Energized Land Is a Credit Product

A data-center parcel starts life as an option on a data center. Its value rises as the developer proves the path from land to electrons to contracted cash flow: utility process, permits, equipment, customer, and a borrower who can hold the pieces together. The acreage is the starting point. The links do the work.

Documentation tells me what may work. Price tells me what I am being asked to pay for that possibility. A project can become more credible while the return gets worse if land sellers and lenders have already bid up the same queue position, substation, and customer contract.

I keep seeing the same underwriting problem from three directions. At SpotWire, a GPU mark matters only after I can explain what kind of price it represents. In the data-center registry, a nearby substation matters after I trace the path from the grid to the proposed load. In North Texas records, permits, utility work, and local approvals give an announcement weight.

The headline starts the inquiry. The links decide the credit.

What has to be true before land earns the label

I start with the failure case: what breaks first?

Grid. Show the load path and an interconnection process beyond a preliminary request.

Facility. Show water, fiber, cooling, switchgear, transformers, and civil infrastructure sized for the proposed rack density.

Revenue. Show a customer obligation a lender can rely on, plus capital that can carry the project from construction spend to customer cash flow.

The gates arrive in sequence. A failed grid connection strands the facility. A working grid with no customer leaves the lender financing an expensive waiting room. A customer contract without a deliverable load turns revenue into an IOU for power.

A substation next door can have no available capacity. A signed memorandum can have no financeable customer behind it. A large load request can sit years from energization.

For one proposed campus, I want to tie a queue filing to a named substation, a transformer order, the permit record, and the customer contract—and identify the link that remains an assertion. A map shows proximity. A credit file shows obligation.

SpotWire uses the same discipline on prices. An OEM quote gives me a reference point. A transacted price gives me a market observation. A lender needs the number that survives a mark and a recovery analysis.

GPU racks changed the facility

At high rack density, the building is the wrapper. The business is power in, heat out, compute delivered.

A conventional rack can be evaluated through floor area, electrical service, and cooling capacity. A high-density GPU deployment raises the stakes on all three. Rack power rises. Liquid cooling becomes part of the facility design. Busway, switchgear, backup systems, heat rejection, and water strategy become part of the compute contract.

NVIDIA’s PORTS filing shows the chip vendor participating in that contract’s credit structure. The filing describes approximately 4.25 GW of IT load across nine phases at the Ohio campus, an option for another 3.8 GW, and a first phase expected in 2028. It also discloses contingent residual-value guaranties capped at $105 billion, tenant reimbursement and indemnification obligations, and OpenAI in the lease structure.

That support shifts the question. Who absorbs the loss if the tenant fails, the equipment ages faster than expected, or the campus cannot be repurposed? The guaranty moves risk through the structure; it does not erase the risk.

GPU collateral has a resale market, subject to a mark. A powered site has a bundle of grid rights, equipment, permits, contracts, and an operator who has to turn all four into uptime. Before I call the site collateral, I want to know who would buy it after default.

Batch Zero is a queue step

Batch Zero assigns study and allocation treatment to qualifying 75-MW-plus loads. Energization comes later, after the rest of the project survives.

ERCOT’s June explainer set July 10, 2026 for the initial documentation package and August 7 for classification. The original framework contemplated a later Q2 2027 commitment and security milestone. On August 3, ERCOT said it would not classify projects by August 7 while it carried out a verification process under the governor’s directive and sought a good-cause exception from the PUCT.

Eligibility leaves plenty of work ahead: transmission upgrades, equipment, permits, financing, and customer commitments. Queue position is evidence. Cash flow arrives later.

The phrase “energized land” compresses five milestones into one adjective. I would underwrite the verbs:

1. control the land;
2. connect the load;
3. build the facility;
4. secure the customer; and
5. collect the revenue.

The August verification pause makes the distinction concrete. Large-load projects are being asked to establish that the information supporting their requests is real and current. A request that fails verification loses its claim to capacity. The diligence step is part of the credit.

DDTL bridges capex timing

A DDTL addresses a timing mismatch. The lender advances capital before the project produces customer revenue, so repayment depends on customer acceptance and performance.

The borrower draws as eligible infrastructure and equipment costs arrive. Land control, interconnection work, transformers, switchgear, cooling, servers, and construction all run on different schedules. “Eligible” and “customer” carry most of the underwriting weight.

The acronym is doing too much work.

CoreWeave’s DDTL 4.0 was described as an $8.5 billion facility supporting capital expenditures required to perform customer contracts, including GPU servers and related infrastructure. It included floating-rate and fixed-rate tranches, was available into 2027, and matured in 2032. The structure reads like platform finance: broad enough to fund a portfolio of deployments over a longer draw and repayment horizon.

DDTL 5.0 carried a different repayment story. It was a $3.1 billion facility tied to infrastructure for two large, non-investment-grade customers, with a shorter availability period and a 2031 maturity. The smaller amount comes with greater customer concentration. The lender is underwriting specific contracts, deployment schedules, and counterparty risk.

Same company, different answer to “what pays us back?”

“CoreWeave has a DDTL” leaves the important questions unanswered:

• Which legal entity borrowed?
• Which customer contract supports the draw?
• Is the collateral the GPU equipment, the lease receivable, the facility, or a combination—and who absorbs the loss as the GPU generation changes?
• What has been funded, what remains committed, and can the equipment move to another site or customer?

Those answers place the facility somewhere among equipment finance, project finance, receivables finance, and combinations of the three. The label comes last. Collateral and repayment come first.

Sponsors fund the expensive uncertainty at the front end: site control, development, and the risk that the project never reaches a financeable state. Customer contracts unlock asset-level capital later. Equipment-backed loans follow servers and related infrastructure; construction debt and leases follow the shell, electrical systems, and power buildout.

NVIDIA’s PORTS support shows another layer. Its contingent residual-value guaranties support a leased campus and tenant obligations rather than purchasing GPUs. Guarantees and residual support appear where the asset value and customer credit leave a funding gap.

Five lenders can still share one repayment source. If one customer pays them all, the credit remains concentrated.

The risk application I am designing around SpotWire turns that concentration into marks and stress cases: residual value, current LTV, SKU and vintage concentration, and recovery in an orderly or distressed sale.

A rising GPU mark can hide a bad loan. The lender gets paid in cash, while yesterday’s resale price belongs to the rear-view mirror.

The $/MW question

There are three businesses here, each with its own denominator: rent compute, host equipment, and build power. Put them together and the result answers no borrower’s question.

For a retail compute proxy, AWS says a P5.4xlarge contains one H100 and a P5.48xlarge contains eight. AWS’s current US pricing snapshot is approximately $4.326 per H100-hour, although the pricing page is dynamic. NVIDIA’s H100 datasheet gives a configurable SXM power level up to 700 watts.

Using those figures as an upper-bound conversion, a 1 MW facility load and an assumed 1.3 PUE support roughly 1,099 H100-equivalents at 700 watts each:

Assumed accelerator utilization	Gross GPU rental revenue per facility MW-year
50%	~$20.8M
70%	~$29.2M
85%	~$35.4M
100%	~$41.6M

The 70% case is a useful middle line. It represents gross rental revenue for inference-capable H100 capacity. CPU, network, storage, support, discounts, power, cooling, financing, taxes, and server-level draw sit outside the calculation. A token-level inference revenue figure has no public disclosure that supports a clean $/MW conversion.

The public hosting contracts describe a different business. Core Scientific’s CoreWeave contract adds approximately $2.0B of projected cumulative revenue over twelve years for 120 MW of critical IT load. That works out to approximately $1.39M per critical-IT MW-year before capex credits. Across the full 500 MW and $8.7B of projected cumulative contracts, the equivalent is approximately $1.45M per critical-IT MW-year. Core Scientific says critical IT load excludes ancillary cooling and chiller power. This is contracted hosting revenue for the facility owner; the GPU rental table describes the compute operator.

Construction brings a third set of numbers. Turner & Townsend’s 2025 index puts US air-cooled construction at roughly $9.5M to $13.3M per MW across Charlotte, Phoenix, Columbus, Atlanta, New Jersey, and Silicon Valley. It reports a 7% to 10% premium for similarly sized US liquid-cooled facilities, implying roughly $10.2M to $14.6M per MW before rounding. Turner excludes land, utility works and interconnection, site works, active IT equipment, and professional fees.

JLL’s 2026 global outlook gives another benchmark: approximately $11.3M per MW for shell-and-core construction, with AI tenant technology fit-out up to $25M per MW. Adding those figures produces a modeled upper-bound stack of roughly $36.3M per MW. Land, interconnection, and the scope difference between shell/core and tenant fit-out remain outside that arithmetic.

The financing disclosures need their own denominator. CoreWeave’s DDTL 4.0 is an $8.5B facility; DDTL 5.0 is $3.1B. The public announcements supply no matching MW figure, so they describe facility size rather than unit economics.

The cleanest financing-adjacent datapoint is again Core Scientific: its $180M estimated owner infrastructure investment for the incremental 120 MW equals $1.50M per critical-IT MW. CoreWeave funds the required modifications, and the owner investment is credited against hosting payments, capped at $1.5M per MW. That is project-specific infrastructure capex. It gives no generic debt-per-MW answer.

If I apply a financing scenario to the JLL upper-bound stack, 60% to 70% debt implies approximately $21.8M to $25.4M of debt per MW. Applying the approximately 5.9% fixed tranche reported for DDTL 4.0 implies roughly $1.29M to $1.50M of interest-only carry per MW-year. Treat those as sensitivities. The actual answer depends on the legal borrower, collateral package, contracted customer, draw schedule, and whether the MW means critical IT or gross facility load.

The figures belong to different layers: GPU rental to the compute operator, hosting revenue to the facility owner, and construction costs to portions of the physical build. Put one denominator across the stack and the credit memo starts telling stories instead of measuring cash flow.

Dollar-per-MW sources: AWS P5 instances, AWS EC2 pricing, NVIDIA H100 datasheet, Core Scientific/CoreWeave 120 MW contract, Turner & Townsend construction index, Turner methodology, JLL Global Data Center Outlook, CoreWeave DDTL 4.0, CoreWeave DDTL 5.0.

What I would underwrite first

I would begin by trying to break the site. Show me the load path, study status, transmission and substation work, and the post-pause ERCOT process. Then show the rack density, cooling, electrical systems, and procurement schedule.

Next comes the payer. Is the capacity supported by a take-or-pay customer, a lease, a reservation, or a memorandum? What happens to the draw if that customer misses acceptance?

Then I want the failure map. Can the equipment move to another site or customer? How specialized is the facility? What happens to recovery value when the next GPU generation arrives? Which claims come from an executed filing, an ISO record, a permit, a vendor quote, a broker deck, or a press release?

That is why my North Texas intelligence layer starts with permits, municipal actions, infrastructure, and parcels. Announcements describe intent. Independent systems record work.

What I am watching

The number I want is the conversion chain:

announced load → verified load → financeable load → energized capacity → paying load

Four things will move that chain:

1. ERCOT’s Batch Zero verification backlog. The market needs a clean distinction between study eligibility, approved interconnection, and commercial energization.
2. The spread between GPU equipment debt and facility debt. Collateral, draw mechanics, and recovery assumptions belong in separate boxes.
3. Customer concentration inside DDTL structures. A facility tied to two large non-investment-grade customers carries a different credit profile from a diversified platform facility.
4. Whether energized land becomes a separable asset. Grid rights, permits, equipment procurement, and customer demand would give the land option value beneath an infrastructure platform. A speculative queue position gives the broker a phrase.

The headline pipeline is measured in megawatts. The credit is measured in paid invoices.

Sources: ERCOT June Batch Zero explainer, ERCOT August 3 market notice, ERCOT large-load integration materials, NVIDIA FY2027 Q2 Form 10-Q, NVIDIA August 17 filing, CoreWeave DDTL 4.0, CoreWeave DDTL 5.0, SpotWire project methodology, Data-center registry.

Nothing here is investment advice.