Start with the customer. Trace the system.
An AI application delivers a task, but its operation depends on a larger system. It may pay a model provider, which may buy cloud compute. A compute operator pays for equipment, facilities and electricity. Some firms own several stages, and open models or on-premises deployments create other routes. The OECD describes an interconnected compute supply chain spanning chips, data centers and cloud services. [2]
The six layers below are an AI Keeper teaching framework, not a formal market definition. They overlap. They complement the five analytical lenses elsewhere on this page: infrastructure, energy, economics, geopolitics and business application.
| Layer | Typical sale | Typical buyer |
|---|---|---|
| Energy & grid | Electricity and connection capacity | Data-center operators |
| Facilities | Powered, cooled and connected space | IT owners and cloud operators |
| Chips & systems | Accelerators, servers and networking | Clouds and compute operators |
| Compute & cloud | Access to computing resources | Model developers and business users |
| Models & tooling | Model access and development tools | Application builders and businesses |
| Applications | A task or business outcome | End users and organizations |
A payment is not a margin.
Revenue records what a business receives. To understand its economics, identify what remains after the costs of delivery, what it owns and which risks it carries. Trace the customer's payment back through the inputs needed to fulfil the promise.
Find the constraint.
Ask for dated, specific evidence. For power, request a connection offer and complete tariff schedule. For the building, check cooling, rack density and connectivity against the intended workload. For hardware, confirm compatibility and delivery. For compute, distinguish installed capacity from paid use. For an application, define the task, acceptance criteria, human review and operating cost.
The IEA identifies grid constraints and energy-sector bottlenecks as issues for data-center expansion. [1] Our practical inference is to assess when usable power will be available alongside its price.
A worked example, with the assumptions exposed.
Assume 1,000 kW of installed IT capacity, an average electrical load factor of 80%, and 8,760 hours in a non-leap year. The load factor measures average power draw; it does not measure occupancy or billable utilization. Power Usage Effectiveness (PUE) is total facility energy divided by IT equipment energy. [3] Here we assume a constant annual PUE.
Annual facility energy = IT capacity × electrical load factor × 8,760 × PUE
| Input / result | Scenario A | Scenario B |
|---|---|---|
| IT capacity | 1,000 kW | 1,000 kW |
| Average electrical load | 80% | 80% |
| Annual PUE | 1.30 | 1.60 |
| Assumed energy rate | $0.06/kWh | $0.09/kWh |
| Annual facility energy | 9,110,400 kWh | 11,212,800 kWh |
| Annual energy charge | $546,624 | $1,009,152 |
The difference is $462,528 a year. Both the tariff and PUE change, so this comparison does not isolate either effect. It excludes capital spending, financing, fixed and demand charges, taxes, backup fuel, maintenance, connectivity and revenue. It cannot establish profitability or the preferred location.
Ten questions before the next commitment.
- Customer: Who pays, for which outcome, and how often?
- Dependency: Which external input can stop delivery?
- Power: How much usable power is documented, and from what date?
- Cost boundary: What is included in the quoted price?
- Workload: What does the customer task require technically?
- Demand: What evidence supports usage or purchase commitments?
- Delivery: Which permits, equipment and connections set the schedule?
- Quality: How will useful performance and errors be measured?
- Alternatives: What happens if a supplier, model or tariff changes?
- Decision: Which unresolved fact would change the next commitment?
Record each answer as documented, assumed or unresolved, alongside its evidence, owner and date. Resolve the issue most likely to change delivery, timing or unit economics before expanding the model.
Primary sources and editorial scope
The map, questions and examples are original AI Keeper educational synthesis. Background references were checked on 19 October 2026. The cited reports identify their publication editions; this briefing is not a live market-data feed.
- IEA: Energy and AI, Executive summary (2025) — electricity supply and grid constraints.
- OECD: Competition in artificial intelligence infrastructure (2025), supply-chain overview — interconnected compute infrastructure.
- The Green Grid: Power Usage Effectiveness — the PUE definition.