2026-11-16
AI and the Transport Of Energy Value
Zem Gao
A region can possess abundant renewable energy and remain economically distant from the people who could use its benefits. A business can have customers, capital, and valuable AI applications yet struggle to obtain power where its computing facilities are concentrated. These are usually treated as separate problems. They can also be understood as two sides of an opportunity.
AI allows part of the work demanded in one region to be performed using electricity in another, with the results delivered through communications networks. A renewable resource can therefore serve markets beyond the practical reach of its electrical grid. The strategic significance extends from the location of a data center to the financing of generation, the development of a region, and the design of an energy ecosystem.
AI can become a major means of transporting the economic value of energy. Realizing that possibility requires decisions at three levels: a sound business model, an operating system suited to the resource, and public institutions that permit the parts to work together.
The underlying principle has a long history. A region can export a resource, or it can use that resource locally and export the resulting product. Aluminum production is a familiar example of the second arrangement. Iceland’s electricity industry has supported export-oriented smelting for decades. Computing extends this principle into services whose delivery does not require shipping the finished product as physical cargo. The World Trade Organization already recognizes cloud computing among the services supplied across borders through computer networks. 1 2
In Bit & Coin, I described Proof-of-Work as a “superconductor” of value transport and identified cloud computing as a closely related mechanism. Work could take place where energy was advantageous, while its economic effect became available elsewhere. That earlier example helps identify the principle. AI now gives the principle a wider strategic application through training, inference, and services sold to businesses and individuals. 3
What travels is the economic effect of work. Electricity powers the machines at their location. Models, analyses, predictions, and other outputs become usable at another location. Payments flow through the ownership and contractual arrangements connecting the participants. The same process can both transport value and create additional value by making exchanges possible that geographical separation previously prevented.
These effects occur at different layers. An energy supplier provides electricity. A computing operator turns that electricity, together with equipment and other inputs, into a service. A customer uses the service for another purpose. A completed, economically valuable service remains a successful instance of transport even when its customer’s subsequent activity proves disappointing.
The distinction matters for business judgment. We should separately assess whether transport occurred, whether the operator earned an adequate return, whether the customer used the service productively, and whether society benefited. Demand at the higher layer influences prices and future viability below it. Nevertheless, the collapse of a customer’s business does not erase work previously delivered. Equally, a profitable transport business does not by itself establish that every activity it supports is beneficial.
The phrase energy value transport also requires care in financial analysis. AI revenue reflects the contribution of chips, software, data, skills, intellectual property, and distribution as well as electricity. The entire selling price of an AI service cannot be attributed to its power input. The strategically relevant fact is that local power can participate in a distant market through the service it helps produce.
This changes the geographical problem facing an energy owner. Traditionally, the owner seeks a buyer within electrical reach, builds transmission to a more distant buyer, or attracts an industry that can ship its output. AI adds a growing class of industries whose customers can remain far away while a substantial part of production moves toward the power supply.
Its importance becomes particularly clear with renewable energy.
Wind, sunlight, water, and geothermal resources are distributed according to physical geography. Demand is concentrated according to settlement, industry, wealth, and history. The two maps overlap imperfectly. Renewable generation also has a time dimension: the available resource and the customer’s preferred consumption schedule may differ. Distance, congestion, connection delays, and timing together limit how much of the resource can be put to economic use.
The scale of the constraint is substantial. In its 2023 study of electricity grids, the International Energy Agency identified grids as a potential bottleneck to the energy transition and emphasized new corridors connecting renewable resources to distant demand. Berkeley Lab’s 2026 queue assessment reported more than 2,060 GW of proposed generation and storage seeking U.S. grid connections at the end of 2025. These are proposals, many of which will be withdrawn, rather than operating plants waiting to sell electricity. They nevertheless show how consequential connection processes have become. 4 5
Operating systems reveal another part of the problem. According to the U.S. Energy Information Administration, CAISO curtailed 3.4 million MWh of utility-scale solar and wind output in 2024, 29 percent more than in 2023. Congestion and oversupply both contributed. The implication is that installing generating capacity and creating the conditions for its full economic use are different tasks. 6
Locating suitable AI workloads near such resources creates another route to market. Some work that would otherwise consume power in a constrained metropolitan region can be performed near renewable supply. The distant customer receives the computational service, while the energy-producing region gains a buyer.
This extends renewable energy’s economic reach. It does not supply electricity to every load at the customer’s location. A factory motor or a household still needs its own electrical connection. Transmission and storage retain their essential roles, and efficient long-distance technologies such as high-voltage direct current remain important. Computation adds an option for the portion of demand whose work can move. 7
Three commercial situations should be distinguished. One involves energy already being curtailed: a flexible buyer may capture an otherwise lost selling opportunity. Another involves a persistent regional imbalance in supply and demand, where transmission constraints make local production attractive even though the electricity has a positive price. The third involves a resource that has not yet been developed: a credible computing customer may help justify building generation in the first place.
These situations require different contracts and different capital commitments. Sporadic surplus may support an opportunistic workload. A new power project requires a sufficiently dependable revenue base to recover investment. Cheap power available for a few hours cannot be evaluated as though it were a continuous supply.
AI also cannot be treated as a single kind of electricity demand. Its geographical flexibility depends on the work being sold.
| AI workload | Opportunity for locating near energy | Operating requirement |
|---|---|---|
| Large-model training | The training site can be distant from eventual users. | Reliable power, sustained equipment utilization, and fast communication within the computing cluster. |
| Fine-tuning and independent experiments | Separate jobs can be assigned to suitable locations. | Data access, compatible equipment, checkpoints, and completion deadlines. |
| Batch inference | Large volumes of non-urgent work can be scheduled and delivered later. | Clear deadlines, adequate throughput, and secure handling of inputs and outputs. |
| Interactive inference | Some services can tolerate regional distance; others need proximity. | Response-time targets, network reliability, and adequate capacity during demand peaks. |
| Real-time control and device inference | Tight physical or response constraints often favor local execution. | Dependable performance close to the user, machine, or process. |
The distinction between distance from users and distance between cooperating processors is especially important. A training cluster can sit far from its customers while keeping its GPUs closely connected. NVIDIA’s documentation emphasizes high-bandwidth, low-latency networking between training nodes. A cheap-energy location therefore still needs a well-engineered computing facility; ordinary wide-area connectivity cannot automatically substitute for the cluster’s internal network. 8
Inference offers a broader range of choices than the familiar image of a chatbot suggests. Classifying documents, creating search embeddings, analyzing an archive, and preparing reports can often be performed asynchronously. Google’s batch-inference service explicitly serves these non-urgent tasks. Interactive services have tighter requirements, but their tolerances also vary. The IEA observes that some generative AI tasks take long enough to compute that moving a facility farther from users adds relatively little to total response time. 9 10
This suggests a practical division of the AI industry by the flexibility of its workloads. Regions with dependable power may compete for sustained training and serving operations. Locations with variable surplus may serve work that can wait, pause, or be divided into independent jobs. Facilities close to users can concentrate on tasks whose response requirements justify that proximity.
There are also three different decisions hiding inside the phrase “move the computation.” A business can choose where to build capacity. It can assign new jobs among existing sites. Or it can move a job already in progress. The third is usually the most demanding because software state, model data, and sometimes large checkpoints must move with it.
An operator considering migration should compare the size of a job’s saved state, available network bandwidth, and the duration of the renewable-energy window. Recent modeling of self-contained, single-GPU jobs illustrates how these factors determine feasibility. The business implication is to qualify particular workloads for migration before building a facility around the assumption that all work can move. 11
For an operator, the most useful starting point is often simpler: place appropriate capacity well, bring the required data to it, and schedule eligible jobs intelligently. This can create a workable energy advantage before attempting sophisticated migration.
The commercial comparison should be made in terms of the full cost of delivering an accepted service at the required quality and deadline. Electricity price is one component. Equipment recovery, utilization, network access, cooling, maintenance, financing, data movement, and reliability determine whether the energy advantage survives delivery.
A simple illustration makes the discipline clear. Assume two facilities each draw 1 MW while operating and both operate at 90 percent annual utilization. Each consumes 7,884 MWh in a year. An electricity-price advantage of five cents per kWh saves $394,200 annually per MW under those assumptions. That saving is available to cover the remote location’s additional costs and generate a return. If extra costs exceed it, cheap electricity has not produced a competitive service.
Utilization can reverse an apparently attractive comparison. With identical equipment and throughput while running, reducing utilization from 90 percent to 40 percent increases the fixed annual capital burden per unit of output by 2.25 times. This arithmetic does not disqualify intermittent operation. It establishes the size of the advantage that lower energy costs, cheaper equipment, flexible pricing, or other benefits must provide.
The distinction between power capacity and energy consumption also matters. A contracted megawatt describes capacity; megawatt-hours describe energy over time. Neither states how much paid work the facility will deliver. A business needs to know which machines will run, their useful throughput, the hours customers will purchase, and the service obligations attached to those purchases.
The customer contract is part of the energy architecture. A customer paying for immediate response creates a different power requirement from one paying for completion by the next morning. An operator can offer different prices for different degrees of flexibility, provided customers receive the service they actually contracted for.
The same principle applies to the power contract. A facility seeking uninterrupted operation must arrange dependable supply through the available combination of generation, grid access, storage, and backup. A facility accepting interruption needs defined limits, notice arrangements, operating procedures, and a business model that can absorb lost hours. The computing and power contracts must fit each other.
This is where the ecosystem thesis becomes more consequential than a simple search for inexpensive electricity.
A credible portfolio of AI customers can support a power-purchase commitment. That commitment can help finance generation and infrastructure. Suitable computing loads can improve the use of available energy. Additional infrastructure may support local consumers, businesses, and further investment. Under the right conditions, each part improves the commercial position of the others.
This is a possible investment cycle, not an automatic consequence of building a data center. Its strength depends on the quality of the demand, the allocation of risks, and the compatibility of the assets. A long-lived generator and rapidly changing computing equipment do not share the same investment horizon. A renewable project backed by one weak computing customer may acquire a new concentration risk rather than a dependable market.
Operators can address that mismatch through customer diversification, credible counterparties, equipment replacement plans, and facilities capable of serving successive generations of hardware. Policymakers can help by making network access and contracting arrangements predictable. The objective is a productive relationship among businesses with different technologies and different lifetimes.
Over time, competition may reduce the initial location advantage. New demand can raise local power prices. Better transmission can connect other buyers. More efficient hardware and models can change the value of a computing fleet. A successful project should be able to adapt as the market develops; permanent dependence on unusually cheap surplus is a fragile strategy.
There are already practical examples of this geography.
In September 2025, Iceland’s Landsvirkjun announced a five-year agreement to supply up to 12 MW to atNorth’s Akureyri data center, effective from the second half of 2026. The utility explicitly connected the opportunity to transmission limitations between northern and southern Iceland, and described AI and high-performance computing as part of the industry’s changing focus. This is a concrete instance of computing capacity being developed where the electrical system creates a regional advantage. 12
Finland’s LUMI system illustrates a broader combination. Its operator describes hydropower supply in Kajaani, access through European research networks, and an arrangement to recover heat for the surrounding city. The operator reported heat recovery equivalent to up to 20 percent of Kajaani’s annual district-heating needs. Computing services can serve distant users while another useful output serves nearby people. The opportunities depend on the surrounding infrastructure as much as on the machines. 13 14
Flexibility is also moving from a general proposal into operating practice. Google reported utility agreements in 2025 that target machine-learning workloads for demand response. Separately, a published preprint described a Phoenix field demonstration in which a 256-GPU cluster reduced power use by 25 percent for three hours while maintaining its defined service guarantees. The result concerns that cluster and workload mix, not an entire industry, but it demonstrates a business-relevant capability. 15 16
The capability needs to be designed into the service. Work that can finish later can contribute flexibility. Work that must respond immediately may need to continue operating. Scheduling, power management, and contract design determine how much of the facility can respond to the electricity system without failing its customers.
A region should therefore think in terms of a portfolio of uses. Local households and industry, firm computing, flexible computing, storage, and physical power exports have different requirements and benefits. Their proportions should reflect the resource and the market. Flexible computing can be valuable as a buyer during surplus periods, while dependable local demand and appropriately contracted firm loads help support longer-term investment.
A computing facility also offers no general guarantee of environmental improvement. The relevant comparison is what would happen with and without it. Using electricity that would otherwise be curtailed differs from taking renewable power that would otherwise displace fossil generation for another customer. New generation, backup operation, network energy, water use, and equipment manufacture all affect the result. Annual renewable procurement alone does not describe the location and timing of actual electricity use.
The ecosystem objective should consequently be to increase useful economic output while improving the use and development of clean energy. That calls for measuring actual changes in generation, curtailment, reliability, emissions, and local costs. A low electricity tariff by itself establishes none of those outcomes.
The potential also has a scale limit. The IEA’s 2026 central projection places total data-center electricity consumption at roughly 950 TWh in 2030, about 3 percent of global demand; that includes much more than AI. Computing is therefore a strategically important additional market for energy, especially in particular locations, rather than an unlimited buyer for every renewable resource. 17
For business operators, the discipline begins with demand. Identify the customers and the workloads before sizing the computing plant. Establish which jobs need immediate responses, which can wait, which data can move, and which service quality customers will accept. Those decisions define the facility’s useful flexibility.
Next, characterize the power supply hour by hour and season by season. Establish whether the advantage arises from persistent regional abundance, occasional curtailment, a favorable contract, or a new generation project. Include the cost of making the promised service dependable. A power offer becomes commercially meaningful only when its availability matches the business.
Then evaluate the location as a complete operating environment. Network resilience, cooling conditions, maintenance access, skilled personnel, physical security, financing, and permission to handle customer data can determine whether the project succeeds. The cheapest generator may be some distance from the best site. What matters is the advantage of the complete system.
Deployment should demonstrate accepted customer output and operating economics before expansion. Useful measures include revenue per paid computing hour, cost per completed task, service performance, power availability, actual utilization, and the cost of interruptions. Energy and regional outcomes require their own measures. Keeping these accounts distinct makes it possible to identify which part of the system is succeeding and which part needs adjustment.
For policymakers, the central implication is that the location and character of demand deserve attention alongside the location of supply. An energy strategy should identify where generation is constrained, which constraints are spatial or temporal, and which productive loads can respond. A digital strategy should identify which computing services the region can reliably supply and which infrastructure makes those services accessible to customers.
The two strategies belong in the same room.
Government can contribute by coordinating power and communications planning, making connection requirements clear, and allowing technically suitable arrangements for flexible demand. Where local generation and computing can operate together, the regulatory treatment should recognize their actual interaction with the grid. Where they rely on shared infrastructure, their obligations should reflect the costs and reliability requirements they impose.
Public incentives should reward outcomes that can be demonstrated. A project’s nameplate capacity or announced capital expenditure says little about whether it will reduce curtailment, expand clean generation, improve electricity access, or create durable income. Incentives tied to verified infrastructure contributions, operating performance, and local benefits are more informative than a general preference for anything labeled AI.
Cost allocation deserves particular care. A private customer can rationally seek a subsidized tariff or a grid upgrade paid for by others. That does not establish a regional economic gain. Governments should assess who pays for new infrastructure, who carries the risk if the computing load fails to arrive or later departs, and how existing electricity users are affected.
At the same time, regions should recognize the value of outside capital, technology, and customers. International participation can supply capabilities a local economy does not yet possess. The strategic task is to structure that participation so that the region gains durable income, infrastructure, expertise, and opportunities for its own businesses.
Those gains need to be measured at the appropriate layer. The power producer earns from electricity. A hosting business earns from facilities and service. A computing provider earns from usable machine capacity. Model and application businesses earn from further capabilities and customer relationships. Each activity has different capital needs, skills, and risks. Hosting a data center does not automatically confer ownership of the models running inside it or of the applications sold elsewhere.
Regional prosperity therefore requires more than counting the gross revenue passing through a facility. It requires tracking local wages, local suppliers’ value added, public receipts, and income retained by resident owners, while accounting for infrastructure obligations and forgone alternatives. Wider benefits may include better connectivity, stronger engineering capabilities, and access to computing for local businesses. These should be developed deliberately.
A region need not enter every layer at once. It may begin with generation and reliable hosting, then build capabilities in operations, managed computing, or specialized applications. Shared training programs, accessible digital infrastructure, and commercial partnerships can help local firms participate. Ambition becomes more credible when it follows the capabilities the region can actually build.
A resource can be physically near an electrical network yet commercially distant from useful demand. Another resource can be electrically remote while participating in global production through computing and communications. AI expands the set of services for which the place of energy consumption and the place of economic benefit can be separated.
That separation creates an opportunity for both value transport and value creation. It offers energy owners additional buyers, AI businesses additional production locations, and regions additional paths into international trade. Properly organized, it can also help renewable projects obtain the demand and infrastructure that make them viable.
The business strategist’s task is to connect those possibilities to an executable arrangement. Policy must permit and support the right connections. Operators must deliver services that customers will buy. Investors must understand where the returns arise and who bears the risks. Communities must receive benefits commensurate with the resources and infrastructure they contribute.
AI can give renewable energy access to markets beyond the reach of power lines. The regions and businesses that understand this will plan generation, computing, communications, contracts, and local development together. That is the opportunity: to extend the economic reach of energy by organizing where work is performed and how its benefits are delivered.
Sources
Landsvirkjun, Landsvirkjun and Norðurál sign a 7-year power purchase agreement, March 24, 2026.
World Trade Organization, Digitally Delivered Services Trade Dataset.
ZeMing M. Gao, Proof-of-Work (PoW) is the Only Way to Prevent Corruption, Bit & Coin, section on energy and value transfer.
International Energy Agency, Electricity Grids and Secure Energy Transitions, 2023.
Lawrence Berkeley National Laboratory and GridTracker, Queued Up 2026 Edition, data through the end of 2025.
U.S. Energy Information Administration, Solar and wind power curtailments are increasing in California, May 28, 2025.
U.S. Department of Energy, Connecting the Country with HVDC, September 27, 2023.
NVIDIA, Multinode Networking, NeMo Helix documentation.
Google Cloud, Batch inference with Gemini.
International Energy Agency and Korea Energy Economics Institute, Energy and AI in East Asia, revised April 2026, particularly pp. 15, 71–72, and 93–94.
Giuseppe Tomei et al., Green Distributed AI Training, extended version revised May 27, 2026; feasibility modeling and simulation.
Landsvirkjun, Landsvirkjun and atNorth Akureyri Sign New Green Firm PPA, September 2, 2025.
LUMI, LUMI data center receives the Green Data Centre of the Year award, March 9, 2023.
LUMI, Creating a unique AI strategy with supercomputing.
Google, How we’re making data centers more flexible to benefit power grids, August 4, 2025.
Philip Colangelo et al., Turning AI Data Centers into Grid-Interactive Assets, July 1, 2025, field-demonstration preprint; see also its stated workload and deployment limits.
International Energy Agency, Key Questions on Energy and AI, April 2026.