Anthropic says artificial intelligence is beginning to weaken one of Nvidia’s biggest competitive advantages: its CUDA software ecosystem. The AI company claims that its frontier AI model, Mythos, helped port Claude to AMD graphics processing units (GPUs) in just three days, a process that traditionally required dozens of engineers working for six to 12 months. The development could make it easier for AI companies to switch between chip manufacturers and reduce their dependence on Nvidia’s hardware ecosystem.
The claim was highlighted in an Anthropic investor presentation discussed on November 11, 2026. According to Anthropic co-founder and Chief Compute Officer Tom Brown, the company used Mythos to write or port almost 100% of the required computing kernels for the AMD implementation, with virtually no human intervention. The achievement comes as Anthropic pursues a multi-chip strategy involving Nvidia GPUs, AMD accelerators, Google TPUs and Amazon’s Trainium processors. However, the reported three-day result has not been independently verified and applies to a specific model-porting task rather than a complete replacement of Nvidia’s infrastructure.
Anthropic Uses Its Own AI Model to Port Claude to AMD Chips
Anthropic’s experiment demonstrates how AI models can help solve one of the most difficult problems in moving software between different hardware platforms.
The company used Mythos, its frontier AI model, to adapt Claude’s inference software for AMD GPUs.
Inference is the process through which an AI model generates responses after receiving a prompt. Running an advanced AI model efficiently requires specialised software that manages memory, computing operations and communication between processors.
Different chip manufacturers use different architectures and software environments. As a result, software optimised for Nvidia GPUs cannot necessarily run efficiently on AMD hardware without modifications.
Anthropic says Mythos handled almost all the work required to port or write the computing kernels for the AMD implementation.
Tom Brown explained that the company previously needed dozens of engineers and several months to bring a new chip platform into operation. With Mythos, the company reported getting Claude running on AMD GPUs in approximately three days.
The result suggests that AI models could increasingly assist with the engineering work required to support different hardware architectures.
Why Nvidia’s CUDA Software Ecosystem Matters
Nvidia’s dominance in AI computing is not based entirely on the performance of its graphics processors.
One of its strongest competitive advantages is CUDA, a parallel computing platform that allows developers to write software capable of using Nvidia GPUs.
CUDA provides programming tools, libraries and optimisation capabilities that have become deeply integrated into the AI development ecosystem.
Many AI frameworks, research projects and enterprise applications have been designed around Nvidia’s software infrastructure.
This creates a significant switching cost for customers.
Even when another manufacturer offers competitive hardware, companies may need to rewrite software, optimise computing kernels, test performance and resolve compatibility problems before moving workloads.
For large AI companies operating thousands of GPUs, the engineering effort can become substantial.
This software advantage has helped Nvidia maintain a strong position in the AI accelerator market, even as competitors introduce alternative processors.
Anthropic’s experiment suggests that AI-assisted software development could reduce some of these switching costs by automating the work needed to adapt applications to different chips.
AMD Could Benefit From Faster AI-Powered Chip Porting
AMD is one of Nvidia’s most important competitors in the AI accelerator market.
Its Instinct GPU family is designed for large-scale AI training and inference, while its ROCm software platform provides the tools needed to run workloads on AMD hardware.
However, Nvidia’s established software ecosystem has historically made it difficult for AMD to attract customers whose applications were already optimised for CUDA.
Anthropic’s reported experiment could help address this problem.
If AI models can automatically adapt computing kernels and improve compatibility between different hardware architectures, AMD may become easier for companies to adopt.
The potential benefits include:
- Lower migration costs: Companies could spend less time adapting software to AMD GPUs.
- Faster deployment: AI-assisted porting could reduce the time needed to bring new hardware into production.
- Greater hardware flexibility: Businesses could distribute workloads across different chip manufacturers.
- More pricing competition: Chip providers may face greater pressure to compete on performance and cost.
- Reduced vendor dependence: AI companies could avoid relying too heavily on one hardware supplier.
However, faster software porting does not automatically guarantee equivalent performance, reliability or operational efficiency across different GPU platforms.
Anthropic Is Building a Multi-Chip Computing Strategy
Anthropic is increasingly treating computing capacity from different manufacturers as part of a shared infrastructure pool.
The company says it now trains and runs Claude across four major chip families:
| Chip manufacturer | Computing platform |
|---|---|
| Nvidia | GPUs |
| AMD | Instinct GPUs |
| Tensor Processing Units (TPUs) | |
| Amazon | Trainium AI chips |
The strategy allows Anthropic to select computing resources according to availability, performance and cost.
Different hardware platforms may be better suited to different workloads. Some processors may offer advantages for training, while others could be attractive for inference.
By reducing dependence on a single supplier, Anthropic wants greater flexibility when securing the enormous computing capacity required to develop and operate its AI models.
The company has described this approach as an unconventional decision that required substantial investment in software engineering.
It believes that developing the ability to operate across multiple hardware platforms can improve the amount of computing performance it obtains from every dollar spent.
Anthropic and AMD Already Have a Major Chip Agreement
The reported software breakthrough comes alongside a significant commercial relationship between the two companies.
In July 2026, AMD announced a strategic investment of up to $5 billion in Anthropic.
The companies also agreed to work towards deploying up to two gigawatts of AMD Instinct MI455X GPU capacity for Anthropic.
The first phase, involving approximately one gigawatt of computing capacity, is expected to begin in the first half of 2027.
The agreement gives Anthropic access to additional computing resources as demand for Claude increases.
For AMD, the partnership provides an opportunity to demonstrate that its processors can support workloads operated by one of the world’s leading AI developers.
The relationship is also strategically important because Anthropic can help identify software improvements and performance requirements that could strengthen AMD’s AI computing platform.
The commercial agreement and the Mythos experiment therefore reinforce each other: AMD provides hardware capacity, while Anthropic’s AI models may help reduce the engineering work required to use that hardware efficiently.
AI Could Become a Tool for Improving AI Hardware Software
The significance of Anthropic’s experiment extends beyond switching between GPU manufacturers.
AI models are increasingly being used to write code, identify performance bottlenecks, optimise algorithms and test software.
In chip development, these capabilities could help engineers produce specialised computing kernels that take advantage of different processor architectures.
A computing kernel is a small program that performs a specific operation, such as matrix multiplication or data processing. AI workloads depend on thousands of such operations working together efficiently.
Traditionally, engineers have spent considerable time optimising these kernels for individual hardware platforms.
AI coding systems could automate parts of this process by generating code, testing different implementations and refining performance.
This creates a potential feedback loop: AI systems help improve the software infrastructure used to run AI models, which could make future AI development and deployment more efficient.
However, the quality of generated code still needs to be tested. Performance optimisation must also account for memory management, numerical accuracy, multi-GPU communication and reliability under production workloads.
The three-day demonstration is an important reported result, but it does not establish that all these challenges can be solved automatically.
Nvidia’s CUDA Advantage Is Not Disappearing Overnight
Despite the development, Nvidia retains several important advantages.
CUDA has a mature ecosystem supported by years of engineering investment, extensive documentation, optimised libraries and a large developer community.
The platform is also integrated into many established AI frameworks and enterprise computing environments.
Even if AI reduces the time needed to port software, companies must still evaluate hardware availability, performance, reliability, networking, memory capacity and total operating costs.
AMD also faces its own software-development challenges. Semiconductor industry research published by SemiAnalysis in July 2026 highlighted improvements in AMD’s ROCm ecosystem while identifying continuing gaps in automated testing, infrastructure stability and software compatibility.
This demonstrates that reducing the cost of writing software is only one part of competing with Nvidia.
Nvidia’s advantage also includes its hardware roadmap, networking technologies, developer tools and relationships with major cloud providers.
Consequently, Anthropic’s experiment should be viewed as evidence that one important barrier may be weakening, not proof that Nvidia has lost its competitive position.
Important Questions About the Three-Day Claim
The reported result has attracted attention because it represents a dramatic reduction in the time needed to adapt AI software to alternative hardware.
However, several qualifications matter.
First, the three-day figure comes from Anthropic’s own investor presentation and has not been independently verified.
Second, the reported task involved getting Claude’s prompt-processing workload running on AMD MI355X GPUs. It should not be interpreted as proof that Anthropic migrated its entire training and inference infrastructure in three days.
Third, successful software porting does not necessarily mean that performance, reliability and operating costs immediately match those achieved on Nvidia hardware.
The claim is therefore best understood as a demonstration of AI-assisted hardware adaptation rather than a complete solution to the industry’s multi-chip engineering challenges
The Bigger Picture
Anthropic’s experiment highlights a potentially important change in the economics of AI computing. Nvidia’s CUDA ecosystem has historically created high switching costs because developers must invest substantial engineering resources to optimise software for alternative hardware. If AI models can automate much of this work, chip manufacturers may find it easier to compete for customers that previously preferred to remain within Nvidia’s ecosystem.
The development could also accelerate the shift towards a more diversified AI infrastructure market. Companies may increasingly select processors based on availability, performance and cost rather than relying on a single manufacturer. Nevertheless, Nvidia’s software ecosystem remains a significant competitive advantage, and AMD must continue improving its hardware, software tools and production reliability to convert easier migration into sustained market share gains.
Looking Ahead
Anthropic’s multi-chip strategy is likely to become increasingly important as the company expands its computing infrastructure and prepares to deploy additional AMD accelerators. The first phase of its planned AMD capacity expansion is expected in the first half of 2027. Further progress will depend on whether the company can use AI-assisted engineering to improve performance and reliability across different processor architectures, not simply get software running on new hardware.
For Nvidia, the challenge is that AI itself could help competitors reduce the software barriers that have protected its market position. CUDA is unlikely to lose its importance immediately, but the cost of moving workloads between chip manufacturers could decline as AI coding systems become more capable. If Anthropic’s results can be reproduced at a larger scale, the competitive advantage in AI computing may increasingly depend on hardware performance, software quality, supply and total cost rather than ecosystem lock-in alone.
SEO Metadata
SEO Title: Anthropic Says AI Could Erode Nvidia’s CUDA Lead as AMD Chip Switch Takes 3 Days
Meta Description: Anthropic says its Mythos AI model helped port Claude to AMD GPUs in three days instead of six to 12 months, potentially weakening Nvidia’s CUDA advantage.
URL Slug: anthropic-ai-erodes-nvidia-cuda-lead-amd-chip-switch
Focus Keyphrase: Anthropic Nvidia CUDA lead
Secondary Keywords: Anthropic AMD GPUs, Mythos AI model, Nvidia CUDA ecosystem, AMD Instinct GPUs, AI chip competition, Claude AMD migration, AI software porting, ROCm, multi-chip AI strategy
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
