"Is AI a Bubble?" Is the Wrong Question

From investment committees to CNBC talking heads, YouTube personalities to LP letters, conference panels to the hallway chat between them, and even your aunt who forwards Minion memes: everyone is asking some version of the same question.

"Is AI a bubble?"

It's a reasonable anxiety, given how thoroughly this technology has embedded itself in our daily lives, and it is especially pertinent at the intersection of finance and infrastructure. It is also unanswerable, because the simple question bundles three different risks that share a headline.

First is stock value multiple compression: the repricing of public stocks carrying the AI narrative, the moment P/E ratios decline. Second is capex digestion: the hyperscalers cutting AI infrastructure spending to protect and justify their earnings. Third is a capability plateau, pushing back on logic that launched the current craze, more GPUs equals smarter AI, a narrative whose future depends on that trendline continuing. This trendline, current murmurs say, may be flattening. The popular imagination has all three muddled into one mystical bubble on the precipice of popping. In reality, they are separate.

Each can fire without the others. Each is carried by different assets. And each transmits into the real economy, into power, land, silicon, copper, and concrete, through a different means.

Stock Valuations Tumble: Multiple Compression

Let's start with the stock market, public equities, because that is where much of the anxiety lives. The absolute numbers became extraordinary this decade. Seven US companies, the Mag 7, sit at one-third of the total S&P 500, about $22 trillion of market value.¹ And for two years the group was valued on price-to-earnings (P/E) multiples 35x-50x+, assuming the future would arrive early. If your gut says this is unsustainable, it is. The market this year has agreed with you. This group has underperformed the larger S&P, its June downturn alone erased over $2 trillion of value,² and their valuation premium over the rest of the market has compressed to its lowest level in more than a decade. Nvidia, the Wall Street darling at the center, now sits near 18 times forward earnings against its own long-run average around 36.³

So stocks are down. Why is that a risk to the buildout, rather than just a bad year for anyone holding tech? Because this cycle wired the stock prices into the machine itself, in two ways.

First, the multiple is the permission slip for the spending on datacenter buildout, the capex. Corporate boards commit hundreds of billions to data center capex because shareholders pay for the spend in the share price; every quarter the market rewards the buildout is a quarter the buildout continues. The same report card worked in the other direction in 2022, when the market punished Meta's capex spending: the board cut it within quarters, and the stock's reward for cutting it was one of the largest single-day rallies in its history.⁴ The share price is how owners vote on capex, and corporate boards obey it in both directions.

Second, this public equity is now the deal currency. OpenAI's payment from AMD is warrants over roughly ten percent of AMD, denominated in AMD's stock. Nvidia's commitment into OpenAI is an equity investment.⁵ SoftBank sold its entire Nvidia stake to fund its OpenAI bet and, per recent reporting, borrows against its OpenAI shares to fund its Stargate commitments.⁶ Private lab valuations are marked against public AI comps. When multiples compress, the currency all of this is denominated in shrinks mid-transaction.

The report card flipped six months ago. Earnings kept growing,⁷ and the capex accelerated anyway. We are living in the gap between a compression that has happened and the digestion it has historically preceded. Whether the gap closes with Mag 7 prices and multiples recovering, or with capex being cut, is an open question.

For real assets, there is another distinction matters: infrastructure is exposed to the business, not the stock, unless its financing was built assuming the stock.

How does a public equity repricing actually reach an infrastructure deal?

Not through vibes. Through debt capital markets, where spreads and new-issue windows reprice first and fastest, through infrastructure funds, which run on their own cycles and are often surprisingly detached from public-equity volatility, though a deep enough drawdown drags it in through the denominator effect, and through private credit, its own ball of wax, wired simultaneously to the debt markets and to risk appetite. These channels are slower than the daily stock movements, and they are partially pre-buffered, which is the point: a public equity event has to travel through credit and fund allocation machinery before it touches a single project, and it can arrive diminished or not at all. What it cannot do, by itself, is switch off demand for deployed compute or the electricity that feeds it.

Hyperscale Dollars for Data Centers Dwindling: Capex Digestion

The four largest hyperscalers, Amazon, Microsoft, Google and Meta, are pacing for roughly $700 to 725 billion of combined capital expenditure (capex) this year, up about seventy percent from around $410 billion in 2025, with around three-quarters of it tied to AI infrastructure.⁸⁹ Amazon alone is at nearly $200 billion.¹⁰ To fund the buildout, the group raised over $100 billion of debt last year,¹¹ free cash flow is falling sharply, and Amazon is now projected to go free-cash-flow negative this year.¹² Capital intensity has reached forty to fifty-plus percent of revenue at companies where fifteen would have been notable five years ago.¹³

Every dollar of that spending is a line item defended to boards and public shareholders, and the defense gets harder each quarter the returns stay unproven. The pressure does not require the technology to fail or anyone to stop believing in AI. It requires only math: free cash flow bleeding into construction while the revenue attached to it stays a forecast. Sold against that simple math is a dream: the race to general intelligence, even superintelligence, that the hyperscalers sell to the institutional investors who make up most of their cap table. The dream buys patience that the numbers alone would not. But narrative and arithmetic compound at different rates, and digestion begins as soon as the math outlasts the story. When the patience runs out, capex is the lever that gets pulled, because it is the biggest one and the fastest.

"Sold against that simple math is a dream: the race to general intelligence, even superintelligence, that the hyperscalers sell to the institutional investors who make up most of their cap table."

In early 2025, supply-chain checks by TD Cowen found Microsoft had canceled hundreds of megawatts of US data center leases with private operators, they slowed the conversion of pre-lease agreements (that historically all converted), they paused construction in Wisconsin on a data center earmarked for OpenAI workloads, and they pulled a considerable portion of its international spend back to the US.¹⁴ The shareholders funding today's buildout retain both the power and the incentive to force its pause.

This means multiple compression and capex digestion are separable but not independent: they cause each other, in both directions. Compression leads to digestion when the report card flips and boardrooms obey, as Meta's did. Digestion leads to compression when a cut signals doubt and the market reprices the stock on the news. When TD Cowen reported Microsoft walking away from those leases, the sector traded off on it,¹⁵ and when a Chinese lab suggested frontier AI might need less compute than assumed, Nvidia lost nearly $600 billion in a single session, the largest one-day loss in market history at the time.¹⁶ Some companies will digest before their multiple cracks. Some will be forced to after it does. Lead or lag, the chain is the same.

The media and politically paraded Project Stargate by OpenAI prices their data center at nearly fifty billion dollars per gigawatt of compute all-in.¹⁷ At that price, this year's AI spending is trying to energize on the order of ten to fifteen gigawatts of new load, and the committed deal stack behind OpenAI alone implies roughly twenty-five to thirty gigawatts over its life,¹⁸ rivaling the total peak demand of the entire state of New York.¹⁹ These mind-boggling capex headlines are load forecasts denominated in dollars.

The money is arriving faster than the megawatts.

Interconnection timelines, transformer lead times, and turbine backlogs mean capex converts into operating capacity on the grid's schedule and the equipment queue's schedule, not the board's, and capital parked in half-energized shells earns nothing while the free-cash-flow clock runs.

The workaround of the moment is behind-the-meter thermal generation, and it is not a fix. Gas turbines are sold out years ahead, into 2029 at the largest manufacturer,²⁰ so bolting gas generation onto the site does not escape the procurement queue; it joins a different one. The stopgap version is already visible at scale in Memphis, where xAI surrounded its data centers with dozens of trailer-mounted gas turbines, ran most of them without federal air permits by classifying them as mobile equipment, drew an EPA violation finding and a federal lawsuit from neighboring communities, and then disclosed plans to buy another $2.8 billion of turbines anyway, at least $2 billion of it in the same mobile class at the center of the litigation.²¹ The economics are poor, the emissions land on communities that already carry more than their share, and none of it moves the underlying constraint: the machines that make megawatts are on backorder just like the megawatts themselves.

The longer power delays revenue, the shorter shareholder patience gets. Digestion, when it comes, will arrive not because AI belief failed but because training deployment did: it's a cut made out of necessity. The electricity needs to come from somewhere, or the capital will get cut.

The Trendline Flattens: Capability Plateau

Scaling laws may only go so far. Fourteen years ago, a Toronto graduate student named Alex Krizhevsky, working with Ilya Sutskever and Geoffrey Hinton, trained a neural network on a gaming PC with GPUs and crushed the early AI field's benchmark for image recognition.²² The discovery that graphics chips instead of CPUs made AI dramatically better, and that adding more of them continued the trend, commenced the intelligence and capital explosion that has chased that trendline higher ever since. The last couple of years, fear has been mounting that the predictable "up and to the right" of adding more GPUs for better intelligence is flattening out. It is frontier progress decelerating, the next model arriving late and unimpressive.

But "scaling laws" is not one law. It is four different curves wearing one name. Pre-training, the original: more GPUs and more data into a bigger base model. Post-training: reinforcement learning that teaches a finished model to reason and use tools by trial and error. Test-time compute: the model thinking longer on your specific problem, burning tokens when you ask instead of months before. And agentic scaling: chaining those thinking runs across long tasks, where the model tries a tool, reads the result, and tries again. Four curves. One of them bent.

The one that bent is pre-training, the curve the more-GPUs-equals-smarter-AI story was built on. Ilya Sutskever, one of the industry's pioneers, has said publicly that results from scaling up pre-training have plateaued and that the industry has hit peak data: the internet's supply of high-quality human text is effectively spent.²³ The precision matters, though. The law itself has held across thirteen orders of magnitude of compute, and the researchers closest to it mostly expect it would keep holding.²⁴

When OpenAI shipped GPT-4.5, its biggest base model, smaller models that think longer were already outrunning it at a fraction of the cost.²⁵ The scaling law didn't break; the bill did. And the labs' own budgets moving to the other three curves is the confession.

Those three are early and climbing. The reinforcement learning that turned models into tool users took about five percent of DeepSeek R1's total compute, and grew more than tenfold between OpenAI's o1 and o3:²⁶ Test-time compute is already in the products; the pause before a reasoning model answers is capability being bought per query, and a hard problem can burn a hundred times the tokens of a simple one.²⁷ Agents multiply that again, five to thirty times the tokens per task.²⁸ Notice what these three scaling laws have in common: they spend their compute dollars at serving time, not training time. The industry's answer to the pre-training plateau is to buy more inference.

Honestly, this is the least knowable of the three. The public evidence is genuinely mixed, the labs' own statements are not disinterested, and anyone projecting confidence about frontier capability in either direction is selling something. I don't have privileged insight here, and neither does anyone in the market.

For our decisions, however, we don't need to have an answer. A plateau is not a collapse. A world where frontier capability stops improving is not a world where the technology gets uninstalled. It is a world where attention and capital move from training the next model to deploying the current one. And, counterintuitively, a plateau fixes a blocker to enterprise adoption, which is the fear that whatever you deploy this year is obsolete next year. Boring technology is deployable technology. A capability plateau is a tailwind for deployment.

Holding these three side by side looking at the past year, each risk is visibly rearing its head: multiples compressing, capex accelerating, some players fumbling, and the capability question still unanswered while caution grows. The past seven months have demonstrated that the bubble is three.

The loop that connects them

So what actually does connect them?

Lab valuations (OpenAI, Anthropic, xAI), enable fundraising. Fundraising enables compute commitments. Compute commitments justify hyperscaler capex. That capex becomes Nvidia's revenue. Nvidia's revenue sustains the S&P index concentration. The S&P index sustains the equity environment in which lab valuations are set, around and around.

The Structural Weakness

It sits at one node: the hyperscalers can broadly fund the buildout of datacenters from their massive operating cash flow, and the labs cannot, not close.

One analysis of OpenAI's 2026 cost base found internal revenue covering roughly 47 percent of operating expenses, with vendor financing and external capital covering the rest.²⁹ The labs are the node that requires continuous external belief.

Training versus Inference: The Distinction that Matters

AI capex is not one thing. Frontier training capex is a handful of labs and hyperscalers making the same bet at the same time: that the next order of magnitude of compute produces the next order of magnitude of capability in the race for general intelligence or superintelligence (AGI/ASI), and that someone else will pay for it later. It is concentrated, correlated, and it lives inside the loop above.

Inference and deployment capex are the opposite shape. Inference capex is a long tail of enterprises paying to run capability that already exists, inside products and workflows that already have budgets. Distributed, diversified, and paid for by customers outside the circle.

And the inference bill is already the bigger line item. Across the fleets serving the world, inference took over. Inference crossed half of all AI compute last year and is heading for two-thirds this year, up from one-third in 2023.³⁰ The split that matters for capital is who pays. The experiments are paid for by the loop above. The bill is paid by customers outside it.

Training demand is a bet on the future. Inference demand is a bill for the present.

Let's put the three risks through that split.

• Multiples compress: the training bets get rationed first, because they are the discretionary experiments, while deployment continues wherever the unit economics already clear, because turning it off means turning off revenue.

• Capex digests: the speculative clusters pause, as Meta's did in 2022 and Microsoft's did last year, while the inference fleet serving paying workloads keeps its budget, because it is the budget.

• Capability plateaus: deployment accelerates rather than stalls as the technology stabilizes and globally companies look to implement the proven technology.

The popular intuition says Pandora's box is open, the technology is irreversibly developed, and there is no going back. That intuition is right. But it is a claim about inference. The irreversibility lives in the millions of workflows that now assume the capability exists, not in the training runs that produced it. Deployment is the ratchet. The frontier is the gamble.

This doesn't require optimism about any particular company, model, or valuation. Only the observation that the resilient half of the demand and the fragile half currently trade, get financed, and get built against as if they were the same thing. They are not, and the gap between them is where the next several years of infrastructure capital will be made and lost.

The Questions Worth Asking

Whether AI is a bubble was never answerable. A useful question is:

"Which kind of AI capex are you exposed to?"

For anyone investing or lending to real assets, the discipline is simple. Exposure tied to inference survives the three risks. Exposure tied to training survives none of them.

Most portfolios currently hold both without knowing the ratio, because both were bought under one word, 'AI'.

The real second question is hiding underneath. Someone still has to serve this demand with electrons, and the system that is supposed to do that, the grid, has a timeline problem that the demand does not respect.

"Where is the electricity going to come from?"