Nvidia's most valuable CUDA moat was pried open by a single Anthropic PPT

📡 AI News 2026-11-16 4 min read

NVIDIA's most valuable moat, CUDA, was pried open by a single Anthropic PPT. A few days ago I came across a rather interesting slide. Anthropic, the company behind Claude, is sprinting toward an IPO, and this PPT page is the introductory material they're showing investors before going public—put plainly, it's their self-introduction to drum up investment. Just one page, but it packs a lot of information: their Claude Mytho

Nvidia's Most Valuable CUDA Moat Got Cracked by a Single Anthropic Slide

A few days ago I came across a rather interesting slide. Anthropic, the company behind Claude, is racing toward an IPO, and this slide is part of the pitch deck they're showing investors ahead of the listing—basically a self-introduction to raise money. Just one page, but it packs a lot: their Claude Mythos Preview model took roughly 3 days to port one of their large models' prompt-processing stage onto AMD's latest MI355X data center GPU, reaching a usable level of performance.

And according to Anthropic's own account, adapting this kind of workload to a new accelerator used to take 6 to 12 months.

3 days versus 6 to 12 months—two orders of magnitude apart. My first reaction was disbelief. I dug into the details, and found the story is more exaggerated than the headline—and also more restrained than the headline.

Where it's exaggerated. According to details 36Kr dug out of the roadshow materials, AMD provided Anthropic with an MI355X server. Engineers hooked the machine's control interface directly up to Claude, assigned the hardware-adaptation task, and went home. Nobody touched it all weekend. When they came back Monday and checked the monitoring, the performance curve on the new hardware was already climbing steadily under the AI's autonomous tuning. Kuai Technology's report added a hardcore number: in this adaptation, nearly 100% of the GPU kernel code was ported or written by Claude, with very little human intervention. GPU kernels are the layer of code that actually does the work—different vendors' architectures, programming environments, and optimization approaches are all different, and cross-platform porting has always been grueling work, typically taking an engineering team months to grind through.

Where it's restrained. The slide's footnote spells it out clearly: these 3 days covered prompt processing for a single model—that is, the Prefill stage of the inference pipeline—not the token-by-token Decode phase, and certainly not a full port on the training side. The "ran another model" in wccftech's headline is actually a loose paraphrase. So "the CUDA moat has collapsed"—anyone saying that now is jumping the gun.

What this slide pries at is precisely the root of that moat.

For the past decade, Nvidia's position has rested half on chip performance and half on the lock-in of the CUDA ecosystem. The world's existing code is all written around it, and switching to another vendor carries migration costs so high they're financially unjustifiable—so customers don't move. The core of that lock is really human labor: you need a batch of engineers who understand that vendor's hardware, spending massive person-months grinding away. And once AI can take on that layer of work itself, the root of the lock loosens. Anthropic's framing of GPUs in the roadshow materials as interchangeable, general-purpose compute is exactly this point: from now on, buying cards is just about compute and price—no more worrying about the engineering cost of switching platforms.

Some background is in order. In July this year, AMD and Anthropic announced a partnership, planning to deploy up to 2GW of AMD Instinct MI450-series compute, with the first 1GW set for the first half of 2027. So this material is both a technical showcase and a story told to investors: look, I can buy cards from whoever I want—the bargaining power is in my hands. Pre-IPO materials always carry a bit of marketing, so take the numbers with a grain of salt—but I believe the direction is real.

Zooming out one level, this should be a wake-up call for every compute buyer: the wall that is the software ecosystem is being dismantled brick by brick by AI. AMD's ROCm and the open-source Triton have been paving the way for years, solving "the code compiles"; what AI is now taking over is the last, most labor-intensive expert work—deep tuning. Once this capability becomes widespread, competition among GPU vendors will return to the hardware itself.

Here's my take: Nvidia's moat won't collapse anytime soon—CUDA's decade-plus of optimization accumulation isn't something 3 days can catch up to, and NVLink's advantage in multi-GPU interconnect for the Decode phase still stands. But this is the first time anyone has put up a single slide proving that the cost of switching platforms can be compressed from years to days. Once the anchor changes, the rest is just a matter of time.


Source notes

Suggested figures: figure-1 (3 days vs 6–12 months comparison chart) after the first paragraph; figure-2 (four-step weekend unattended workflow) after the "Where it's exaggerated" paragraph.

Related Articles
2026-08-14
OpenAI is still wrestling with valuation, Anthropic is already heading to the public market
2026-08-25
Today Changxin Technology went public, surging 471% at the open, directly becoming the new top stock on A-shares
2026-09-06
The Codex harness has long been open source; what OpenAI did this time is something else
2026-07-11
Anthropic quietly dropped Claude 4 Sonnet — and it's actually great
2026-11-01
Teknium of Nous Research: Hermes 4.3 Sees 28,972 Quant Downloads Against 2,007 for the BF16 Original
2026-07-28
JD.com Open-Sources Real-Time Video AI Model — Deployable on a Single RTX 4090

Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.

💬 Comments (0)

No comments yet. Be the first!

Login to comment