AI demand does not become revenue merely because customers want more compute. Increasingly, analysts need to understand whether power, grid access, permitting, flexibility, and geography allow that demand to become energized capacity, and at what cost.
For several years, the analytical boundary around the AI trade was relatively intuitive.
Start with model capability and application demand. Translate that into accelerator requirements. Follow foundry capacity, advanced packaging, high-bandwidth memory, networking, servers, and cloud capex. Estimate how quickly new compute can be installed and how efficiently it will be used.
That boundary is becoming a risk in its own right.
An AI accelerator does not generate inference revenue because it exists. A cloud provider does not monetize a data-center shell because construction is complete. A leased megawatt is not necessarily an energized megawatt. And a signed utility request is not the same as a credible energization date.
Electricity is moving from an operating input in AI models toward a causal variable in the investment thesis.
The International Energy Agency now expects global data-center electricity consumption to roughly double from about 485 TWh in 2025 to 950 TWh in 2030, with consumption at AI-focused facilities growing considerably faster than the overall data-center total. At the same time, its 2026 update says infrastructure and supply-chain bottlenecks are reducing the probability of the most aggressive near-term demand scenarios.
The United States makes the interaction especially visible. The IEA forecasts U.S. electricity consumption rising by close to 2% annually through 2030, more than twice the pace of the previous decade, with data centers accounting for about half of the increase. FERC, meanwhile, has ordered all six regional grid operators under its jurisdiction to justify or reform the rules governing how data centers and other large loads connect to the transmission system, including rules for flexible loads, co-location, and nearby generation.
That does not establish a deterministic “power shortage” thesis.
It establishes something more useful for investors:
The conversion of AI demand into revenue increasingly runs through the electricity system.
The analytical task is to map that transmission company by company.
The coverage boundary is now the analytical risk
Traditional technology coverage is optimized for technology supply chains.
Power markets are optimized around a different set of variables: generation adequacy, transmission topology, interconnection studies, local congestion, peak demand, reliability standards, rate structures, permitting, and regulatory cost allocation.
For an AI analyst, many of those variables can look several steps removed from earnings.
The mistake is assuming that distance means immateriality.
Consider a simplified chain:
AI application demand → compute demand → accelerator order → data- center installation → energization → utilization → billable compute.
A semiconductor analyst may concentrate on the first three links.
An electricity constraint acts farther down the chain. But depending on contract terms and where the bottleneck sits, it can feed back into equipment acceptance, deployment schedules, customer capex, cloud availability, depreciation, power costs, and future ordering.
The effect is rarely uniform.
A grid delay may postpone revenue for a data-center developer while leaving a chip supplier's near- term shipment unchanged.
It may cause a hyperscaler to redirect equipment to another geography.
It may increase demand for more efficient accelerators because compute per megawatt becomes more valuable.
It may support gas generation, batteries, transformers, or utility investment without reducing aggregate AI infrastructure spending.
Or it may do very little because the site already has firm power.
That is why “AI electricity demand is rising” is not yet an investment conclusion.
The relevant question is:
Where, in this company's economic chain, does electricity become binding?
The IMF's 2026 scenario work makes essentially this macroeconomic distinction. It identifies energy and grid capacity alongside data-center infrastructure as physical constraints that can slow the diffusion of AI even if frontier model capability continues advancing rapidly. The IMF does not claim electricity will always be the binding constraint; its point is that technological capability and economic deployment are separate variables.
Equity research increasingly needs the same separation.
Translate megawatts into company assumptions
The first discipline is to stop treating megawatts as an endpoint.
A headline that a company has secured or announced several gigawatts of capacity sounds substantial. But “gigawatt” can describe very different stages of economic reality.
Is the power operational?
Contracted but awaiting transmission upgrades?
Supported by a utility study but not yet a final service agreement?
Dependent on generation that has not been built?
Behind the meter?
Interruptible?
Available continuously or only under defined operating conditions?
Expected in 2027 or 2031?
The distinction already matters in official load forecasting.
PJM's 2026 long-term forecast tightened how it treats large-load requests: near-term projects require firmer commitments, while less-certain projects farther out are discounted. The resulting forecast was lower than the previous year's in the near term in part because of changes to how large-load adjustments were assessed.
For an analyst, this suggests a basic rule:
Discount power pipelines the way you discount revenue pipelines.
Do not assign the same probability to a proposed 500 MW campus, a signed interconnection arrangement, infrastructure already under construction, and an energized facility.
Then connect the power status to the relevant financial variable.
For a cloud or hyperscale operator, the chain may run:
energization timing → available compute capacity → ability to serve backlog → revenue timing → utilization → depreciation and energy expense → cloud margin.
Alphabet's second-quarter 2026 results illustrate why this translation increasingly belongs in technology models. The company said Q2 capex reached $44.9 billion, raised its full-year 2026 capex guidance to $195–205 billion primarily because it was accelerating capacity delivery, and said growing technical infrastructure would create P&L pressure through depreciation and data-center operating costs including energy.
For an accelerator or equipment supplier, the path can be different:
energized customer capacity → deployment schedule → acceptance or installation cadence → mix of products favored by constrained power → future order timing.
For a data-center developer or operator:
connection probability → development timetable → leased versus energized capacity → occupancy and utilization → return on invested capital.
For a utility:
credible large-load additions → generation and network investment → regulatory treatment → customer contribution and cost allocation → rate-base growth and earnings.
For an equipment supplier to the power system:
new generation and transmission requirements → project sanction → transformer, switchgear, turbine, cable, storage, or power-electronics demand → backlog conversion.
The point is not to force every AI company into an electricity model.
It is to identify the specific financial assumption through which the electricity system can enter the thesis.
If that connection cannot be articulated, a megawatt statistic is probably context rather than a model input.
Five signposts matter outside the tech stack
A useful electricity map for AI-exposed companies can be built around five variables.
1. Connection quality, not connection headlines
The first signpost is the maturity of the grid connection.
FERC's June 2026 intervention is important because large-load integration has become sufficiently consequential that the Commission is asking regional grid operators to revisit transmission-service applications, cost allocation, co-location, flexibility, and processes for nearby generation.
The analyst should therefore track more than “power secured.”
Ask:
What stage of the utility or grid process has been completed?
What upgrades are necessary?
Who bears their cost?
What is the firm energization date?
What contractual commitment supports the forecast?
Can the load be curtailed?
Does the site's economics depend on a regulatory reform that has not yet taken effect?
This is especially important because announced load can greatly exceed realized load.
Generation queues provide a useful cautionary analogue. Lawrence Berkeley National Laboratory's 2026 queue study found more than 2,060 GW of generation and storage capacity seeking U.S. grid connection at the end of 2025, yet only 13% of capacity entering queues from 2000 through 2020 had reached commercial operation by the end of 2025. Projects that did reach operation in 2025 had spent more than five years between their initial interconnection request and commercial operation in regions where the relevant data were available. The dataset covers generation rather than data-center load, so the percentages should not be transferred directly to AI projects. The lesson is narrower: queue volume and executable capacity are different quantities.
A model should reflect that difference.
2. Local adequacy and electricity economics
A terawatt-hour is a global statistic. AI infrastructure is built locally.
Power availability in northern Virginia, Texas, Ireland, France, or the Nordics is not interchangeable simply because all are connected to sophisticated electricity systems.
The IEA explicitly warns against assuming that large new loads have one universal price effect. In systems with tight supply-demand balances, additional load can trigger investment and increase costs; in systems with excess capacity, it can improve utilization of existing generation and networks and potentially lower average costs.
For investors, this means power prices need to be understood locationally and causally.
Is the relevant price an energy-market price, a regulated tariff, or a long-term power-purchase agreement?
Does the customer pay for network upgrades?
Is the business exposed to congestion?
Is firm, around-the-clock capacity the scarce product rather than annual energy?
Do higher local costs flow through to cloud customers, compress provider margins, or change where workloads are placed?
The answer can create a competitive advantage for one geography and a delay for another.
Ireland provides a useful European boundary case. Its utility regulator said data centers accounted for 22% of national electricity demand in 2024 and introduced a new connection policy that explicitly considers grid constraints, security of supply, and generation adequacy. New data-center connections must support their requested demand with generation and/or storage under the policy framework.
That does not imply Europe generally has an “Ireland problem.”
It demonstrates why an AI-capacity model that treats Europe as one electricity market can miss material local differences.
3. The build rate of the power system
The next signpost sits on the supply side.
If load is growing faster than existing infrastructure can accommodate, investors need to know what is being built to close the gap, and whether it is likely to arrive on time.
The U.S. Department of Energy's draft 2026 National Transmission Needs Study identifies data centers and other new large loads as contributors to a pressing need for additional transmission capacity.
The IEA similarly points to the tension between fast-moving data-center development and slower electricity infrastructure, and notes that rising AI rack density is testing supply chains for power electronics and transformers as well as generation and grid capacity.
This creates second-order investment questions.
A semiconductor model may need a transformer lead-time assumption without the semiconductor company ever buying a transformer.
A data-center model may depend on transmission construction controlled by a utility.
A utility thesis may depend on whether regulators permit investment costs to enter rate base.
A generation-equipment supplier can benefit from the same constraint that slows a hyperscaler's campus.
The electricity thesis is therefore not simply a bearish overlay on technology.
A bottleneck destroys value for some parts of the chain and transfers it toward others.
4. Flexibility and substitution
The most important counterweight to a deterministic power-shortage thesis is adaptation.
Data centers do not have to consume electricity in exactly the same way forever.
Some workloads may become more geographically mobile.
Training may tolerate more temporal flexibility than low-latency inference.
Batteries can alter peak demand.
On-site or nearby generation can reduce dependence on traditional grid timelines.
Companies can choose different markets.
Chips and models can become more energy efficient.
Pricing can encourage customers to shift workloads.
FERC's current large-load proceeding explicitly includes flexible transmission services, co-located load, behind-the-meter generation, and generation serving electrically proximate large customers. ENTSO-E's 2026 work likewise treats European data centers as potentially flexible grid participants rather than purely inflexible demand.
The supply response is already visible in corporate strategy.
Google has described a data-center energy portfolio spanning geothermal, advanced nuclear, and gas generation with carbon capture as it works to secure future power. OpenAI and SB Energy announced an integrated infrastructure partnership around a 1.2 GW Texas data-center project that includes plans for new generation to support the site's power needs.
These examples do not prove that every constraint can be bypassed economically or quickly.
They show why the investment question cannot be reduced to “Is grid power scarce?”
It must include:
What substitutions are technically available, how much do they cost, and how quickly can the company deploy them?
5. Geography and permitting
The final signpost is location itself.
AI workloads vary in their dependence on latency, data sovereignty, customers, connectivity, labor, water, and regulation. Those requirements determine how substitutable one prospective data-center market is for another.
A training cluster that can move across several states is exposed differently from an inference facility that must sit near a particular metropolitan market.
An enterprise workload subject to data-residency requirements has fewer options than a globally portable batch workload.
A proposed campus whose power requires a new transmission project is different from a site beside surplus generation.
The United States and Europe illustrate different architectures rather than a simple ranking.
U.S. federal regulators are actively redesigning rules around the interconnection of very large loads. Europe operates across national systems in which data-center connection policies and constraints can vary materially, even while ENTSO-E coordinates at the continental transmission level.
An AI analyst therefore needs a geographic layer in the capacity model:
Where is the planned compute, and what has to happen locally before it can turn on?
Build scenarios around substitution and delay
The electricity thesis becomes most useful when treated as a set of scenarios rather than a single bottleneck forecast.
Consider three.
Scenario 1: Power constrains timing more than demand
AI demand remains robust, but grid interconnection and power-system construction prevent some facilities from energizing on schedule.
The likely consequences differ by company.
Cloud backlog stays elevated.
Data-center activation shifts right.
Some accelerator deployment is delayed or redirected.
Cloud providers carry more infrastructure spending before full utilization.
Grid equipment and generation investment accelerate.
Utilities face rising load forecasts but also regulatory questions about who pays for system expansion.
This is the scenario in which the difference between ordered compute and productive compute matters most.
The IMF's 2026 work is broadly consistent with this type of pathway: infrastructure bottlenecks slow diffusion without requiring a collapse in frontier AI capability.
Scenario 2: Capital routes around the grid
Demand remains strong and developers overcome constraints through new geography, dedicated generation, storage, flexible connection agreements, or infrastructure partnerships.
Aggregate AI capex remains high.
But the winners can change.
Regions with credible power paths gain share.
Developers capable of integrating energy and data-center construction become more valuable.
Gas, nuclear, renewables, storage, and grid equipment can all benefit depending on local conditions.
Traditional utility interconnection becomes one route among several rather than the sole route.
Shell's 2026 Energy Security Scenarios should not be treated as forecasts, but their range makes a useful analytical point: AI-intensive futures can be supported by materially different energy mixes depending on technology, geopolitics, security priorities, and policy.
The electricity thesis survives in this scenario, but principally as a capital-allocation and geography thesis, not a scarcity thesis.
Scenario 3: Efficiency outruns the bottleneck
The third scenario is the one an electricity-heavy AI thesis must take most seriously.
Compute becomes substantially more efficient.
Models require less computation for equivalent capability.
Purpose-built accelerators improve performance per watt.
Inference economics improve.
Workloads are scheduled more intelligently.
Underutilized infrastructure is used more effectively.
Application demand may still grow, but electricity consumption rises more slowly than today's extrapolations imply.
The IEA's scenario work has long treated efficiency as a major uncertainty. Its 2025 Energy and AI analysis found that a high-efficiency case produced data-center electricity demand in 2035 roughly 20% below its base case. Its 2026 update continues to emphasize uncertainty from efficiency, AI economics, supply constraints, and the speed with which bottlenecks are removed.
This possibility is not peripheral.
If compute per watt improves faster than economically useful AI demand expands, some of today's proposed power infrastructure could turn out to be excessive.
That would reverse parts of the causal map.
A causal indicator map for the investment model
The practical discipline is to connect every energy indicator to a modeled investment assumption.
For each AI-exposed company, identify five things.
Power assumption: What amount of energized capacity does the current model implicitly require?
Do not begin with industry TWh. Begin with the company's own deployment path.
Transmission mechanism: How does a change in that capacity affect the company's economics?
For example:
energization delayed six months → fewer available accelerators → slower cloud revenue recognition.
Or:
local power price rises → higher data-center opex → lower incremental cloud gross margin unless passed through.
Or: grid constraints accelerate utility capital programs → higher equipment orders subject to regulatory approval and manufacturing lead times.
Observable indicator: Which external evidence sits closest to the causal mechanism?
Possibilities include executed utility agreements, firm versus speculative load forecasts, transmission construction milestones, generation procurement, grid-equipment backlogs, permitting decisions, local capacity prices, or new connection tariffs.
Evidence threshold: How much does that indicator need to move before the financial model changes?
“Grid constraints are worsening” is not sufficient.
A defensible threshold might look like:
Two major planned campuses representing more than 15% of modeled 2028 compute capacity move energization beyond the revenue forecast period.
Or:
Contracted electricity cost for new capacity rises enough to reduce modeled incremental gross margin by 150 basis points if unpassed through.
Or:
A new flexible-connection structure allows 1 GW of previously delayed capacity to energize a year earlier, changing the deployment assumption.
Review date: When will the evidence next become decision-relevant?
Power-system information moves on different clocks from quarterly semiconductor results. Utility planning, regulatory decisions, interconnection agreements, and transmission construction need their own monitoring cadence.
That is the map.
It prevents the analyst from turning every electricity headline into an AI estimate change while making it harder to ignore the power developments that genuinely matter.
Watch the denominator: AI is not the only load on the grid
Another reason to avoid simplistic causal claims is that AI does not operate in an empty electricity system.
Electric vehicles, heat pumps, industrial reshoring, semiconductor manufacturing, cooling, electrification, and ordinary economic activity are also changing demand.
The IEA expects global electricity consumption to grow by an average 3.6% annually between 2026 and 2030, with data centers only one contributor among several. Even in the United States, where data centers are expected to account for about half of incremental demand, other buildings, industry, and transport still matter.
For investors, attribution matters.
A utility's load growth is not automatically an AI exposure.
A transformer backlog is not necessarily caused primarily by data centers.
An increase in power prices cannot be attributed to AI without considering weather, fuel prices, generation retirements, transmission constraints, industrial load, and the broader market design.
The causal map should therefore contain a denominator:
How much of the relevant power-system change is actually associated with the AI capacity in the company model?
Otherwise the electricity thesis becomes a thematic story that explains everything after the fact.
The trade is becoming cross-sector before it becomes purely electrical
Power is also not the only external constraint.
The AI buildout still depends on semiconductors, memory, networking, cooling, construction labor, financing, data availability, trade policy, application demand, and the willingness of customers to pay for AI services.
WTO data underline the scale of the technology investment cycle: its current trade outlook says unexpectedly strong 2025 merchandise trade was supported by surging trade in AI-enabling products, while warning that 2026 trade growth remains exposed to geopolitical and energy-price risks.
That is an important boundary for the thesis.
Electricity should not replace semiconductor analysis.
It should extend it.
The analytical advantage comes from knowing where the company sits across multiple interacting bottlenecks.
A GPU supplier may remain constrained by advanced packaging even in a region with abundant electricity.
A fully energized data center may remain underutilized if AI demand disappoints.
A hyperscaler may have abundant accelerators and demand but struggle to energize the building.
A utility may prepare for extraordinary load growth only to discover that developers submitted overlapping projects in several jurisdictions and ultimately build only one.
The investor's job is not to nominate the single “real” bottleneck.
It is to identify which constraint is most likely to bind for this company, in this period, under this demand scenario.
What would falsify the electricity thesis?
A useful thesis needs evidence that would weaken it.
At least five developments should cause analysts to reduce the weight placed on power constraints.
First, AI efficiency consistently outruns workload growth.
If equivalent AI output requires dramatically less energy, and demand does not expand enough to consume those efficiency gains, electricity becomes less important at the margin.
Second, announced power projects repeatedly exceed realized AI demand.
The industry may be double- or triple-counting prospective data centers across competing locations. PJM's decision to distinguish firmer from less-firm large-load requests is one institutional response to precisely this forecasting problem.
Third, flexible load becomes commercially routine.
If training and other workloads can shift geographically or temporally without significant performance or economic penalties, existing grid capacity becomes more usable and connection constraints become less binary. Both FERC and ENTSO-E are actively examining this possibility.
Fourth, behind-the-meter and dedicated power scale faster than grid delays.
If integrated energy-and-data-center development reliably produces capacity faster and economically, the grid may stop being the principal gating mechanism in the most important markets.
Fifth, AI application economics weaken.
This is the largest falsifier because electricity demand is downstream of economic demand for compute. The IEA itself notes that future data-center electricity consumption is sensitive to market expectations about returns on AI and data-center investment as well as macroeconomic and financing conditions.
Those conditions are why the right statement is not:
AI is power-constrained.
It is:
> The probability distribution for AI deployment increasingly depends on power-system variables that many technology models still treat as exogenous.
That is a much more durable investment proposition.
Where Enquire fits: research across the coverage boundary
The practical challenge is organizational.
Technology analysts understand accelerators, cloud architecture, and model economics. Utility analysts understand generation, grids, tariffs, and regulatory mechanisms. Infrastructure specialists understand construction and financing. Policy analysts understand permitting and regional implementation.
The AI thesis now crosses all four.
Enquire's current capital-markets offering is explicitly designed for questions that extend beyond an analyst's formal coverage, including adjacent value chains, emerging themes, and second- and third- order risks. It combines public information, structured AI research, and expert perspective while preserving context as an investment debate evolves.
For this use case, the value is not producing another forecast of global data-center electricity consumption.
It is connecting the forecast to operating evidence.
A semiconductor analyst can maintain the core technology thesis while adding structured research on the utility systems serving major deployment regions. A utility or data-center operator can test whether the assumed energization schedule is realistic. Power-market specialists can explain which
congestion, flexibility, or cost-allocation issues are actually material in a particular jurisdiction. Semiconductor-supply-chain experts can test whether higher efficiency meaningfully changes the demand for power rather than simply enabling more compute.
Enquire's current product materials describe structured AI research that identifies gaps, direct input from vetted specialists, and an evolving research context that preserves prior inquiries as new evidence arrives.
That structure is useful precisely because this is a cross-coverage problem.
Expert opinion should not substitute for system-operator data, company filings, or regulatory records. But it can help determine which causal assumptions deserve further investigation, and where a neat top-down model is colliding with operating reality.
The relevant unit is not electricity demand. It is deployable compute.
The AI power debate can easily become a competition in enormous numbers.
Gigawatts announced.
Terawatt-hours forecast.
Billions of dollars of grid investment.
Nuclear reactors proposed.
Gas plants ordered.
Those numbers establish scale. They do not tell an investor what to change in a model.
The more useful analytical unit is deployable compute.
How much compute can actually be installed, energized, cooled, connected, and economically utilized in the geography and period assumed?
What limits that quantity?
What does the constraint cost?
Who pays?
How quickly can it be substituted?
And which financial line changes when the answer moves?
For the technology analyst, that may mean adding interconnection milestones beside accelerator supply.
For the cloud analyst, it may mean separating server capex from energization and utilization.
For the utility analyst, it means distinguishing serious large-load commitments from speculative requests.
For the infrastructure analyst, it means understanding which shortages produce pricing power and which simply delay projects.
The AI trade is not becoming an electricity trade because chips and software have stopped mattering.
It is becoming an electricity thesis because the economic value of those chips increasingly depends on whether the physical system around them can turn demand for compute into operating compute.
The hardest question for an AI-exposed model is therefore moving one step outside the traditional stack:
What power-system assumption is embedded in our earnings forecast, and what evidence would make us change it?
Sources and further reading
- International Energy Agency, Electricity 2026 iea.org
- International Energy Agency, Key Questions on Energy and AI iea.org
- International Monetary Fund, Global Economic and Financial Implications of Artificial Intelligence elibrary.imf.org
- Federal Energy Regulatory Commission, large-load interconnection actions ferc.gov
- Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition emp.lbl.gov
- ENTSO-E, “Data centres and the power system” entsoe.eu
- Commission for Regulation of Utilities, Ireland data-centre connection policy cru.ie
- Alphabet, Q2 2026 earnings call abc.xyz
- Shell, 2026 Energy Security Scenarios shell.com
- Enquire, Capital Markets enquire.ai