Current Location:Home > News > Company News
Anthropic has disclosed that the Mythos model is compatible with AMD MI355X within 3 days.
Release Time:2026-11-13 2:05:54

In just 3 days, this AI completed half a year's work. It moved itself onto the AMD platform.


On November 10, 2026 local time, Anthropic officially launched its IPO roadshow. It plans to issue 825 million Class A common shares, with the price range set at $70 to $78 per share. Based on the upper limit, the maximum fundraising amount could reach approximately $64.4 billion. It has applied for listing on the Nasdaq Global Select Market, with the stock code "ANT".


According to the disclosure in Anthropic's IPO roadshow materials, the company's cutting-edge model Mythos completed the adaptation of the model running on the AMD Instinct MI355X GPU in about 3 days. In the traditional engineering path, completing similar adaptations for a new computing platform typically takes 6 to 12 months. Anthropic's statement in the prospectus is more cautious: With the help of Claude, the engineering team can now put new chip platforms into production at a much faster speed than in the past.


AI writes its own kernel, and the migration cost is recalculated.

According to the disclosure in the IPO roadshow materials, Mythos independently migrated or wrote nearly 100% of the kernel code - the underlying programs required for the model to run on the chip, with almost no human intervention. It should be noted that this data is based on the time required for Anthropic models to achieve usable performance during prompt processing on the MI355X, which belongs to the inference-side adaptation, and does not cover training migration. The relevant figures come from Anthropic's investor materials and have not been independently verified.

In the past, moving a large model to new hardware usually required dozens of compilers and high-performance computing engineers to work for 6 to 12 months: sorting operators, rewriting the CUDA interface corresponding to the implementation, aligning numerical precision, and layer-by-layer optimization. In the procurement assessment, the migration man-months, the number of optimization rounds, and the reconstruction of the operation team's knowledge system all need to be included in the total account. When the migration cost reaches a certain level, changing the hardware supplier becomes financially unfeasible. NVIDIA's 20-year accumulation of CUDA libraries and toolchains is the main cost source for customers to change hardware suppliers. If AI-assisted adaptation can be replicated in batches, the threshold of "can it run" will be significantly lowered, and the room for buyers of computing power to compare and switch suppliers will expand accordingly.


Four chip families, one computing power pool

Anthropic's co-founder and chief computing officer Tom Brown stated in the roadshow that the company has bet on the multi-chip route since 2021. At that time, the industry generally believed it was not worth it. Now Claude has been trained and run on four families of NVIDIA GPUs, AMD GPUs, Google TPUs, and AWS Trainium. The computing power from different manufacturers, architectures, and regions is regarded as an interchangeable resource pool, uniformly scheduled based on research, customer needs, and internal use. When one platform's production capacity is tight, the inference load can be shifted to other hardware. The company also plans to reduce the total cost of ownership through leasing chips and the Anthropic Direct model, which involves building its own data center.

For AMD, this directly affects its ROCm software ecosystem. On July 22, 2026, AMD officially announced a strategic partnership with Anthropic: Anthropic will deploy up to 2 gigawatts of MI450 series GPUs in the Helios rack-level solution, with the first 1 gigawatt starting in the first half of 2027; AMD committed a maximum $5 billion equity investment, and the engineering teams of both parties will use Claude to optimize Instinct GPU workloads and accelerate ROCm development. The three-day adaptation results disclosed this time are highly consistent with the direction of both parties in promoting the ROCm ecosystem and cross-platform deployment capabilities.


However, what was completed in three days was only the "runnable" part. The actual throughput gap between the same model on different hardware can reach several times. Performance tuning, numerical stability verification, and continuous maintenance after the version iteration of the software stack still require engineering investment. The long-term effectiveness of automatically generated code also needs to be observed.


Disclaimer: This article is from the Internet. If there is any dispute, please contact customer service.


Recommend News

Copyright © UDU Semiconductor