$125B
AI data-center financing, 2025
Up from $15B in 2024. The buildout stopped being a technology story and became a capital-markets story.
AI, debt, private credit, consumer fragility, and the shrinking middle of the economy.
AI is real. The risk is that Wall Street is financializing the buildout faster than end-customer cash flows, labor markets, and consumer demand can prove the economics. The dots do not look dangerous one by one. The danger is the circle they form together.
Each datapoint below has a reasonable explanation in isolation. Read together, they describe a system that is becoming more interconnected, and more dependent on a narrow set of supports. Every figure is current as of June 2026.
Up from $15B in 2024. The buildout stopped being a technology story and became a capital-markets story.
Morgan Stanley, June 2026. AI-linked debt is now the largest single segment of US investment-grade credit.
FSB estimate, end-2024. Risk has shifted outside regulated banks into more opaque markets.
The consumer layer is already stretched before any major labor shock fully lands.
Headline consumption is increasingly dependent on asset owners and high earners.
22% of all 2026 cuts, already past the full-year 2025 total. AI is measurable in headcount decisions.
Since 2023, vs 32% equal-weighted. The index can rise while the average company gets squeezed.
These are usually discussed separately. The danger is the pressure they pass back and forth. Slow fractures, sudden breaks.
They are built on things that are completely real.
AI is not fake. That is not the argument, and anyone making it has already lost the plot. The internet was not fake either. Neither were railroads, radio, electricity, or housing. Every one of them was real. Every one of them changed the world. And several of them still produced some of the largest wipeouts in financial history.
Because a technology being real has never been the question. The question is what gets built on top of it while everyone agrees it is real.
Real and overpriced are not opposites. They are usually the same story, told one chapter apart. The technology delivers. The financing around it gets ahead of what the technology has yet to prove.
Not whether AI works. Whether the financing structures being built around AI are already pricing in more economic value than the end market has actually proven.
There is a second mistake, quieter than the first. We examine each pressure point alone. Layoffs, alone. Corporate debt, alone. Consumer delinquencies, alone. Private credit, alone. Narrow market leadership, alone. Each one looks survivable in isolation, so each one gets explained away, on its own page, by its own expert.
But cycles do not break because one number suddenly goes wrong. They break when manageable stresses start handing pressure to each other until the whole thing becomes a single circuit.
Bankruptcy works this way. So does divorce. It is never one event. It is a sequence of small fractures that each seem survivable in the moment. One missed payment. One bad quarter. One layoff. One downgrade. Then people call the collapse sudden, even though the pressure had been building for years.
Markets are no different. The risks below are usually drawn on separate pages. The rest of this letter is about the lines between them.
Wall Street rarely creates complex vehicles for things nobody cares about. It creates them after a story has already worked, after investors are hungry for exposure, and after the obvious equity gains have already been made. Financial engineering is usually not the spark. It is the accelerant.
In the 1920s, brokerage houses, investment trusts, and margin accounts let ordinary people buy stocks with borrowed money. In the 1960s, conglomerates used acquisition math to make growth look more durable than it was. In the 1980s, junk bonds funded LBOs and hostile takeovers. In the 2000s, securitization and derivatives made mortgage risk look diversified until investors could no longer locate the losses.
Today the vehicle is different: private credit, AI infrastructure debt, data-center project finance, GPU-backed financing, capacity pre-purchase agreements, and asset-backed securities tied to digital infrastructure. The instruments are different. The incentive structure is familiar.
The most dangerous part of financial engineering is not always the leverage itself. It is that it creates hidden connections. A risk that appears contained in one corner of the market can become a funding problem, then a liquidity problem, then a labor problem, then a demand problem. By the time the connections are obvious, the market has usually already repriced them.
For the first few years, AI was mostly a software story. Models improved. Products launched. Enterprises experimented. That was the innovation phase. Now the center of gravity has shifted to chips, power, cooling, land, data centers, cloud capacity, fiber, debt issuance, private equity, and private credit. That does not make the story wrong. It makes it more fragile.
By October 2025, AI-linked debt had reached roughly $1.2 trillion, the single largest segment of the US investment-grade bond market, overtaking US banks. Because major bond indexes are weighted by market value, every new AI bond increases its issuer's share of the index, and passive funds must buy proportionally more. Target-date funds, which held about $4.8 trillion at the end of 2025, own those index funds.
Many retirement savers now carry indirect AI-infrastructure debt without ever choosing it. The exposure reaches well beyond Wall Street credit desks and into ordinary 401(k) accounts.
The most important question in this cycle is not whether AI usage is real. It is whether the current infrastructure demand is paid for by durable end-customer cash flow, or pulled forward by investors, debt, vendor financing, and fear of missing out.
When the original version of this letter was drafted, the clearest example was a single relationship: Nvidia agreed to purchase any unsold CoreWeave cloud capacity through 2032 under a $6.3B order, while also being an investor in CoreWeave and its chip supplier. One overlap. By mid-2026, that overlap had become a web.
None of this proves anything fraudulent. None of it proves AI is fake. It does mean some of the demand signal is financially supported by the same ecosystem reporting the growth. When Goldman Sachs analyzed OpenAI's finances including its capital commitments, internal revenue and vendor financing covered only about 17% of operating costs, while external funding swelled to 75%.
Bears call this an Ouroboros, the snake eating its own tail, and compare it to the vendor-financing arrangements that preceded the dot-com collapse. Bulls, including asset managers like Janus Henderson, call it a virtuous circle: in a supply-constrained market, locking in chips, builders, and customers with long-term commitments and financing is rational, not fraudulent. In February 2026, a single Wall Street Journal report that Nvidia's $100B OpenAI investment had "stalled" briefly rattled all three names — a live stress test of how tightly the chain is wired.
Some portion of AI demand is real customer demand. Some portion is strategic overbuild. Some portion is circular ecosystem demand. The mistake is treating all three as equally durable.
The US fiscal picture matters because it tells us how the broader system is being held together. The debt is not expected to be repaid the way a household pays down a mortgage. It is expected to be rolled, serviced, and diluted through nominal GDP growth. That can work for a long time. But it is not the same thing as health.
The likely exits are growth, inflation, austerity, default, or some politically engineered mix. The political system has strong incentives to choose inflation and financial repression over explicit default. That means the dollar does not need to collapse for purchasing power to erode. The burden can be shifted slowly, quietly, and unevenly. This is another domino that looks survivable in isolation — until fiscal pressure reduces policy flexibility at the same moment private credit tightens and consumers weaken.
After 2008, risk did not disappear. It moved. Banks became more regulated, and credit creation migrated toward private markets. Private credit is no longer a niche.
The Financial Stability Board has warned that private credit contains vulnerabilities around bank–nonbank interconnections, borrower credit quality, leverage, valuation opacity, data gaps, and liquidity mismatch. That is regulator language for: we do not fully know where all the risk goes when things get bad.
Private credit is useful. It finances companies that banks may not serve efficiently. But useful is not the same as safe. The issue is not that it exists. The issue is that it may be tied to the same borrowers, data-center projects, consumer-credit structures, and refinancing assumptions that depend on the rest of the cycle continuing. If the cycle weakens, the private-credit layer may not be the cause. It may be the amplifier.
The US economy ultimately depends on spending. Consumer spending is roughly 70% of output. That makes the consumer the transmission mechanism between layoffs, credit stress, corporate revenue, and market earnings. The consumer enters this AI/labor transition with an already large debt stack.
None of this means immediate collapse. It means the cushion is thinner than the headline economy suggests. And the clearest evidence that separate dots are already merging into one pressure system comes from the borrowers who recently defaulted on student loans after the pandemic pause ended.
Among post-pandemic student-loan defaulters, the New York Fed found that nearly 40% are now past due on auto loans and 56% are past due on at least one credit card. A single missed obligation rarely stays single. It spreads across every other line on the same balance sheet.
A rising credit-card balance is one data point. Auto delinquencies are another. Student-loan stress is another. A layoff is another. Each can be explained away separately. But when the same household faces higher debt service, weaker job security, higher living costs, and fewer replacement opportunities, the separate dots become one pressure system.
The phrase "the consumer is strong" has become increasingly misleading. Which consumer?
A note on rigor: this Moody's figure (Zandi) is widely cited but also contested — some economists, including UC Berkeley's Antoine Levy, argue the methodology overstates the top 10% share. Even discounted, the direction is not seriously disputed: spending power has migrated upward, and the economy leans harder on asset owners than it did a generation ago.
This is one of the most important facts in the economy, and it connects directly to the AI thesis. The top 10% own the bulk of equities. The AI boom has inflated those equities. That paper wealth funds their spending. So the same force lifting AI valuations is propping up headline consumption — a K-shaped economy where one group climbs on asset prices while the broad middle merely keeps pace with inflation.
A wealthy consumer can spend a lot. But wealthy spending is narrow: luxury travel, premium brands, high-end real estate, bespoke services. Most businesses were not built to serve only rich people. They were built for broad recurring demand: groceries, cars, restaurants, subscriptions, insurance, childcare, gyms, local services, repairs.
A billionaire can buy a $50M house. He does not buy 50,000 burritos. Broad economies need broad purchasing power.
This is why inequality is not only a moral or political issue. It is a demand-architecture issue. If spending power keeps migrating upward, more businesses become dependent on fewer customers, fewer regions, and fewer asset-price cycles. That makes the whole system more sensitive to a shock at the top.
This cycle differs from many prior automation cycles because AI targets cognitive labor. The exposed groups are analysts, marketers, recruiters, support teams, developers, designers, consultants, paralegals, junior finance workers, and administrative staff. These workers are not just payroll costs. They are the demand base. They buy homes, travel, start companies, own stocks, and support local economies.
A note on rigor: Challenger tracks employer self-attribution, not independently verified causation. Some "AI" cuts may be ordinary cost-cutting wearing a fashionable label — the "AI washing" concern. That caution cuts both ways: it means the labor signal is noisier than the headline, but also that the trend toward citing AI is itself a real shift in how companies frame their decisions.
This does not mean AI is causing a labor-market collapse. The BLS still reported 172,000 payroll jobs added in May 2026 and unemployment at 4.3%. But the composition matters. Gains came in leisure and hospitality, local government, and health care, while financial-activities employment declined — down 107,000 from its May 2025 peak.
The economy can add jobs while the white-collar income layer weakens. A new data-center construction job is not economically interchangeable with a laid-off software, finance, sales, or marketing employee. Different wage. Different geography. Different spending profile. Different wealth effects.
The same concentration is happening in equities. Since the start of 2023, the cap-weighted S&P 500 gained 67%, more than double the equal-weighted index's 32%, reflecting the growing dominance of a small group of mega-cap companies.
This is not automatically bearish. The largest companies are large for a reason: real revenues, strong products, global distribution, cash flow. But concentration creates a different kind of risk. Indexes can rise while the average business, worker, and consumer gets squeezed. Three of the concentrations in this letter — debt, spending, and market leadership — are the same shape pointed at three different layers of the economy.
Put the pieces together and the loop becomes clearer. The market treats these as separate headlines because the connections are not visible day to day. Layoffs are a labor story. Debt is a capital-markets story. Consumer stress is a household story. Private credit is institutional. Fiscal deficits are governmental. Concentration is an equity story. But in the real economy, they are all connected through spending, credit, confidence, liquidity, employment, and asset prices.
This is not a prediction that the system breaks tomorrow. It is a framework for why it becomes more fragile even while the headlines look good. Every individual stress can be explained away. The risk is that the explanations are correct in isolation and still wrong in aggregate, because the transmission happens between the categories.
In 2008, the too-big-to-fail institutions were banks. In the next cycle, the systemically important layer may be cloud, chips, data centers, power infrastructure, cybersecurity, and AI compute. If that layer gets into trouble, the government response will likely not be called a bailout. It will be called national security, infrastructure resilience, industrial policy, energy security, or strategic competitiveness.
Support would arrive through government contracts, tax credits, emergency liquidity, regulatory forbearance, chip subsidies, energy policy, forced mergers, or public-private financing. The label changes. The economic effect can be similar. This matters for the middle class because the benefits of the boom are increasingly captured by capital owners, while the costs of system support tend to be socialized through inflation, taxes, debt, weaker real wages, and reduced mobility.
The goal is not to be bearish. The goal is to be honest. A good thesis should be falsifiable, so here is the evidence that would weaken it, and the evidence that would deepen it.
| If this happens… | …it weakens the thesis because |
|---|---|
| AI revenue turns into durable free cash flow at scale | The buildout was pulled forward, not overbuilt. |
| Enterprise AI lifts productivity without broad layoffs | Productivity growth can offset labor compression. |
| Private credit losses stay contained through a real downturn | The shadow-credit layer is less fragile than feared. |
| Consumer spending broadens beyond high-income households | Demand concentration is less dangerous. |
| Data-center utilization stays high and GPU rental economics hold | The capacity was economically justified. |
| New job categories absorb displaced white-collar workers quickly | AI becomes a broad-participation technology, not a compression one. |
| Warning sign | Why it matters |
|---|---|
| More vendor financing and supplier-backed customer deals | Demand gets harder to separate from ecosystem support. |
| AI infrastructure loans securitized broadly | Risk moves into the wider investor base. |
| Private credit redemption gates spread | Liquidity mismatch becomes visible. |
| Hyperscalers cut capex guidance suddenly | It resets the entire AI infrastructure chain. |
| Startup revenue grows while cash burn worsens | It suggests investor-funded demand, not durable economics. |
| White-collar layoffs spread into finance, legal, consulting, sales, ops | An AI story becomes a consumer-demand story. |
That ambiguity is the honest center of this whole letter. The same hardware that depreciates fast enough to scare lenders is currently scarce enough to command rising rents. Both facts are true today. Watch this number, data-center utilization, private-credit marks, redemption gates, hyperscaler capex guidance, white-collar unemployment, consumer delinquencies, and market breadth. Those are the scoreboard. Watch the indicators, not the narrative.
A thesis like this is only useful if it helps define the range of possible futures. I do not want to write a one-way doom letter. Both worlds can begin the same way — rising AI spending, strong mega-cap earnings, heavy infrastructure investment, productivity headlines. The difference only becomes obvious when cash flows, labor markets, and credit quality either validate the buildout or expose the weak links.
| Dimension | If I am right | If I am wrong |
|---|---|---|
| AI economics | Adoption is real, but monetization lags the capex, debt, and capacity commitments built around it. | Revenue and productivity compound fast enough to justify the infrastructure spend. |
| Labor market | White-collar compression spreads beyond tech into finance, legal, consulting, sales, and ops. | AI eliminates tasks but creates enough new roles to absorb displaced labor. |
| Consumer demand | Spending becomes too dependent on affluent households and asset prices while the middle weakens. | Lower costs and new income streams broaden participation over time. |
| Private credit | Opaque structures and redemption limits become the transmission mechanism for stress. | It absorbs losses without systemic contagion and keeps financing productive infrastructure. |
| Mega-cap tech | The largest platforms become de facto utilities: privately owned upside, publicly protected infrastructure. | Scale produces cheaper compute and broader diffusion rather than permanent concentration. |
| Government | The state increasingly backfills demand and protects infrastructure through deficits and subsidies. | Policy bridges the transition without permanently socializing losses. |
| Market outcome | The technology survives, but the financing cycle reprices. Damage shows first in weak AI borrowers, private credit, and overbuilt capacity. | The buildout looks aggressive but rational. Earnings grow into valuations and the market broadens. |
If I am right, AI still changes the world, but the current financing cycle creates a painful misallocation before the benefits fully diffuse. If I am wrong, the market is correctly discounting a productivity step-change that traditional valuation models are too slow to recognize.
I am not arguing that AI is fake. I am arguing that AI is entering the phase where a real technology gets wrapped in financial structures that can distort the signal.
I am not arguing that a crash is imminent. I am arguing that the economy is becoming more dependent on a narrow set of supports: mega-cap tech, AI capex, private credit, government deficits, asset prices, and affluent consumers.
I am not arguing that productivity is bad. I am arguing that productivity without participation creates a different kind of economy. A more efficient one, maybe. But also a more fragile one.
The danger is that we keep explaining away each stress because each one has a reasonable isolated explanation. Then if the cycle turns, everyone acts surprised that the dots connected. That is how slow problems become sudden crises. That is how bankruptcy works. That is how divorce works. That is how credit cycles work. The story feels manageable while the fractures are separate. Then the connections reveal themselves. Slowly, then all at once.
AI may be real enough to transform the economy and financialized enough to create the next major misallocation of capital. The thing to watch is not whether AI works. The thing to watch is who pays, who earns, who borrows, who loses their job, who gets protected, and who is left buying.
I am not worried about one domino. I am worried about a market that keeps insisting every domino is unrelated. Layoffs, debt, private credit, consumer stress, geopolitical tension, fiscal deficits, AI capex, and market concentration do not need to be individually catastrophic to become collectively dangerous. The future may be extraordinarily productive. The harder question is whether enough people get to participate in it before the system discovers that efficiency without broad participation is not resilience.