The real cost of financing AI compute: WACC from hyperscalers to tier 3 neoclouds

The same GPU costs about 11% a year to fund at a hyperscaler and roughly 23% at a small neocloud. The gap is not about the hardware. It is about who can borrow, from whom, and how much equity has to sit underneath the debt.

In 2026, AI compute is financed in four distinct tiers. Each tier pays a different price for debt, a different price for equity, and ends up with a different asset-level weighted average cost of capital (WACC). Knowing where an operator sits on that ladder explains its pricing, its growth limits, and which financing actually makes sense for it.

Three findings stand out:

• Debt cost more than doubles from the top of the market to the bottom, from about 5% for hyperscalers to 15 to 20% for small operators shut out of bank and OEM programs.
• Asset WACC roughly doubles too, from about 11% at hyperscalers and tier 1 neoclouds to about 23% at tier 3.
• At tier 3, borrowing barely helps. Debt costs half as much as equity but must be repaid in three years, so moving asset risk off the operator’s balance sheet can matter more than finding the cheapest loan on offer.

Four segments, defined by access to capital

We segment the market by what an operator can borrow against, not by how many GPUs it owns. Access to capital sets the cost of capital, and the cost of capital shapes everything downstream.

Exhibit 1. Four segments, by access to capital

Segment	Financing access	Typical offtakers	Examples
Hyperscalers	Investment-grade corporate balance sheet and bonds	Their own products, frontier AI labs	AWS, Microsoft Azure, Google Cloud, Oracle OCI
Tier 1 neoclouds	Rated, asset-backed GPU debt in ring-fenced SPVs	Hyperscalers and frontier labs on multi-year contracts	CoreWeave, Nebius
Tier 2 neoclouds	Syndicated bank and private-credit facilities, OEM financing	Labs, hyperscaler overflow, large enterprises	Nscale, Lambda
Tier 3 operators	No bank or OEM access; specialist lessors, on-chain lenders, equity	AI startups, enterprises, marketplaces on short terms	Regional and independent GPU clouds

Source: GPU Capital.

Our working definitions: tier 1 operators hold rated GPU-backed facilities sized in the billions. Tier 2 operators raise unrated syndicated facilities of roughly $100 million or more and qualify for OEM financing. Tier 3 operators finance clusters under about $50 million with no bank facility or OEM program.

GPU financing rates: who lends to whom

Each step down the ladder loses one source of cheap capital, and each lender prices the risk it can actually underwrite.

Exhibit 2. Who lends to whom

Segment	Lenders	Offtakers
Hyperscalers	Investment-grade bond market, mostly own cash flow	Own products and frontier AI labs
Tier 1 neoclouds	Banks and institutions via rated, asset-backed GPU facilities	Hyperscalers and frontier labs on multi-year contracts
Tier 2 neoclouds	Bank and private-credit syndicates, OEM financing programs	Labs, hyperscaler overflow, large enterprises
Tier 3 operators	Specialist lessors, on-chain lenders, founders’ equity	AI startups and enterprises on 12 to 18 month terms

Source: GPU Capital.

Hyperscalers fund GPUs mostly from operating cash flow, topped up with bonds. Amazon, Alphabet, Meta and Oracle had issued nearly $223 billion in bonds by August 20, 2026, more than double all of 2025.

Tier 1 neoclouds borrow against GPUs and customer contracts inside ring-fenced SPVs. CoreWeave’s cost of debt fell from about 15% on its first facility to a 5.9% fixed tranche on its $8.5 billion DDTL 4.0, rated A3 in March 2026, the first investment-grade GPU-backed financing.

Tier 2 neoclouds raise syndicated senior secured facilities. Nscale’s $1.4 billion GPU facility bears interest at Term SOFR plus 5.00%, and Lambda raised about $1 billion of private debt to buy Nvidia infrastructure that Microsoft will lease. Those financings differ greatly in size, structure and credit quality from ours.

Tier 3 operators sell to startups and enterprises whose contracts are short. Most neocloud offtake agreements with AI developers run 12 to 18 months, and rating agencies heavily discount revenue beyond that term. Without a creditworthy, long-dated customer, the cheap rungs of the ladder are out of reach.

The OEM financing cliff

The sharpest break in the ladder sits between tier 2 and tier 3. Operators that lose access to bank syndicates and OEM financing programs see their cost of debt roughly double.

Source: Damodaran (January 2026), CoreWeave and Nscale financing disclosures, AltStreet; tier 3 pricing from GPU Capital deal flow. As of September 2026.

Show data

The cost of debt roughly doubles between tier 2 and tier 3

Segment	Low	Central estimate	High
Hyperscalers	4.75%	5.3%	8%
Tier 1 neoclouds	5.9%	7%	11%
Tier 2 neoclouds	8%	8.9%	12%
Tier 3 neoclouds	15%	17.5%	20%

In our deal flow, smaller neoclouds typically cannot access OEM financing at all. They are pushed to specialist lessors and to on-chain lenders, where pricing reflects both the collateral and the thin credit file behind it.

Across our deal flow, tier 3 debt from specialist lessors and on-chain lenders typically prices at 15 to 20% APR, and often above 20% once fees are included. By comparison, private-credit facilities for established operators price at SOFR + 400 to 700 bps, or 8 to 12% all-in.

Headline APR also understates the true cost. Upfront fees and cash parked in reserve accounts raise the effective rate, and short amortization schedules pull cash out of the business early.

Debt is only half the stack: asset WACC

Comparing interest rates alone flatters the smaller operators, because every dollar of debt sits on top of equity that costs far more.

The right yardstick for a GPU deployment is its asset-level WACC: the blended cost of the debt and equity that fund the GPUs inside the SPV, weighted by loan-to-value.

WACC of the asset equals D over V, times the cost of debt k d, times one minus the tax rate t, plus E over V, times the cost of equity k e.WACC_asset = (D/V) × k_d × (1 − t) + (E/V) × k_e

Here D/V is the loan-to-value, k_d the cost of debt, t the tax rate and k_e the cost of equity. For loss-making neoclouds the tax shield is close to zero, so we set t to zero for tiers 1 to 3.

• Debt
• Equity

WACC = (D/V) x cost of debt x (1 - t) + (E/V) x cost of equity. Debt shares: 11%, 70%, 65% and 70%. Costs of equity: 11.5%, then 20%, 30% and 35% (assumed). Hyperscaler debt is 4.0% after tax; t is zero for tiers 1 to 3.

Source: GPU Capital calculation from Damodaran (January 2026), company filings and GPU Capital deal flow.

Show data

Asset-level WACC roughly doubles from hyperscalers to tier 3

Segment	Debt	Equity	Total
Hyperscalers	0.4	10.2	10.7%
Tier 1 neoclouds	4.9	6.0	10.9%
Tier 2 neoclouds	5.8	10.5	16.3%
Tier 3 neoclouds	12.3	10.5	22.8%

Exhibit 5. Inputs to asset-level WACC by segment, 2026

Segment	Debt share (LTV)	Cost of debt	Cost of equity	Basis
Hyperscalers	11%	5.3% pre-tax, 4.0% after tax	11.5%	Damodaran, Software (Internet), January 2026
Tier 1	70%	~7%	20% (assumed)	60 to 70% LTV playbook; blend of IG and older facilities
Tier 2	65%	~8.9%	30% (assumed)	SOFR + 5.00% with SOFR near 3.9%
Tier 3	70%	~17.5%	35% (assumed)	50 to 70% advance rates; 15 to 20% APR from GPU Capital deal flow

Source: Damodaran (January 2026), Nscale Form S-1, AltStreet and Medium as linked; GPU Capital deal flow and estimates.

Two patterns stand out. Tier 1 reaches roughly the hyperscaler’s WACC by a different route: cheap, customer-backed debt instead of cheap equity. From tier 2 down, the cost of equity drives the result, and at tier 3 the debt itself becomes the largest cost.

One caveat: tier 1 WACC is measured at the asset (SPV) level. Corporate-level WACC is higher, because the parent’s equity sits beneath the SPV debt.

Rising rates hit the ladder unevenly

The Fed raised its target range to 3.75 to 4.00% on September 16, its first hike since 2023, and held it there on October 28 by a 10 to 2 vote. Both dissenters wanted another increase. In the September projections, 16 of the 18 officials expected at least one more increase by the end of 2026.

The 10-year Treasury yield was 4.98% on October 29, up from 4.19% on January 2, and in late September spreads on AI-related bonds were reported near 115 bps versus about 78 bps for corporates broadly.

The impact depends on where an operator sits:

• Hyperscalers and tier 1 reprice with the rate curve on every new bond or facility. Because hyperscaler inputs above use January 2026 averages, we estimate the latest moves add roughly a point to their asset WACC.
• Tier 2 carries floating, SOFR-based debt that reprices immediately with each Fed move.
• Tier 3 debt barely moves, because it is priced on scarcity and credit risk rather than Treasuries.

The bigger effect is on access. Private-credit investors reportedly expect some neocloud projects to become harder to finance, because operators have less cushion to absorb higher costs. For tier 3, a tighter credit market matters more than a higher rate.

When debt stops helping

At tier 3, leverage buys less than it appears. Debt at 15 to 20% costs about half as much as equity, but it must be repaid fast, and every dollar borrowed still needs expensive equity underneath it.

The cash burden is the real constraint. On $100 of GPUs financed at 70% LTV and 17.5% over three years, year-one debt service is about $36: roughly $23 of principal plus $12 of interest. That cash leaves the business in the same months the operator is still ramping utilization.

Exhibit 6. Borrowing versus full-stack ownership for a tier 3 operator

Borrow: typical tier 3 stack	Full-stack ownership: one source for the whole asset
Loan: 70% of capex at 15 to 20% APR, often more with fees	100% of the hardware funded
Short amortization, reserve accounts	No loan amortization or reserve account
Operator keeps residual and obsolescence risk	Provider owns residual and obsolescence risk
Fixed payments whether GPUs are busy or idle	Paid from the cash the GPUs generate
	No interest-rate exposure for the operator
Equity: 30% of capex, highest-cost capital	Operator equity stays free for growth

Source: GPU Capital.

Full-stack ownership funds 100% of the hardware, so it replaces the whole capital stack, not just the loan. More importantly, it moves risk. With a loan, the operator keeps residual-value risk, obsolescence risk, re-leasing risk after the contract ends, and a fixed payment due whether the GPUs are busy or idle. With full-stack ownership, the capital provider owns the hardware and those risks, and its return comes from the cash the GPUs actually generate.

So the right test is not which option carries the lower headline rate. It is total cost against the risks shed: a loan leaves all of them with the operator, while full-stack ownership takes them off its balance sheet and frees its equity for the next cluster.

At GPU Capital we screen deployments with an internal measure, the GPU Capital Yield Score, and we finance deployments of $1 million to $5 million, up to the full capital stack, that are often too small and too contract-specific for programmatic lenders.

Why each segment’s financing makes sense

None of these structures is irrational. Each lender prices the risk it can actually underwrite, and the core rule is simple: lenders underwrite the customer, not the GPU.

• Hyperscalers fund GPUs as part of a diversified cloud business with its own demand. Bond investors lend against the whole company, so the GPUs inherit the cheapest capital in the market.
• Tier 1 neoclouds borrow cheaply because their collateral is really a contract. A multi-year commitment from a hyperscaler or frontier lab lets rating agencies look through the neocloud to its customer.
• Tier 2 neoclouds have real scale and credible offtakers but thinner contract coverage. Lenders charge a spread for renewal and refinancing risk, and OEMs extend financing to keep their own hardware moving.
• Tier 3 operators sell to startups and enterprises on short terms. Banks cannot hold fast-depreciating collateral cheaply, and large private-credit funds seldom write tickets of $1 million to $5 million. Scarce capital commands the highest price.

For lenders, the flip side is the opportunity. The same scarcity that raises tier 3 costs creates the widest spread for a lender that knows how to underwrite hardware, operators and offtakers at small ticket sizes. It also carries the highest risk of loss.

What this means for neoclouds and investors

For neocloud operators: know your asset WACC before comparing financing offers. A loan quote tells you only half the cost; the equity it requires, the amortization schedule and the reserve account tell you the rest. If you are shut out of bank and OEM programs, compare every offer on total cost and on the risks it leaves with you, not against the headline rates larger players pay.

For lenders: spreads in GPU financing are widest in the segment that large lenders find hard to serve efficiently, and so are the risks: short contracts, thin credit files and collateral that loses value every year. Lending there takes hardware-level underwriting, disciplined structures and fast-payback deployments, not just capital.

Operators: can’t get bank or OEM terms? Apply for financing for your next cluster.

Methodology

All figures are as of September 2026 unless dated otherwise, and illustrative. Hyperscaler inputs use Damodaran’s January 2026 industry averages as a proxy and predate the latest rate moves. Debt costs for tiers 1 and 2 come from disclosed facilities and published market data; tier 3 debt pricing reflects GPU Capital’s deal flow. Costs of equity for tiers 1 to 3 are GPU Capital estimates reflecting the returns growth and venture investors typically require; readers should substitute their own. SOFR is approximated from the Federal Reserve’s 3.75 to 4.00% target range set on September 16, 2026 and held on October 28, 2026. Tax shields are set to zero for loss-making neoclouds.