Anthropic Says Its AI Did Months of Chip Work in Three Days
A roadshow claim about AMD GPUs, a 26 billion dollar side deal with Nvidia, Microsoft, Google, AMD and TSMC, and a week of very large AI checks.
Three days. Versus six to twelve months, with dozens of engineers. That's the claim Anthropic made to IPO investors this month. And I'm going to read you the fine print, because the fine print is where the real story is.
The claim
Some background. AI models are written to run well on one kind of chip, mostly Nvidia's. Moving a model to a different company's chips, like AMD's, has historically been a huge, slow software job.
In a roadshow video dated November 10, Anthropic's co-founder and chief compute officer, Tom Brown, said the company's model, Claude Mythos, did that software work for AMD's MI355X chips in about three days. Mythos ported or wrote, quote, "almost 100 percent of the kernels." Kernels are the small, heavily optimized programs that do the actual math on the chip.
The fine print
Now the footnote. According to Tom's Hardware, the slide limits the claim to one model and one part of the job: prompt processing, the step where the model reads what you typed. Not the whole thing. Still impressive. But one model, one stage.
And remember why they're saying it. This is an IPO roadshow. Every company on a roadshow is putting its best foot forward.
Follow the money
Here's the wild part. Anthropic has a deal with AMD for up to two gigawatts of computing. And separately, a 26 billion dollar private placement of Anthropic stock at the IPO price includes Nvidia, Microsoft, Google, AMD and TSMC. Nvidia and AMD, the two fiercest chip rivals in the world, are both buying in.
The rest of the week's checks
And it's not just Anthropic. On Monday alone: a defensive cybersecurity company called Corma raised 125 million dollars at a one billion dollar valuation. It came out of hiding in August and has about 25 employees. Twenty-five people. One billion dollars.
A new company called Noeri, started by Stanford professors Yejin Choi and Carlos Guestrin, launched with 110 million dollars. Guestrin co-created XGBoost, a tool data scientists use every day. And a London investment firm, Plural, closed a one billion euro fund.
So what do you do with that? When this much money moves this fast, it means the people with the most information think the next few years matter a lot. Whether they're right is the bet. And Thursday, regular people get to place it.
Disclosure, as always: this site drafts stories with help from Claude, which is made by Anthropic.