In research published in April 2026, the International Energy Agency projects that global data centre electricity consumption will rise from 485 TWh in 2025 to around 950 TWh in 2030.[1] This is a global forecast, not a measure of a local supply shortfall or a confirmed purchase order for a particular site. Applying the headline growth rate directly to local planned capacity overlooks competition between locations, grid constraints and differences in equipment delivery schedules.

Project screening should begin by separating why the customer needs electricity from why it would choose this location. Demand may reflect the expansion of an existing operation, a signed computing-services contract, or an alternative site being considered to secure land and grid access. All three warrant attention, but they should not enter a revenue forecast with the same level of confidence. The commercial team should establish whether the computing customer, facility operator and electricity bill payer are the same entity. Where they differ, it must explain how commitments pass through the contractual chain.

ZHICE recommends turning the opportunity pipeline into a traceable chain of evidence: business demand, site-selection status, equipment plans, funding, procurement authority and conditions for construction to start. Each item should identify its provider, date and period of validity. Alternative requests from the same end customer at different sites must retain their competitive relationship rather than simply being added together. If a customer will not provide material evidence, technical discussions can continue, but those discussions should not justify ordering dedicated equipment with long lead times.

The aim is not to eliminate all uncertainty at an early stage. It is to control the cost of carrying that uncertainty. Reusable design work, surveys and connection studies can proceed; assets that are difficult to redeploy should wait for stronger commitments. The investment proposal should explain which distinctive project condition could secure the customer’s final selection and what the company could recover if that condition failed—not merely describe the size of the market.

Lawrence Berkeley National Laboratory’s United States Data Center Energy Usage Report: 2025 Update, published in June 2026, builds a bottom-up model using factors including equipment shipments, unit energy consumption and cooling performance, and tests the results under different assumptions.[2] The methodological lesson is to build forecasts from equipment and operating patterns. The US estimates themselves cannot be directly extrapolated to demand at a Chinese project.

For an individual project, first establish consistent boundaries for three quantities: computing-equipment power, maximum load at the facility’s grid-connection point, and electricity consumed over the settlement period. These determine equipment configuration, connection capacity and energy procurement respectively, yet are frequently conflated in promotional material. Cooling, power-distribution losses and auxiliary facilities must be included within the relevant boundary. Energy-efficiency indicators are comparable only under consistent environmental conditions and load levels; design performance cannot stand in for year-round operation.

Next, develop a phased load profile from first energisation to target scale. Each phase should specify rack and equipment arrivals, the migration of customer workloads, load ramp-up and seasonal variation. A base case coordinates plans; a low-load case tests cost recovery; a high-load case tests supply capability. These are management scenarios, not forecasts with precisely measurable probabilities. The customer’s option to expand should also be recorded separately from the minimum obligations it has already accepted.

Where operating data are unavailable, customer equipment inventories, comparable facility samples and manufacturer specifications can establish an initial range, with differences and outstanding validation needs clearly identified. Commissioning data should then be used to revisit the original assumptions, not merely redraw the curve in a presentation. Unexplained deviations should trigger a review of procurement plans and contractual boundaries. Otherwise, engineering may build to the upper limit while finance assumes full-load revenue, leaving the same project dependent on two incompatible sets of premises.

Funding and a customer do not guarantee energisation on schedule. Equally, grid and land availability do not guarantee a load that will keep paying. ZHICE assesses these as independent dimensions. Customer commitment rests on authorisation, enforceable obligations and payment security; delivery readiness rests on connection arrangements, critical equipment, site infrastructure and operating preparation. Neither substitutes for the other. Combining them into a weighted average can conceal a critical gap behind an apparently acceptable overall score.

Where customer commitment is strong but delivery readiness is weak, the priority is to negotiate an acceptable phased service rather than repeat a promise of full capacity. Establish which time periods, capacities and workload types can be served first, and whether deferring the remainder would prevent the customer’s overall launch. Phasing is a solution only if the reduced service retains commercial value for both parties. If the business must launch as a whole, connection or equipment availability should become a condition for the relevant obligations to take effect.

Where delivery conditions are favourable but customer commitment is weak, favour assets and contract structures that can serve several potential users and expand in stages. Available resources should not pressure the commercial team into accepting any expression of demand. Projects with strong evidence on both dimensions may proceed to detailed investment review, but still require stress testing and an assessment of residual risks. Where neither dimension is well supported, preserve the opportunity at low cost.

The matrix classifies actions; it is not a probability model. A project’s position should be agreed through an evidence review and updated as customer plans, connection approvals and equipment lead times change. Record the triggers for moving between quadrants: which document justifies a larger commitment, and which change requires a pause. Stage reviews can then alter resource allocation rather than merely complete the paperwork for decisions already taken.

The Special Action Plan for Green and Low-Carbon Development of Data Centres, issued in 2024, calls for integrating data centres with renewable energy on the basis of stable supporting supply and flexible balancing capability, while considering constraints including energy and water resources.[3] It establishes a principle that remains useful for analysis: abundant renewable resources alone do not demonstrate that a particular data centre has a viable supply arrangement for sustained operation.

Start by identifying the business functions that cannot be interrupted, then work back to connection, supplementary supply, backup and recovery arrangements. Renewable electricity volumes, reliable capacity and capability available on demand answer different questions. Matching annual energy consumption does not establish deliverability in every critical interval. The project team should examine these requirements on a common timeline, specifying who fills a shortfall, at what price, and what happens if the shortfall cannot be covered.

Engineering deliverability extends beyond connection approval. Critical-equipment production slots, site construction, communications and protection coordination, commissioning windows and operational staffing may all determine the actual service start. For interdependent works, the responsibility schedule should cross organisational boundaries and identify which downstream commitments would be affected by an upstream delay. Unconfirmed external conditions must remain explicit constraints, not silently become assumed on-time inputs in an integrated schedule.

ZHICE recommends a joint review organised around five service boundaries: business constraints, physical supply, flexibility and recovery, metering and verification, and operating responsibilities. Each should produce an acceptance criterion, and the review should test for contradictions between layers. For example, has a promised recovery time been translated into an operating procedure that can be rehearsed? Can a claim about renewable attributes be traced to settlement records and supporting documentation? An indicator without an evidence trail should not become an unconditional commitment.

The IEA’s 2026 research discusses the potential for load flexibility and non-firm connection arrangements to ease connection constraints, while noting continuing uncertainty around data centre peaks and ramp-up.[1] For an energy company, flexibility is not a customer’s verbal assurance that it can adjust demand. It is the ability to change electricity use lawfully, promptly and repeatedly under specified conditions, while accepting the resulting business consequences.

All AI workloads should therefore not be treated as one adjustable resource. Customers must identify which jobs may be deferred, moved or slowed, and which face strict service deadlines. The energy team must define notice periods, event duration, call frequency and recovery limits. An unknown parameter should constrain the flexibility included in the supply model, rather than allowing equipment savings to be calculated from an idealised response.

Recovery also belongs inside the assessment. Reducing demand may be followed by a concentrated burst of deferred computing, shifting pressure rather than removing it. A flexibility call that consumes emergency resources may also alter the customer’s existing reliability arrangements. Begin with a mutually agreed validation under controlled conditions, recording the actual response, business impact and recovery trajectory. Actions affecting critical systems require customer authorisation and operational approval; production must not be disrupted simply to demonstrate the solution’s value.

The commercial arrangement should cover, at a minimum, the response baseline, measurement method, availability declarations, treatment of non-performance and both parties’ exit conditions. Compensation should reflect verified service, not merely declared capacity paid for indefinitely. If the customer retains an unrestricted right to refuse a call, the supplier should not depend on that capability alone to meet a firm supply obligation. Treat it as an additional optimisation opportunity, increasing the committed share only as operating evidence accumulates.

Berkeley Lab’s August 2026 study of electricity tariffs for large loads reviews arrangements including minimum contract terms, financial security, ramp-up periods and exit charges, with a focus on demand–supply mismatch and stranded-investment risk. It covers enacted and proposed US rules.[5] These mechanisms can inform responsibility design, but they are not tariff standards or a contractual basis for projects in China. Each mechanism should address a real risk rather than simply add another protective clause.

China’s 2026 notice on multi-user direct renewable-power connections requires an identified lead entity and agreement on responsibilities including balancing, costs, settlement and default. The public grid’s supply responsibility is linked to connection capacity and the relevant agreement.[4] Projects within the notice’s scope cannot assume that the public grid will unconditionally cover every internal shortfall. Single-user direct connections and other supply models require separate checks of the rules that apply; the regimes should not be conflated.

Begin negotiations with the matters each party can control. Customer-controlled equipment migration and load ramp-up should connect to minimum obligations, adjustment windows or proportionate security. Supplier-controlled procurement and operations should connect to delivery standards and remedies. Third-party grid connections, policy changes and extreme events require conditions precedent and change procedures that reflect the extent of control. Events outside a party’s control should not all be presented as guarantees that it can fulfil.

For the customer, security should correspond to the dedicated costs genuinely borne by the supplier, with provision to release protection as the underlying risk falls. For the energy company, price adjustments, delayed energisation, capacity reductions and termination cannot be negotiated in isolation. Otherwise, a customer may reduce capacity and shed its minimum obligations while dedicated assets remain on the supplier’s balance sheet. Map the resulting clauses back to cash-flow stress scenarios before closing the agreement.

Investment review should show more than a base case of completed construction and on-time load ramp-up. ZHICE recommends testing customer delays, delivery delays, loads below plan and procurement-cost changes, including combinations of these factors. Stress testing should establish how much funding, time and contractual exposure the company can withstand. Decimal-point precision must not obscure uncertain inputs.

Different stages warrant different levels of commitment. Start with demand validation and solution studies. Secure necessary resources once material boundaries are defined. Expand dedicated capital expenditure only when customer obligations and delivery conditions align. Acceptance results, not the passage of time, should determine progression between stages. Prior expenditure does not entitle a project manager to treat subsequent investment as inevitable.

Compare the value of redeploying assets as well. At the same expected return, a design that can be built in sections, serve other users or change its service scope may preserve more room to respond to changing conditions. That adaptability can carry additional cost. Specify how much is being paid for it and the scenarios in which it has value; flexibility is not a benefit that can be asserted without analysis.

Management should ultimately approve a bounded set of commitments: the expenditure ceiling for the current stage, permitted operating exposure, the next review conditions and changes that require renewed approval. Each enterprise must set its own return thresholds and risk tolerance; this report prescribes no universal rate of return. If a project works only at full customer load, stable energy prices and complete on-time delivery, revise the transaction structure before using larger scale to make risk appear more manageable.

The following ninety-day sequence is a suggested management cadence, not a construction deadline or a promise of returns. A later stage should not start automatically when the calendar advances if an earlier evidence threshold remains unmet. In the first phase, commercial and finance teams jointly establish end demand, procurement authority and payment paths, while engineering maps external dependencies. The output should be a shared assumptions register, not separate optimistic versions maintained by each department.

The second phase should validate the gaps most likely to change the decision. For delivery questions, create an integrated schedule covering connection, equipment, commissioning and operational readiness. For flexibility, obtain a customer-accepted validation plan and allocation of responsibilities. Do not spread research expenditure evenly. If an unresolved connection condition could invalidate the project, resolve it before refining financial parameters that do not affect the go/no-go decision.

In the third phase, a cross-functional committee should review one consistent set of demand, delivery and contractual assumptions. Every material number should trace to evidence or a clearly labelled management assumption; every risk should have an owner and a next review date. The decision may be to advance, phase, seek further evidence or stop. Stopping a project whose conditions do not hold is also a sound outcome. Retain versions and reasons for changes so that new information can be routed back to the appropriate decision gate.