Anthropic began its IPO roadshow yesterday, and the roadshow video contains operational details about how the company trains and serves its Claude models across multiple chip families and how it automates platform porting.
Chip platforms and the multi‑chip approach
In the video’s "Approach to compute" segment, co‑founder Tom Brown explained that Anthropic now trains and serves Claude across four major chip families and platforms: AWS Trainium, Google TPU, NVIDIA GPU, and AMD GPU. Brown said this has been feasible because the company early on made a deliberate decision to pursue a multi‑chip strategy.
Anthropic noted the challenge: each chip family requires its own tailored software and inference stack, which is costly and complex to develop. Their position is that after years of investment, they have become highly proficient at operating across multiple hardware platforms.
Automated porting and the Mythos example with AMD
A transcript extracted with MacWhisper from the video reports that this year Anthropic started working with AMD GPUs for the first time. Historically, enabling a new chip platform took dozens of engineers months to complete. This time, they tested a system called Mythos, which they say got Claude running on AMD in just three days with almost no human intervention. According to the video, "Claude ported or wrote almost 100% of the kernels."
Such automation, if accurate, reduces the time and engineering effort needed to bring new hardware online.
Compute multipliers and effective compute
The video also emphasized Anthropic’s focus on so‑called "compute multipliers." The company describes total effective compute as the product of physical capacity and the compute multipliers they discover. Each new multiplier increases the company’s total effective compute and thus its computational advantage.
They illustrated with an example: if training a model of a certain capability requires 100 units of compute, a 4× compute multiplier would allow the same model to be trained with 25 units; alternatively, the same 100 physical units could be used to train a model that previously would have required 400 units.
Anthropic said these multipliers stack and reinforce one another. The company further suggested that, as systems approach recursive self‑improvement, Claude itself could eventually discover compute multipliers autonomously, accelerating the rate at which such multipliers accumulate.
Earlier indications: TPU‑specific issues
Some of the operational context in the roadshow video echoes a September postmortem from Anthropic that described three recent incidents and bugs that affected their TPU deployments specifically.
Why this matters
The details in the IPO roadshow video indicate Anthropic is investing in hardware‑agnostic operations and automation to improve scalability and cost‑efficiency. The compute‑multiplier framing shows how the company is attempting to quantify and amplify its effective compute resources, including a stated aspiration that models may eventually contribute to discovering their own optimizations—an element the company sees as part of its future competitive edge. These operational claims provide investors with additional insight into how Anthropic plans to manage and scale compute across diverse hardware platforms.
