An AI data centre can have land, buildings and a customer waiting, yet remain unable to operate because the electricity cannot be delivered. Generation is only the beginning. Transmission, grid connection, transformers, switchgear, backup systems and rack-level conversion must all work together. Power is the point where digital growth meets physical infrastructure, and where a delay can leave expensive computing assets idle.
The investment question
The investment question is which part of the power chain earns durable returns from data-centre demand. Utilities, generators, grid-equipment suppliers and in-facility electrical vendors have different economics and risk exposures. Increased electricity consumption does not translate into the same profit opportunity for every participant.
Our view is that delivery capability and time to service are often more important than a low quoted energy price. Customers need reliable electricity at the right site, with equipment that supports the intended load. Suppliers that shorten commissioning or solve difficult distribution requirements can create substantial value, but project execution and capital intensity remain central to returns.
How electricity reaches the accelerator
Electricity moves from generation through transmission and distribution infrastructure to the data centre. Transformers change voltage; switchgear controls and protects electrical circuits; backup and uninterruptible-power systems support continuity; power distribution and conversion then bring electricity to racks and individual components. The exact architecture varies by site and equipment generation.
Power and energy must be distinguished. Megawatts measure the rate at which electricity is used or supplied. Megawatt-hours measure energy over time. A project described as 100 MW does not necessarily consume 100 MW continuously, and its stated capacity may refer to IT load, total facility load or a contracted connection. Those definitions materially affect comparisons.
Power usage effectiveness, or PUE, compares total facility energy with IT energy over the same measurement boundary and period. A PUE of 1.2 means total facility energy is 20% above IT energy. It is a facility-efficiency measure, not a measure of useful AI output. A low-PUE site can still run poorly utilised processors or inefficient software.
Market structure and competitive advantage
| Layer | Main function | Economic issue |
|---|---|---|
| Generation and energy supply | Produce or procure electricity | Fuel, availability and contract structure |
| Grid and connection | Deliver electricity to the location | Capacity, upgrade cost and timing |
| Electrical equipment | Transform, protect and distribute power | Qualification, lead times and service |
| Rack-level conversion | Supply components at required voltages | Efficiency, density and architecture fit |
Eaton, Schneider Electric, ABB and other electrical suppliers participate across different parts of the distribution chain. GE Vernova has exposure through power generation and electrification equipment. These are broad businesses; data-centre orders should be separated from other industrial and infrastructure demand. A supplier’s backlog can include projects with different margins, delivery periods and execution risks.
Customer relationships and installed-base support matter because electrical failures can interrupt valuable workloads. Standardisation can help large operators repeat designs across sites, but equipment still needs to match local conditions and project requirements. A supplier’s ability to manufacture, deliver, commission and service reliably can be as important as the specification of an individual product.
Economics: price, availability and delay
The cost of electricity includes more than the energy price. Connection charges, network costs, demand charges, reliability provisions and the equipment needed to use the power can all matter. Commercial arrangements differ, so comparisons should use the same scope rather than treating a quoted generation price as the complete delivered cost.
Illustratively, a constant 10 MW IT load at a PUE of 1.2 uses 105,120 MWh of facility electricity over a 365-day year. At an assumed energy price of $80 per MWh, that is about $8.4 million annually before other charges. These are illustrative assumptions, not a forecast for any operator. Changes in utilisation, efficiency and delivered pricing can materially alter the result.
Delay can be even more consequential than energy cost. A modestly cheaper power arrangement may be unattractive if it postpones revenue from a large computing investment. The correct comparison includes the financing and opportunity cost of waiting, while recognising that rushed commissioning can create reliability problems that offset the apparent timing benefit.
Equipment suppliers face their own cycle. Strong demand can improve pricing and factory utilisation, but expansion requires investment, skilled labour and supply-chain execution. Backlog should be examined for cancellation provisions, delivery schedules and cost exposure. A growing order book can contain future profit, future work or both; the distinction appears as projects are delivered.
AI and hyperscalers: density changes the architecture
AI raises both total load and, in demanding deployments, power density. More electricity must reach a smaller physical area while responding to changing workloads. This can require different distribution, conversion and buffering arrangements. The site may have sufficient aggregate power but still need substantial upgrades to support a new rack design.
Higher-voltage distribution is one response. For a given power level, increasing voltage reduces current, with implications for conductor requirements and losses. NVIDIA’s 800 VDC technical discussion sets out its proposed transition for future high-density systems. The benefit depends on complete-system design, including conversion, protection, reliability and maintenance.
Hyperscalers influence specifications and can enter long-term power or infrastructure arrangements. Their scale can support new investment, but does not remove construction and connection constraints. Buying renewable energy on an annual contractual basis is also different from obtaining firm, locally deliverable electricity every hour. The distinction matters when evaluating the practical readiness of a site.
Current market debates — September 2026
Recent supplier reporting shows strong demand. Eaton’s second-quarter 2026 results describe accelerating orders and backlog. Schneider Electric’s half-year results identify data centres as an important growth driver, while GE Vernova’s second-quarter release highlights demand in electrification. These results support an equipment opportunity, but do not establish identical exposure or returns across suppliers.
The immediate debate is whether onsite generation can accelerate projects waiting for grid connections. It can offer an alternative in suitable circumstances, but requires its own equipment, fuel supply, operating capability and approvals. The IEA’s 2026 analysis explicitly treats onsite generation as a response with remaining delivery and economic questions, rather than a universal shortcut.
Another debate is the pace of electrical-architecture change. NVIDIA’s August 2026 update discusses approaches for introducing higher-voltage power into existing facilities. Announced availability and ecosystem collaboration should be separated from broad installed adoption. Operators must evaluate retrofit cost alongside the capabilities of the next computing generation.
Structural debates: who pays and who benefits?
One structural issue is the allocation of grid-upgrade costs. Large new loads can require investment that affects utilities, project developers and other customers. The resulting commercial and regulatory arrangements vary by jurisdiction. A supplier thesis should identify who is committed to paying rather than assuming all planned load becomes economically attractive demand.
Another question is how much flexibility AI operators can offer. Some workloads may be shifted in time or location, while latency-sensitive services have less freedom. Storage and scheduling can help manage variability, but flexibility has an opportunity cost if it reduces computing availability. Its value depends on the workload and the commercial arrangement with the power system.
Finally, long-lived electrical infrastructure serves shorter-lived computing equipment. Good design preserves options for future density and voltage requirements. Overbuilding wastes capital; underbuilding can force expensive retrofits. Suppliers that help operators manage this mismatch can create value beyond an individual hardware sale.
What to watch
Track contracted versus energised capacity, connection milestones, equipment delivery, commissioning and actual load. Compare order growth with manufacturing capacity and project completion. Evaluate energy cost alongside reliability and time to service.
The strongest power opportunity is attached to electricity that customers can actually use, delivered through equipment that remains relevant as computing changes. Announced megawatts are the beginning of the analysis, not the evidence of a completed business outcome.
Explore this sector
AI infrastructure & data centres — sector overview
Related sectors: Semiconductors & chipmaking · Data platforms & analytics
