A $35 billion AI financing: who repays the principal within five years?

The central financing question is whether compute can repay its capital before equipment loses economic value. The AI XPV platform from Broadcom, Apollo and Blackstone starts with a $35 billion capital solution supporting Anthropic’s previously announced expansion above 1GW. Apollo management subsequently described the relevant debt as a five-year amortizing instrument.^[1][2][4]

Annual principal recovery reduces reliance on selling equipment at maturity. It also requires customer payments to arrive early enough. Lenders want repayment, customers want affordable compute and suppliers want system sales. Those interests align only if operating cash covers both running costs and capital recovery.

Funding an existing plan is not a second order

The June 9 joint announcement describes deployment at Fluidstack-based sites from mid-2026 using Broadcom custom XPUs and networking. The platform aims to support over 20GW through 2028 for laboratories including Anthropic and OpenAI. That target is not fully contracted, delivered or billable capacity.^[1]

Apollo separately describes committed funding over a multi-year draw schedule and identifies A1 and A2 financing arrangements. The $35 billion is a capital solution, not necessarily cash disbursed on announcement day. Draw timing affects interest, liquidity and coordination with equipment and construction.^[2]

The transaction facilitates Anthropic’s previously announced expansion. Where procurement already sits in supplier expectations, funding chiefly improves execution confidence. Counting customer expansion, financing and chip sales as three independent demand additions overstates the economic activity.

Dividing $35 billion by more than 1GW does not establish standard construction cost. Financing uses, hardware, networking and facilities are not itemized, while capacity is stated as a lower bound. Neither relative cost nor funding for the full 20GW target can be inferred safely.

What stands between customer payments and debt service

Anthropic provides compute demand, Broadcom the custom silicon and network, deployment sites the operating platform, and investors the funding. Asset ownership, direct borrower liability, minimum payments and recourse are not fully disclosed.

Long-term commitments may make cash predictable, but contract value is not cash available for debt service. Energy, operations, maintenance, taxes and required reinvestment precede interest and principal in the economic calculation.

Fixed capacity payments can leave utilization risk with the customer; usage flexibility exposes project revenue more directly. Performance or availability failures can also reduce otherwise fixed payments. Contract terms are needed before describing long duration as credit protection.

Custom XPUs may improve workload efficiency while limiting alternative users. Redeployment value depends on software, operating expertise and demand, not merely whether a chip still runs. Neither zero nor full residual value is assumed.

Site delays separate payment for equipment from service revenue. Financing costs can accrue before acceptance. Depreciation depends on when assets are ready for their intended use, so arrival alone does not establish its accounting start.

Amortization reduces the final balance but demands early cash

In a July interview, Apollo president Jim Zelter described the Broadcom transaction as five-year amortizing debt intended to avoid residual risk. Rates, annual principal schedules and the complete capital structure are not public.^[4]

For an independent illustration, assume principal of 100, interest of 8% on the opening balance and five equal annual principal repayments of 20. Debt service is 28, 26.4, 24.8, 23.2 and 21.6. Total interest is 24 and total repayment 124. These are not transaction terms.

With the same rate but all principal due at maturity, annual interest is 8 and final-year payment 108. Total interest becomes 40. Amortization brings cash requirements forward while reducing cumulative interest and the maturity balance.

For rapidly changing hardware, that reduces the risk of large debt remaining after competitiveness deteriorates. During deployment, however, early repayments can strain liquidity. Drawdowns, grace periods, prepayments and service commencement need to align.

If illustrative first-year cash available for debt service is 35, coverage of 28 is 1.25 times. At 24 following delay, coverage is approximately 0.86 times. A temporary deployment problem can create a funding gap without permanently destroying customer demand.

Better technology can reduce the value of old assets

A new generation can lower customer cost per task while reducing renewal prices for existing equipment. Early amortization reduces reliance on resale, but does not remove price competition during the loan.

Model efficiency can reduce compute per task and also make more usage affordable. Neither falling demand nor complete volume compensation is automatic. Credit analysis needs to establish whether collection is fixed and whether greater usage produces more payment.

Assume annual revenue of 100, variable expense of 40 and fixed cash expense of 20: contribution is 40. A 20% price reduction at unchanged volume and cost cuts contribution to 20. More volume also requires variable expense, so restored revenue need not restore the original cash contribution.

Custom-compute efficiency can cushion that pressure. Hardware refresh, however, creates new capital needs. Reusable facilities are valuable only after considering changes to distribution, cooling and networking.

The favorable case is stable collection during the equipment’s economic life, declining debt and affordable adaptation to later systems. Residual asset value then becomes additional upside rather than the essential repayment source.

The participants earn different kinds of return

Broadcom potentially receives equipment and networking sales plus greater customer funding certainty. Revenue recognition still follows contracts and delivery. Financing does not become Broadcom revenue immediately, and any guarantee or repurchase obligation requires documentary evidence.

Funds and affiliates associated with Apollo and Blackstone receive investment returns; listed managers may earn management, arrangement or other fees. Financed assets are not manager revenue. Balance-sheet exposure, client capital and fees need separate treatment.

Anthropic may avoid funding all hardware upfront and secure capacity, while accepting longer payment obligations or reduced flexibility. Accounting and cash implications depend on its exact contractual role.

Deployment providers create value through timely delivery. Delay can generate claims and rescheduling across the chain. Contracts allocate those losses; it is too broad to say private credit simply assumes all technology risk.

Investment-grade language is not a return guarantee

Apollo describes the opportunity in investment-grade terms. That is the arranger’s characterization of the structure and customer exposure. A formal rating for a particular debt layer, the scope of payment support and investors’ net return require the underlying documents. Less frequent public price discovery in a private transaction does not remove asset losses.

Where financing relies on long customer commitments, lenders need to understand deductions for late equipment, inadequate performance or unavailable service. Gaps between vendor, financing vehicle and user obligations can create cash shortfalls despite a nominally long contract. The party funding a payment dispute bears a real liquidity cost.

Equipment residual value should be a secondary recovery source, not an automatic solution to contract shortfalls. Resale depends on configuration, software, relocation cost and demand. Platform-specific assets can be harder to dispose of in a stressed market.

Successful five-year amortization genuinely reduces outstanding exposure, but requires stable cash and performance during those five years. Actual principal recovery and explicit allocation of risk are stronger structural evidence than transaction size or familiar institutional names.

Equipment financing needs a recovery schedule for each drawdown and delivery cohort, rather than placing a multiyear facility entirely into year one. Later cohorts may use different chips, purchase prices and customer terms. A successful first deployment does not establish an identical return for every subsequent asset financed by the platform.

This is especially relevant when growth is rapid. Aggregate receipts can rise because new assets are added even while older assets recover capital more slowly than expected. Looking only at platform revenue can hide that deterioration. A cohort view asks how much of each original investment has been repaid, how much remains dependent on future contracts, and whether necessary refresh is funded before equity cash is distributed.

The positive case is repeatability: every new commitment finds deliverable equipment, a dependable customer and collections aligned with amortization. That permits capital to be recycled with a clearer record of performance. The adverse case is expanding commitments faster than the evidence of collection, leaving later financing to support earlier cash gaps. Nothing in the announced size alone distinguishes those cases. The 20GW ambition describes the direction of the platform; demonstrated returns require successive funded assets to complete their own cash-recovery cycles.

Replicating capacity must preserve unit economics

Standardized technology, experienced lenders and repeatable documentation can reduce later financing costs. But subsequent sites may face less attractive power, equipment or customer conditions. Volume does not establish a stable return.

The strongest counter-case is that available capital accelerates procurement ahead of useful workloads. Suppliers record sales while customers and project entities accumulate fixed obligations. Falling equipment values or delays may pressure equity and liquidity before senior lenders lose principal.

Over twelve to thirty-six months, drawn funding, sites in service, cash collection and principal outstanding should move consistently. Timely payments, amortization and affordable hardware transitions support replication. Repeated extensions, additional equity or borrowing to cover recurring shortfalls weaken it.

The $35 billion demonstrates willingness to provide capital. The more consequential outcome is recovery of that capital while equipment remains competitive, so growth does not depend entirely on the next, larger financing.