# The Price of Customer Capital

**Publisher: AI Infra Credit**  
**Research cutoff: September 8, 2026; cited disclosures retain their dates**  
**Research type: Original conditional pricing model**

## The thesis: price the advance from both sides

Customer prepayments are becoming part of the financing architecture of AI. The next question is not just how much capital a customer supplies. It is what that capital costs—and whether the transaction leaves both the customer and the compute provider better off than their alternatives.

Our conclusion is constructive but specific: **there can be room for a mutually beneficial exchange between a customer's funding cost and the provider's avoided bank financing. That room can be narrow, and delivery risk can consume it.** A credible route to earlier useful compute can reopen the bargain. Calling the advance “free capital” obscures the work needed to make it repeatable.

In the hypothetical example developed here, a customer needs a discount of at least **2.26% of total contract price**, while the provider can concede up to **3.69%** without reducing its incremental equity NPV. A 3% concession works for both. These are not observed contract discounts, annual interest rates, or a market quote.

## 1. Why the question is timely

IREN's August 27, 2026 update described recent customer prepayments equal to 45–55% of estimated GPU and ancillary capex. Its footnote means amounts contractually payable before service delivery, not necessarily cash already collected. It excludes other campus investment and notes that contract terms vary. [IREN FY26 update, note 4](https://www.sec.gov/Archives/edgar/data/1878848/000187884826000051/irenreportsfy26results.htm)

Nebius's August 12, 2026 shareholder letter reported that roughly 70% of deals closed in Q2 included prepayments, covering 50–60% of associated capex. Most of those contracts concerned capacity coming online later. These are company-reported cohorts, not an industry-wide prevalence estimate. [Nebius Q2 shareholder letter](https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-2.htm)

Neither observation identifies the discount each customer received, its cost of capital, the incremental value of delivery priority, or the expected loss on its advance. A prepayment percentage alone cannot establish who got the better deal.

## 2. Specify the transaction before valuing it

Start with a project investing 100 at inception, financed by 60 of bank debt and 40 of equity. It has three years of services, billed at 60 at each year-end, and annual operating cash cost of 20. Bank principal amortizes equally over three years, with 8% interest on each year's opening balance.

A customer instead advances 40 at inception. Every unit replaces bank debt: bank borrowing falls to 20, equity remains 40, and the construction budget stays 100. The advance is credited equally against the three annual bills. A negotiated discount d applies to the entire original contract price of 180, not just to the advance or the remaining payments. Each subsequent cash receipt is therefore:

\[
\text{Customer cash}_t=R(1-d)-P/N
\]

Here R=60, P=40 and N=3. The provider's future recognized service revenue is 60(1−d); subtracting the advance credits again for cash does not reduce revenue a second time. Cash at inception plus future collections equals the discounted total contract price.

The customer evaluates cash timing at a chosen 5% opportunity cost. We compare incremental equity cash flows at a common 12% valuation rate across the two structures. Holding that equity rate fixed is a sensitivity convention, not a claim that the market prices both capital structures identically. No taxes, fees, upkeep capex, residual value, cash reserves, or loan repricing are modeled.

The comparison presumes an available no-advance alternative. The model checks its stated cash-coverage test, but does not prove a lender would offer it. If the baseline project cannot be funded at all, the alternative is delay, redesign, or cancellation; a different operating counterfactual is then required.

## 3. The customer's minimum compensation

Define A_N(k) as the present value of one unit paid at each of N year-ends:

\[
A_N(k)=\sum_{t=1}^{N}(1+k)^{-t}
\]

Relative to paying as service is delivered, the customer spends P today and subsequently saves P/N + Rd each year. Let V be the inception present value of additional customer benefits, such as earlier productive availability. Let L be the inception present value of additional delivery, refund-recovery, or flexibility burdens not already counted in the discount rate or V. Then:

\[
\Delta NPV_c=-P+(P/N+Rd)A_N(k_c)+V-L
\]

The minimum discount leaving the customer no worse off is:

\[
d_c=\frac{P-(P/N)A_N(k_c)-V+L}{R A_N(k_c)}
\]

With V=L=0 and k_c=5%, the timing cost of the advance is 3.69 in present-value units. Compensating it requires a 2.2583% concession on contract price. A zero discount leaves the customer bearing that timing cost even though the provider's financing position improves.

V and L are values to substantiate, not adjustment knobs that prove a favored transaction works. Future benefits must first be discounted to inception. Do not increase k_c for a particular risk and then charge the same risk again in L.

## 4. The provider's maximum concession

With equal principal amortization, each year's reduction in bank principal repayment is P/N. That precisely offsets the P/N reduction in customer collections from prepayment credits. Initial equity is unchanged. The incremental equity cash flow is therefore avoided interest minus the price concession:

\[
\Delta CF_{e,t}=Pr\left(1-\frac{t-1}{N}\right)-Rd
\]

The maximum concession preserving incremental equity NPV is:

\[
d_e=\frac{\sum_{t=1}^{N}Pr(1-(t-1)/N)/(1+k_e)^t}{R A_N(k_e)}
\]

The three nominal interest savings are 3.20, 2.13 and 1.07. Their present value at 12% is 5.3171; the resulting concession ceiling is 3.6896%. Our previous financing model's 3.56% threshold used undiscounted totals. The difference is timing, not a change in the assumed loan rate or the 6.40 nominal interest saving.

The provider also needs cash to operate and pay its remaining debt. With a 1.2x minimum annual coverage requirement, the first year binds in this level-cash-flow, equal-principal example:

\[
d_{cash}=\frac{R-C-P/N-m(D_0-P)(1/N+r)}{R}
\]

Future bills cannot become negative without adding refund mechanics. This imposes d ≤ 1−P/(NR). The nonnegative discount interval is consequently:

\[
\max(0,d_c)\leq d\leq\min(d_e,d_{cash},1-P/(NR))
\]

A negative upper bound means no feasible nonnegative discount; it must not be replaced by zero. If the advance pays off all bank debt, DSCR is inapplicable. A separate nonnegative-operating-cash test remains. The model requires 0≤P≤D₀≤K: advances replacing equity or funding additional capex are different transactions.

The interactive tool accepts a coverage floor of at least 1x, so passing that test also covers the stated debt payment without an unmodeled cash injection. This is an explicit tool boundary, not a claim that all lending contracts require that threshold.

## 5. Four cases reveal where the bargain breaks

All amounts below are hypothetical units. Only the stated inputs change; the provider's 3.69% ceiling is unchanged across these cases.

| Case | Customer opportunity cost | V | L | Customer minimum discount | Joint nonnegative-NPV interval |
|---|---:|---:|---:|---:|---|
| Funding-cost exchange | 5% | 0 | 0 | 2.26% | 2.26%–3.69% |
| Added customer burden | 5% | 0 | 3 | 4.09% | None |
| More expensive customer capital | 12% | 0 | 0 | 5.53% | None |
| Valuable earlier availability | 12% | 4 | 0 | 2.76% | 2.76%–3.69% |

At a 3% discount in the first case, customer incremental NPV is +1.21 and provider incremental equity NPV is +0.99. First-year debt coverage is 3.01x. This is a two-party participation test relative to the baseline—not proof of a bankable project, adequate total investment returns, or a particular negotiated outcome.

At the provider's maximum concession, the 5% customer has room for only 2.34 of net additional burden, L−V. That is about 5.85% of the 40 advance. A burden of 3 destroys the interval despite comfortable debt coverage. A financially strong provider is not automatically offering an attractive advance to its customer.

At a 12% customer opportunity cost, earlier availability must contribute more than roughly 2.66 of net present value before a positive-width interval can open. The final case assumes 4. It demonstrates a threshold to investigate; it does not establish that earlier compute is worth 4 in any actual deployment.

## 6. Private gains are not a measure of social value

Do not add the customer's NPV to the sponsor's NPV and label the sum value created. Different discount rates prevent even a pure price transfer from canceling in that sum. The bank is also missing: it lends 40 less today and receives correspondingly less future principal and interest. The missing principal repayment is not a lost profit. At the assumed 8% loan rate, the present value of the removed loan payments equals the 40 never lent.

Three tests should remain separate:

1. **Participation:** do the customer and provider improve relative to credible alternatives?
2. **Financeability:** can the project fund construction, operate, service debt, and satisfy its capital providers' total return requirements?
3. **Productive additionality:** does the arrangement bring forward useful output or reduce real resource and failure costs after including every party that bears them?

The model addresses the first test and a limited cash constraint within the second. It identifies where evidence is needed for the third. Customer value might reflect real new production, but it can also reflect priority over other customers or a contractual transfer. Those are not interchangeable.

## 7. Improve the bargain through delivery design

When there is no pricing overlap, raising the customer's advance is not an automatic solution. A stronger design can reduce the time funds are idle, make performance easier to observe, or improve service continuity. Stage payments against procurement and commissioning; specify bill credits clearly; negotiate commercially workable substitution, refund, and step-in arrangements where feasible.

Such protections have costs and counterparties. A refund guarantee is not free if it consumes another balance sheet's liquidity or risk capacity. A staging arrangement can reduce customer exposure while leaving the operator needing a bridge loan. Recalculate both sides after those changes instead of treating contract language as costless risk removal.

For AI finance, the opportunity is to convert a price negotiation into a delivery-and-capital design problem. Lower financing friction, credible performance, and genuinely valuable earlier output can enlarge the set of investments worth making. More customer cash alone does not prove that this has happened.

[Explore the interactive pricing model](#model), or [download the same calculation](https://aiinfracredit.com/downloads/bilateral-prepayment-model.mjs). The model uses no company-calibrated inputs. Bring a concrete question to [research@aiinfracredit.com](mailto:research@aiinfracredit.com): whose capital is scarce, which payment moves, and what measurable outcome improves?