INFRASTRUCTURE STUDY / 001 MODEL V1.0 · OCT 2026

800 V DC.
More compute per megawatt.

Follow the power. Price the capacity. Find out when a better electrical architecture becomes a better compute business.

THE OPERATOR’S LENS

External compute sales
Same hardware. Same workload.
Every assumption exposed.

ILLUSTRATIVE · NOT A FORECAST

Change an assumption to explore the counterfactual.

TOKENS / UTILITY kWh—

Matched workload · electrical gain

ANNUAL REVENUE UPLIFT—

Extra capacity × paying demand

ANNUAL SURPLUS UPLIFT—

After modeled capital recovery

EXTERNAL 800 V CAPACITY—

Scenario stock · selected year

01 / THE POWER-TO-REVENUE CHAIN

One grid connection. Two outcomes.

NATIVE DC

A schematic of the same utility power budget. Moving dots explain the sequence; they do not represent measured electrons or tokens.

Start with the constraint.

The utility connection caps how much power this campus can draw.

00 / 14s

Loading model…

02 / ADOPTION, NOT ANNOUNCEMENTS

A gradual change in the installed fleet.

Year

External-sales IT capacity, end of year. New installations and retrofits accumulate; existing cohorts retire at 2% per year.

SidecarRow DCNative DCAll external AI

External-sales share is an analyst assumption applied uniformly to all architecture cohorts. This chart is a scenario, not a disclosed inventory of 800 V projects.

03 / THE OPERATOR’S ECONOMICS

Capacity is an option. Sales make it valuable.

Annual economics of AC, DC with the same sales, and DC with expanded sales
Annual unless noted	AC reference	DC · same sales	DC · expand

Surplus = revenue less modeled delivery costs, including capital recovery. It is neither EBITDA nor net income. Software, sales, corporate costs and taxes are excluded. Constant 2026 dollars.

04 / READ THE MODEL

What changes. What stays constant.

The electrical bridge

Utility power per IT watt = 1 / electrical efficiency + auxiliary load. The ratio of AC to DC power requirements gives the possible increase in matched-workload tokens per utility kWh, or IT capacity under a fixed power limit.

The expansion case uses the released power. It does not also claim the same-output electricity saving.

The commercial bridge

Revenue = installed GPU equivalents × 8,760 × billable utilization × realized hourly price. Extra capacity has a separate sell-through assumption. GPU equivalents are a normalization, not a specific accelerator specification.

Identical hardware and workload are assumed across AC and DC. Hardware performance advances are not attributed to voltage.

The adoption bridge

The base case starts from 25.75 GW of AI capacity in 2025, reaching 100 GW in 2030 and 180 GW in 2035. New-build adoption, retrofits and architecture mix determine installed 800 V capacity.

Sidecars and row systems count as 800 V-class interfaces; they are not wholly native DC facilities. ±400 V interfaces are included.

The evidence boundary

JLL’s 2025/2030 market endpoints anchor the original global envelope. The 2035 endpoint, external-sales share, adoption rates, prices, efficiencies and cost allocations are modeling assumptions.

Supplier roadmaps indicate technical feasibility. They do not prove commissioned capacity, realized savings or customer demand.

Fixed cost conventions & limitations

Facility maintenance and IT support each equal 2% of their original capital base annually; electrical maintenance savings apply to the electrical budget only. Cooling/building savings apply to 44% of baseline facility capex. Capital recovery uses the chosen asset lives and discount rate. Additional infrastructure per MW is an incremental capacity assumption, not the full new-campus cost.

Electrical capacity is delivered at a common low-voltage IT boundary. The auxiliary term excludes electrical losses and is not PUE. Default power load is 75%; default billable utilization is separately 75%. The default GPU price is calibrated to the earlier $12M/MW-year revenue assumption, not sourced from a cloud price list.

Selected-year price erosion assumes the same hardware cohort and no performance refresh. The adoption mix slider scales native DC’s annual new-build share and reallocates the remainder to sidecar and row systems. Retrofits split 80% sidecar / 20% row. Contracted demand is required for incremental sales. Earlier commissioning displays gross billing shifted into an earlier window, not recurring annual growth or incremental lifetime profit.

NVIDIA · architecture rationale ↗NVIDIA · deployment roadmap ↗JLL · capacity outlook ↗Enverus · electrical economics ↗