Anthropic's new software bridge challenges Nvidia's hardware dominance
Anthropic successfully ported its Claude model to AMD GPUs in just three days, potentially breaking Nvidia's long-standing software moat.
The short version
- Anthropic ported Claude to AMD MI355 GPUs in only three days using its Claude Mythos Preview model.
- The process previously would have required a dozen engineers working for six to 12 months.
- Anthropic's goal is a 'fungible compute pool' where models can run on Nvidia, AMD, Amazon, or Google chips.
- AMD's Helios system may offer a significant price advantage over Nvidia's Vera Rubin platform.
Why it matters
This development could lower the barrier for companies to switch away from Nvidia hardware. It allows AI developers to optimize for the best price-to-performance ratio rather than being locked into a single vendor's software ecosystem.
Breaking the CUDA moat
For years, Nvidia has maintained a dominant position in the AI chip market by creating a software moat called CUDA. This platform made it easy for developers to program GPUs for complex tasks, moving beyond just rendering graphics for video games.
Because early AI research was conducted on CUDA, an entire generation of developers became accustomed to Nvidia's specific software ecosystem. This created a situation where most original AI code was optimized specifically for Nvidia's hardware, making it difficult for competitors to gain a foothold.
However, Anthropic recently revealed that it has built a bridge across this moat. The AI lab successfully ported its Claude model to run on AMD's MI355 GPUs in a timeframe that significantly undercuts the industry standard for hardware onboarding.
How the porting works
The technical achievement involves the Claude Mythos Preview model, which handled the porting and optimization process. Specifically, the model was able to write or port 100% of the kernels required for the operation.
A kernel is a specific piece of AI code that provides instructions to a GPU on how to execute a mathematical operation. By automating this process, Anthropic bypassed the manual labor typically required to make a model compatible with different hardware architectures.
Anthropic stated that under normal circumstances, a team of twelve engineers would have needed between six and 12 months to complete this task. By finishing the work in three days, the company demonstrated a massive leap in software flexibility.
Creating a fungible compute pool
Anthropic's primary objective with this technology is to create what it calls a 'fungible compute pool.' This means the company wants to be able to prepare any AI chip to run its models in just a few days, regardless of who manufactured the hardware.
This strategy allows Anthropic to purchase chips based on the best price-to-performance ratio available at any given time. They can utilize Nvidia GPUs, AMD chips, or custom silicon like Amazon's Trainium or Alphabet's Tensor Processing Units depending on which is best for training or inference.
By removing the friction of hardware switching, Anthropic levels the playing field for the entire industry. If AI models can run on alternative chips without high costs or complexity, the reliance on a single vendor's ecosystem begins to weaken.
Impact on AMD and Nvidia
This shift is particularly beneficial for AMD, which has been seeking to capture a larger share of the AI market. AMD's Helios system is positioned as a more affordable alternative, selling for roughly $20 billion per gigawatt compared to Nvidia's Vera Rubin platform, which can cost around $40 billion per gigawatt.
Even if the Helios system does not match the performance of Nvidia's platform, the price-to-performance calculation could win over many customers. This development comes at a time when Nvidia is facing potential reputational challenges and a shrinking software advantage.
If Nvidia loses market share because its CUDA moat is shrinking, it may be forced to lower its prices. Such a move would likely impact Nvidia's sales, gross margins, and overall profitability as the market becomes more competitive.
The road ahead for hardware
While this news is a significant blow to Nvidia's exclusivity, the company still holds several advantages. It remains the premier AI infrastructure solution and possesses the best overall ecosystem for many developers currently in the field.
Additionally, Nvidia has secured a large amount of supply-constrained high-bandwidth memory, which is a critical component for all high-performance AI chips. This physical supply constraint remains a hurdle for competitors regardless of how easy the software becomes to port.
The next phase will involve watching how quickly other AI labs adopt Anthropic's approach to multi-vendor hardware. If the trend continues, the AI industry may move toward a more diverse hardware landscape where software compatibility is no longer a barrier to entry.
What to watch
- Whether other AI labs will adopt Anthropic's method to use multiple chip types.
- How Nvidia responds to the shrinking of its CUDA software moat.
- The actual market share shift for AMD as more companies seek cheaper compute.
Try it
Check the price-to-performance specs of AMD's Helios system versus Nvidia's Vera Rubin platform to see the current cost gap.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Yahoo Finance · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
