Is AI Infrastructure Losing Cheap Capital? Financing, Midterms and the Next Investment Cycle

As of November 4, 2026, falling share prices, negative free cash flow and a closed financing market must be treated as separate propositions. They can reinforce each other, but each requires evidence. This report examines US-led financing and transmission into 2027–2028, separating recent operating results, completed transactions and conditional policy outcomes.

The tension is tangible: large technology companies maintain substantial investment plans while some are spending more than current operating cash generation. Yet asset-backed structures with customer contracts can still obtain relatively inexpensive funding. For the next uncompleted project, the decisive questions are who supplies equity, who pays for delay and who ultimately pays for compute.

1. Define withdrawal before trying to measure it

A broad retreat should produce several observable changes: fewer new financings, persistently worse terms for comparable borrowers, cancellations and reduced actual investment. One failed transaction establishes a problem for that borrower, not closure of an entire market. A comprehensive series covering private credit, equity flows, repayments and fund drawdowns is not available in this research; there is therefore no purported industry-wide net capital outflow number.

Four funding channels need separate treatment: operating cash from established businesses, parent-company debt or equity, project borrowing against assets and contracts, and investment-fund capital. Money raised through one channel often purchases equipment through another. Adding platform fundraising goals to loan commitments, hyperscaler capex and chip revenue double-counts money moving through the system.

The August NVIDIA partnership with six financial institutions illustrates the distinction. Its more-than-$500 billion mobilization goal was announced through MOUs subject to final agreements. It shows an effort to organize financing, not $500 billion already received, and cannot be added uncritically to transactions inside the participating institutions.^[9]

2. Investment persists as internal funding coverage weakens

Microsoft generated $58.000 billion of operating cash in the September quarter and spent $39.367 billion on cash additions to property and equipment. The $18.633 billion difference compares with $25.663 billion a year earlier. Cash additions grew 103.0%, versus operating cash growth of 28.7%. This calculation is not lease-inclusive investment or a comprehensive measure of cash available to shareholders.^[5]

Meta reported $32.79 billion of quarterly operating cash against $34.91 billion of capex including finance-lease principal, leaving negative $2.12 billion of FCF. Its annual capex range narrowed upward to $135–145 billion from $130–145 billion. Alphabet reported $42.3 billion of quarterly operating cash, $51.4 billion of capex and negative $9.1 billion of FCF, retaining $195–205 billion of annual capex guidance.^[6][8]

Amazon requires a different period: trailing-twelve-month operating cash was $168.300 billion, net property investment $210.657 billion and FCF negative $42.357 billion, versus positive $14.788 billion a year earlier. It retained approximately $220 billion of 2026 cash capex guidance. These are group figures, including activities beyond AI.^[7]

The table is deliberately not summed. Quarterly and TTM periods differ, as do lease treatments. The shared implication is that substantial asset investment continues, while a portion increasingly requires reserves, external capital or partners rather than current internally generated cash alone.

Negative FCF is not proof of uneconomic assets: construction precedes billing. Alphabet also identified a $5.2 billion EC fine payment and TPU inventory accumulation as cash headwinds, while TTM FCF remained positive $19.7 billion. The decisive test is whether operating cash improves after delivery, not whether one quarter is negative.^[8]

3. Inexpensive funding still exists, with conditions

On October 28, the Fed held its policy range at 3.75–4.00%. November 2 nominal Treasury yields were 4.59% for two years, 4.91% for ten years and 5.28% for thirty years. The ten-year yield had fallen from 5.05% on October 28, undermining any claim that all recent funding costs moved inexorably higher. Treasury yields, the policy rate and SOFR are different benchmarks.^[1][2]

CoreWeave’s March DDTL 4.0 is a counterexample to the disappearance of lower-cost finance: an $8.5 billion facility, initially permitting approximately $7.5 billion of borrowing, with a SOFR-plus-225bp floating tranche and an approximately 5.9% fixed tranche, maturing in March 2032. Capacity rises with asset stabilization; a facility ceiling is not cash drawn.^[10]

HPC assets and a customer contract support that project financing. Its A3/A(low) ratings are not the parent’s ratings. The transaction cannot be compared mechanically with another issuer’s unsecured bond at a different date. The lender is underwriting a defined cash-flow and collateral package.

Unenergized projects with weaker customers or contracts shorter than asset payback may not receive equivalent terms. Nor does a low convertible coupon establish low economic cost: conversion rights, dilution, issuance discounts, fees and hedging belong in the assessment alongside cash interest.

4. Deals show higher coupons and shorter money alongside continuing access

Oracle offers a same-issuer comparison. Roughly ten-year bonds issued in September 2025 and February 2026 carried 5.20% and 5.70% coupons; roughly thirty-year tranches carried 5.95% and 6.70%. These increases do not isolate credit deterioration from benchmark-rate changes. February issuance also included floating debt at compounded SOFR plus 111bp, demonstrating continuing access.^[12]

CoreWeave’s August DDTL 5.5 priced at SOFR plus 550bp, 325bp above the floating margin of DDTL 4.0. Its approximately five-year debt finances customer contracts averaging roughly three years. Lenders accepted renewal and re-leasing risk at a different price. These are different risk packages, not a controlled time series of one borrower’s credit spread.^[10][11]

Nebius signed a $3.5 billion secured Citi facility on October 22, with a 364-day maturity. The initial Term SOFR margin is 275bp, rising as high as 350bp and subject to other adjustments. Near-term liquidity is available, but the maturity creates a refinancing task. Management’s plan to obtain future asset financing against investment-grade customer commitments is not completed long-term funding.^[13]

Its August convertible terms make a separate point: the 2030 tranche has a 0.50% coupon but reaches 110% of original principal at maturity; the 2034 tranche pays 4.50% and reaches 125%, unless repurchased, redeemed or converted earlier. Lower current interest can defer rather than eliminate economic cost.^[14]

Partnerships can likewise retain sponsor exposure. In the Hyperion arrangement, Meta’s initial lease is four years, alongside a conditional, capped residual-value guarantee covering the first sixteen operating years. External investment and continuing contingent obligations coexist.^[15]

5. Bank lending is selective, not uniformly closed

The October SLOOS, released November 2 and covering third-quarter changes, reported broadly unchanged large/middle-market C&I standards, tighter small-business standards, and tightening in commercial real estate and nonbank financial-institution lending. Weaker C&I demand partly reflected investment weakness and borrowers choosing other funding channels; it cannot all be attributed to rejection by banks.^[3]

Table 1 makes the scale clear. Four of sixty banks tightened large/middle-market standards and three eased: net tightening of 1.7 percentage points. Small-business figures were four versus one out of fifty-nine, or 5.1 points. Among large/middle-market loans banks were willing to approve, five widened spreads over funding costs while fourteen narrowed them: net negative fifteen points. Better credits can receive cheaper terms while riskier borrowers face premiums.^[4]

This survey is not an AI loan index. Its relevance is the funding chain: private-credit and other nonbank lenders also use bank liquidity. Tighter upstream terms can produce lower project leverage, greater equity requirements or preference for contracted assets before they produce a reduction in headline hyperscaler budgets.

6. How a financeable project loses its cushion

This is an illustrative model, not an issuer forecast. Assume $1 billion of investment, $700 million debt and $300 million equity; stabilized annual cash flow available for debt service, or CFADS, of $150 million; and annual scheduled principal of $70 million. CFADS is assumed to be after operating and maintenance needs, not EBITDA. Taxes, fees, hedges, reserve requirements and terminal equipment value are excluded.

At 6%, interest is $42 million and debt service $112 million: coverage is 1.34 times, leaving $38 million. At 9%, debt service is $133 million, coverage 1.13 times and residual cash $17 million—55.3% lower despite unchanged operations. If CFADS also declines 20%, coverage falls to 0.90 times with a $13 million shortfall.

A lender can tighten without raising rates: reduce debt from 70% to 60% of cost. Required equity rises from $300 million to $400 million, or 33.3%. At 9% interest and annual principal equal to 10% of debt, coverage returns to 1.32 times. The safer loan requires more sponsor cash or potentially greater dilution; identical total project funding can produce different outcomes for existing shareholders.

Timing can dominate a rate cut. With $700 million fully drawn at 9%, a six-month delay adds $31.5 million of interest alone. A 100bp rate reduction saves $7 million annually. Delayed customer billing and possible contractual compensation would add further pressure, but are not quantified here without a specific contract. Energization and acceptance terms can therefore matter more than a modest change in policy rates.

7. Midterms change constraints, not project arithmetic

The source-verification cutoff is November 4, 2026, 09:55 Asia/Shanghai. This is a conditional policy analysis, not live vote reporting, an election forecast or a claim about final congressional control. The transmission chain runs from legislative and regulatory decisions to taxes, delivery dates and cash flows, and then to lending terms. A party label is not a financing rate.

Congress can change tax and spending legislation, but ordinary legislation generally needs both chambers and presidential approval; overriding a veto requires two-thirds support. An opposition majority does not automatically repeal existing incentives. The president and agencies, state regulators and local permitting authorities also have distinct responsibilities.^[18]

Divided government can delay new fiscal measures and increase scrutiny without cancelling private investment already supported by enforceable contracts. Conversely, unified government can ease coordination without ensuring cheap long-term debt or timely electricity connections. Neither outcome directly determines Federal Reserve rates, Treasury term premiums or customer demand.

Fiscal expansion expectations could raise long-term discount rates even under construction-friendly policies; budget friction could raise reserve requirements, while a durable bipartisan energy agreement could reduce waiting times. These are conditional mechanisms, not a prediction that Treasury yields must move in one direction.

The relevant uncertainty is contractual: a longer approval process postpones billing while interest continues, and a changed cost-allocation rule can require additional equity. Those channels can affect investment before a company's headline capex budget changes. Projects dependent on unapproved incentives or unsecured power are more exposed than contracted projects with settled funding; that is a conditional judgment, not a statistical finding across all projects.

8. Tax timing and electricity allocation reach the cash-flow model

The existing tax baseline matters. January IRS guidance describes permanent 100% additional first-year depreciation for eligible property acquired after January 19, 2025. Permanent means no automatic sunset in current law, not immunity from future amendments. Eligibility does not cover every building or expenditure at a data-center campus.^[16]

Depreciation changes tax timing and the present value of deductions; it does not reimburse the entire asset cost. Illustratively, $100 million of fully eligible investment at an assumed 21% rate yields a $21 million current tax shield if taxable income is sufficient and no other restriction applies. Without usable taxable income, the cash benefit can be deferred. The 21% is a model assumption, not an issuer's effective rate. A tax shield cannot substitute for paying customers.

On June 18, FERC required six regional grid operators to justify or reform large-load tariffs. Its categories address study speed, cost shifting, behind-the-meter generation and flexible loads. This is regulatory action, not proof that individual campuses have obtained electricity approval; subsequent regional tariffs and project decisions still matter.^[17]

Making new loads pay more of grid upgrades can increase initial investment and guarantee requirements. If a clear allocation also makes connection dates more credible, financing risk can fall. Both effects can coexist. The investment question is who pays upgrades, whether minimum power payments apply and whether customers accept pass-through costs—not simply whether a politician supports data centers.

Where revenue contracts prohibit electricity-cost pass-through, higher power prices reduce CFADS. Where pass-through is permitted, customer affordability remains relevant. Permitting uncertainty can increase capitalized interest and delay collections even without a change in policy rates.

9. What would change the conclusion over twelve to twenty-four months?

The base interpretation is continuing investment with differentiated financing and tighter return scrutiny. It depends on creditworthy customers performing, capacity being delivered in stages, and project cash gradually reducing reliance on refinancing. It does not require every neocloud to survive or imply that every equipment purchase earns an adequate return.

A broader contraction would become more convincing if several major customers cut capex or cancel committed capacity in 2027 while spreads, collateral demands and equity requirements worsen for comparable risks. Sustained cash growth from delivered capacity, stable renewals and better matching of debt and contract maturities would weaken that concern.

Track actual cash investment and operating-cash coverage rather than budgets alone; drawn amounts, net proceeds, maturities and guarantees rather than facility ceilings; and energization, acceptance, customer payments and renewals rather than announced gigawatts. After elections, follow enacted law, final tariffs and approvals, not campaign statements as though already effective.

The downside feedback loop runs from disappointing customer monetization to shorter or repriced contracts, lower equipment residual values, shrinking borrowing bases, equity calls and construction delays that postpone revenue again. The upside countercase is strong utilization and paid demand encouraging prepayments or more durable contracts, reducing lenders' risk premiums. Both require evidence of actual payment.