Spending on AI infrastructure is growing faster than the revenue that has to pay for it, and both numbers are still climbing. That mismatch is the entire bubble debate. Bulls point to sold-out chips and record earnings. Bears point to debt, circular deals, and cash flow stretched thin. Both sides are citing real data.
Quick Takeaways
- Demand is real today. Nvidia reported $96.2 billion in quarterly revenue, up 106% from a year earlier, with Data Center revenue at $89.0 billion.
- Funding is the weak point. Analysts estimate AI capex rose from 33% of hyperscaler operating cash flow in 2023 to about 93% in 2026.
- Circular deals amplify risk. Vendors invest in customers, and customers buy from vendors. This is manageable in a boom and painful in a downturn.
- Verdict: A bubble is not proven. A financing-fragility problem is visible. The risk sits in debt and concentration, not in the technology.
| Signal | Bull Reading | Bear Reading |
|---|---|---|
| Chip demand | Revenue doubling year over year | Customers fund purchases with debt |
| Capex / cash flow | Giants can afford it | ~93% of operating cash flow |
| Vendor-customer deals | Revenue is real | Circular, hard to value |
| Adoption | Inference workloads scaling | Revenue still trails build-out |
What Does “AI Bubble” Actually Mean?
A bubble is a price detached from the cash flows an asset can plausibly produce. It is not the same as a hype cycle, and it is not the same as a bad quarter.
Real technology can still sit inside a bubble. Railroads, fiber, and the web all changed the world, and all three produced spectacular overbuilding and investor losses. So “AI is useful” and “AI is a bubble” can both be true.
Use three tests:
- Valuation vs. earnings: Do prices assume revenue that does not exist yet?
- Funding quality: Is the build-out paid from cash flow or from leverage?
- Demand durability: Do end users pay for the output at scale?
The Spending Numbers
Estimates differ by source, so treat any single figure with caution. CreditSights projected about $602 billion in 2026 capex for the top five hyperscalers, with roughly 75% going to AI infrastructure. By June 2026, sell-side estimates for the same five companies had climbed to $697 billion, up $173 billion since the start of the year. A broader UBS projection puts total hyperscaler capex at $1.009 trillion in 2026 and $1.619 trillion in 2028.
The cumulative scale matters more than any one year. UBS expects about $4.1 trillion in hyperscaler capex from 2026 through 2028, more than three times the $1.292 trillion spent over the previous six years.
Gartner supplies the demand-side view. It forecasts global AI spending across all categories of $2.52 trillion in 2026, up 44% year over year.
Historical Context Cuts Both Ways
Bulls note that the build-out is not yet extreme relative to the economy. Goldman Sachs found AI capex at roughly 0.8% of GDP, versus peaks of 1.5% or more in earlier technology booms. Bears reply that earlier booms ended in busts.
| Metric | Late-1990s Telecom Boom | 2026 AI Build-Out |
|---|---|---|
| Asset | Fiber, switches | GPUs, HBM, power, data centers |
| Asset life | Decades (fiber) | 3–6 years (accelerators) |
| Primary funders | Debt and equity | Cash flow, bonds, private credit |
| Capex / GDP | 1.5%+ at peak | ~0.8% (Goldman estimate) |
| Customer base | Many unproven carriers | A few giant labs plus broadening enterprise |
The asset-life row deserves attention. Fiber stayed useful for decades after the telecom crash. GPUs lose value as new generations ship. A cash-flow shortfall hurts more when the underlying assets depreciate quickly.
Nvidia as the Scoreboard
Nvidia is the cleanest read on whether demand is real, because every dollar of AI compute spending flows through it.
Nvidia guided to $108.0 billion in third-quarter fiscal 2027 revenue, plus or minus 2%, and assumed no Data Center compute revenue from China. Gross margin was 75.0%.
The customer mix is the bullish detail. Nvidia reported that revenue from AI clouds, industrial, and enterprise customers grew 138% year over year, while hyperscale revenue more than doubled. Demand is broadening beyond a handful of buyers.
That is evidence against a pure bubble. A pure bubble depends on a closed loop of a few players. It does not usually show a widening customer base.
The Circular Financing Problem
Here the bear case gets sharper. In circular financing, a supplier invests in a customer, and the customer spends that money on the supplier’s products.
Nvidia reportedly explored backstopping financing for an OpenAI data center, and the prospect of more circular deals unsettled the debt market. One analysis describes Nvidia as backing $105 billion in OpenAI financing, tied to a single Ohio campus. It also reports that OpenAI agreed to buy about $300 billion of Oracle compute over five years starting in 2027, far beyond its then-reported revenue.
The risk is correlation. If one link weakens, the others weaken together. If monetization disappoints at a major AI lab, the impact cascades to suppliers, cloud providers, and ultimately chip demand.
The counterargument has merit. Economist Noah Smith argues that in the Nvidia-OpenAI case, revenue flows one way, since OpenAI pays Nvidia for chips it needs to sell its services, and that the “circular” label is applied loosely. UBS estimated the OpenAI-Nvidia arrangement at up to 13% of Nvidia’s projected 2026 revenue, and noted that the performance-based structure differs from the fixed commitments of the telecom era.
Institutional warnings are growing anyway. Fitch has named an AI correction a top global credit risk, and Michael Burry has built short positions on AI infrastructure names. Read short positions as one investor’s view, not as evidence.
Debt Is the Pressure Point
Equity bubbles hurt shareholders. Debt bubbles hurt balance sheets, lenders, and sometimes the wider economy.
Morgan Stanley expects hyperscaler debt issuance to exceed $400 billion. Hyperscalers increasingly lean on debt markets to bridge the gap between rising AI capex and internal free cash flow. Leasing adds another layer. CreditSights expects a continued shift toward leasing data centers, which reduces cash capex but moves commitments out of view.
Watch these indicators:
- Capex / operating cash flow above 90%
- Credit spreads on AI-linked bonds widening
- Lease commitments growing off balance sheet
- Customer concentration in a handful of labs
The Revenue Gap
Infrastructure spending only pays off if end-user revenue arrives. Bain’s estimate sets the scale. AI companies would need $2 trillion in annual revenue by 2030 to fund the required infrastructure, and Bain projected a shortfall of roughly $800 billion.
Revenue is growing, but from a small base. One analysis noted that OpenAI’s roughly $20 billion in annualized revenue equals only about 3% of projected 2026 hyperscaler capex. Whether enterprise adoption closes that gap is the central open question.
How to Stress-Test the Bubble Thesis Yourself
Run a back-of-envelope payback model. The script below shows how much annual AI revenue a given capex level must generate. All inputs are assumptions you should replace with current filings.
# payback_check.py — sanity-test AI infrastructure economics
def required_revenue(capex_bn, useful_life_yrs, target_roi, opex_ratio):
"""
capex_bn: total infrastructure spend in $B
useful_life_yrs: GPU depreciation window (3-6 typical)
target_roi: annual return required on invested capital (e.g., 0.10)
opex_ratio: power, staff, networking as a share of revenue (e.g., 0.35)
"""
annual_depreciation = capex_bn / useful_life_yrs # straight-line
required_return = capex_bn * target_roi # cost of capital
gross_need = annual_depreciation + required_return # profit before opex
return gross_need / (1 - opex_ratio) # revenue after opex
# Example: $700B capex, 5-year life, 10% ROI, 35% opex ratio
print(f"Required annual revenue: ${required_revenue(700, 5, 0.10, 0.35):,.0f}B")
With these inputs the output is about $323B of annual revenue for one year’s capex alone. Rerun it with a 3-year life, and the requirement jumps to about $466B. Asset life is the most sensitive variable.
Prompt Template for Earnings Analysis
Use this with any frontier model (set temperature to 0.2 for consistent extraction):
You are a financial analyst. From the attached 10-Q, extract:
1. Capex for the quarter and trailing 12 months
2. Operating cash flow for the same periods
3. Debt issued, lease commitments not yet commenced
4. Revenue concentration (top customers, % of total)
Return a table, then flag any metric where capex exceeds 80% of
operating cash flow. Cite page numbers. Do not estimate missing values.
Boom, Bubble, or Both?
The evidence supports a split verdict:
| Question | Evidence | Reading |
|---|---|---|
| Is the technology real? | Rising enterprise and inference usage | Yes |
| Is demand real today? | Nvidia revenue +106% YoY | Yes |
| Is the funding sustainable? | ~93% of operating cash flow; heavy debt | Uncertain |
| Are valuations tied to cash flow? | Depends on 2027–2030 revenue | Unproven |
| Is concentration risk high? | A few labs and vendors intertwined | Elevated |
Even the bubble-skeptic camp is divided. A Bank of America survey in October 2025 found 54% of fund managers considered AI a bubble. Sam Altman said in August 2025 that he believed an AI bubble exists. Believing in a bubble and believing in the technology are compatible positions.
A reasonable framing: the technology is not the bubble. The financing structure might be. If revenue catches up, the spending looks prescient. If it stalls, the damage lands first on leveraged and concentrated players.
Frequently Asked Questions
Is the AI bubble about to burst?
Nobody can time it. Current data show accelerating chip demand, which argues against an imminent collapse. The vulnerability is funding: heavy debt and cash flow stretched thin. A shock would more likely come from credit conditions or a missed revenue target than from falling demand.
How is the AI boom different from the dot-com bubble?
The biggest spenders today are profitable giants, not cash-burning startups. Nvidia reports 75% gross margins. The differences that worry analysts are shorter asset lives (GPUs vs. fiber) and rising use of debt and circular deals.
What is circular financing in AI?
A vendor invests in a customer, and the customer spends the money on that vendor’s products. Examples include chipmakers taking equity in AI labs that then buy their chips. It can inflate apparent demand and magnify losses if the customer cannot pay.
What would signal that the AI bubble is deflating?
Watch for slowing quarter-over-quarter chip revenue, widening spreads on AI-linked debt, delayed data-center leases, and hyperscalers cutting capex guidance. Any one is a warning. Several together would change the picture.
