August 2026
Disclaimers and Disclosure:
First off, I am no luddite. I use AI as part of my research process, as well as to assist with writing and editing. The image in the header was generated using Grok. I am very in favor of technological progress. This piece is a thought experiment on where the AI Capex boom may run into challenges ahead. I am optimistic on the prospects for AI over the coming decades, but fear too much may have been bet on large incremental progress which may prove elusive in the short-run. My principal concern is that too much of the American economy has been leveraged based on a nascent technology with unclear long-run economics, and that this may ultimately harm the main street economy and the average citizen via the stock market and possibly issues at large financial services firms. As of this writing, I am short Blackstone (BX) and SharonAI (SHAZ) via puts. I intend to continuously transact in these securities based on my investment objectives and risk management principles. Any changes to my positioning do not necessarily reflect a change in our thoughts laid out herein.
The Data Center Dustbowl
The year is 2028. Outside Abilene, Texas, the first families are moving into a converted data center. The building went up in 2026 as a powered shell for Nvidia GPU’s. It never ran above 50% capacity. After the Great Chip Depression of 2027, it sat dark for a year, a windowless shell wired for more electricity than the entire town around it. After months of countless new data center projects sitting idle, the DSA ran a nationwide campaign to convert them to affordable housing. A DSA housing collective bought the Abilene project for pennies on the dollar and converted it to affordable housing. The pitch was almost too perfect. The power and water were already built, and the buildings could be easily divided up into starter housing for young adults and new families. By the summer of 2028, there are hundreds of these across the old data-center belt, from Texas to Louisiana to Ohio, each one a monument to the great capex boom that lasted from 2024-2027, with trillions spent on a nascent technology with an uncertain long-run economic profile.
The converted shells are the most visible part of the wreckage. The convoluted flow of capital that financed them is much harder to piece together; a large portion of it ended up somewhere stranger than a warehouse in Texas, in the retirement accounts of average Americans. How it got there starts with the great financing machine that roared from 2024 to 2027.
The Financing Engine
By some point in 2026, the AI buildout outgrew the balance sheet capacity of the hyperscalers, the neoclouds, and the frontier labs that had been capitalizing it. The replacement was a multi-pronged financing machine, a sort of frankenstein of leveraged GPU’s.
In this specific case, the revenue commitments added up to about $200bn. It was one of the largest infrastructure financing packages of all time.
The financing structure was structured so no one party had the chips entirely on their balance sheet. Morgan Stanley arranged a private-credit vehicle, funded by outside investors, that bought the chips and leased them to Anthropic. This way, neither Google nor Broadcom had to directly hold the assets on their books. The FT compared it to the vendor financing Boeing and GE built to move aircraft and engines.
In June 2026 the first batch of capital ran through the machine when a shell called Compute SPV paid $35bn for about a gigawatt of hardware, roughly a million chips. The cash came from three tranches of debt anchored by Apollo and Blackstone. Broadcom guaranteed the two senior tranches, promising to cover the shortfall if Anthropic stopped paying and the chips couldn't be sold for enough to make the investors whole. That guarantee, called residual value support, covered about $30bn of the $35bn.
The senior investors were told their money was safe. The reason was not the chips. The reason was Broadcom which would cover any potential shortfall if the chips weren’t worth enough. Strip the guarantee and the collateral is a warehouse of hardware and one start-up's IOU. That template was set to finance hundreds of billions more. In April 2026 Google agreed to sell Broadcom another 3.5GW of chips for Anthropic, and Broadcom's own filings disclosed $128bn of purchase commitments, with $55bn of deliveries due in fiscal 2027 and $73bn in 2028.
The chips needed powered shells and power took years to permit so financing the hardware solved only half of Google's problem. It found the other half in an unlikely place, the crypto miners that already had electricity. A small miner called TeraWulf was the first to take a Google backstop and add a 360MW hall in upstate New York. Morgan Stanley packaged Google's guarantee of the lease payments into a construction bond that raised $3.2bn. Google took penny warrants and a stake in the miner, then ran the same play with Cipher and Hut 8 in Texas and Louisiana. By the FT's count Google had backstopped ten developments with 2.4GW of power.
Google's backstops put it on the hook for as much as $44bn if the leases went bad and on its balance sheet it carried the liability at $815mn. Meta had done the same with its own projects. Its Louisiana campus, Hyperion, sat inside a vehicle called Beignet Investor, 80 percent owned by Blue Owl, which in October 2025 sold a single $27.3bn note at 6.58%, rated A+, bought mostly by Pimco and BlackRock. Meta kept the debt off its books by arguing it didn't control the entity, even though it was the only tenant. Ernst & Young flagged that judgment in Meta's own filings as a critical audit matter, with an auditor stating that the judgment was contestable.
One Oracle campus in Michigan carried $16bn of debt a rung lower on the credit ladder, most of it a bond stretched over 19 years and anchored by Pimco after the banks stepped back. S&P cut Oracle to BBB- in July 2026, one notch above junk. A nascent market turning data-center leases into rated bonds grew from about $4bn outstanding in 2020 to more than $60bn by 2026. By one count AI-related debt was about 30 percent of all net investment-grade dollar bond issuance in 2025. The investment-grade market was becoming a concentrated bet on AI.
Data-center projects with a Google backstop borrowed money at about 7.1% by the spring of 2026, while the neocloud operators building around Nvidia's chips paid ~9.3%, a gap Jefferies called a structural cost-of-capital disadvantage for anyone in Nvidia's orbit. All $200bn of the edifice rested on a single question, whether Anthropic could pay. One Jefferies analyst put it plainly, that a whole world had been built underneath a handful of companies, and if their appetite to spend fell, all of it would slow.
The Collateral
The entire financing machinery rested on chips holding their value over a long enough depreciation curve, which didn’t turn out to be the case. Nvidia ships a new flagship chip roughly once every 12 months, and the buyers depreciated them over five to six years. Microsoft's move to a six-year server life in 2022 added $3.7bn of operating income in one year. Alphabet's cut depreciation by $3.9bn. Amazon went the other way in 2025, shortening some servers back to five years and blaming the pace of AI.
Nothing New Under the Sun
In the 1970s a Dallas salesman named Chris Christopher, a high-school dropout who had sold encyclopedias and later insurance, convinced Lloyd's of London to issue a policy that paid computer-leasing firms if their leases were cancelled by customers due to obsolescence. On the strength of that guarantee, banks lent his firm and others hundreds of millions to buy computers to lease out. The machines were leased to corporations and government agencies, and the cashflows paid the debt down over seven years. If a customer canceled between years four and seven, Lloyd's covered the residual loss. The banks would previously only lend for four years, because they knew IBM could make their collateral obsolete overnight. The Lloyd’s implicit financial guarantee rapidly expanded the financing market which powered aggressive growth in the computer leasing business and thereby the sale of computers. The financing engine is always the key variable in a bubble. The party tends to go on unabated until the financing mechanism runs into trouble.
About $1bn of the policies had been written when IBM did the obvious thing and announced the 4300 series, faster and up to 30 percent cheaper than the machines already operating. The leases were quickly unwound and Lloyd’s was stuck with a massive financial loss. Its own adjusters told the underwriters to reserve $220mn, and the feared loss reached half a billion dollars, the biggest in Lloyd's 291-year history to that point. The IBM record had looked stable. Nobody priced IBM cutting prices by 30 percent.
With the benefit of hindsight, so many of the SPV structures custom designed to finance compute capex were eerily similar to the original Lloyd’s residual value guarantee. The core debate was always what would happen when the initial customer contract needed to be renewed – what would the renewal rate be for a 4-5 year old GPU, in a market ostensibly no longer as constrained for compute. In one notable deal in Summer 2026, investors balked at lending to a compute SPV where the renewal risk was particularly high, with renewals coming in just a few short years: “Investors briefed on the deal said they were particularly concerned about the relatively short contract duration underpinning the latest financing, which exposes lenders to renewal risks.” The flaw was that the buildings were leased for 15 years while the customers signed for three to five. The cost to insure that borrower against default rose about a quarter in half a year. The market had slowly started to charge for the risk before analysts admitted the risk was there. It had been hiding in plain sight.
The solution was that a “lockbox” around the SPV was tightened such that no residual cashflow could flow to the neocloud until lenders had received half of their principal back. While this worked as a stopgap to get the deal and subsequent deals like it done, the issue was that when the time came for the neoclouds to repay their corporate debt obligations, they had no firm access to the collateral “locked” away in the SPV’s. The neoclouds ended up looking much more like levered CLO managers than technology firms. Or put differently, they were functionally similar to specialized banks financing compute but showing rapid EBITDA growth to investors without including the principal cost for their main ingredient - interest and financing.
Modern Monetary Theory of Compute
If you analyzed all of the compute leases, the money ultimately flowed at some point through the frontier labs, OpenAI and Anthropic. For many months in the spring and summer of 2026, investors were constantly wowed with leaks and rumors of how quickly the annual recurring revenue run-rates were growing for both OpenAI and Anthropic. On one episode of the All In Podcast, there was a report of Anthropic having surpassed $80bn of ARR. Then one day in August, Bloomberg reported that Anthropic had surpassed $65bn of ARR. An impressive feat no less, but well short of some of the numbers being thrown around by analysts on the street. Then that same week, an article came out in the WSJ, noting: “OpenAI’s Second-Quarter Sales Show Tepid Growth Compared With Anthropic” That same article noted that OpenAI had $6.7bn of revenues in the 2nd quarter of 2026. Again a very impressive feat, but short of some of the numbers being thrown around by analysts. OpenAI saw approximately 18% sequential growth in the 2nd quarter. By all accounts impressive financial performance, but surely this left some analysts wondering if maybe AGI was coming a little more slowly than advertised. But then again, maybe the issue was that the company’s operating loss including stock-based compensation widened from $9.3bn in the first quarter to $12.3bn in the second quarter. So losses were growing almost as fast as revenues, for an annualized rate of almost $50bn of losses.
The issue was not the absolute size of the revenues or even the losses per se, but rather that OpenAI had signed over $1tn of compute commitments, including $300bn to Oracle and $250bn to Microsoft. The gap between its purchase commitments and its ability to pay was an order of magnitude. The same applied to Anthropic as well. Anthropic had initially been much more conservative than OpenAI with aggressive compute commitments, but then once it faced compute shortages in 2026, it too began to aggressively commit to large scale compute. The market was eagerly awaiting an Anthropic IPO in the fall, as were many venture capitalists eager to lock in once in a generation returns. When the long awaited Anthropic prospectus was finally released publicly, investors were shocked by the scale of operating losses and that without aggressive and numerous adjustments, the business was massively unprofitable. Some graybeards went as far as to compare it to WeWork with its infamous community adjusted EBITDA. Anthropic preferred to show Adjusted EBITDA Gross of Training costs.
Anthropic ultimately did IPO after having to cut the size and valuation of the IPO multiple times, and only with Nvidia ultimately backstopping the IPO with a $25bn commitment. The IPO was a disappointment and the stock traded down 50% in the very first month. OpenAI never even went through with filing its prospectus, preferring to wait for more favorable market conditions, which never came.
The AI Maxis hoped that “Scaling Laws” would fix the “AI Gully” as it was being called, and that revenue would compound until the losses turned into businesses big enough to fully fund the purchase commitments. One AI researcher, Andrew Ho, called the labs overvalued in a viral post after he resigned from a frontier lab in 2026. His analogy was a treadmill. A lab couldn’t stop training the next model, because the moment it stopped, its product became a commodity, and a cheaper open model from Qwen or Kimi would do the same job for less. So the spend on the next model had to keep rising, faster than the revenue the last one earned. The losses were not one-time in nature at all. They were the cost of staying at the so-called “frontier”, and they only grew as the models got more expensive.
Nvidia invested in OpenAI, which bought Nvidia chips, so the demand underneath the whole thing was never as independent as it looked. It took stakes in the neoclouds that bought its chips and backstopped the capacity they couldn't sell. It held 9 percent of Nebius and a piece of xAI. Microsoft owned a quarter of OpenAI and sold it Azure. Oracle bought Nvidia chips to lease to OpenAI. Money left Nvidia, ran through a start-up and a neocloud, bought Nvidia chips, and came back as Nvidia revenue. At each hop someone booked a number that read as fresh demand.
The Life Insurers and the Central Bank of Nvidia
The senior slices of this debt orgy had to go somewhere and they went to the one buyer built to hold long-dated paper and never mark it to market, the life insurers. Over the previous decade, the major private capital firms had all delved deeply into insurance asset management. Apollo had Athene. KKR had Global Atlantic. Blackstone managed Corebridge. Blue Owl had Kuvare. Brookfield, Carlyle and Ares ran versions of the same playbook.
An annuity is close to a perfect funding source for a credit business, because the saver hands over the money now and can't easily take it back for years unless they pay a hefty surrender charge.
When Apollo led a $3.5bn slice of an Nvidia-chip deal for xAI, or the Broadcom machine that fed Anthropic, Athene was one of the places the senior paper found its home.
Structured securities made up about 27 percent of private-equity-linked insurers' portfolios by 2026, as compared to 12 percent for other large insurers, and in some cases the sponsor managed those securities itself, according to the Financial Stability Board. Those insurers controlled nearly $900bn of insurance liabilities, up from $67bn in 2012, and captured about 35 percent of new US annuity sales in 2023. A Fitch study found 97 percent of privately rated securities carried higher grades than the regulators' own analysts would give, a third of them by more than three notches.
One night in August 2026, Jensen Huang, CEO of Nvidia, announced a grand financing plan with the heads of some of the largest financial services firms in the world, including Apollo, BlackRock, Blackstone, Brookfield Asset Management, Goldman Sachs and KKR. The plan was for $500bn of financing for even more AI Capital Investment, creating a new asset class to standardize chip financing. The WSJ compared it to financing airplanes, credit cards, and mortgages. The main difference that went undiscussed was that airplanes and homes have usable lives in the decades, and GPU’s only in the single digit to low double digit years. To some skeptical observers, this half-baked plan with limited specifics seemed like a last-ditch effort to keep the music from stopping. Alongside the announcement, Jensen published a largely AI written note which made little sense explaining the move.
The Trigger
Throughout the summer of 2026, chatter started to grow about a billionaire financier facing issues with his complex financial empire. Mark Walter was a multibillionaire and owned multiple premier sports teams including the LA Dodgers, the LA Lakers, and Chelsea FC. Reports had started coming out that Walter’s insurance firms may have significantly misstated the portion of their investments that were affiliated. Reports highlighted over $20bn of investments that should have been marked affiliated but were not. There were also numerous reports of potential ongoing criminal investigations tied to this matter. One day a shock hit both the sports and financial worlds simultaneously: Walter sold the LA Lakers to Josh Kushner and Bob Iger, less than a year after finalizing the team’s purchase. Sales of marquee sports and entertainment assets were rare as is, let alone less than a year after buying one of the best ones of all. Immediately people began to suspect that Walter was in dire need to raise liquidity. Reports also circled of Walter offering double digit yields for short-dated loans with his stake in Guggenheim as potential collateral. Onlookers noted that he was showing the telltale signs of financial distress.
Months later, Walter’s empire would be caught up in an array of lawsuits as well as civil and criminal investigations. Several units of his financial empire would ultimately be forced to file for bankruptcy. Walter’s close associate Todd Boehly who had operated a fairly similar insurance strategy soon found himself in the regulators’ crosshairs and had to also aggressively firesale many of his key assets. The importance of this “insurance showdown” was what came next. Treasury Secretary Scott Bessent had grown increasingly concerned about risks building in the life insurance system. After the Walter and Boehly debacles, Bessent addressed a joint session of Congress to give his assessment on how bad things could get. He warned that if aggressive measures were not taken urgently, there was a risk of a Global Financial Crisis 2.0, only bigger this time because the numbers were an order of magnitude larger. Bessent asked for immediate and unprecedented authority for the Treasury to supersede State Insurance Commissioners for so-called “problem insurers”. He also asked Congress to approve a $100bn “Insurance Stability Fund” to help orderly liquidate and wind-down troubled insurers. Lastly, he asked for the authority to issue a moratorium on so-called “affiliated investments”, as well as higher risk “AI Linked” investments which had started to show some early signs of trouble on several insurance balance sheets. This meant that the turbine of the financing engine immediately ground to a halt. All the major insurance companies immediately begin detailing their exposure to “affiliated investments” and “AI Linked investments” in presentations designed to reassure customers, investors, debtholders, and regulators. They all announced that they would suspend all new AI investments in their insurance units until further notice.
The Aftermath
Suddenly other debt investors began to reassess some of the risks they were taking with their AI investments. Some realized they had been taking equity risk for debt upside. They too began to slow their pace of funding. Deals that would have closed in weeks just months earlier began to drag on for months and months or not get done at all.
Cracks started to show in the demand side as well. Anthropic’s growth started to level off, until it showed a sequential decline in revenues in the fourth quarter. Many enterprises that had spent money hand over fist on AI to wring out efficiencies just months prior began second-guessing new AI projects after all that money had failed to show meaningful improvements to the bottom line. Then one day Bloomberg reported that several neocloud compute projects had fallen substantially below planned renewal rates and missed several key covenant tests, and lenders were negotiating with them to potentially seize the GPU collateral. Then suddenly, analysts began predicting that Nvidia would see a previously unheard of slowdown in growth and then even a decline in revenues in the following year. And then the bad news really started to roll.
Several insurers began aggressively raising capital by selling whatever assets they could for fear they would lose access to the institutional funding markets which were significant portions of their liabilities. Whereas previously they could continue to take in new policyholder funds and use some of that liquidity to pay existing claims, suddenly the new insurance money flow dried up. Retirees began to question whether annuities were safe. Bessent took the extraordinary step of extending the FDIC guarantee to all fixed annuities. Nonetheless, several large insurers prepared to enter into insolvency proceedings. The stock market was in freefall for months on end, aided only temporarily by short-lived and violent “bear market rallies” which brought false hope of light ahead. President Trump ultimately nationalized some of the frontier labs and key AI infrastructure builders including chips companies, power companies, and several neoclouds.
The main street economy felt the spending slowdown as well. AI capex had become one of the largest single drivers of US growth in 2026 and 2027, with over $700 billion dollars of annual spending. Suddenly, when the trend reversed, all the companies that had benefited from a once in a generation demand boom saw their business dry up completely. Companies that had begun retrofitting jet engines to power data centers suddenly found themselves with no buyers at all. Chipmakers, power-equipment suppliers, construction crews, and neoclouds were hit much as the fiber builders had been when the telecom boom unwound in 2001. The financing had, in fact, been circular, and circles run in both directions.
It’s still 2026
None of this has happened, yet. The AI capex machine is still humming, the ratings have not been downgraded, no data-center bonds have defaulted, and the surface of the water has never looked better. It causes a visceral angry reaction to even dare question any of the math. What if demand for compute and intelligence is not, in fact, infinite? Doesn’t a shortage of compute imply that some day there may be a glut of compute? Questions like these receive angry, terse replies, or none at all.
There is no clear evidence that the combination of capital, chips, and power turbocharged by leverage won’t work out as planned. But then again, there was no evidence that home prices would do anything but go straight up either. Perhaps this time is different.
Additional disclaimer:
This communication is for informational purposes only and is not a recommendation, an offer to purchase, or a solicitation of an offer to sell shares.
This communication contains our current views. Our views are based on our own analysis of publicly available information and assumptions we believe to be reasonable. There can be no assurance that the information we considered and analyzed is accurate or complete. Similarly, there can be no assurance that our assumptions are correct. Our views and our holdings could change at any time. We may sell any or all of our holdings or increase our holdings by purchasing additional securities. We may take any of these or other actions without updating this communication or providing any notice whatsoever of any such changes (except as otherwise required by law).
Sources:
FT: crypto-miner data centers built for AI in TX and LA, Janus Henderson: 157GW announced vs ~85GW deliverable
FT, Ryan McMorrow, "Inside Google's $200bn Wall Street finance machine for Anthropic," 3 Aug 2026
Apollo press release, 9 Jun 2026
CNBC: Meta/Blue Owl $27bn Hyperion, Moody's flags ~$662bn off-balance-sheet leases
S&P cuts Oracle to BBB-, Forbes: data-center ABS growth
CNBC: Burry on AI depreciation
Amazon (10-K), Silicon Data: H100 rental index
Washington Post, John F. Berry, "Lloyds Facing Huge Loss Over Computer Leasing," 2 Jul 1979, archive copy
Fortune: OpenAI $14bn loss, HSBC shortfall, CNBC: OpenAI resets to ~$600bn by 2030, Altman: $1.4tn / 30GW,
https://www.ft.com/content/9d2117af-b3ec-4ca4-b00f-6813ab5075ec?syn-25a6b1a6=1
https://x.com/andrewho03/status/2082786931419812338
BIS 2026 annual report on circular financing
NAIC: private-equity-owned insurer investments, FSB: vulnerabilities in private credit, May 2026
Athene Holding 2025 annual report (10-K), Apollo: $3.5bn for xAI's Valor project
FSB: private credit, May 2026, NAIC scrutiny, Fitch 97% notch gap, FSOC 2024: offshore reinsurance and private credit
Treasury FSOC: ~$1tn private credit in life insurers, Hyperscaler 2026 capex ~$725bn, up ~77%
Chanos: AI build echoes the fiber bubble
https://finance.yahoo.com/markets/stocks/articles/mark-walters-insurer-cutting-6-133456851.html
