Corrie Driebusch and Kate Clark, writing for The Wall Street Journal:

Anthropic has officially started the clock on what could be the biggest initial public offering ever. The maker of Claude unveiled its IPO paperwork Monday morning, putting it on track to go public as soon as the week before Thanksgiving. 

The 400-page prospectus offered a mix of promise and peril. Artificial intelligence, it said, could usher in a new era of progress, transforming how people code, conduct research, build financial models and automate routine tasks, and even help cure major diseases like cancer. At the same time, AI could lead to the development of new bioweapons, “destabilize the geopolitical order” and evolve beyond the ability of “humans to understand and control.”

It’s not a good start to your IPO when you position your company’s technology as destabilizing geopolitics and overthrowing humanity, but that has always been what Anthropic has been about. I’ve spent a lot of time over the last few weeks and months criticizing the company’s biggest adversary: OpenAI, which has lost its product, model, and market-share advantages in the back half of this year. But Anthropic is not anodyne or faultless. While I am more bullish on its prospects than OpenAI’s or even Google’s, the company’s public relations strategy can be succinctly summed up as: “Our technology that we are actively building may destroy the world, but don’t blame us. We’re just the messengers.” That is not a winning strategy, not now and not ever. Either the company should develop a solution to the grand problems it posits will occur, or it should stop asking for funding altogether.

Its operating loss, which measures the profitability of its core operations, was much smaller: $693 million, down from $2.1 billion a year earlier.

For all the talk about how much money AI loses, Anthropic has a lot of it. The company said that it had $112.4 billion in cash and cash equivalents as of June 30. That is 80% of the $141.1 billion it has raised in the six years since its founding. Anthropic also had $19.3 billion of short-term investments in a variety of securities, including money-market funds.

Anthropic is the least unprofitable AI start-up due to those enterprise relationships it forged early on after the launch of Claude Code. OpenAI tried to play catch-up for a while over the summer but seems to have largely fallen back on consumer products, such as Dots and the merging of ChatGPT Work and Chat. Anthropic, while obscure overall compared to ChatGPT among consumers, is heavily favored among software engineers, students, professors, mathematicians, and other academics. Anyone who has been to a university in the past year knows how popular Claude is among researchers, in part due to its heavy use in enterprise. The company has unequivocally nailed the token subsidization strategy in the past year, and unlike model advantages, that (i.e., loyalty from power users stemming from enterprise contracts) tends to hold over time.

Anthropic is on the hook to pay more than $475 billion in compute capacity agreements over the next seven to 10 years. The bulk of these deals are noncancelable, regardless of how much it uses.

Among the biggest beneficiaries are Google and Amazon, who are each set to receive around $110 billion in the coming years. Elon Musk’s SpaceXAI is also set to receive $120.5 billion through 2029, though the deal can be canceled by SpaceXAI or Anthropic with 90-days notice.

The most pressing concern for any technology company in 2026 is not profitability or investment, but compute (as a noun). Whether it’s Apple margins nosediving due to increased memory costs, OpenAI’s rebranding of unpopular data centers ahead of Tuesday’s midterm elections, or Anthropic’s unlikely deals with other “Big Tech” companies, compute will probably define who wins the AI wars over the next few years. If Anthropic wants this IPO to be successful, it must convince Wall Street that it won’t hemorrhage money just to serve its existing customers. Wall Street is already showing signs of frustration with the tech industry — any AI company that can’t source inexpensive compute for its operations is almost certainly dead long-term. We’re past the point now of thinking that balanced token subsidization is the sole roadblock to AI profitability.

Unfortunately, compute is the hardest obstacle to overcome in the modern political climate, and rightfully so. I keep turning back to the public sentiment issue because it is so urgent — data centers are extraordinarily unpopular, and yet they’re the only solution to Wall Street’s AI problem. (And yes, this is Wall Street’s problem for bankrolling unprofitable start-ups for far too long.) The opposition to AI data centers is less about pollution, water use, noise, etc., and more of a (generally justified) vitriol against Silicon Valley and AI companies. Anthropic’s biggest problem on paper might be compute, but really the most immediate concern ought to be its very first point about AI destroying the world. Anthropic cannot get very far with this “warning signal” attitude. Anthropic is an AI company, not a group of hippie altruists; it wants to turn a profit selling access to large language models. People want to see real proof that AI companies are putting solutions to these concerns ahead of profit, and that proof is nonexistent.

Until the public is assured that the AI companies are not entirely self-interested, the backlash against AI development will continue. And that will unquestionably precipitate real compute constraints.