NVIDIA’s moat may not be as deep as everyone thought. In just 30 minutes, Claude Code took a real CUDA project and moved it to AMD ROCm. No human wrote a single line of code. The wall that kept AMD out of AI for years came down in half an hour.
Overnight, CUDA went from unbeatable to just another target.
A few days ago, a developer named johnnytshi posted something shocking on Reddit.
He used Claude Code for 30 minutes. He fed it a real CUDA project. It successfully ported the code to AMD ROCm.
During the process, not a single line was written by hand.
More importantly, it skipped the translation layer that has blocked most ports between the two ecosystems.
The key point is that the port had no traditional middle layer. No Hipify tool. No manual rewrite. Just a CLI command.
AMD engineer Anush E. was amazed. He said the future of GPU programming has been rewritten by AI.
Once the news spread, the tech world went silent for a moment. Then everyone started saying the same thing. NVIDIA’s CUDA moat might not hold anymore.
How did this happen?
ai erotic smut
Claude Code is not just a chatbot. It is a system that can read, write, and port code on its own.
In this case, the key was not just translating keywords. It was understanding the bottom-level logic of both platforms.
According to johnnytshi, during the port, even the data layout and memory access patterns were handled by AI. The core logic of the kernel files matched perfectly.
best free ai porn
What is even more shocking is that johnnytshi completed the full CUDA to ROCm port in just 30 minutes. There was no manual translation step at all.
The tool that used to take weeks or months, Hipify, was bypassed. The complex feedback loop was gone. A single CLI command did the job.
NVIDIA’s CUDA moat suddenly looked like a small fence.
After all, NVIDIA has spent years building CUDA into the industry standard. AI frameworks, deep learning tools, and scientific computing all depend on it.
AMD’s ROCm has always been strong, but it suffered from a weak software ecosystem and the high cost of porting code.
Now, Claude has removed that cost in 30 minutes. The future where CUDA code runs on AMD GPUs is no longer a dream.
The real details
On GitHub, johnnytshi shared the full log. He said:
“This enables ROCm support for AMD GPUs, specifically RDNA 3.5, and brings AMD gaming laptop APUs into the AI training and inference world.”
GitHub: https://github.com/LeelaChessZero/lc0/pull/2375
He also said:
“Tested on Strix Halo – Radeon 8060S, gfx1151.”
Of course, there are still limits.
For simple to medium complexity kernels, Claude Code performed very well. The key was that it wrote the core logic and handled memory layout and hardware optimization.
But there is one area where Claude Code is still weak. For kernels that require deep hardware-level tuning, specific memory patterns, and complex optimizations, AI cannot fully replace human experts yet.
Still, the signal this event sends is strong enough.
In the past, projects like ZLUDA tried to run CUDA on AMD through internal efforts. They aimed to break the CUDA lock-in.
But these efforts relied on manual translation layers. The level of automation was limited.
Claude Code changed the game. It bypassed these layers entirely. It used a “reasoning plus tool” approach to flatten the ecosystem gap.
As AMD engineer Anush E. said, the future of GPU programming has been rewritten by AI.
By the way, Claude Code has already eaten the coding world. They call it “Claude-Pilled.”
Earlier, CEO Dario Amodei said at a conference that engineers who do not adapt will have no place. In 6 to 12 months, AI will replace many of them.
Inside Anthropic, engineers are already fully AI-coded. It is all Claude.
Not 90 percent. 100 percent.
In a Wired interview, Claude Code lead Boris Cherny admitted that 100 percent of his code is written by AI.
Even Anthropic engineers did not expect that a business project could be taken over by AI so quickly.
Boris Cherny recalled that when they first released Claude Code, they were not sure if it would work. “It seemed too crazy.”
Cherny watched the numbers change in real time.
At launch, only 5 percent of code was written by Claude Code.
By May, with Opus 4 and Sonnet 4, the share reached 30 percent.
Now, with Opus 4.5, over 100 percent of code is generated by Claude Code.
Inside Anthropic, full AI coding is the new normal.
Over 100 percent of employees use Claude Code. And the Claude Code team itself writes 95 percent of its own code with AI.
Stanford AI lab enters the second era
We have to say that the speed of AI coding evolution is stunning.
From 2021 to 2024, advanced developers used auto-complete and code suggestions. Junior developers wrote boilerplate code.
In 2025, tools like Cursor and Windsurf changed the game. Agentic coding products turned the industry upside down.
Now, only the framework and architecture need human input. The rest is fully generated by AI.
Claude Code was not the first, but it is the fastest.
Boris Cherny admitted that the current version is just the beginning. In the next cycle, Anthropic has one goal: to build the next generation of AI coding products for AI itself, not for humans.
One footnote: Anthropic’s flagship Claude Opus 4.5 release was AI’s “D-Day” moment.
Stanford AI professor and Workera CEO Kian Katanforoosh moved his entire company to Claude Code.
He admitted that for senior engineers, Claude Code, Cursor, and Windsurf are all useful.
But Katanforoosh sighed and said: “The only model I see that makes a real leap is Claude Opus 4.5.”
“It feels like the model is not just writing code. It is designing a new way to think about software.”
Even Microsoft is building its own version of Claude Code.
A 10 billion dollar coding business
Claude Code’s success has made Anthropic’s revenue explode.
Last year, the AI coding market was turned upside down. In November, Anthropic launched Claude Code for developers. In just one week, annual recurring revenue crossed 1 billion dollars.
By 2025, ARR had grown to over 1 billion dollars.
At the time, the product made up about 90 million dollars of Anthropic’s ARR, roughly 12 percent. For a company that provides AI systems to enterprises, this is a massive win.
Now, Anthropic’s AI coding products are moving faster. The halo effect of Claude Opus 4.5 has made the numbers go even higher.
Cursor also hit 1 billion dollars ARR in November. OpenAI, xAI, and others are fighting for market share. The model war has become a product war.
But Anthropic is not stopping.
Just a few days ago, they released Cowork, an AI office worker that can read files, write code, fix bugs, and handle tasks without human input.
What is being replaced is not just coding
When talking about Cowork, Cherny revealed that he has already let go.
He said that on a recent project, a Cowork engineer did not know who to contact. The AI automatically sent a message on Slack and found the right person.
Cherny watched the whole thing happen while drinking coffee. He thought: “I do not need to do these boring things anymore.”
These tasks used to be done by mid-level engineers who were about to be laid off. Cherny gave a new definition:
“The industry has always changed. First we wrote machine code, then assembly, then C, then Java, then Python.”
“Now, on top of all that, AI is just another layer.”
Now, Cherny wakes up every morning and checks 3 to 4 projects on his phone. His company runs on autopilot.
At any time, he can see ten to twenty projects running at once.
Cherny concluded that AI coding will speed up software creation by ten times. “The bottleneck is no longer writing code. It is deciding what to build.”
In other words, Anthropic’s core job is to make Claude easier to use.
