A full set of AMD compatibility achieved in one weekend: Anthropic uses Claude bootstrapping instead of Nvidia hardware, breaking the CUDA "human resource barrier"

At this year's AdvancingAI conference by AMD, Anthropic not only officially announced its computing power deployment plans, but also revealed a technical detail that could potentially reshape the rules of competition for AI chips: its engineers fully adapted and optimized the performance of the AMD Instinct MI355 chip with the ROCm platform over a single weekend, using only the Claude model in an automated fashion.

Tom Brown, an Anthropic executive, publicly stated at the conference that the company plans to deploy 2GW of AMD Helios computing facilities and that it clearly favors the MI355 chip. This statement upgrades previous market rumors of a partnership to an official confirmation at the executive level. AMD’s official account subsequently retweeted a third-party post containing Tom Brown’s remarks, adding a thank you note for his participation.

Even more impactful was Brown’s disclosure of the adaptation process. The team originally expected that deploying the model on new hardware would be “a major project,” but the actual experience turned out completely different: an engineer assigned the adaptation task to Claude, let it run over the weekend, and by Monday, the team received a real performance curve of Anthropic’s leading model “steadily climbing over the weekend” on the hardware.

Market observers immediately commented: “We are past the CUDA moat era.” For investors, this breakthrough means the hardware lock-in logic of the AI computing supply chain is facing a counterattack from the AI’s own abilities. When AI itself can replace engineers in the most labor-intensive part of hardware migration, the fundamental moat of the Nvidia CUDA ecosystem—its high migration costs—is being eroded at its core by AI.

From Rumor to Official Announcement: Anthropic’s AMD Deployment Surfaces

Previously, market recognition of Anthropic’s partnership with AMD was only at the level of industry speculation. Tom Brown’s public appearance at the AMD technical conference raised the credibility of the partnership from an indirect signal to a formal confirmation.

A deployment scale of 2GW means that AMD chips play a substantial, not just symbolic, role in Anthropic’s computing architecture.

As an AI giant with $47 billion ARR, a $965 billion valuation, and a formal IPO application already filed, Anthropic is actively building a diversified computing supply system—for example, it previously procured chips and cloud services from Google, signed a nearly $45 billion collaboration with SpaceX, and established an $1.8 billion agreement with Akamai. AMD’s official involvement further enriches its sources of computing power.

AI Bootstraps Hardware Adaptation: The “Achilles’ Heel” of the CUDA Barrier

Even more consequential for the long term than order size is the AI-driven automation capability for adaptation demonstrated by Anthropic.

Traditionally, migrating large models to new hardware platforms requires numerous engineers to manually adapt low-level operators, optimize performance, and verify stability—this has been the core source of Nvidia CUDA’s ecosystem barrier. Developers’ reliance on CUDA is not just technological inertia, but also dictated by migration costs: switching platforms means months or even years of engineering resources.

But when Claude can accomplish the entire adaptation process within a single weekend, the foundation of this logic starts to loosen. Brown’s description is straightforward: one engineer starts Claude, “let it run the machines,” and the process operates itself over the weekend. By Monday, the team receives a performance chart showing “continuous improvement.” The whole process involves only one engineer and a server rack provided by AMD.

“We thought this would be a major project,” Brown said, but the actual experience was “completely different.”

Why This Signal Has Substantial Significance for the AI Chip Landscape

For the market valuation of Nvidia, the CUDA ecosystem barrier is a core premium support. The essence of this barrier is not the irreplaceability of the technology, but the “human cost barrier” of ecosystem migration—even if competitors’ hardware matches the performance, companies still need to invest a large number of engineers to re-adapt the software stack. This sunk cost is the highest switching threshold itself.

AI-driven automation for adaptation targets exactly the cost side of this barrier. If leading-edge labs can use their own AI models to finish adaptation and optimization for new hardware within a few days, hardware procurement decisions will depend more on performance, price, and supply availability, rather than ecosystem lock-in. As analyst Austin Lyons noted: “We are past the CUDA moat era.”

For Anthropic, this ability gives its computing procurement strategy much greater flexibility. The latest analysis from SemiAnalysis points out that Anthropic’s inference infrastructure gross margin has jumped from 38% to over 70%, and reached operating profit in Q2. Including AMD as a computing supplier not only diversifies supply chain risks but also provides more cost-effective hardware options for the rapidly expanding inference demand.

For AMD, gaining a formal deployment commitment from a cutting-edge lab like Anthropic is a key validation of its data center GPU roadmap. With AI-driven automated adaptation tools, the engineering barrier for more prospective customers to try the AMD platform is lower—which may be a strategic asset with more long-term value than a single order.

SemiAnalysis analysts pointed out that Anthropic’s aggressive investment in programming data previously gave its models a leading edge, and now this capability is in turn empowering its own flexibility in hardware selection.