Nvidia, Meta, Tesla are worth trillions. But who is actually funding the AI boom?
AI moved from hype to habit. Valuations exploded. Chipmakers and hyperscalers pulled markets higher. Yet the more important question sits underneath the charts: who is financing the physical backbone of AI-and how secure is that funding?
Regulators are starting to say the quiet part out loud. The Federal Reserve flagged that AI infrastructure is being financed in opaque private markets. The Financial Stability Board (FSB) is probing risks in private credit. And the US Treasury is looking upstream at who the money ultimately comes from.
The signal from regulators: opacity, liquidity, and valuation risk
Fed minutes (Jan 2026): Officials linked the AI buildout to "opaque private markets," alongside broader concerns about private credit's ties to banks and insurers. Translation: big financing flows are off-screen, hard to price, and can be sticky under stress.
FSB (2026 agenda): The watchdog is assessing vulnerabilities in private credit and drafting sound practices for AI use in finance. It highlights third-party concentration, correlated exposures, cyber risk, model risk, and data quality-plus major data gaps across private finance that make risk assessment harder.
US Treasury (CFIUS): A proposed "Known Investor Program" aims to improve visibility into foreign limited partners (LPs) behind private funds-especially relevant if those funds finance sensitive AI infrastructure.
A boom built off-balance sheet
Capital needs are massive. S&P points to a bond and loan surge from tech and communications issuers, with maturities of US corporate debt rated B- and below jumping from about $56.6B in 2026 to ~$215B in 2028-right as AI capex crests. That's a refinancing wall staring straight at higher-for-longer funding costs.
Private credit has become a core funding channel for lower-rated borrowers. Lending to B- and below reached roughly $146B in 2025 versus ~$85B in broadly syndicated loans-and has topped syndicated issuance for four straight years. Much of this sits in closed-end funds with model-based marks and limited price discovery.
Performance that leans on unrealised marks
Peer-reviewed research on 262 North American private-credit funds shows a heavy tilt toward residual value (RVPI) vs. actual cash returned (DPI). Even 2015-2016 vintages still carry large unrealised components; newer vintages are 80-90% unrealised. If marks are optimistic, losses surface late.
Net of fees, senior and mezzanine funds have barely outperformed (or underperformed) liquid floating-rate ETFs in several comparisons. Meanwhile, nearly half of direct-lending borrowers have negative free operating cash flow, and payment-in-kind (PIK) interest reached ~8% of BDC interest income in 2024-both features that can smooth reported results without improving cash health.
"Little secondary market for AI loans so far"
That line matters. If AI-infrastructure loans don't trade, you lose release valves: selling down risk, hedging exposures, or discovering prices quickly. Stress becomes bilateral-extensions, covenant relief, or PIK-pushing recognition of losses into the future while IRRs hold up on paper.
UBS has already warned that AI-driven business disruption could add $75-$120B in new defaults across leveraged loans and private credit by end-2026, with private-credit default rates rising toward ~4% in baseline and higher in severe scenarios. That's the demand side. The Fed's concern is the supply side: data centers, compute, and digital energy-all funded in markets that don't clear daily.
Who is ultimately providing the money?
Follow the LP capital. Treasury's questions go to the heart of private-fund opacity: foreign LPs can back vehicles that finance sensitive infrastructure without triggering full review if they lack control rights. The proposed Known Investor Program would push for more visibility into that chain of capital.
For finance leaders, this isn't a geopolitical footnote. LP provenance affects refinancing options, deal governance, and workout behavior under stress-especially if policy risk rises around AI compute, advanced chips, and critical energy interconnects.
What finance teams should do now
For CFOs and Treasurers (issuers and sponsors)
- Map your refinancing wall: 2026-2028 maturities, rate resets, and covenant bandwidth under downside cases. Include off-balance-sheet SPVs and JV facilities tied to AI capex.
- Interrogate valuation practices: How frequently are marks refreshed? Which comps and discount rates are used? What's the policy for moving to PIK or extending maturities?
- Demand liquidity proof: Evidence of real secondary bids, historical trade prints, and time-to-cash under stress-not just manager narratives.
- Stress test energy and construction risk: Delays on power hookups, equipment lead times, EPC overruns, and change orders can push loans into amendment cycles.
- Check counterparty concentration: Single servicer, single cloud customer, single colocation market, or single insurer channel increases tail risk.
For LPs and Investment Committees
- Disaggregate DPI vs. RVPI by vintage: Track cash realization pace versus model uplift. Tie carry and incentives to DPI thresholds, not just IRR.
- Review PIK and amendment usage: Rising PIK and repeat extensions are early smoke. Ask for cohort-level migration data (performing → amended → watchlist → NPL).
- Examine sector overlap: AI exposure may hide in "data infrastructure," "digital energy," or "mission-critical real assets." Quantify aggregate AI-linked share.
- Source-of-capital diligence: Understand upstream LPs in co-mingled vehicles and sidecars. Note any constraints from evolving CFIUS policy on follow-on capital.
- Liquidity planning: Match fund cashflows to your own liabilities. Include gates, NAV financing terms, and potential margin calls in downside cases.
Banks and insurers with private-credit links
- Trace interconnections: Warehouses, total-return swaps, first-loss tranches, and reinsurance channels can transmit stress.
- Collateral realism: For AI loans, how resilient are residual values if compute pricing compresses or power costs spike?
- Hedge efficacy: If loans don't trade and indices don't track, hedge slippage is your baseline. Calibrate limits accordingly.
Key risk markers to track
- Refinancing ratios: Share of maturing AI-linked loans refinanced on time versus extended or toggled to PIK.
- Secondary activity: Actual trade volumes, bid-ask spreads, and settlement times on AI-linked credits.
- Mark dispersion: Same-asset valuation differences across managers; widening gaps signal model strain.
- Concentration metrics: Top-5 obligor, top-5 servicer, single market, and top energy provider exposures.
- Cash conversion: DPI progression by vintage; rising RVPI without DPI is a warning sign.
Bottom line
The AI buildout is real, large, and financed increasingly in places that do not trade daily. That doesn't mean crisis; it does mean slower price discovery, stickier losses, and higher reliance on bilateral fixes when deals wobble.
If you manage capital or liquidity, assume limited secondary exits, model delayed loss recognition, and scrutinize who sits behind your lenders. In a market built on unrealised marks, cash is the tell.
Further reading and tools for finance leaders
Your membership also unlocks:
