Full AMD Adaptation Achieved Over a Weekend: Anthropic Uses Claude for Self-Bootstrapping Instead of NVIDIA Hardware, Breaking CUDA's "Labor Barrier"

At AMD's Advancing AI conference this year, Anthropic not only officially announced its computing power deployment plan but also disclosed a technical detail that could rewrite the rules of competition in the AI chip market: its engineers used the Claude model to automatically complete the full adaptation and performance tuning of the AMD Instinct MI355 chip with the ROCm platform within a single weekend.

Tom Brown, an executive at Anthropic, stated publicly at the conference that the company plans to deploy 2GW of AMD Helios computing facilities and explicitly favors the MI355 chip. This statement upgraded previous market rumors of cooperation to a formal confirmation at the executive level. AMD's official account subsequently retweeted a third-party post containing Tom Brown's remarks, accompanied by a message thanking him for his attendance.

Even more impactful was the adaptation process disclosed by Brown. The team originally expected that launching models on new hardware would be "a major project," but the actual experience was entirely different: after an engineer assigned the adaptation task to Claude, the process ran autonomously over the weekend. By Monday, they had obtained the actual performance curve showing "continuous improvement throughout the weekend" for Anthropic's leading model on this hardware.

Market observers promptly commented, "We have passed the era of the CUDA moat." For investors, this breakthrough means that the hardware lock-in logic of the AI computing supply chain is facing a reverse impact from AI's own capabilities. When AI itself can replace engineers to complete the most expensive labor-intensive aspects of hardware migration, the fundamental moat of the NVIDIA CUDA ecosystem—high migration costs—is being eroded from its roots by AI.

From Rumors to Official Announcement: Anthropic's AMD Deployment Emerges

Previously, market perception of the collaboration between Anthropic and AMD remained at the level of industry speculation. Tom Brown's public endorsement at the AMD technology conference elevated the credibility of the cooperation from indirect signals to formal confirmation.

The deployment scale of 2GW implies that AMD chips play a substantive role in Anthropic's computing system, rather than being a symbolic procurement.

As an AI giant with an Annual Recurring Revenue (ARR) of $47 billion, a valuation of $965 billion, and formally filed IPO application documents, Anthropic is actively building a diversified computing supply system. It has previously purchased chips and cloud services from Google, signed a nearly $45 billion cooperation agreement with SpaceX, and reached an $1.8 billion agreement with Akamai. The formal addition of AMD further enriches its sources of computing power.

AI Self-Bootstrapped Hardware Adaptation: The "Achilles' Heel" of the CUDA Barrier

More impactful in the long term than the order size is the AI automated adaptation capability demonstrated by Anthropic.

Traditionally, migrating large models to new hardware platforms requires a large number of engineers to manually complete low-level operator adaptation, performance tuning, and stability verification—this constitutes the core source of the barrier for the NVIDIA CUDA ecosystem. Developers' reliance on CUDA is not just due to technical inertia but is determined by migration costs: switching platforms means investing months or even years of engineering resources.

However, when Claude completed the entire adaptation process over a weekend, the foundation of this logic began to loosen. Brown's description was clear and straightforward: an engineer started Claude, "letting it get the machine running," and then the process ran autonomously over the weekend. By Monday, the team received a chart showing a performance curve that was "continuously rising." The entire process involved only one engineer and one rack provided by AMD.

"We thought this would be a big project," Brown said, but the actual experience was "completely different."

Why This Signal Has Substantial Implications for the AI Chip Landscape

In the market's pricing of NVIDIA, the CUDA ecosystem barrier is the core support for its premium. The essence of this barrier is not the irreplaceability of the technology, but the "labor cost barrier" of ecosystem migration—even if competitors' hardware performance catches up, companies still need to invest a large number of engineers to re-adapt the software stack. This sunk cost itself is the highest switching threshold.

AI automated adaptation directly attacks the cost side of this barrier. If frontier laboratories can leverage their own AI models to complete the adaptation and optimization of new hardware within a few days, hardware procurement decisions will depend more on performance, price, and supply availability rather than ecosystem lock-in. In the words of analyst Austin Lyons: "We have passed the era of the CUDA moat."

For Anthropic, this capability grants greater flexibility in its computing power procurement strategy. SemiAnalysis's latest analysis points out that Anthropic's inference infrastructure gross margin has jumped from 38% to over 70%, achieving operating profit profitability in the second quarter. Incorporating AMD into its computing supplier system not only disperses supply chain risks but also provides more cost-effective hardware options for rapidly expanding inference demands.

For AMD, securing a formal deployment commitment from a frontier laboratory like Anthropic is a key validation of its data center GPU roadmap. Furthermore, the support of AI automated adaptation tools lowers the engineering threshold for more potential customers to try the AMD platform—this may be a strategic asset with greater long-term value than a single order.

SemiAnalysis analysts pointed out that Anthropic's previous aggressive investment in programming data gave its model capabilities a lead, and now this capability is inversely empowering its own freedom in hardware selection.