Anthropic began its IPO roadshow on November 10, and its video features co-founder Tom Brown saying the company trains and serves Claude across four chip families.
Brown's central claim is that "a dollar of compute at anthropic goes further than a dollar at any other frontier lab," according to a transcript of the video's "Approach to compute" segment1. It is a company assertion, and the segment offers no cost comparison with other labs. Anthropic expects3 to set the final IPO share price on November 18 and begin trading on Nasdaq under the ticker "ANT" on November 19, subject to regulatory clearance.
Simon Willison, who published the transcript, wrote2 that the video contains "previously unreleased information about how their business works." He identified the speaker as co-founder Tom Brown; the transcript itself does not name him. Willison said he did not know why there is no official transcript and produced his own with MacWhisper. The quotations here come from that machine-generated text and have not been checked against the video.
The roadshow is the company's marketing material. The IPO page describes the prospectus as the regulatory filing with detailed information on the business, financial statements and risk factors. The figures below come from the video transcript and the IPO page, not the prospectus.
A fungible pool across four chip families
Brown said Anthropic trains and serves Claude "across all four major chip families" and described the approach as "an unconventional bet on a multi-chip strategy." Willison's post lists the four as AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs.
Brown acknowledged the cost: "Each chip needs its own tailored software stack and inference stack, which is expensive and difficult to build." He gave two advantages. In what he called a supply-constrained compute market, the strategy lets the company "focus on the best price performance regardless of the provider." And it lets Anthropic match each chip to a job: "Some are better at training, some are better at inference." Together, he said, this allows it to treat all its chips "across providers, architectures, and locations as one fungible pool of compute."
Anthropic's September 2025 postmortem of three infrastructure bugs4 said it deploys Claude on "AWS Trainium, NVIDIA GPUs, and Google TPUs." It did not list AMD. Willison called the postmortem "a hint at" the multi-chip strategy and noted that it described bugs affecting TPU deployments. Two of the three bugs, the output corruption and the XLA:TPU compiler error, were TPU-specific. The context-window routing error was not; it affected traffic on Anthropic's own API, Amazon Bedrock and Google Cloud's Vertex AI. The postmortem said that "any infrastructure change requires careful validation across all platforms and configurations." That sentence suggests a validation burden from running several platforms, though this reading is the writer's, not Anthropic's or Willison's.
As an example of the multi-chip capability, Brown said Anthropic "started working with AMD GPUs for the first time" this year. He said bringing on a new chip platform used to take "dozens of engineers months of work." This time, he said, "we had Mythos try it out. It got Claude running on AMD in just three days, with basically no human intervention. Claude ported or wrote almost 100% of the kernels." The transcript does not say what Mythos is. It contains no benchmarks of AMD performance, though it covers only the compute segment of the video.
Anthropic Direct and compute multipliers
Brown said Anthropic historically bought compute as a cloud service and has "now added a second approach" in which it leases chips and runs them in data centers "designed for us." The company calls this Anthropic Direct. He said it "helps to lower the total cost of ownership of compute and allows us to separate the commitment to data centers from the chips." The segment gives no figures on how much capacity runs under each model.
The second driver he cited is the rate at which Anthropic discovers "compute multipliers." He defined "total effective compute" as "the multiplicative product of physical capacity times the compute multipliers that we discover." His example: with a 4x multiplier, a model that needs 100 units of compute can be trained with 25, or the same 100 units can train a model that would previously have needed 400. He said multipliers "stack," and that as the company approaches "recursive self-improvement," Claude "will be able to discover compute multipliers entirely by itself." The segment gives no timeline for that and does not say how many multipliers exist or what they add up to.
What the IPO page reports
The IPO page3 lists a ">4x compute efficiency gain since Q1'25." A footnote defines it as revenue divided by compute and infrastructure expenses in Q3'26, divided by the same ratio in Q1'25. That is a financial ratio, not a measured training multiplier. In the compute segment, Brown's 4x example is not tied to this figure in the transcript, and the page does not present it as confirming his illustration.
The page's Q3'26 financial highlights also show more than 11,900 customers with at least $100,000 in trailing-twelve-month revenue, up more than 12x year over year. They list a net loss of $23.3 billion. Anthropic attributes it to an increase in its valuation and says it includes a $21.9 billion non-cash charge related to a change in the fair value of convertible notes, derivative liabilities and SAFEs. The page shows adjusted operating margin and operating margin as headings, but the retrieved text includes no values for them.