Why the AI boom is hitting a 'reset wall'
If I stumbled upon a magic lamp, a genie appeared, and he offered to answer three questions, somewhere near the top of my list would be: "How will the AI-infrastructure boom play out? Is it a bubble? If so, when will it burst?"
As I've written many times, I do think it's a bubble. But I don't say this with a high degree of certainty – there's a wide range of possible outcomes.
And I have even less certainty about the timing. There were many smart people who correctly identified the Internet and housing bubbles, but got destroyed shorting them because they were a year or two early.
This topic was on my mind over the weekend because I came across this article on a blog called Groundbreaker: The Teaser Period: Why the AI Boom Is Hitting a Reset Wall.
In it, the anonymous author compares the AI-infrastructure boom to the subprime mortgage market in 2006. At that time:
Nothing looked wrong in the summer of 2006. Home prices had risen for the better part of a decade. Delinquencies were near historic lows. Credit spreads were tight, the ratings held, and the securitization machine hummed. If you had asked a hundred people on a trading desk whether the American mortgage market was months from seizing, most would have laughed.
Millions of subprime borrowers were, at that moment, paying the low introductory rate on a two-year adjustable rate mortgage – the 2/28 ARM [adjustable-rate mortgage]. A low fixed-rate for two years, then the rate reset to a payment 30% to 50% higher. During those first two years the loan performed beautifully: the borrower paid, the servicer collected, and the bond paid its coupon. Nothing looked wrong because the whole complex – housing, mortgages, securitization – was sitting inside the teaser period.
However, a "reset wall" loomed:
Roughly a trillion dollars of adjustable-rate mortgages were contractually set to reset across 2007 and 2008 – thirty to forty billion dollars a month at the peak. Credit Suisse published the chart in March 2007. The [International Monetary Fund] reprinted it. It circulated on every trading floor in New York and London.
Despite the fact that "every reset on that chart was contractually inevitable from the moment of origination," nearly all market participants ignored it because:
... every participant believed the exit would arrive before the reset: home prices would keep appreciating and the borrower would refinance into a fresh teaser before the old one expired.
I saw what was happening and did my best to warn people – such as in the Emmy-winning 60 Minutes episode that aired on December 14, 2008 (transcript here and video here).
But everyone was making so much money that a handful of other canaries in the coal mine (Mike Burry, Steve Eisman, John Paulson) and I were ignored.
The market's failure to anticipate the consequences of the reset wall led to the biggest market crash since the Great Depression.
Groundbreaker makes a compelling argument that similar dynamics are in play today:
The AI boom has rebuilt this exact structure, and the market is once again underwriting the teaser.
It has a reset wall of its own – a schedule of dated, contractual, non-negotiable payment shocks – hiding inside the trillions of dollars of compute contracts signed by OpenAI and other frontier labs since 2024.
The take-or-pay compute contract – the instrument at the center of the AI build-out – has a structural feature that almost no one prices: its payments do not begin at signing. They begin at delivery. A lab signs a multi-year capacity commitment today, but the payments do not start until the data center is energized, the capacity is accepted, and the contractual ramp schedule commences – an interval set not by finance, but by construction: siting, powering, and filling a gigawatt-scale campus takes 24-to-36 months from signature – mirroring the two-to-three-year teaser of a subprime ARM.
More than $2.3 trillion of compute contracts now sit on the books of the four largest American cloud providers as remaining performance obligations and contracted backlog – signed, celebrated, capitalized into equity prices, and, critically, not yet billing.
As with the subprime bubble:
During the teaser period, everyone wins. The seller reports backlog growth that compounds at rates no operating business has ever sustained – Oracle's [remaining performance obligations] grew 363% in a single fiscal year. The buyer – a frontier lab burning cash at historic rates – books no expense because the capacity does not yet exist. The market capitalizes the booked number as if it were revenue and ignores the billed number as if it were a technicality. And then, on a schedule fixed at signing, booked compute becomes billed compute. The take-or-pay clock starts. From that day forward, the frontier labs and the hyperscalers incur those costs regardless of utilization. The invoice is a function of the contract, not of demand. That is the reset.
The parallel to 2006 is exact and it explains the single most-cited absurdity of this cycle: How does OpenAI, a company with some $40 billion of run-rate revenue, sign $1.4 trillion of compute commitments? The same way a household with $60,000 of income signed a $600,000 mortgage: because the terms at signing do not require the payment yet, and because everyone at the table – borrower, lender, and the market – believes the growth will arrive before the payment does.
The 2/28 borrower's defense was always the same: by the time the reset arrives, my house will be worth more and I will refinance. The frontier lab's defense is structurally identical: by the time the capacity commences, my revenue will have grown into the obligation.
Interestingly, the size of the "compute commencement wall" today is very similar to the subprime one in 2006:
The compute commencement wall can be made visible in exactly the way the reset wall was visible in 2007 – from disclosed contracts and delivery schedules. The only question is whether the market listens this time.
The looming wall doesn't necessarily mean that the AI-infrastructure boom is a bubble and will bust. As Groundbreaker acknowledges, it's possible that demand for the compute being built will grow so rapidly – and pricing will remain strong enough – that profits will cover the performance obligations. But Groundbreaker doesn't think it's very likely:
For committed compute merely to equal revenue in 2027 – not to be comfortably covered, simply to reach parity, before a single dollar is spent on wages, research, sales, or tax – revenue would have to compound at 217% annually off the 2025 base... [No] company at this scale of revenue has ever done [this] for a multi-year stretch. The obligation is accelerating at more than double the rate of the best case for the cash flow meant to cover it...
You do not need demand to fall. You need it only to decelerate below the rate the booked compute was underwritten to.
There's another way companies like OpenAI can avoid a bust: They can continue raising huge amounts of money to cover their cash burn. But Groundbreaker doesn't think this is likely either:
The thesis for OpenAI requires capital markets to fund roughly half a trillion dollars of cumulative operating deficit at a single pre-profit counterparty, at non-punitive terms, through a window in which that counterparty's compute costs are contractually rising faster than any plausible revenue path.
That may happen. But it should be named for what it is: a refinancing assumption rather than an operating plan, and one that depends on the collateral – the valuation – exactly in the period in which the true cash cost of the build becomes visible for the first time.
Groundbreaker concludes:
The bull case wins if – and it is a real if – demand scales into the committed supply before the reset wall lands, and the counterparties stay funded through any air pocket in between. The bear case in this piece is not that artificial intelligence will fail, or that the demand is fake, or that the technology disappoints. It is narrower: that the financing structure can break before the demand arrives, because the obligations are fixed and front-loaded in commencement while the revenue is variable and back-loaded in adoption – and a fixed obligation meeting a lagging revenue stream is a solvency problem regardless of how transformative the underlying technology turns out to be.
The industry will spend the next eighteen months debating whether artificial intelligence is a bubble, which is the wrong question, asked at the wrong layer. The technology is real; so were the houses. The question is narrower: what happens when instruments underwritten at the teaser meet their reset schedule, and who is holding the paper when the obligations cannot be met as written. The reset wall is published above and the AI boom sits in a period of fiction.
Groundbreaker's essay triggered a great deal of discussion in a WhatsApp group I'm part of. Here was my initial comment:
This is a must-read. I was wrong: The AI bubble doesn't most closely resemble the Internet bubble, but rather the housing bubble. OpenAI will likely never go public and end up a zero (for the equity).
One of my friends responded:
While the remaining performance obligations of the hyperscalers are dominated by the frontier labs, look through that. What is the end customer paying for? Inference. The end customer wants intelligent tokens.
If the frontier lab were to disappear, the demand would still be there. It just wouldn't be routed through OpenAI or Anthropic!
As I replied, isn't that like saying: "Who cares if WorldCom or Global Crossing goes bankrupt because if the demand for Internet bandwidth is sufficiently great, then someone will pay – and someone will profit?"
That's Groundbreaker's entire point: If AI demand rises at around 200%-plus per year and pricing remains strong, then all is well – it's not a bubble, and everybody wins.
But this sure reminds me a lot of projections for the Internet – which was one of the most important, transformative technologies of all time. However, that didn't prevent the mother of all bubbles – and busts.
My friend replied:
The overbuild argument is more persuasive, I agree. That is a better analogy than housing.
Clearly, to me, current demand is real and insatiable. But we very well could overbuild capacity if models become way more efficient and there's more inference at the edge and adoption is slower.
The dark fiber problem happened because the bubble burst in March 2000 before the Internet had a chance to diffuse to large scale use cases in enterprises and for consumers. Cloud didn't appear until 2006. Mobile didn't take off until the iPhone. Facebook took off in the late 2000s and early 2010s. YouTube was tiny in 2006, etc. So you couldn't have demand for bandwidth.
However, now we have the rails (aka the distribution): enterprise software building in AI, AI on consumer phones and PCs, and the Internet is built out.
So then the question is, can useful applications that demand a lot of AI compute be created such that we don't run into an overbuild situation?
So far it seems the answer is an emphatic "yes."
I responded that I think the answer is an emphatic "nobody knows" – because it's unknowable with any degree of certainty.
My friend's Internet examples above are correct. They were six to 10 years after the peak of the Internet bubble and four to eight years after the crash.
Whereas, as Groundbreaker notes, the entire AI ecosystem is going to hit a cash wall in less than two years. It's not at all clear that revenue growth (while maintaining pricing) will be sufficient to support trillions of dollars in investments.
In fact, every prior investment boom (all much smaller) has ended in a catastrophic bursting of the bubble. Maybe this time will be different. But I don't think it's likely...
That said, while investors will get burned, consumers will be huge beneficiaries.
I consider myself an early tech adopter. My entire life, I've always had to have the latest gadget, I go to the Consumer Electronics Show every year for fun, etc.
And I find AI to be a very useful tool – mostly via better search results from Google Gemini (I usually just type a query into my Google search bar, as I've done for decades). I also pay $20 a month for Anthropic's Claude, use it roughly once a day, and have found more uses for it over time.
But if you told me the government was going to ban AI and I'd have to go back to how it was before, I would only be mildly annoyed.
My WhatsApp group friend, on the other hand, is an even earlier tech adopter. He's clearly incorporating AI into his professional and personal life in a huge, game-changing way. I think this is great, and I'm curious to learn from his experience.
So here's the big question that will determine if this is a bubble that's going to burst in an ugly way (which nobody knows the answer to): In short order, is every person and business going to be like my friend – or like me?
Best regards,
Whitney
P.S. I welcome your feedback – send me an e-mail by clicking here.