Anthropic’s AI Chip Breakthrough Challenges Nvidia’s Software Advantage

Foto de Milad Fakurian na Unsplash

Anthropic’s AI Chip Breakthrough Challenges Nvidia’s Software Advantage

Anthropic has demonstrated a potentially significant advance in artificial intelligence infrastructure, using its Claude Mythos Preview model to accelerate the adaptation of AI software to processors manufactured by Advanced Micro Devices (AMD).

The development could reduce one of the principal technical barriers that has helped Nvidia maintain its dominant position in the AI accelerator market: the time, cost and engineering expertise required to move sophisticated AI workloads between different chip architectures.

According to information presented by Anthropic during its November investor roadshow, Claude Mythos completed a specific software adaptation task for AMD GPUs in approximately three days. The company estimated that comparable work using traditional engineering methods had previously taken six to twelve months.

The demonstration raises an important question for the semiconductor industry: Could artificial intelligence itself make the hardware used to develop and operate AI systems easier to replace?

How Anthropic Reduced Months of Engineering Work to Three Days

Anthropic’s co-founder and chief compute officer, Tom Brown, described how the company used Claude Mythos to prepare an AI model to operate on AMD’s Instinct MI355X accelerators.

According to the company’s presentation, Mythos wrote or adapted nearly all the specialized software kernels required for the demonstrated task, with minimal human intervention.

Kernels are low-level software routines that allow AI models to perform calculations efficiently on particular processors. Because chip architectures differ, these routines often require extensive engineering work when developers move applications between hardware suppliers.

The demonstration suggests that advanced AI coding systems could automate a substantial portion of this process.

However, the reported three-day result covered the initial processing of user inputs, known as prefill, for one Anthropic model on AMD hardware at a useful level of performance.

Why Nvidia’s CUDA Platform Matters

Nvidia’s competitive advantage extends well beyond manufacturing high-performance processors.

Its CUDA software platform, introduced in 2006, has become a foundational technology for accelerated computing, providing developers with programming tools, libraries and optimization capabilities.

Over nearly two decades, the ecosystem has encouraged researchers, software developers and technology companies to build applications optimized for Nvidia hardware.

Moving those applications to alternative processors can require extensive code modification, testing and performance tuning.

That switching cost has helped reinforce Nvidia’s market position, alongside the performance of its chips and the capabilities of its networking and data center systems.

AMD offers a competing software environment known as ROCm, which supports the development and deployment of accelerated computing applications.

If AI tools can substantially reduce the effort required to adapt software between these ecosystems, customers could gain more flexibility in selecting hardware suppliers.

For Nvidia, that would introduce additional competition around a part of its business that has historically been difficult for rivals to replicate.

Anthropic Expands Its Multi-Chip Strategy

Anthropic’s approach is part of a broader effort to diversify the computing infrastructure supporting its Claude models.

In its November 2026 IPO prospectus, the company described a multi-chip architecture incorporating Nvidia GPUs, AMD GPUs, Google’s Tensor Processing Units and Amazon Web Services’ Trainium accelerators.

Rather than relying exclusively on one hardware supplier, Anthropic aims to distribute workloads across different computing platforms according to availability, technical requirements and economic considerations.

The company has also entered into a hardware supply agreement with AMD involving an initial commitment to obtain computing capacity in 2027.

According to its prospectus, the full initial procurement could represent expenditures exceeding $20 billion, subject to contractual and deployment conditions.

The strategy reflects the growing importance of securing sufficient computing capacity as AI developers expand their products, research activities and infrastructure commitments.

Facebook
Twiter
LinkedIn
Picture of Newsroom

Newsroom

More News