ARTIFICIAL intelligence (AI) has a maths problem.
Anthropic’s initial public offering (IPO) filing shows revenues rose nearly 12-fold last year but operating losses ballooned as well, raising an uncomfortable question: will leading AI labs ever turn a profit?
The prospectus, obtained last week by Reuters, showed that while Anthropic’s revenue jumped to US$4.6bil in 2025, it spent US$7.3bil on compute and infrastructure.
The result was an operating loss that widened from US$2.98bil in 2024 to US$8.06bil in 2025.
Dario Amodei’s firm is certainly not the only AI-focused company to be in the red. Elon Musk’s SpaceX saw its AI-related capital expenditure dwarf its revenue in the second quarter (2Q26).
The question hanging over Anthropic’s potential US$2 trillion valuation and those of its similarly unprofitable peers is whether such losses will ever shrink.
The core argument made by AI firms is that revenues will grow extremely fast as more people adopt large language models (LLM), and that this will result in greater spending on tokens.
But Anthropic and its competitor OpenAI have recently released new versions of older models offering cheaper token prices.
Take a look at the LLM Token Expenditure Index, created by Silicon Data, which estimates the average price of LLM tokens. The index has declined by over 40% since June 30 as AI labs have significantly reduced the prices for models just behind the frontier.
How does all this square with the supposed paths to profitability?
The likely retort from AI firms would be a version of Jevons’ paradox, the idea that lower costs will spur enough additional use to increase overall consumption.
Based on this theory, token volumes should grow fast enough to offset any price declines. But, for now, there is little to suggest this is the case.
Token usage may have surged temporarily in the 2Q26, attributed by some to so-called “tokenmaxxing”, where workers, mostly in the tech industry, sought to maximise their token usage.
If the index is accurate and “tokenmaxxing” has dissipated, as many believe, overall spending on LLMs by businesses and consumers could be lower than it was three months ago.
Even if usage is still rising, that may not be enough to offset the significant drop in the average price of a token.
To be fair, the index is a new tool that could be missing segments of the market – and it also doesn’t distinguish between models.
But if its findings are borne out, this could become a significant obstacle for Anthropic’s IPO ambitions.
The company’s current IPO prospectus uses 2Q26 results, the latest quarterly data available if the company had gone public in October.
But given that Anthropic is now expected to delay its filing until November, it would instead have to show 3Q26 results, which may not look quite as attractive.
Other publicly available indicators also raise questions about the company’s revenue outlook.
OpenRouter, which helps users direct their requests to the model that can reliably provide an answer at the lowest available cost, currently sends only 2.6% of its requests to Anthropic’s models. This is down from 5.2% at end-June and a peak of 6.4% in May.
Growing doubts
I have said before that I think the future of AI is not in LLMs running in data centres but in smaller, cheaper models running locally on desktops, which, if proven correct, would dramatically undercut the valuations of today’s AI leaders.
Others seem to have doubts about the true value of AI labs as well. While many analysts and investment bankers tout the growth of AI and remain bullish on the companies in that space, some loan officers appear more cautious.
For example, when SoftBank in August took out a US$10bil margin loan against its stake in OpenAI in order to invest more in the ChatGPT maker, among other corporate purposes, it valued its position at more than US$89bil.
Yet the tech investment conglomerate has spent several months trying to replace a US$40bil bridge loan used primarily to fund its investment in OpenAI, and when it did finally manage to place an US$11.1bil high-yield bond offering in the United States, it had to back this unsecured bond with other investments as well.
Bond investors are notoriously risk-averse and generally demand large margins of safety when lending against existing investments, so it is no surprise that SoftBank could not borrow the desired US$40bil against its OpenAI stake alone.
What matters here is the scale. The margin loan is less than 11% of the value SoftBank assigned to its OpenAI stake at the time. This suggests the risk-averse lenders likely assigned a much lower valuation to the holding.
A bit too bullish
All US AI leaders are facing questions about current valuations and future profitability, as there appears to be an internal inconsistency in Wall Street’s hyper-bullish AI forecasts.
As Apollo’s chief economist Torsten Slok has highlighted, US technology companies are projected to grow their operating cash flow by more than US$1.2 trillion from 2025 to 2028, much more than the other sectors combined.
The two forecasts are difficult to reconcile. If US tech companies generate more than US$1 trillion in additional operating cash flow, their customers in the rest of the economy must either generate much more cash themselves or find other ways to finance that spending.
The implications here go far beyond any one IPO.
Much of the current equity boom on Wall Street and around the world is being driven by the belief not only that AI will change the world but that today’s tech leaders will be the ones to profit from that. Undercut that narrative, and investors could get a rude awakening. — Reuters
Already a subscriber? Log in
Get 20% OFF The Star Digital Access
Cancel anytime. Ad-free. Unlimited access with perks.
