AI is working great. And yet, the AI bubble will burst.
The technology is real, and the demand for computing power is also very real. Businesses are spending heavily on AI, consumers are using it at unprecedented scale, and the capabilities of the models continue to improve at a dizzying pace. All good on that front.
The problem though, is the financial structure being built around that demand. It’s showing up in two main ways.
The first is capital. The same handful of companies increasingly sit on both sides of the industry’s biggest transactions: investor, supplier and customer at the same time. Nvidia has committed up to $100 billion to OpenAI while supplying the chips that power its models, and holds a stake in CoreWeave, which sells that same computing power back to both OpenAI and Microsoft — while Microsoft is one of OpenAI’s largest shareholders and one of Nvidia’s biggest customers.
The list goes on and on, but I think you get the gist.
The second is the product itself. ChatGPT, Claude, Gemini and their competitors are becoming dramatically cheaper to use with every new generation, while open-source models are closing the gap with the frontier labs. Capability is rising, but the price of intelligence is falling.
That combination creates an awkward tension (for the frontier labs at least). The industry is spending extraordinary amounts of money building infrastructure for a product whose marginal price is moving steadily towards zero. Will the likes of Anthropic and OpenAI eventually struggle to sell their models at a high enough price to justify current expenditure levels?
Don’t get me wrong, the demand for computing power is undeniable, the technology is improving rapidly, and many of the applications being built on top of it will be enormously valuable. There is no question about that. But the financing structure underneath the boom is becoming increasingly worrisome.
The bubble, in our view, is not AI itself. It is the assumption that the current economics of AI can continue indefinitely.
Here’s the case for why we think that assumption breaks and why it doesn’t change how we invest at Momentous.
The self-feeding loop
The circular spending which is increasingly being covered publicly is built on a very narrow customer base.
That structure can work totally fine while growth remains strong. The problem is what happens if and when one of the central customers slows down.
OpenAI and Anthropic are expected to account for an increasingly large share of cloud revenue, despite neither company being consistently profitable. UBS estimates the pair will make up 27% of Google Cloud’s revenue this year, rising above 48% next year. Microsoft’s cloud division, meanwhile, has already become heavily exposed to OpenAI. In fact, 69% of its growth in 2025 came from that one customer alone, and growth would have been a meager 8% without it. And Google has built a very similar relationship with Anthropic. In effect, some of the world’s largest technology companies are relying on a small number of AI labs to justify an enormous amount of infrastructure spending.
Let’s play that forward. If OpenAI ever pulled back (i.e. scaled down its compute spending, lost ground to a rival, or simply grew slower than forecast) Google Cloud and Microsoft Azure would see a meaningful chunk of their revenues evaporate. Public markets, which price the stocks on the assumption that growth continues at pace, would see a considerable correction and broader market sell-off on fears of a bubble.
Under scrutiny, the numbers quickly become difficult to reconcile. One widely cited estimate put the data-centre capacity now being built at around 190 gigawatts, requiring something like $1.6 trillion in annual revenue to justify the investment. These figures aren’t audited, but they give you an idea of the scale we’re talking about (and betting on). OpenAI itself reportedly posted $13.1 billion of revenue in 2025 against a $38.5 billion net loss, while committing hundreds of billions of dollars to future computing capacity.
And much of the financing isn’t immediately visible. A lot of it runs through leases and separate entities that build and own data centres on a hyperscaler’s behalf — arrangements that work more like renting than borrowing, so the obligation is just as real but doesn’t always show up as debt on the books. Add bonds, leases and this kind of hidden borrowing together, and the industry’s total obligations come out to somewhere near $1 trillion, per reporting from Nikkei Asia and Moody’s — a meaningful chunk of it ultimately funded by pension and insurance money, in what the Bank for International Settlements calls “shadow borrowing.”
This doesn’t prove that the spending is wasteful by any means. There are very legitimate reasons to believe demand for compute will continue to grow. There’s also a structural advantage specific to the hyperscalers that doesn’t apply to OpenAI or Anthropic: Google and Microsoft aren’t betting the house on AI working as a standalone product. Rather, they’re layering it onto businesses that are already enormous and already profitable. In 2025, Google and Microsoft posted net profits of $132B and $102B respectively. Google’s roughly 90% share of global search isn’t going anywhere either; Gemini and AI Overviews are being used to cleverly defend that position. If the AI trade corrects, both companies still own the businesses AI was built on top of.
Simply put, the point is that the financing structure leaves very little room for error. When that error eventually comes (in the shape of something subtle like major AI labs missing expectations, or cutting capex), investors will question the assumptions supporting the companies around it. And spoiler alert… the losers won’t be the hyperscalers (Amazon, Google, etc.).
The model layer has no moat
This becomes more interesting when you look at what is happening to the economics of the models themselves.
The best AI models used to come in one flavour: closed. You paid a company like OpenAI or Anthropic for access, and the model itself stayed private (centralised and controlled by the company that owns it). Increasingly, that’s no longer the only option. “Open-weight” models — versions anyone can download, run on their own servers, and modify freely, from labs like China’s DeepSeek and Moonshot, or tech companies like Meta and Nvidia — have closed the performance gap dramatically, especially on the tasks businesses actually pay for, such as coding. For many buyers, whether a model is “open” or “closed” now matters less than whether it’s good enough and cheap enough. Add the appeal of running your own data privately and securely on your own servers (rather than sending it to someone else’s through the cloud), and open-source starts looking like the more attractive option. Particularly for businesses.
When models become interchangeable on quality and performance, price starts to matter more, and the price gap is enormous as it stands. AI companies charge based on how much text a model reads and writes, measured in “tokens” (roughly chunks of a word). DeepSeek’s models cost as little as $0.28 to $0.87 for every million words of output; the best Western models charge $15 to $25 for the same amount — 20 to 30 times more, for performance that Epoch AI, an independent AI research group, puts only about four months behind. For the vast majority of use cases, that remaining gap doesn’t really matter though — it’s only relevant at the frontier, for things like drug discovery or advanced cybersecurity research that 99% of people and businesses will never ever touch. For everything else, the models are already interchangeable, and price will decide the winner.
Across the board though, the technology keeps improving fast, and every generation makes it cheaper to actually run these models day to day. More competitors enter the market. Open-source models get easier to deploy. Hardware improves. And the result is that the cost of intelligence keeps falling.
That is fantastic for businesses building on top of AI. It is much less attractive for the companies whose valuations depend on maintaining pricing power at the model layer (i.e. Anthropic, OpenAI).
The AI trade is often compared against the dot-com boom or even cloud computing. But there is a more useful historical analogy here: Mobile data. A sharp point Benedict Evans made well in this piece.
TL;DR: As cellular networks improved, the amount of data people consumed exploded. The underlying network became enormously important, but much of the economic value accrued to the businesses built on top of it (Instagram, Uber, Netflix, …) rather than to the network itself. Telcos built the infrastructure and captured comparatively little of the upside; their stocks have been notoriously flat for two decades. The network simply became infrastructure.
AI could follow a similar trajectory where intelligence becomes abundant and cheap, while the value accrues to the software layer (specialised workflows, proprietary data sets, vertical business solutions). A handful of companies are currently aggressively financing the infrastructure before we know where the value will settle.
Public and private markets running one bet
The Magnificent Seven — the seven biggest US tech companies, including Nvidia, Microsoft, Google and Amazon — now make up roughly a third of the value of the entire S&P 500, the index tracking America’s 500 largest public companies. Nvidia alone accounts for around 7.5% of it. AI-related investment accounted for the large majority of US GDP growth in the first quarter of 2026, while spending on data centres, hardware and networking reached around 1.4% of GDP. The productivity gains that are supposed to justify this haven’t shown up at anything close to that scale, and companies are massively exceeding their AI budgets rather than saving on them.
Public markets are also suddenly being asked to absorb an unusual amount of new AI-linked equity at once. SpaceX’s IPO — the largest in history, at a $1.77 trillion valuation — will likely be followed by even larger Anthropic and OpenAI listings, all while the Mag 7 simultaneously ask investors to finance their own buildout. Everyone is drawing from the same pool of capital, which is okay while expectations stay high but could lead to a steep correction once that stops.
South Korea just showed us a miniature version of what that might look like. Samsung and SK Hynix, the country’s two big AI chip suppliers, came to make up more than 60% of the Kospi (South Korea’s main stock market index). Leveraged retail money — ordinary investors borrowing to buy more than they could otherwise afford — continued to pile in, and when US chip stocks wobbled in late June (purely on investor sentiment, nothing more), the index shed 16% in two days…
Private markets carry the same concentration. AI accounted for 61% of global venture funding in 2025, rising to 79% by Q1 2026, with three deals — OpenAI, Anthropic and xAI — taking two-thirds of it. At the seed stage, AI companies commanded a 42% valuation premium over non-AI peers, even as deal counts fell and capital concentrated with a smaller group of brand-name managers.
When the top of the market reprices, it resets the comparable valuation every investor uses underwriting a seed or Series A round — whether or not that company has anything to do with AI.
That’s why we think a correction is due. Even though the demand is very real, the financing structure around it has become too dependent on a small number of companies continuing to grow at extraordinary rates, and we don’t need to know the exact trigger to see how little room for error that leaves.
And that brings us to the part of the market where we think the implications are most interesting.
What doesn’t compress
The Human Potential Stack sits outside this particular loop. Momentous’ performance, wellness and longevity sectors are not priced on the assumption that AI infrastructure spending needs to compound indefinitely, or that frontier models need to command a price premium. In fact, they benefit from AI becoming cheaper.
As digital intelligence becomes abundant, anything that cannot be digitised becomes relatively more scarce. Sport is the clearest example. The LA Lakers recently sold to Bob Iger and Josh Kushner at a record $12.5 billion valuation — barely a year after Mark Walter bought control at $10 billion. Meanwhile, Liverpool has just sold about 30% of the club to the 1892 Holdings consortium at a record $7B+ valuation, with Jeff Bezos as lead investor. Investors haven’t suddenly woken up and decided to pay that price because these sports are becoming more technologically sophisticated. They’re paying for something technology cannot reproduce.
You cannot download a live sporting event, or automate the experience of being in a stadium with thousands of other people. Technology can make digital content abundant, but it cannot make scarce physical experiences abundant. And we believe AI will only increase the value of that scarcity.
The same is true for wellness and longevity, arguably with an even more durable underlying demand driver. The global wellness economy reached $6.8 trillion in 2024 and is forecast to approach $9.8 trillion by 2029. These markets are driven by demographics and behaviour (growth that is completely separate from a hyperscaler’s next earnings call). People want to perform better. They want to feel better. They want to live longer. And they want to do it with other people, in the real world.
That said, AI will still play an enormous role here as the underlying technology, improving diagnostics, personalising interventions, automating monitoring and accelerating discovery. Businesses operating in health and sports will use cheaper intelligence as an input to drive efficiency gains, while selling something much harder to commoditise.
AI is incredibly useful. Clearly. The question is whether a small group of companies selling compute to each other can continue growing fast enough to justify the enormous amount of infrastructure being built around them. We don’t think they can forever, and for Momentous, that is not a problem we need to solve.
Our portfolio is built around the stadium seat, the healthier body, the stronger performance, the extra year of life. In other words, the things that people actually get to experience and enjoy.
As artificial intelligence becomes cheaper, being human becomes more valuable.