatibility like it seems that kind of thesis is is changing… There's on the order of tens maybe model companies.” In that world, leading labs can justify writing custom kernels and targeting several backends at once, while common runtimes, harnesses, and orchestration layers absorb more of the portability work that once reinforced a single-vendor software moat.
That portability is now visible in the stack itself: TechSurge described NVL72 as being “fine-tuned” for specific Nvidia platforms even as Nvidia explores letting other GPUs plug in, framing a pathway toward heterogeneous deployment via interoperability rather than strict vendor lock-in, a sign that interoperability is becoming a product requirement rather than a concession. Outside the GPU world, Alibaba ran its open-source Qwen 3.8 27B model natively on a 64-core C950 RISC-V server at 30+ tokens per second with 1.9 seconds to first token, and because “C950 is fully compliant with the RVA23 profile… RVA23 mandates vector extensions (RVV 1.0), hypervisor support, and native handling of FP8 and BF16 data types,” developers get a concrete unified-stack target for cross-architecture deployment and generative AI workloads across heterogeneous systems; the same logic appears in inference orchestration, where “This uses llm-d, one of the largest ecosystem-based production inference kubernetes orchestrators,” and, during the recent Kimi K3.1 Pro launch, they had Day 0 interoperability, while “llm-d is the open-source K8s router that coordinates them, each host is 4 TPUv7 chips, and each result uses 24-32 chips.”
Switching Costs Collapse Overnight
Anthropic’s three-day Claude port shattered the months-long onboarding barrier, signaling an irreversible breakdown of Nvidia’s traditional CUDA lock-in.
The clearest sign that CUDA lock-in is weakening is not a theoretical promise but Anthropic’s November 2026 result: according to The Motley Fool, Claude Mythos Preview “took just three days to successfully port and optimize Claude to run on AMD's MI355 GPUs,” with Claude handling the porting and “writing 100% of the kernels.” That is a direct practical breach of Nvidia’s long-standing software advantage, which rested on the idea that CUDA code did not run on rival chips and was costly and difficult to adapt once a model stack had been built around it.
What makes the episode so consequential is the collapse in switching time. Anthropic said that “under normal circumstances, it previously would have taken a dozen engineers between six and 12 months to onboard the use of new AI accelerators,” while Trends reported that Anthropic told investors Claude Mythos Preview got one of its models running on AMD’s chips in about three days, contrasted with the same six-to-12-month benchmark. When the practical cost of moving off CUDA falls from months of specialist labor to days, the lock-in mechanism is plainly eroding fast.







