Anthropic Taps Google And Broadcom For Hardware

All of the news around Anthropic's recent work doesn't really fit well into one short headline, so: essentially, the company is preparing to roll out 3.5 gigawatts of compute, nested in various data centers, using TPUs from Google and Broadcom, according to new reporting at TechCrunch (and some details from a recent SEC filing by Broadcom.)

"This reworking of Anthropic's compute deals comes as demand for its AI models continues to soar," writes veteran IT reporter Rebecca Szkutak April 7.

But there's also other context to this: namely, a plan by Anthropic to dot the U.S. landscape with data centers in order to support the kinds of operations that it is seeking the new hardware for.

Apparently, many of these facilities will be located in the states of New York and Texas. Another TechCrunch reporter, Russell Brandom, covered this last November, noting a partnership with Fluidstack, a "neocloud" company, and estimating that Anthropic will spend around $50 billion on the effort.

"We're getting closer to AI that can accelerate scientific discovery and help solve complex problems in ways that weren't possible before," said Anthropic co-founder Dario Amodei at the time. "Realizing that potential requires infrastructure that can support continued development at the frontier."

Why TPUs?

Some of you might be wondering why the purchase agreements in Anthropic's ledger are for Tensor Processing Units or TPUs, and not the GPUs popularized by chip front-runner Nvidia.

The answer, in a nutshell, is that the TPUs have a more efficient build for specialized workloads. The GPUs are more practical for broader uses, and so, if a company like Anthropic has highly defined AI processes, a TPU might be a better match. However, the industry as a whole is very much geared toward the likes of the Grace Hopper and Vera Rubin lines of Nvidia GPUs named after famed female computer scientists.

Anthropic's Ascendance

All of this also underscores the dramatic rise of Anthropic as a top U.S. AI company. Famously, Dario Amodei co-founded this company with his sister just a few years ago, and now, they're handling quite a lot of money.

"Anthropic basically went from startup to money printer in 12 months," writes Dani at Humai.blog, citing estimates of new Anthropic funding. "$30 billion revenue run rate is the kind of number that makes OpenAI executives wake up in cold sweats, and Google just became their best friend with the biggest compute deal in AI history."

Here's a new quote from Anthropic CFO Krishna Rao, on the Google-Broadcom deal:

"This groundbreaking partnership with Google and Broadcom is a continuation of our disciplined approach to scaling infrastructure: we are building the capacity necessary to serve the exponential growth we have seen in our customer base while also enabling Claude to define the frontier of AI development. We are making our most significant compute commitment to date to keep pace with our unprecedented growth."

Inner - and Outer - Conflict

Of course, amidst all of this, the big news about Anthropic a couple of months ago was its tiff with the Department of Defense over autonomous lethal weapons and mass spying on U.S. citizens, in which the AI company was on the "no" side. But there's also a critique of the company's internal motivations, where some argue that Anthropic's trajectory seems a little schizophrenic.

"Anthropic is at odds with itself -- thinking deeply, even anxiously, about seemingly every decision," writes Matteo Wong at The Atlantic in an article titled, melodramatically: Anthropic is at War with Itself.

The basic idea seems to be that, even though Amodei has publicly acknowledged the dangers in moving ahead quickly with AI, it can't really seem to slow down. That's a piece of critical context for a big TPU purchase, and a national data center plan.