AI's Billions Are Real. The Cash Flows Behind Them Are Fragile.
Markets love a clean story: endless demand, compounding productivity, and a tidal wave of profits. That's the bull case many AI insiders are selling. Nvidia's Jensen Huang says there's no bubble. Investors like David Sacks, Ben Horowitz, and JPMorgan's Mary Callahan Erdoes call this an "investment super-cycle."
Pull back the curtain, and the financing tells a different story. Useful tech, yes. But the growth assumptions, debt structures, and circular deals now propping up AI capacity demand a sober risk review from anyone in finance.
Demand Story vs. Usage Reality
OpenAI says it's at $20 billion in annual revenue and planning roughly $1.4 trillion in data center spend over eight years. That path assumes rising, durable end-demand. Yet research shows most firms aren't seeing chatbot-driven profits, and only about 3% of people pay for AI tools. MIT economist Daron Acemoglu calls current expectations "exaggeration."
Venture investor Paul Kedrosky adds that model improvements have slowed versus the hype. If performance gains taper while costs explode, the industry's top-line narrative begins to wobble.
The Capex Wave Meets the Balance Sheet
Amazon, Google, Meta, and Microsoft are set to pour roughly $400 billion into AI this year-largely data centers. Some are dedicating about half of current cash flow to this buildout. To avoid draining cash, they're turning to private equity and debt.
Enter special purpose vehicles (SPVs). These arrangements let companies expand compute capacity while keeping much of the obligation off the parent's balance sheet. Example: Blue Owl Capital and Meta structured a Louisiana data center where a $27 billion loan is backed by Meta's lease payments. Meta owns 20% of the entity and effectively has a mortgage on the facility-but the big loan doesn't sit on Meta's balance sheet.
SPVs aren't inherently bad. But they've carried baggage since the Enron era. For a refresher on how SPVs work, see this explainer on special purpose vehicles. The key takeaway for finance teams: off-balance-sheet ≠ off-risk.
Circular Money Flows Inflate Apparent Demand
One of the more concerning patterns: circular deals that boost near-term sales but muddy true end-demand. A reported $100 billion setup between Nvidia and OpenAI effectively routes capital from Nvidia into OpenAI to build data centers that are filled with Nvidia chips. It's a subsidy loop that scales revenue optics.
CoreWeave adds another layer. Once focused on crypto mining, it now rents GPU capacity to AI companies. Deals include stock-for-capacity arrangements with OpenAI, and Nvidia has a backstop to consume unused capacity through 2032. These structures can smooth utilization-but they also risk masking weak organic demand.
What Breaks If Growth Cools
Morgan Stanley estimates Big Tech will spend about $3 trillion on AI infrastructure through 2028, with internal cash flows covering only half. If demand stalls or even steadies, capacity gets ahead of revenue, and lenders will be left holding paper tied to underutilized data centers.
We've seen rhymes before. The dot-com era built fiber for a future that hadn't arrived. If AI capacity overshoots, impairments and debt repricing follow-along with funding pressure across suppliers, landlords, and credit markets.
Early Market Signals
Some large holders are trimming. Peter Thiel reportedly exited his Nvidia stake (~$100 million). SoftBank sold nearly $6 billion of Nvidia shares. Michael Burry says the industry is leaning on "fancy accounting tricks," calling out circular customer funding and questioning "true end demand."
Even insiders are hedging their tone. Sam Altman says investor excitement is overheated. Sundar Pichai notes "elements of irrationality" and admits Google wouldn't be immune if sentiment turns.
Finance Checklist: Questions to Pressure-Test Your Exposure
- Utilization economics: What is revenue per GPU-hour and the trend? What utilization rate is required to hit target ROIC? What's your downside at 60-70% utilization?
- Customer quality: How much demand comes from counterparties funded by vendors or strategic partners? What percent of revenue is tied to stock-for-capacity or pre-buys?
- Payback math: What are payback periods by generation (H100 vs. H200/GB200, etc.)? Are assumptions realistic on model efficiency gains and cooling/energy costs?
- Cash burn vs. funding mix: How dependent is the plan on SPVs, leases, and non-recourse debt? What are the hidden recourse triggers or take-or-pay clauses?
- Pricing durability: Are discounts or credits embedded in "ARR"? How sensitive are unit economics to GPU pricing normalizing or supply loosening?
- Obsolescence risk: What's assumed salvage value if next-gen chips compress prior-gen economics faster than planned?
- Energy and latency constraints: Can sites secure long-term power at modeled rates? How do grid delays affect ramp schedules and penalty clauses?
- Off-balance-sheet obligations: Inventory every SPV, lease, and throughput guarantee. Model worst-case utilization and termination fees.
- Audit and governance: Who audits key vendors and SPVs? Are related-party and circular deals fully disclosed?
- Scenario tests: Run base, steady, and downside (flat demand) cases. Track liquidity runway, covenants, and debt service under each.
Portfolio and Credit Actions
- Watch insider and strategic selling: Large share reductions by early backers can signal a turn in risk appetite.
- Scrub "AI revenue" quality: Separate usage-based revenue from subsidized or circular flows. Re-rate names heavily reliant on the latter.
- Credit screens: Flag issuers with heavy SPV exposure, take-or-pay contracts, or backstops that can boomerang on liquidity.
- Supplier risk: Chip, colocation, and cooling vendors with concentrated customers are most exposed if utilization dips.
- Debt maturity ladders: Rising refi costs plus delayed ramps can force restructurings even in "growth" stories.
Contracts and Controls Worth Negotiating
- Pre-commit thresholds: Tie capex tranches to verified customer pre-commits and minimum revenue density per MW.
- Clawbacks and step-downs: Insert pricing step-downs if new chip generations reset cost curves faster than modeled.
- Termination flexibility: Avoid rigid take-or-pay without utilization outs. Align capacity timelines with credible demand signals.
- Disclosure riders: Require full visibility into SPVs, related parties, and any vendor-financed customer deals.
- Covenant protection: Hard guardrails on leverage, fixed-charge coverage, and capex-to-cash-flow ratios.
Bottom Line
AI is useful and will stick. But usefulness isn't the same as endlessly accelerating demand. Today's growth optics lean on debt-heavy capex, off-balance-sheet structures, and circular funding loops that can make demand look deeper than it is.
If you own, fund, or buy from this buildout, your edge won't come from belief. It will come from underwriting discipline, ruthless disclosure standards, and a willingness to say "not yet" until the unit economics stand on their own.
Further Reading and Tools
- Primer on SPVs: Investopedia: Special Purpose Vehicle (SPV)
- Curated AI tools for finance teams: Complete AI Training: Tools for Finance
Your membership also unlocks:
