When does a megawatt become economically useful?
AI infrastructure is often discussed in terms of chips, models, capital, and computing capacity.
But another constraint is becoming increasingly difficult to ignore.
Power.
Not power in the abstract. Not generation capacity announced in a presentation. Not a future transmission project on a planning map.
Power that can actually reach a facility, on a credible schedule, with sufficient reliability, at a cost the workload can support.
That distinction matters.
A megawatt on a slide is an option. A megawatt delivered to a paying customer is a business.
The Timing Problem
Compute can scale remarkably quickly.
Electrical infrastructure cannot.
A data center can move from design to deployment far faster than a new transmission corridor, substation, generating plant, or major grid upgrade can typically be planned, permitted, financed, constructed, and connected.
The International Energy Agency estimates that global data-center electricity consumption could roughly double from 485 TWh in 2025 to around 950 TWh by 2030, with electricity consumption from AI-focused data centers growing even faster. At the same time, the IEA notes that grid connection queues are already long and that new transmission lines in advanced economies can take roughly four to eight years to build. IEA
The U.S. Department of Energy reached a similar conclusion in its 2026 National Transmission Needs Study: load growth from data centers, manufacturing, and other large users is creating a pressing need for additional transmission infrastructure. The Department of Energy’s Energy.gov
That creates a mismatch:
Compute demand can arrive in months.
Infrastructure can require years.
And once time becomes scarce, time itself acquires economic value.
Time-to-Power
That may be where the AI infrastructure conversation starts to change.
The question is no longer simply:
Where can we build a data center?
Increasingly, it becomes:
Where can we obtain reliable power soon enough for the investment to make sense?
That is why developers are exploring alternatives that would once have looked unusual.
Behind-the-meter generation — producing electricity directly at or near the data center rather than waiting entirely for utility-grid capacity — is moving from contingency planning toward a development strategy in constrained markets. S&P Global describes grid delays as driving data-center projects toward on-site gas generation in search of faster access to electricity. S&P Global
Some temporary-power providers already advertise natural-gas bridging systems measured in tens of megawatts for the gap between immediate AI requirements and permanent utility service. And developers have even used mobile gas generators mounted on semi-trailers where conventional generation equipment or grid connections cannot arrive fast enough. Aggreko
That tells us something.
The scarcity is not simply energy.
It is deliverable energy within the required timeframe.
But Finding Power Does Not Solve the System
Behind-the-meter generation may shorten one part of the timeline.
It does not remove infrastructure.
It simply changes which infrastructure matters.
Instead of depending primarily on the electrical grid, a project may now depend on gas pipelines, fuel contracts, turbines, engines, emissions permits, storage, cooling systems, and local infrastructure.
One constraint has been exchanged for another.
That is why the real chain looks more like:
Generation → Fuel → Transmission or Pipeline → Interconnection → Site → Cooling → Reliability → Cost → Compute
A megawatt somewhere along that chain is not necessarily a usable megawatt at the rack.
And a usable megawatt is not automatically an economic megawatt.
Start With the Units
The terminology is simple but important.
Megawatts, or MW, measure power — the rate at which electricity is being delivered or consumed.
Megawatt-hours, or MWh, measure energy over time.
A facility operating continuously at 10 MW consumes:
10 MW × 24 hours = 240 MWh per day.
Simple arithmetic.
But the economics begin after the arithmetic.
What did those 240 MWh cost?
- How much infrastructure was required to deliver them?
- How much backup capacity was necessary?
- How often can the facility operate at full utilization?
And, ultimately:
How much economic value can the compute produced by that electricity generate?
AI Also Changes the Nature of the Load
There is another complication.
Large AI workloads do not necessarily behave like traditional industrial electricity demand.
Microsoft researchers have documented substantial power swings during large AI training workloads. Thousands of GPUs may simultaneously move between compute-heavy and communication-heavy phases, producing rapid changes in electrical demand. Microsoft argues that these fluctuations increasingly require stabilization techniques across software, hardware, and data-center infrastructure. Microsoft
Grid operators are paying attention as well.
NERC has identified rapidly growing computational loads as a reliability challenge, and in September 2026 foundational computational-load reliability standards passed their initial ballot. NERC
So the challenge is not merely supplying enough electricity.
It is supplying electricity reliably to enormous, concentrated and increasingly dynamic loads.
From Expansion to Economics
For several years, much of the AI conversation has centered on scale.
- More GPUs.
- More models.
- More data centers.
- More megawatts.
But scale eventually encounters economics.
If power becomes more difficult or expensive to obtain, simply adding compute capacity may no longer be sufficient.
The question shifts from:
How much AI infrastructure can we build?
to:
How productively can we use the infrastructure we can actually power?
That brings utilization, workload scheduling, model efficiency, cooling efficiency, power management, location, and revenue generation into the same equation.
A very expensive AI facility operating below its intended utilization because electricity is unavailable is not merely an engineering problem.
It is an economic one.
The Competitive Advantage May Be Changing
Historically, data-center location decisions emphasized land, connectivity, taxes, labor, and proximity to customers.
Power was important, but often treated as an available input.
That assumption is becoming harder to make.
The IEA estimates that grid constraints could delay around 20% of global data-center capacity planned through 2030. IEA
If that proves directionally correct, then access to deliverable power is no longer simply an operating expense.
It becomes part of competitive strategy.
But the advantage does not necessarily belong to whoever announces the largest number of megawatts.
It may belong to whoever best manages the entire chain:
Power availability.
Time-to-power.
Reliability.
Infrastructure cost.
Operational efficiency.
And the economic value produced by the resulting compute.
Which returns us to the original question.
When Does a Megawatt Become Economically Useful?
- Not when somebody announces it.
- Not when generation capacity exists somewhere on the system.
- Not even necessarily when a data center can connect to it.
A megawatt becomes economically useful when it can be converted reliably into productive work at a cost that creates sufficient value.
That is a much higher standard.
And perhaps it is the distinction that matters most as the AI infrastructure race moves into its next phase:
A megawatt on a slide is an option.
A megawatt delivered to a paying customer is a business.
Power may be the constraint.
Turning power into economic value is the opportunity.