Where the AI Bubble Breaks First

We wrote about the 18x revenue gap, the GDP dependency, the capex-driven layoffs. This piece is about something more specific: where the stress shows up first. Not in theory. In the actual financial structures that are already bending.

$660 Billion in Capex. $30 Billion in Revenue.

The gap between what’s being built and what’s being earned

The five largest hyperscalers are on track to spend $660–690 billion on AI infrastructure in 2026. That’s nearly double 2025 levels. Triple what they spent two years ago. Amazon alone plans roughly $200 billion, and Morgan Stanley projects the company will burn through its entire free cash flow and then some, running negative FCF of $17–28 billion for the year.

JPMorgan’s math says the industry needs $650 billion per year in AI revenue to justify current spending. Generative AI revenue today sits around $30–37 billion. The number you’re doing in your head right now is correct. That’s an 18x gap.

For context: during the telecom bubble, companies laid hundreds of thousands of miles of fiber based on demand projections that took 15 years to materialize. By 2005, 85% of that fiber was still dark. The AI capex-to-revenue gap today is wider than the telecom equivalent was in 1999.

NVIDIA Is Still Growing. The Growth Rate Isn’t.

Revenue up 62%. But the slope is flattening.

Q3 FY2026 revenue: $57 billion. Up 62% year-over-year. Data Center alone pulled $51.2 billion. Gross margins above 73%. By any normal standard, these are spectacular numbers.

Now look at the trajectory. Full year FY2025: +114%. Q1 FY2026: +69%. Q2: +56%, the slowest in nine quarters. Q3 bounced to +62%, but still well below the trend line from a year ago. The second derivative has gone negative. Revenue keeps climbing, but the rate at which it’s climbing has bent downward.

This pattern has a name in market history. Cisco’s revenue growth decelerated for two consecutive quarters before the dot-com crash. The company was still printing record numbers. Analysts were still raising price targets. The stock hit $80.06 in March 2000, then lost 89% over the next two years. It only passed that price again in December 2025. Twenty-five years.

NVIDIA isn’t Cisco. Real margins, real products, real demand. But the market prices trajectory, not absolute level. When that trajectory bends, the repricing has historically been severe regardless of how good the underlying business is.

Credit Markets Are Doing That Thing They Do Before Corrections

AI infrastructure debt issuance, 2023–2026

Credit spreads have been the most reliable early warning signal of financial stress for three decades. They widened before the dot-com crash, before the financial crisis, before COVID. They typically lead equity selloffs by 6–18 months.

The hyperscalers raised a record $108 billion in debt in 2025, more than 3x the nine-year average. Bank of America calculated that hyperscalers would need to spend 94% of their operating cash flow to fund AI buildouts. Ninety-four percent. That’s why they’re turning to debt markets. They literally cannot fund this from operations.

AI-linked firms now account for 14% of the investment-grade bond index. Data center ABS hit $13.3 billion across 27 transactions in 2025, up 55% from 2024. UBS projects up to $900 billion in new corporate debt globally in 2026. TD Securities expects investment-grade spreads to reach 100–110 basis points, up from 75–85, partly driven by this issuance surge.

And the structures are getting creative: off-balance-sheet SPVs, GPU-collateralized loans where the hardware has already lost 50–70% of its rental value, synthetic leases. Bloomberg reports that AI bubble fears are spawning entirely new credit derivatives products for hedging big tech exposure.

CoreWeave: $29 Billion in Debt on a Company That Didn’t Exist Three Years Ago

The most leveraged bet in AI infrastructure

If you wanted to build a company that would be the first to crack in an AI downturn, you’d build something that looks a lot like CoreWeave. $29 billion in total liabilities. $9.7 billion due within 12 months. Interest expense of $311 million in Q3 alone, nearly triple the year-ago figure. Five-year CDS spreads above 640 basis points. The market-implied default probability over five years: 42%.

The company IPO’d at $40 in March 2025, below its own indicated range. Hit $187 by June. Crashed 46% in November after Q3 earnings showed widening losses and data center delivery delays. Morgan Stanley downgraded it to Equal-weight on February 21, 2026, citing execution concerns. HSBC maintains a Reduce rating with a $41 target, estimating a $9.8 billion liquidity shortfall this year. The stock sits around $88.

CoreWeave’s business model is straightforward: borrow money, buy NVIDIA GPUs in bulk, install them in data centers, lease compute back to customers. When demand is high and rising, this prints money. The 2031 bonds yield 11.5%. S&P rates the company B+, Moody’s Ba3. Speculative grade. A DA Davidson analyst wrote that the equity “may ultimately lose all its value since the entire value of the enterprise is owned by debt holders.”

CoreWeave matters beyond CoreWeave. Its debt has been packaged into asset-backed securities and sold to institutional investors. When one company with this risk profile represents a growing share of a new asset class, its distress doesn’t stay contained. It reprices the entire category.

Spending Ratios Last Seen at the Peak of the Telecom Bubble

Capex as a percentage of EBITDA

Big tech capex as a share of EBITDA is running at 45–57% for the hyperscalers. Oracle hit 86% of sales in capex for 2026, and its CDS widened above 125 basis points, levels not seen since 2009. The company plans to raise $45–50 billion in combined debt and equity this year to fund it.

For historical reference: AT&T at the peak of the 2000 telecom bubble ran capex at 72% of EBITDA. Exxon at the 2014 energy peak: 65%. These used to be asset-light companies with enormous free cash flow going back to shareholders through buybacks. Now they’re turning into asset-heavy infrastructure operators, and Barclays projects Meta’s free cash flow will drop roughly 90% in 2026. Microsoft’s FCF projected to slide 28%.

BCA Research calculated that the five hyperscalers plan to add $2 trillion in AI-related assets to their balance sheets by 2030, with annual depreciation of $400 billion. That figure exceeds their combined 2025 profits.

The Power Stocks That Priced in a Decade of Demand in 18 Months

AI energy plays vs. utilities index, 2024–2026

Vistra was the #1 stock in the entire S&P 500 in 2024. Up 258%. Constellation Energy was #3, up 130%. NRG gained 78%. These aren’t utility returns. These are stocks priced for a structural break in power demand that assumes every announced data center gets built and powered on schedule.

Constellation is down 23–24% year-to-date in 2026, hit by the Trump administration pushing tech companies toward lower electricity prices. NRG is down roughly 6%, with Zacks assigning a Strong Sell rating and expected EPS falling 22% year-over-year. Vistra has held up better, signing a long-term deal with Meta, but Morningstar pegs its fair value at $52 against a current price around $172. That’s a lot of air.

The structural problem is asymmetry. When a hyperscaler signs a 15-year power purchase agreement with an energy company, both sides are betting on the same future. But if demand slows, Microsoft can delay a data center, write down some costs, and absorb the hit against $80 billion in cash. The energy company on the other side of that PPA has sized its debt, its capex, and its generation capacity against revenue that may not arrive. They don’t have $80 billion in cash. They have debt covenants.

These companies went from boring defensive plays to momentum trades. If hyperscaler capex guidance gets cut, or even just grows slower than expected, the power stocks have the most air under them. Not NVIDIA. Not Microsoft. The companies building generating capacity for demand that a single customer controls.

What Would Change This Picture

Counter-signals worth watching

Enterprise AI producing measurable returns. Not press releases and pilot programs. Actual revenue growth at the hyperscaler level tied to AI workloads. Right now, an MIT study found 95% of companies see zero return on generative AI investments. Most of the “AI revenue” being reported is generated at a loss.

Credit spreads stabilizing. If lenders get more comfortable with AI infrastructure debt, that would be a genuine counter-signal. CoreWeave CDS at 640+ bps and Oracle CDS at 2009-crisis levels say they are not getting more comfortable.

NVIDIA growth re-accelerating. If the second derivative turns positive and growth speeds up again, that means demand is outpacing the buildout. Q4 guidance of $65 billion implies +14% sequential growth, which would be a strong signal. Watch the February 26 earnings.

GPU spot prices firming up. Falling GPU cloud prices mean oversupply. This is the most direct physical measure of whether infrastructure matches actual usage. If spot prices stabilize or rise, demand is absorbing supply.

JPMorgan argues the AI sector does not meet classic bubble criteria using a five-factor diagnostic framework, pointing to genuine structural utility. That’s worth considering. But a JPMorgan report also said you need $650 billion in revenue to make the math work, and the industry is at $30–37 billion. Both things can be true: the technology is real and the financial structure built around it is fragile.

“May ultimately lose all its value since the entire value of the enterprise is owned by debt holders.”– DA Davidson analyst on CoreWeave, November 2025

The technology gets used eventually. The fiber got used. The question is whether the capital structure built to fund this cycle survives long enough to see it happen. History’s answer, across telecom, railroads, and energy, has consistently been: the builders get wiped out, the infrastructure gets bought at pennies on the dollar, and the next generation of companies builds on top of it.