GRID AI — Infrastructure case
The arithmetic
of one Node,
before the network.
One Node defines the economics. This page is the underwriting case for that Node, the model today, and the measurements MF01 is designed to produce.
The market GRID AI plugs into.
AI compute is already traded by the GPU hour. GRID AI is designed to connect distributed compute capacity to this growing market. Pricing below is read from official external sources server-side every 30 minutes and quoted in USD.
Runpod
Reading official source
Vast.ai
Reading official source
Pricing types differ by platform. No range or average is derived across sources.
Reference GPU · RTX 5090. GRID AI Nodes are designed around modular compute capacity; hardware may vary by workload, location and commercial requirements.
From building
to compute market.
- 01
Property
Existing building and available electrical capacity.
- 02
GRID AI Node
Compute installed inside the building.
- 03
GRID AI Orchestrator
Coordinates Nodes as one intelligent network.
- 04
Compute capacity
Distributed infrastructure becomes usable AI compute.
- 05
AI customer
AI demand consumes that compute capacity.
- 01
Property
Existing building and available electrical capacity.
- 02
GRID AI Node
Compute installed inside the building.
- 03
GRID AI Orchestrator
Coordinates Nodes as one intelligent network.
- 04
Compute capacity
Distributed infrastructure becomes usable AI compute.
- 05
AI customer
AI demand consumes that compute capacity.
GRID AI is designed to turn suitable electrical capacity inside existing buildings into usable AI compute capacity, the economic bridge between the physical GRID AI network and AI demand. Compute revenue is intended to reward hosts and fund new Nodes, so each turn of the wheel adds capacity to the next.
Market price is not GRID AI revenue.
External platform prices indicate what compute customers may pay. GRID AI commercial revenue will depend on workload, utilization, commercial terms and operating performance.
One Node.
One economic case.
A Node converts available electrical capacity into AI compute. Compute revenue is the only revenue in the base case, and it carries everything below it.
Underwriting assumption — compute rate
$0.45 / GPU hour
Model input, stated in USD because compute capacity is quoted in USD. Not a live market price.
Model EUR equivalent
€0.39 / GPU hour
Model FX assumption — EUR 1 = USD 1.14. Fixed for modelling, not a live rate. Every economic output on this page is presented in EUR.
Underwriting assumptions are model inputs and are not live market prices. Current external GPU market observations are shown only in the live compute market section above.
Drivers
Prototype capex
MF01 — 4 × NVIDIA RTX 5090, approx. 2.55 kW full Node power under load.
Compute utilization
Share of time compute is genuinely paid for, within the capacity available at the location.
Compute rate — underwriting assumption
Illustrative underwriting assumption of $0.45 / GPU hour at the model FX assumption
Model input for what a rented GPU hour earns after idle time, discounting and demand cycles. Not a live market price.
Economic bridge — one Node, one year, in EUR
Annual compute revenue
The only revenue in the base case. Recovered heat is carried at €0.
Node electricity
At €0.28 / kWh. A GRID AI operating cost, not a host cost.
Household electricity benefit
Target host model. MF01 measures whether it holds.
Other Node operating cost
Software, monitoring, maintenance reserve.
Node contribution
Indicative operating contribution, before hardware financing, tax and central GRID AI overhead.
Simple prototype payback
On the €25.500 prototype investment. Excludes financing, tax, replacement capital and hardware lifetime.
Model — Node contribution
approx. €3.418
/ year
Measured — Node contribution
€—
Where one Node’s revenue goes
- Node electricity
- €4.90647.3%
- Household electricity benefit
- €1.05010.1%
- Other Node operating cost
- €1.0009.6%
- Node contribution
- €3.41833.0%
Compute pays for the system.
The variable that changes everything
Assumed compute rate.
At approximately 75% compute utilization and the €25.500 prototype cost structure
Every rate below is an illustrative underwriting assumption, shown in EUR first and in USD at the model fx assumption — eur 1 = usd 1.14. Underwriting assumptions are model inputs and are not live market prices.
Model rate / GPU hour
$0.30 underwriting assumption
- Node contribution
- approx. zero/ year
- Simple payback
- —
No meaningful contribution after the modelled costs. Not attractive for residential deployment under the current prototype cost structure.
Model rate / GPU hour
$0.45 underwriting assumption
- Node contribution
- approx. €3.400/ year
- Simple payback
- approx. 7.5 years
The scenario the rest of this page is modelled on.
Model rate / GPU hour
$0.60 underwriting assumption
- Node contribution
- approx. €6.900/ year
- Simple payback
- approx. 3.7 years
Illustrative upside, not an achieved price.
Payback thresholds in the current model
approx. €0.46
Model rate / GPU hour
$0.52 underwriting assumption
approx. 5.0 year simple payback
approx. €0.51
Model rate / GPU hour
$0.58 underwriting assumption
approx. 4.0 year simple payback
The hardware does not define the business case.
Modelled compute revenue does.
These outputs move with
Stated as simple payback, excluding financing, tax, replacement capital and hardware lifetime.
Where the electricity figure comes from.
Compute utilization is not the same variable as electrical load. The model carries a Node that draws power beyond paid compute hours, so the electricity cost is derived from an average continuous draw across the full year, not from peak power multiplied by utilization.
Full Node power under load
Implied average continuous draw
Derived from the model electricity cost, not measured. MF01 measures continuous draw.
Hours in a year
Implied annual consumption
Electricity price assumption
Node electricity
Approx. 17.500 kWh × €0.28. A GRID AI operating cost, not a host cost.
For reference: 2.55 kW × 8.760 h × 75% would give approximately €4.691. The higher figure in the model implies base draw outside paid compute hours. The real split between base draw and GPU load is unresolved until MF01 is measured, and the average draw above is a model input, not a measurement.
Three filters decide how much compute becomes revenue.
Utilization is not a single switch. It is what is left after demand, power, and reliability each take their share.
Market demand
Is a paying AI customer ready to use the Node right now? No demand, no revenue.
Power availability
Can the building safely draw the extra power at that moment? Grid, fuse, and host limits apply.
Technical uptime
Is the Node hardware and software online and healthy? Downtime is lost hours.
Demand × availability × uptime = what can be sold.
Property first.
Compute follows capacity.
What if compute could pay for your power?
The host benefit is not a separate business model. It is one use of the economic value compute already creates in the case above.
Node electricity
GRID AI operating cost
The electricity consumed by GRID AI infrastructure. Intended to be carried by GRID AI and not to economically burden the host.
In the model
approx. €4.906Cost / year
Household electricity
Target host benefit
The normal electricity consumed by the household itself. How far Node economics can support it is what MF01 measures.
In the model
€1.050Target cost / year
Host the compute. The target: compute covers your power.
The prototype is not the product.
MF01
€25.500
Production Node
€18.000
Illustrative production model
Assumed GPU hour revenue — underwriting assumption
€0.48/ GPU hour
Compute utilization
75%
Annual Node contribution
approx. €5.700/ year
Simple payback
approx. 3.2 years
€18.000 is a production cost target, not an achieved cost. The production architecture will be informed by MF01 and subsequent deployments.
Reducing unit cost changes the economics materially.
Prove one.
Then multiply.
1 Node
Prove the architecture
10 Nodes
Prove repeatability
100 Nodes
Prove operations
10.000 Nodes
Mathematical illustration below
The intermediate steps are proof stages, not forecasts. No financial projection is attached to them.
Mathematical illustration — 10.000 production Nodes at €18.000, 75% utilization, €0.48 assumed GPU hour rate
Illustrative underwriting assumption of $0.55 / GPU hour at the model FX assumption, not a live market price. All outputs in EUR.
- Hardware capital
- approx. €180 million
- Annual compute revenue
- approx. €127 million
- Annual Node electricity
- approx. €49 million
- Annual household electricity benefits
- approx. €10.5 million
- Other annual Node operating costs
- approx. €10 million
- Annual Node level operating contribution
- approx. €57 million
Mathematical translation of the current unit model. Excludes central GRID AI overhead, financing, tax and replacement capital. The economics of one Node remain the case that matters.
From model to measured.
MF01 does not exist to prove that a GPU works. MF01 exists to replace model values with measured values.
Assumed compute rate
$0.45 underwriting assumption
Effective utilization
Node electricity
Node contribution
Economic model disclosure
Figures on this page are GRID AI modelling assumptions and scenario outputs unless explicitly labelled measured. Underwriting assumptions are model inputs and are not live market prices. Current external GPU market observations appear only in the live compute market section, read from third-party platforms. Model outputs guide MF01 testing and infrastructure design.
Actual performance will depend on realized compute pricing, utilization, electrical availability, uptime, electricity costs, hardware costs, installation, maintenance, financing and other operating factors. The host electricity benefit is a target model in development, not a current offer, and no eligibility or fair use terms have been set. Measured values will only be shown once generated by GRID AI infrastructure.
The question is not whether AI compute has value.
The question is how much of that value can be captured behind an existing electrical connection.
MF01 is how we find out.