Anthropic Is Lapping OpenAI, But At What Cost?
A bonkers Axios report, and new tax proposals from Rep. Greg Casar and Sen. Elizabeth Warren, signal an end to runaway AI spending.
Anthropic is now the world’s most valuable artificial intelligence startup, according to The New York Times. Dario Amodei’s company announced tens of billions in new investments on Thursday, and is allegedly worth $965 billion. That’s significantly more than OpenAI’s $730 billion valuation. Anthropic’s glow-up is owed to the relative popularity of Claude Code and Claude Cowork. “Hundreds of businesses have signed up to pay for the software” over the last six months, the Times reported, and Anthropic says its run-rate revenue for the year has skyrocketed to an impressive $47 billion.
These bullish numbers coincide with Anthropic’s (undeserved, in my opinion) culture war victories. The Pentagon is still using the company’s models, despite loudly objecting to Anthropic’s perceived wokeness, and AI-wary Pope Leo XIV recently allowed Anthropic co-founder Chris Olah to give a speech at the Vatican.
It’s not at all inconceivable that in the near-future, Anthropic consistently turns a profit and solidifies its place atop an otherwise-volatile AI industry. But if I were part of the Anthropic C-suite, I’d be worried about an impending squeeze on both ends: business clients drawing down their usage, and legislators clawing back revenues.
A pair of news reports indicate that the company’s golden goose, its coding software, might peter out as Claude-curious businesses burn through cash and tokens (which are units of text/data processed by AI models). And if that doesn’t happen, meaning Claude Code eventually cements Anthropic’s place among Big Tech behemoths like Apple, Alphabet, and Microsoft? Well, Anthropic might get dinged by a “token tax,” a financial measurement of token usage, a concept that’s increasingly being bandied about by progressive politicians—and has a realistic shot at becoming part of the national discourse about AI.
Last week, Uber COO Andrew Macdonald was brutally honest about the lack of return on investment that his company has seen from Claude Code. (If any company isn’t going to mess with its bottom line, it’s Uber, which was famously in the red for more than a decade, until 2023.)
“That link is not there yet,” Macdonald said on the “Rapid Response” podcast, where he also acknowledged previous comments by Uber’s CTO that the company has already blown through its AI budget for the year. Macdonald later added, “We’re going to have to start talking about token consumption and the associated cost versus headcount.” Other companies, including Microsoft, are reportedly pulling back on internal Claude Code licenses for financial reasons.
“Whiplash” isn’t a strong enough term for what’s happening here; the tech industry trend of laying off workers and replacing them with AI tools/agents just kicked off in earnest. By one estimate, there have been approximately 144,000 tech layoffs this year, which is on track to surpass 2025’s figure. The only way those layoffs “pencil out” (my scare quotes; the layoffs suck!) is if the automation tools actually cost less than the workers they replace. That’s the entire premise!
Cue the other big, blinking warning sign, courtesy of a new Axios article about AI sticker shock. “Corporate leaders are starting to question whether soaring AI spending is delivering meaningful returns,” wrote Senior AI Reporter Madison Mills. “Companies that rushed to embrace AI are now confronting ballooning IT costs, uncertain productivity gains and growing employee skepticism.”
Mills cited a truly unbelievable statistic from an anonymous AI consultant, who said that one of their clients “spent half a billion dollars in a single month after failing to put usage limits on Claude licenses for employees.” Another source, an anonymous CTO, told Mills that their employees “were using AI models to check the weather.”
The glass-half-full take, as presented to Axios by Micro1 CEO Ali Ansari, is that businesses are course-correcting and learning better habits that will prevent them from continuously burning through way too many AI tokens. I get Ansari’s point—it’s certainly plausible. But even if that comes to pass, one would assume Anthropic’s revenues are due for some massive fluctuations. Worse yet, what happens if companies like Uber just abandon Claude Code entirely? Might they… return to relying on real-life workers who are actual human beings?
Perhaps that’s wishcasting. Fine. Fair. If businesses opt for a middle-ground where they modulate their AI usage, then I’d expect more and more buzz around taxing tokens. California gubernatorial candidate Tom Steyer has floated this idea on a statewide level; his proposal calls for the creation of a sovereign wealth fund based on “on corporate AI use—a fraction of a cent for every unit of data processed by Big Tech.” Alex Bores, a congressional candidate in New York City, has similarly proposed a “token tax.” Sen. Elizabeth Warren just wrote an op-ed about overhauling the tax code as a response to the “looming AI crisis.” And on Thursday, Rep. Greg Casar wrote up his own “token tax” proposal, which was published by The American Prospect.
“Currently, usage of AI products is measured in units of text called ‘tokens,’” Casar wrote. “It is possible to levy a tax on that token. To avoid companies gaming token counts, the tax should measure both the number of tokens and the underlying computing power used to train and use AI models. Taxing AI directly ties the solution directly to the problem. If AI use grows quickly, driving layoffs alongside it, the revenue from an AI tax would go up too. Unlike traditional corporate taxes, an AI tax like the one I am proposing works even if employers fire workers before AI companies show a profit.”
There are open questions about which companies Casar, Bores, and Steyer have in mind for a “token tax.” Casar mentions “providers,” but that’s pretty vague. These details need to be worked out within the next two years. Regardless, there’s no question that Anthropic, specifically, would be considered a “provider,” and thus, subject to a “token tax” under the loose definition offered by Casar.
Any AI taxes are dead-on-arrival so long as Donald Trump is the president and anti-regulation stalwart David Sacks has his ear. But the 2026 midterms are right around the corner, and after that, the leading presidential contenders for 2028 are going to offer up their visions for an AI industry that will either be bubble-bursting or proving to be somewhat useful. Normally, I would assume the worst—that basically all of the major candidates will be inclined to let AI super PACs speak for them. But AI is polling so poorly, and killing off so many jobs, that I actually see an opening for a “token tax.” It would, I presume, be very, very popular—and serve as a backstop against tech worker job losses.
Anthropic has previously claimed to be open to such taxes, but reader, I will be honest with you: I do not believe them. Anthropic is well-aware of its perilous position, and will do whatever it takes to keep its revenues flowing. The good news is those efforts might not matter. “Tokenmaxxing,” as it’s unfortunately known, seems to be on the way out; its enchanting spell among tech CEOs is starting to break. If Anthropic wants to stick around and maintain its newfound lead over OpenAI, I suspect it’s going to have to pay up sooner rather than later.
Here’s what else we’re reading this week:
A Google security engineer who allegedly used internal company data to place bets on Polymarket was charged with wire fraud, commodities fraud, and money laundering on Wednesday. According to a federal criminal complaint, first reported by ABC News, the engineer pocketed more than $1 million on his bets. Last November, Google announced partnerships with Polymarket and Kalshi, so that its users can “harness the wisdom of the crowds.”
Over at Blood in the Machine, Brian Merchant has an encouraging update about University and Professional Technical Employees (UPTE), the largest tech workers union in the country. They’ve grown their ranks to 8,400 workers!
The “Here’s How San Jose Mayor Matt Mahan Screwed Up His Run for California Governor” post-mortems have already begun, even though the primary isn’t until next week. Y Combinator CEO Garry Tan offered an especially funny preemptive defense of the tech industry’s losing Mahan bet: “This is our education,” he told Politico. “This is first grade, second grade for me, personally. We won’t be first and second graders forever.” I’ll likely have more reflections on the governor’s race next week.
Peter Thiel has relocated to Buenos Aires, Argentina, where he is entering local chess tournaments and rambling to libertarians about his favorite topic, the antichrist, according to The New York Times.
Plugging my own reporting for New York magazine: I spoke to a dozen sources about John Fetterman’s dwindling inner circle, and how his closest confidant is now an unpaid, little-known adviser named David Safier. Jokingly referred to as “chief” by staffers, Safier has coordinated calls with Benjamin Netanyahu and sits in on sensitive meetings with diplomats. Fetterman’s actual chief of staff quit the day before my article was published. I wrote up a Bluesky thread that summarizes the reporting, if you can’t get around the New York paywall.