The Week AI Blinked?

This week, Character.AI, a former unicorn startup that allowed users to create virtual companions, announced it was filing for bankruptcy while selling its remaining assets to Disney. Bankruptcy is a convenient way to dodge accountability for the company’s several active lawsuits and investigations into allegations that its chatbots posed as doctors and encouraged children into suicides. The asset purchase is also amusing, since Disney had previously sued Character.AI for deploying its characters for their chatbots.

But bankruptcy also gives a look under the hood at the AI ecosystem, and it’s not a pleasant picture. Character.AI had attracted 12 million active users, but computing costs to handle that user base were exorbitant. The company was losing between $8 million and $9 million a month. Without the capital to finance even more money-losing growth, it resorted to corroding the user experience to reduce costs, both by using cheap substitute models that reduced engagement and by placing usage restrictions on free users. It also introduced advertising and then purchases into its app in a bid to huckster its way out of trouble. The primary effect of all this was turning off core users, a bad sign for a company desperate to increase revenue.

While Character.AI’s scale wasn’t near the levels of the industry’s giants, the problem it encountered is familiar. The industry is saddled with a conundrum: Increasing user bases correlates with worsening financials. Unlike the software industry, where each additional copy of Photoshop or Microsoft Word takes almost nothing to produce, AI models are very resource-intensive to run. The more people use the product, the more money it takes to run the system, and in this case at least, setting prices high enough to cover costs craters the customer base. As a result, corners are cut, revenue fillers like ads are introduced, the product becomes crappier and more predatory, users quit anyway, and the company goes under.

This financial mismatch would be one thing if we were talking about a network of gaming apps for phones, or a series of niche hobby shops. It’s another if the industry with this catch-22 is the one that America and the world is reliant on for virtually all economic growth.

The hyperscale giants would insist that they are not trapped by this problem. At a recent investor day event in advance of its anticipated November initial public offering, Anthropic asserted that third-quarter revenue increased 45 percent to $16.8 billion. But even this did not necessarily translate into profits; Anthropic was cagey about it, claiming that it is “profitable after making certain adjustments,” a massaged accounting that simply isn’t credible. The S-1 prospectus that would lay out the actual numbers has yet to be released publicly.

The most equivalent comparable company to Anthropic, OpenAI, last week had to acknowledge that its annualized revenue was $20 billion less than what it had been touting to investors, suggesting that numbers from these companies could be subject to change. The disappointing number for OpenAI came even after the recent addition of ads in ChatGPT, which sounds like the Character.AI pattern.

There’s also no accepted definition of revenue from these companies: Anthropic is including sales by its cloud partners into its own revenue model, but OpenAI isn’t. Given that this is the key figure to determine the viability of the enterprise, this seems like something you’d want to iron out.

OpenAI, for what it’s worth, expects to lose $280 billion between now and 2030, and that’s even after projecting revenue to increase tenfold in that time period. Bain Capital’s analysis, which is likely more conservative than OpenAI’s, shows a $4.2 trillion revenue gap by 2030 across the entire industry. How can that be made up? Integrating ads into chatbots, autonomously operating virtually everything on Earth, using robots for just about every manufacturing process, and coming up with some untold number of new product developments, according to Bain. This is your reminder that 2030 is just over three years away.

By the way, here’s one such new product: OpenAI’s answer to Meta’s Muse companion, which failed almost immediately at the product launch. Muse became a leading app download upon its launch, but it has cooled significantly since then.

Then there is competition from open-source models, which have been consistently only three to six months behind the frontier labs, and can be run on local hardware—a major advantage for any industry with security worries. Open-source competition also weakens the prices AI firms can charge for usage, which was originally their primary strategy to boost revenue.

And the above discussion is focused on the business financials, not the threat to destroy humanity (which is in Anthropic’s IPO prospectus, reportedly) or the industry’s significant legal risk. Eight states—including ones not normally associated with aggressive enforcement like Idaho, Nebraska, West Virginia, and Kansas—just sued OpenAI over releasing ChatGPT models without safeguards to prevent harmful circumstances. The Third Circuit Court of Appeals just rejected a fair use defense for mass AI training on copyrighted material. And the Federal Trade Commission opened an investigation and on Thursday issued subpoenas to OpenAI and Anthropic over safety incidents like hacking into private and government websites.

I have less than no expectations for the FTC probe, considering that the same team ended an OpenAI investigation begun during the Biden administration. But there are legitimate risks to these firms and the AI industry in general, and even if you toss them entirely aside, it’s a struggle to make the economics work.

And to be clear, if the economics don’t work for the top frontier model makers, that trickles down to every company reliant upon them. That explains the sinking of AI stocks across the board in recent days, including a 15 percent drop in one session for “neocloud” company CoreWeave, whose debt load has close to doubled to $50 billion in just one quarter, and whose bonds have been beaten down beyond distressed levels. A neocloud is a data center developer that supplies the raw materials, hosts tenants, and manages the financing; if AI demand is high and development is brisk, these should be soaring. But according to Reuters, default insurance for CoreWeave implies “a nearly 63 percent chance of default in five years,” despite an effective guarantee from Nvidia, one of its biggest investors.

To be clear, CoreWeave doesn’t have some special problem relative to other parts of the industry but a version of the same problem: There isn’t enough revenue to go around to make up for the astronomical expense in maintaining the system. Even at current or increased demand levels, it’s hard to pencil out to profitability. One ominous research note from the firm Gartner this week estimated that 80 percent of neoclouds like CoreWeave could fail by 2030. When you think about the ways in which the major tech firms hide AI debt from their own balance sheets, a fall of the neoclouds would be poised to push upward over time.

Perhaps the biggest indicator of a downturn in the industry is the sudden failure of the IPO for Firmus, a $30 billion Australian neocloud company led by someone who was sent to prison for insider trading (!) but was nonetheless backed by some of the biggest companies in AI. It would have been the largest IPO on Australian markets in 30 years, but it vanished in a matter of days. It used to be that you could just throw the name “AI” onto a company and soar on the exchanges; one neocloud called SB Energy filed for an IPO in August without a single data center in operation. But that boom mentality may not be the reality anymore. The Wall Street Journal reported that a recent panel at a data center development conference had the title “Are We Screwed?”

Sam Altman said in an interview last week that “the world should accept some bad things happening for the benefits of this technology.” It appears that those bad things are already here; we’re still waiting on the benefits.