AI's Financial Circle Game: What Finance Leaders Need to See Now
The AI trade is breathtaking. A handful of names carry most of the narrative. Nvidia's market cap recently towered over the GDP of every country except the US and China. That concentration-and the way these firms buy from and invest in each other-has investors asking hard questions.
"Circular" is the word CEOs hate and analysts keep using. Because it fits. Supply commitments, equity stakes, warrants, leases, and pre-purchases are looping together. That boosts momentum-until it amplifies risk.
The Web of Circular Commitments
Nvidia's proposed 10-year plan to invest $100 billion in OpenAI and supply 10 GW of infrastructure set the tone. Then Nvidia told investors there was "no assurance" they'd reach definitive agreements. Confidence didn't crack, but doubt crept in.
Microsoft and Nvidia put roughly $15 billion into Anthropic, which plans to spend about $30 billion on Azure and Nvidia chips. AMD granted OpenAI warrants to buy AMD stock; OpenAI would run on AMD silicon. OpenAI and Oracle announced a $300 billion plan to build up to 4.5 GW for Stargate, followed by Oracle's deal to supply $300 billion of compute to OpenAI over five years.
Nine-year-old CoreWeave is 5% owned by Nvidia, buys Nvidia chips, and counts OpenAI as a customer and investor (~$350 million stake). Nvidia and xAI joined a consortium to buy Alligned Data Centers for $40 billion; alongside that, Nvidia struck a $20 billion lease-to-own chip deal with xAI funded in part through a special purpose vehicle that Nvidia also backs.
Why Markets Are Tense
Michael Burry disclosed short positions in Nvidia and Palantir, flagging customer financing tactics and invoking Enron comparisons from the dot-com era. Oracle's stock fell 30% in Q3 on concerns about delivering its five-year commitment to OpenAI-and whether OpenAI can pay for all that compute.
Interconnected balance sheets raise contagion risk. As Gregory Blotnick noted, the "Big Four" (Microsoft, Alphabet, Amazon, Meta) generated $451 billion in operating cash flow in 2024, yet a miss in AI monetization at one could ripple into Azure spend, Nvidia revenue, CoreWeave valuation, and, ultimately, OpenAI's funding capacity.
Is the Spend Worth It-or Just Well-Funded?
In early 2025, AI drew roughly half of US VC commitments. A McKinsey estimate pegs data center buildout needs at $5.2 trillion by 2030 (source). That's a lot of capex chasing uncertain returns.
Accounting muddies the water. Chips wear out or become obsolete within one to three years, but many firms depreciate them over five to six-an "accounting mismatch" flagged by Princeton's Technology Policy Clinic (analysis). Costs are often lumped with broader construction, hiding true depreciation and ROI.
Financing, for now, is plentiful. Compared to the 1990s, there's more private capital, and suppliers understand the tech, so they write creative deals. The hard part is timing and flexibility. Chips, data centers, and energy must scale in sync, regulations keep shifting, and long-term purchase agreements can lock you into yesterday's tech. If LLM growth slows, order books can snap. Creative finance stops helping and starts hurting.
What Finance Teams Should Do Now
- Map your circular exposure:
- List counterparties that are both customers and vendors. Include equity stakes, warrants, SPVs, and off-balance-sheet leases.
- Tie each to purchase obligations, compute commitments, and revenue concentration. Model a cascade: cloud spend → chip supplier → infra provider → model developer.
- Run stress scenarios: 20-40% drop in AI unit pricing, slower model adoption, or delayed data center energization.
- Fix your accounting lens:
- Align depreciation with economic life: accelerators at 1-3 years unless proven otherwise. Add obsolescence reserves.
- Break out chip capex from data center construction. Track GPU-hours deployed, utilization, and revenue/GPU-hour by cohort.
- Set impairment triggers tied to utilization, price per GPU-hour, and model efficiency gains that cut compute needs.
- Contract for flexibility, not just capacity:
- Add tech-switch and downshift options, MFN pricing, and quarterly ramp gates. Avoid unconditional take-or-pay.
- Link purchase volumes to monetization KPIs (gross profit per GPU-hour, model-driven ARR). Bake in service credits with teeth.
- Hedge power and supply risks. Keep a credible second source at 20-30% to avoid lock-in.
- Fund with eyes open:
- Vendor financing can align incentives; circular financing can inflate demand. Separate commercial need from financial optics.
- For SPVs, define recourse clearly. Cap loss exposure, require cash sweeps, and secure audit rights.
- Stage capital. Release on utilization, not promises.
- Tighten risk governance:
- Set concentration caps by vendor, model provider, and data center operator.
- Monitor early-warning signals: Azure/Cloud AI spend growth, chip lead times, average GPU-hour pricing, model efficiency improvements, and policy shifts.
- Score counterparties on liquidity, commitment load, and related-party revenue reliance.
Red Flags Worth Your Time
- Useful lives extended beyond three years for leading-edge accelerators without support.
- Large "other construction" buckets masking chip spend and depreciation.
- Customer financing that props up near-term revenue recognition.
- Big related-party transactions driving growth.
- Massive pre-purchase obligations without opt-outs tied to tech milestones.
Decision Framework: Greenlight, Stage, or Stop
- Stage-gate by 90-120 days. Release spend only when utilization and gross margin thresholds clear.
- Underwrite each capacity cohort with a payback under 24-30 months and a hurdle rate above WACC plus a volatility premium.
- Comp tie-in: reward monetized GPU-hours and model ARR, not provisioned capacity.
- Maintain renegotiation options every 6-12 months to reflect chip and model shifts.
The Bottom Line
The AI buildout is capital intensive and winner-take-most. Circular deals can grease the skids-or magnify shocks. Your edge is disciplined underwriting: real unit economics, honest depreciation, flexible contracts, and clear exposure maps. As one observer put it, buying a tech titan today means buying a slice of its AI bets. Price that risk, set gates, and keep exit ramps open.
If you're shaping the finance strategy for AI investments, this can help: AI Learning Path for CFOs
Your membership also unlocks:
