Just before the year turned, Japan-based holding company SoftBank Group completed a $41 billion investment in generative AI pioneer OpenAI, boosting the ChatGPT creator’s market valuation to $500 billion.
The deal, which closed at the end of December, was the multinational conglomerate’s fourth multi-billion-dollar investment in OpenAI. Just two weeks later, OpenAI contracted with AI chipmaker Cerebras for $10 billion in compute power to fuel the inference and performance of OpenAI’s generative AI models.
All of that activity came after a dizzying year in which OpenAI figured prominently in five of the 10 biggest deals in the seemingly booming AI industry: the $500 billion Stargate project with SoftBank and software-hardware giant Oracle; a separate $300 billion long-term deal with Oracle; a $100 billion investment in OpenAI by AI chip market leader Nvidia; and a partnership between OpenAI and AWS valued at about $38 billion. Meanwhile, tech vendors and investors spent more than $61 billion last year to build hundreds of AI data centers around the world to deliver the compute to run AI models now and into the future.
OpenAI Seen as Vulnerable
And yet, despite confidence in its future displayed by some of the biggest players in tech, countless critics say OpenAI is bound to fail.
Notwithstanding its grand ambitions, the former nonprofit research lab’s expenditures dwarf its revenue — estimated at about $20 billion in 2025 by OpenAI’s CFO. In the meantime, as OpenAI struggled to make serious inroads into the enterprise AI market, its rival, Anthropic, was becoming the generative AI vendor of choice for business users and coders, though it, too, has registered underwhelming revenue. OpenAI continues to subsidize most of its operations with private investment subsidies.
“The amount of money that is needed to justify the fixed costs that are being incurred downstream is because the various AI companies do not have a business model today and do not have a cost structure today that economically works,” said Andy Wu, associate professor of business administration at Harvard Business School. “So across the board, most AI companies lose enormous amounts of money.”
“The problem is the commitments for infrastructure by all companies combined are on the order of $1 trillion,” said Johna Till Johnson, CEO of research and advisory firm Nemertes. “So, you would need to pay about $800 billion per year just to cover the interest on your loan to get your infrastructure. That’s not going to work.”
Visions of Collapse
In the eyes of AI skeptics, OpenAI’s collapse would trigger or contribute to a massive bursting of what they see as a fragile AI bubble, a term that has become part of the everyday AI vernacular, even amid the outward signs of an AI boom.
That bubble has been created by wildly over-anticipated demand for AI services, according to the AI skeptics. Add to that circular financing in which vendors and customers exchange money, overstated revenue, overvalued tech stocks, unaccounted for depreciation of AI chips and unsustainable debt incurred by big AI infrastructure spenders, notably Oracle and neocloud GPU-as-a-service vendors such as CoreWeave.
In that scenario, the bursting of the AI bubble will throw financial markets into chaos and devastate market values, with effects cascading throughout tech and society. The bubble hasn’t burst yet. Indeed, AI vendors keep spending and borrowing to build infrastructure and create new, bigger and smarter generative AI models and agentic AI systems that are on the rise in the enterprise world.
AI Boom Outlook
Those who see AI as a transformational technology akin to the electric grid and the internet tend to shoot down most arguments about an AI bubble.
They argue that, unlike the dot-com bubble of the late 1990s, most of the big AI vendors, particularly the hyperscalers, are in sound financial condition with multiple revenue streams beyond AI.
As for Nvidia — which, along with OpenAI, is at the center of the financing circle — it has cash to burn as the world’s most valuable company and has been smartly seeding its ecosystem by investing in customers, they said. Some point out that Amazon ran similar losses in its early years, but withstood shareholder demands for profits as it built its customer base and demand.
“We’re saying this is not a bubble. That doesn’t mean to say we don’t think that there are any areas of hype,” said Alison Porter, portfolio manager at Janus Henderson Investors, who runs several funds heavily invested in AI. “We’re moving from a position of demand being driven by this generative AI with large language models and just asking questions of ChatGPT to wider spread industry use of agentic AI. And when we move to agentic AI, we see this exponential increase in the number of tokens and the amount of compute supply needed.”
Other drivers of demand for AI beyond consumer use are automating and making corporate processes more efficient, including industrial and physical AI in manufacturing, and scientific and drug discovery and medical research, according to those who see an AI boom rather than a bubble.
As for Oracle and its commitment to build AI data centers to provide OpenAI with $300 billion worth of compute over about five years, Porter said the deal involves risk — with Oracle expected to borrow $100 billion to make the deal work — but won’t bring down the whole AI sector if it doesn’t succeed. “Is there incremental risk in Oracle? Yes. Is that systemic risk? No,” she said. While Oracle has a strong software and database business, it’s not nearly as diversified as the other tech giants, many have noted.
Porter said she sees the same dynamic at work with the neocloud vendors, some of which are heavily weighted with debt as they build data centers to meet anticipated demand. “Is there a risk in the stock prices of many of those new clouds? Yes. Does that mean there’s a systemic risk to overall AI development and capex? No,” she said.
Porter also sees a runway for OpenAI and Anthropic to meet revenue goals, noting that both vendors have made progress over the last three years. Although are still far from profitable, OpenAI went from $2 billion in revenue in 2023 to $20 billion in 2025, a 1,000% increase. Anthropic’s revenue jumped from less than $1 billion in 2023 to $5 to 7 billion in 2025, also a steep growth curve.
“But they are burning cash at a rapid pace and are reliant on large capital infusions. We see a tension between real technology breakthroughs and skyrocketing valuations,” said Kashyap Kompella, CEO of RPA2AI Research. “AI valuations are in speculative territory which only the leading companies may be able to sustain in the long term.”
Much of the bubble-boom debate hinges on how much enterprise AI demand exists and whether businesses are seeing an ROI in what is still an early implementation phase of generative and agentic AI.
A recent survey of 350 mid-sized to large enterprises by Omdia, a division of Informa TechTarget, found that organizations are starting to see that ROI. The vast majority of companies surveyed in October reported “very good” to “extraordinary” returns on their AI investments, said Mark Beccue, an Omdia analyst.
“Everybody talks about a bubble,” Beccue said, referring to an influential MIT study in August that essentially said the AI boom was pure hype and most organizations were seeing zero ROI from generative AI technology. “Our data begs to differ.”
Amid the debate, some on both sides see AI at a stage in which elements of both outlooks contribute to a kind of middle ground as the torrid rate of investment and R&D in generative and agentic AI technology continues unabated, at the same time as fear of a crash appears more acute and widespread than ever.
“There is a duality to the current AI wave,” Kompella said. “It exhibits both boom and bubble dynamics. The bubble elements — such as the sky-high startup valuations, the fact that some AI investment demand is fueled by the investment money itself in a circular loop and the signs of adoption friction suggest that a correction may happen.”
“Not all players will survive the hype, and some capital will be misallocated,” he added.
