The B-word is on everyone's lips: Are we in a bubble? As there have been in recent months, there were all kinds of signs this past week that investors are becoming less enthusiastic about the generational investments into artificial-intelligence data centers. And while the AI bubble doesn't look ready to pop yet, 2027 is shaping up as a make-or-break year for the sector.
CoreWeave, which invented the "neocloud" business model -- speed running to becoming a hyperscaler -- has become an AI battleground stock, especially this past week. The company is built on a cycle of debt, capex, and very rapid sales growth, with a foundation of a nearly $130 billion backlog. It's a tightrope walk across a canyon. If it keeps moving forward, there's a pot of gold on the other side, but if it stops for a second, it could all come crashing down.
Since the market close on Oct. 6, the PHLX Semiconductor index has been down seven days in a row, falling 12% all told. CoreWeave has seen its stock decline by 24% in that period. Its publicly held debt has sold off, trading at roughly 82 cents on the dollar for its 2032 senior notes. The yield to maturity is 14.4%, up from 9.6% at offering in June.
In another blow to sentiment, last week there was a kerfuffle over AI-model leader OpenAI's revenue, which was either at a $70 billion annual run rate in September or $50 billion, depending on whether gross or net sales were used for the calculation. Then, this week, Anthropic released quarterly numbers at a private investor day that failed to quell the market's concerns.
OpenAI had already decided to postpone its initial public offering while it works out the security issues that its new models are causing. It isn't alone: Australian AI cloud provider Firmus decided to delay its IPO after investors balked at the proposed $30 billion valuation, up from $11 billion in August. Anthropic, however, despite facing security issues of its own, is charging ahead with a fall offering.
But it's OpenAI's long-term commitments of over a trillion dollars -- including large cloud contracts with Oracle, Microsoft, CoreWeave, and Cerebras -- that remain central to everyone's worries. Many wonder if it can earn enough revenue to pay those bills, which only grow over the next five years. The bigger risk for OpenAI isn't the AI investment boom rolling over, but rather cheaper competition eating into its gross profit.
And rising long-term interest rates are a developing issue for the cloud business, which is very dependent on debt.
Investor sentiment took a downturn in September. But back on the ground, it's still full steam ahead. My No. 1 indicator is graphics processing unit rental prices, and according to data from Silicon Data and Ornn, they have been rising all year. CoreWeave has even rented out its Nvidia A100 GPUs, first released in 2020, through 2029. Even as the so-called hyperscalers add hundreds of billions of dollars worth of new capacity, demand growth continues to outpace supply.
This is being driven by several concurrent factors. In the first place, AI models took a big leap in capabilities this year, and they are finally ready for real business work, beginning with writing software code. The focus is around agents -- software that can use AI models to accomplish a complex series of tasks from simple conversational commands. Agents use AI models at a pace that no human could possibly match, and it has led to a demand explosion.
In the U.S. Census Bureau's Business Trends and Outlook Survey, carried out in September and October, only 28% of companies polled said that they would be using AI in the next six months. But that is up from 21% at the end of last year, and leaves a lot more room for growth.
Cheap Chinese models are putting pressure on OpenAI and Anthropic, and both companies have instituted price cuts in the past two months. Lower prices typically lead to more demand. Meta Platforms' new Wonder models put the company in the same category, and now there are three frontier AI labs instead of two, adding more pressure.
One reason the models are getting so much better this year is because the GPU clusters they are being built on are getting massive. Meta Wonder was trained at Meta's sprawling Prometheus data center. OpenAI's best model, Astra, was trained on 100,000 Nvidia GPUs at Oracle's Abilene, Texas, data center. These clusters will have to get larger with each generation.
The biggest near-term danger to the AI boom isn't demand; it's capital and politics. Nvidia, Broadcom, Microsoft, Cisco, and others have worked with lenders like BlackRock and Apollo Global Management to create investment vehicles totaling hundreds of billions of dollars to fund AI infrastructure. But as the price of CoreWeave debt shows, some lenders are getting more selective, while others are looking for the exit.
There is also growing local political opposition to data centers, one of the few issues that can cross party lines in the U.S. Data centers have already been canceled, delayed, or downsized, and state and local regulation may follow.
The classic sign of a bubble is concentrated investment in a single asset class, in this case AI data centers. In the past year, domestic investment into data center buildings, computers, networking equipment, software, and research accounted for nearly all the growth in U.S. fixed investment. Just five companies are spending up to $750 billion on capital expenditures this year, with likely more coming in 2028.
For the investment boom to continue, that has to change next year. Meanwhile, investors should remember that the late stages of stock market bubbles can be the most profitable. In 1999, before the implosion of the dot-com bubble the following year, the Nasdaq Composite index was up 86%, after nearly tripling over the previous four years.
Until it's your enemy, the bubble is your ally.
Write to adam.levine@barrons.com
Comments