# From Technical Performance to Financeable Cash Flow

AI Infra Credit · September 24, 2026  
Research extension · Public companion edition, September 24, 2026

## What this adds

The maturity study took monthly operating cash as an input. This extension makes that cash traceable to billed capacity, realized price, allocated IT power, facility overhead, collection timing and economic service life. It preserves the earlier model and its numerical baseline.

The central result is conditional but useful: **a technical improvement is not automatically an improvement in the cash supporting a loan.** Who pays, what the billing unit is, and when collections arrive can matter more than the engineering gain alone.

In a hypothetical uniform fleet, the same lower-power assumption raises monthly cash from 3.1610 to 3.2218 when billed capacity stays fixed. If an alternative commercial arrangement halves billable hours as well, cash falls to 1.1968. Neither scenario estimates a real efficiency improvement or a provider's current economics.

All money below is **USD millions** except explicitly labeled unit prices.

## Separate the variables before varying them

| Layer | Model variable | What it is not |
|---|---|---|
| Commercial activity | Fraction of slot-hours legitimately billed | GPU busy time or proof of contractual payment |
| Engineering load | A factor interpolating allocated IT power between two assumed endpoints | A direct conversion of a GPU utilization counter |
| Facility energy | IT energy multiplied by an assumed period facility-overhead factor | A measured annual PUE or a test of peak electrical capacity |
| Billing | Capacity charges plus any power-cost reimbursement | Cash already collected or a GAAP revenue conclusion |
| Collection | Stated fraction of bills arriving after an explicit lag | Automatic financing from the customer |
| Economic service life | Months during which the modeled fleet provides service | Book depreciation life, guaranteed resale value, or a legal release from obligations |

DOE defines PUE using total facility energy relative to IT equipment energy over the measurement period, conventionally annual. Our fixed monthly multiplier is an assumption for cost calculation, not an annual efficiency observation. [DOE FEMP definition](https://www.energy.gov/cmei/femp/cooling-water-efficiency-opportunities-federal-data-centers)

NVIDIA documents GPU utilization separately from measured power draw and power limits. A kernel-active-time reading is not a billable-capacity metric, and a power limit is not average monthly consumption. This model's linear power proxy therefore needs actual telemetry before an issuer-specific application. [NVIDIA-SMI documentation, utilization and power sections](https://docs.nvidia.com/deploy/nvidia-smi/index.html)

No source above calibrates the prices, power levels, occupancy, costs or financial terms in the experiment.

## Explicit illustrative inputs

The equipment-funding date is month zero. Contractual service dates remain fixed when commissioning or economic life changes.

| Input | Base assumption |
|---|---:|
| Uniform billable compute slots | 2,500 |
| Normalized hours per month | 720 |
| Actual commissioning / planned contract start | End of month 6 / month 7 |
| Planned customer-contract end | Month 42 |
| Economic service life after commissioning | 48 months, ending in month 54 |
| Contract slot-hour price / billed fraction | $2.50 / 90% |
| Renewal price / billed fraction | $1.75 / 75% |
| Contract / renewal power-load factor | 0.60 / 0.45 |
| Allocated idle / loaded IT power per slot | 0.30 / 1.20 kW |
| Facility energy multiplier / effective electricity cost | 1.25 / $0.10 per kWh |
| Other monthly cash costs during service | 0.7000 |
| Power-cost reimbursement / collection lag | None / zero months |
| Collectible share of bills | 100% |
| Capex / debt / initial equity budget | 100 / 90 / 25 |
| Base loan term / effective annual rate | 42 months / 7% |
| Lender hurdle / equity discount rate | 7% / 12% |
| Post-commissioning cash-coverage screen | 1.20x |

The other-cost input must include the rent, staffing, upkeep, cash taxes and other obligations an analyst intends to count. The example does not claim to estimate them. Pre-operation operating costs and post-retirement obligations are zero in the base, but explicit inputs can add both. No asset resale proceeds, replacements, damages or accelerated debt following non-delivery are invented.

The per-kWh input prices the variable energy charge. Fixed demand charges and take-or-pay purchases must remain in the applicable pre-operation, service-period or wind-down cost inputs. Scaling an all-in historical cost per kWh by lower consumption could otherwise erase obligations that have not changed. Full power pass-through here applies only to the modeled variable energy charge.

An allocated IT slot includes its assigned host, network and storage energy, not only GPU board power. Normalized 720-hour months are planning periods, not a reconstruction of calendar billing. Actual month lengths, power curves, cooling seasonality, intra-month payment sequencing and grid capacity need separate treatment.

## The engineering-to-cash bridge

```text
Capacity billings = units * hours * billed_fraction * slot_hour_price / 1,000,000
IT kW = units * (idle_kW + (loaded_kW - idle_kW) * power_load_factor)
Facility kWh = IT kW * hours * facility_energy_multiplier
Power cost = facility_kWh * electricity_price / 1,000,000
Total billings = capacity_billings + power_cost * pass_through_fraction
Collections at month t+lag = billings_at_t * collection_fraction
Operating cash at t = collections_at_t - power_cost_at_t - other_cash_costs_at_t
```

At the base contract inputs, allocated IT load is 2.1 MW and facility energy is 1.89 million kWh per normalized month. Capacity billings are 4.0500, power cost is 0.1890, and other costs are 0.7000. With immediate collection, operating cash is 3.1610.

Renewal cash falls to 1.5039 under the separately assumed lower price, billed fraction and power load. This is a scenario, not a prediction that renewal prices must fall.

The monthly cash path then enters the earlier financing logic: reserve funding before operations, initial equity budget, monthly cash coverage, equity participation and scheduled lender present value. The original 36/41/42/48-month constant-cash results are reproduced by the adapter; the September 22 files are unchanged.

A separate delivery flag detects a late start or retirement before the stated contract end. It does not silently move the contract dates to make a scenario work.

## Results: the contract can reverse the apparent efficiency benefit

All rows below use a 42-month loan. “Cash in month 7” is deliberately a dated measure; in lagged-collection cases it is not steady-state cash.

| Scenario | Bills in month 7 | Cash in month 7 | Minimum cash coverage while debt remains | Later support required | Combined screen |
|---|---:|---:|---:|---:|---|
| Base | 4.0500 | 3.1610 | 1.3098x | 0.0000 | Pass |
| Lower power load; billed capacity unchanged | 4.0500 | 3.2218 | 1.3350x | 0.0000 | Pass |
| Lower power load; billable hours halved | 2.0250 | 1.1967 | 0.4959x | 43.7986 | Fail |
| Facility multiplier 1.25 to 1.10 | 4.0500 | 3.1837 | 1.3192x | 0.0000 | Pass |
| Contract unit price down 10% | 3.6450 | 2.7560 | 1.1420x | 0.0000 | Fail |
| Electricity cost doubles | 4.0500 | 2.9720 | 1.2315x | 0.0000 | Pass |
| One-month collection lag | 4.0500 | -0.8890 | -0.3684x | 3.3024 | Fail |

Reducing only the power-load factor from 0.60 to 0.30 saves 0.06075 per month in the contract period. If bills stay unchanged, the saving accrues to operating cash. If an alternative billing arrangement halves paid hours, the lost billing is much larger than the saving.

This second case changes a commercial assumption as well as an engineering one. It is not a claim that ordinary on-demand GPU services charge by kernel-active time, or that AI efficiency necessarily cuts industry revenue. Demand rebound, quality, throughput, customer value and repricing are not generated by this model.

Lowering the facility multiplier alone helps. If that improvement also requires 5 more of capex while debt remains 90, initial equity required rises to 29.4803 and breaches the 25 budget despite better operating cash. The investment cost of efficiency must accompany the operating saving.

## A debt-service threshold is not the equity threshold

For the base 42-month loan, the chosen monthly coverage screen requires operating cash of 2.8961.

Holding the other contract-period inputs fixed, its isolated steady-state thresholds are:

- slot-hour price at least **$2.3365**;
- billed fraction at least **84.1123%**;
- electricity cost no greater than **$0.240183 per kWh**.

These solve one monthly inequality, not the whole transaction. They cannot be used as stationary thresholds with a collection lag.

Electricity cost affects renewal cash too. At the isolated coverage ceiling, equity NPV is -0.8946, so the parties do not both pass. In this particular zero-fee, par-priced base case, equity participation binds earlier, at about **$0.226688 per kWh**. This is a model frontier, not an electricity-price forecast, credit rating or market lending limit.

## Pass-through protects a margin, not necessarily a payment date

At full collection and zero lag, changing electricity cost from $0.10 to $0.20 with full contractual reimbursement leaves steady contract-period operator cash at 3.3500. Physical energy is unchanged; the customer's bill rises. This transfers the cost exposure, rather than eliminating it. The customer's participation and any offsetting capacity-price concession must be evaluated separately.

With the higher tariff and a one-month collection lag, the first operating month instead has cash of −1.0780 while the 2.4134 debt payment is due. Required later support is **3.4914** under the specified payout and prefunding policy. The reimbursement receivable is not cash available on that date.

The model pre-funds only pre-operation deficits. It does not assume a working-capital line, operating-period coverage grace or additional equity. Those are candidate financing changes to price and verify, not automatic resources. Failing this screen does not prove no alternative cash-management arrangement could finance the project.

## Service life and contract duration are separate clocks

At the weak-renewal base inputs, tested integer loan terms from 24 to 60 months pass jointly only at **41–42 months**. With renewal economics unchanged from the stronger contract period, the same scan admits **41–54 months**.

A 48-month loan in the weak-renewal base has minimum cash coverage of 0.7007x and requires 3.8539 of later support. A genuine improvement in renewal economics or useful life can support a longer loan; it should not be inserted merely to make the term look safe.

If economic service life ends after 36 operating months, the fleet fulfills the planned customer contract through month 42. Debt coverage still passes, but equity NPV falls to **−3.0225** without renewal income. An earlier 24-month retirement leaves twelve contracted service months undelivered; the model flags the mismatch without relabeling it as contract expiry. It does not estimate the resulting damages or acceleration.

The positive counterpart is explicit too: a 60-month service life and assumed renewal price of $2.75 with 90% billed capacity allow the 48-month structure to pass. That is an upside scenario, not assigned likelihood. And switching off equipment does not terminate remaining leases or other obligations: a separate wind-down-cost case exposes those cash calls.

## How technical routes enter the financial study

| Technical or operating change | Variables needing evidence | Inference to avoid |
|---|---|---|
| Model efficiency, quantization, sparsity | Paid workload/billing unit; billed hours; measured power; quality; demand response | A throughput gain automatically becomes the same percentage cash gain |
| GPU versus ASIC design | Capex, workload-normalized paid capacity, flexibility, power and renewal economics | Comparing different chips using an unadjusted slot-hour |
| Memory and interconnect improvements | Delivered service quality, bottleneck relief, power and achievable pricing/occupancy | TFLOPS alone establish monetizable capacity |
| Cooling and power distribution | Facility energy overhead, required capex and commissioning dates | Lower PUE means equally higher free cash flow |
| Electricity procurement or cost pass-through | Effective tariff, settlement dates, collection lag and counterparty obligations | Reimbursement means no liquidity requirement |
| Hardware refresh and reuse | Economic service life, renewal cash, replacement cost and remaining obligations | Depreciation life proves future cash or automatic contract renewal |

The model traces one financing-relevant path. It does not resolve site interconnection, peak power, hardware supply or model-quality feasibility. Those must be checked before treating a cash-positive scenario as deployable capacity.

## Reproduce and calibrate

Run from this directory:

```sh
node verify-technology-cashflow.mjs
```

Files: `technology-cashflow-model.mjs`, `verify-technology-cashflow.mjs`, `technology-cashflow-results.json`. The model imports the frozen September 22 payment function. It passes **84 grouped checks across 20 scenarios**, including units, cash conservation, collection timing, source-baseline preservation, delivery dates and threshold counterexamples.

Before a company-specific application, obtain aligned billing records, payment terms, measured whole-IT/facility energy, non-energy obligations, actual debt service and plausible service-life/renewal evidence. The public documentation used here establishes measurement distinctions, not numerical calibration.

Publication note: this study is included in the September 24 public Compute Financing Research Kit. The dated model and results are unchanged.