US Power Market Investment Notes: Project Kilby and the AI Electricity Gap

What Microsoft and Chevron’s gas-fired power partnership reveals about the U.S. electricity shortfall

Analysis Date: 2026-06-29

Chapter 1: Project Kilby Puts the U.S. Electricity Gap on the Table

Microsoft and Chevron’s Project Kilby in West Texas is a highly representative sample of the current shift in the U.S. power market. In June 2026, Microsoft announced that it would build a new data center campus in Pecos, Texas, with roughly 2GW of planned incremental capacity. Chevron will organize about 2.67GW of natural-gas-fired generation capacity. The two sides signed a 20-year power agreement, with the early phase using a behind-the-meter structure, meaning the power source directly serves the campus and reduces immediate reliance on the existing public grid. The main equipment includes large gas turbines from GE Vernova and equipment from Solar Turbines, a Caterpillar company.

This project is not yet a fully landed revenue number, and first power is broadly expected after 2028. But it already makes the problem visible: compute expansion can no longer bypass reliable power. In the past, discussions about data centers often started with land, buildings, fiber, servers, and GPUs. Now the questions of whether electricity can be connected on time, whether it can be supplied stably after connection, and who is responsible if something goes wrong have moved to the very front of project planning.

The interesting point about Microsoft’s Pecos project is not that the company suddenly abandoned renewable energy. It is that Microsoft acknowledged a change in delivery sequence. Microsoft remains a major buyer of renewable power and will continue to pursue low-carbon electricity. But when an AI factory needs to come online on schedule, the scarcest near-term inputs are often reliable capacity and connection speed.

Chevron is not merely selling natural gas here either. It is trying to sell a power-delivery arrangement that lets the campus start on time: fuel, generation equipment, on-site electrical systems, and a long-term offtake contract bundled together. Traditional oil and gas companies have often moved with commodity price cycles. Now some of that capability can potentially become a longer-duration, more stable infrastructure service. The value of natural gas in this chain is not just the fuel molecule itself; it is the ability to fill part of the reliable-capacity gap before new transmission lines and large substations are completed.

Chapter 2: The Shortfall Is Not National Energy Volume, But Local Reliable Capacity

The U.S. electricity shortfall over the next several years is unlikely to show up as a nationwide blackout, nor is it simply a lack of total generation. It is more likely to appear first in specific regions, specific years, specific voltage levels, and high-reliability loads. AI data centers, advanced manufacturing, industrial electrification, and extreme-weather peaks are lifting demand together, while aging grids, substations, transmission lines, and key equipment are not expanding fast enough.

Over the past two decades, U.S. electricity demand did not grow at a consistently high rate. Manufacturing offshoring, tighter efficiency standards, and improved lighting and appliance efficiency offset part of the electricity growth that would otherwise have come from economic expansion. Utilities, equipment makers, and engineering contractors therefore planned capacity around low growth and replacement demand. Transformers, switchgear, distribution equipment, and construction capacity were not built for a sudden wave of large-scale, highly concentrated loads. Demand is now steepening, while the supply side still carries the cadence of the previous era. That is the deepest background behind the electricity gap.

Capital is also starting to value power assets in a new way. The Financial Times has noted that AI and data center electricity demand pushed U.S. power and utility M&A to a record level, with roughly $203.6B of deals in the first five months of 2026 and a clear increase in data-center-related investment. The number does not need to be over-interpreted, but it shows that utilities, once treated largely as defensive, dividend, and rate-sensitive assets, are starting to become land, interconnection rights, and capacity gateways for AI factories in some transactions. The scarce asset is not always inside the data hall. Very often, it sits outside the building.

LBNL data makes this transition clearer. U.S. data centers consumed about 176TWh of electricity in 2023, or roughly 4.4% of total U.S. electricity use. By 2028, the figure could rise to 325-580TWh, reaching 6.7%-12.0% of national electricity consumption. The range is wide, which itself shows forecast uncertainty, but the direction is hard to ignore. Data centers are no longer a small subcategory of commercial load. They are starting to reshape grid planning, capacity markets, transmission and substation investment, and local politics.

Chapter 3: AI Data Centers Are Not Ordinary Office Loads

AI data centers have high load factors, high power density, and high sensitivity to outages and power quality. They also bring heat rejection, water resources, and backup power into site selection. A rack drawing 5-10kW used to be sufficient. Today, 50-150kW per rack is increasingly part of the discussion. OCP, NVIDIA, and Schneider are already discussing 500kW and even 1MW class racks, which shows that the issue is no longer a marginal increase in electricity use. Power architecture and thermal architecture both have to be redesigned.

The same 500MW data center has very different implications depending on where it is located. In PJM, especially Northern Virginia, it first runs into capacity-market, transmission, and local distribution pressure. In ERCOT, it is more like a stress test of real-time prices, summer peaks, natural gas, and storage flexibility. In vertically integrated utility territories in the Southeast, it may enter the regulated asset base and state-level rate cases. In Arizona, Nevada, and Utah, water, heat, and cooling may become the first bottlenecks. In the Pacific Northwest, low-carbon power and cooling conditions are better, but communities, transmission, and local rates can also become constraints. Electricity demand can rise nationally, but money and projects are realized region by region.

PJM can be understood as a large wholesale electricity market and grid dispatch system covering much of the eastern and mid-Atlantic United States. It is not a company, but a set of rules and platforms for wholesale power markets, capacity markets, transmission planning, and dispatch. ERCOT is the Texas electricity market and grid dispatch system, where real-time prices, summer peaks, gas-fired assets, and flexible resources matter more. Mixing the two together leads to bad conclusions about how data center load is monetized in different regions.

Versions of the same problem are visible outside the United States as well. Europe is more constrained by interconnection, permitting, land, communities, and electricity-price politics. Ireland, the U.K., Germany, and the Netherlands, all data-center-dense markets, have long been debating data center connection and grid carrying capacity. China is more inclined to use national planning to move compute toward energy-rich regions, with western renewables and ultra-high-voltage transmission as another answer. The Middle East combines natural gas, solar, water resources, and sovereign capital, trying to turn compute into a new form of energy export. Institutions differ, but the pressure is similar: AI is not happening only inside cloud companies. It has already passed the burden into the power system.

Chapter 4: Natural Gas Fills the Gap First, While Nuclear and Renewables Shape the Long-Term Structure

Back in the United States, the most realistic near-term source of incremental power remains natural gas. It is not the perfect answer, and it does not solve every carbon constraint, but it can provide dispatchable, scalable, and relatively controllable reliable capacity. That is why arrangements like Project Kilby are appearing.

Nuclear power is the more strategic line. Life extensions for existing plants, restarts, and small modular reactors (SMRs) will all matter to major cloud companies because they are closer to 24/7 low-carbon power. But the incremental scale they can deliver within five years is limited. Wind, solar, and storage will continue to grow and remain indispensable over the long term, but intermittency means the system also has to pay for more storage, frequency regulation, reserves, and transmission.

The answer to energy mix will not be a single-choice question. A more realistic combination is that gas first fills reliable capacity, nuclear carries the long-term imagination of low-carbon baseload, wind and solar plus storage continue to expand, transmission and substations move power closer to load, and behind-the-meter systems deliver stable power into racks. Short-term, medium-term, and long-term rhythms are different, and project owners will choose solutions by delivery timeline. Solutions that can power up first will receive a near-term premium. Solutions that can reduce carbon and system costs over the long run will continue to expand later.

Chapter 5: The Real Bottleneck Is Often Not Generation, But Delivery

The easiest place for this cycle to get stuck is often not generation, but delivery. From announcement to energization, a project has to pass through land and community issues, transmission interconnection, substations, long-lead equipment, power mix, behind-the-meter supply, cooling, and water resources. If any one of these gates is blocked, the GW in a press release will not become real MW. When observing a data center project, announced capacity is only the starting point. The more important question is where the project actually is: merely in an interconnection queue, or already through engineering studies, capacity procurement, an energy services agreement (ESA) or long-term power purchase agreement (PPA), equipment orders, substation construction, or real energization.

Equipment is the most practical gate in this cycle. Large power transformers, generator step-up transformers (GSUs), medium-voltage switchgear, high-voltage breakers, and prefabricated electrical houses (E-houses) cannot be replenished like generic industrial products. They require grain-oriented electrical steel, copper, insulation materials, high-voltage testing capacity, skilled installation and commissioning labor, and utility certification. Many projects are not missing an abstract “grid.” They are missing a qualified transformer arriving on time, a certifiable switchgear lineup, or a substation design that can pass testing and protection coordination. Equipment delivery is not a background condition. It is the energization schedule itself.

Transmission is even slower. Reports from Grid Strategies and ACEG show that new high-voltage transmission lines completed in the United States in 2024 were far below long-term demand. A commonly cited public figure is roughly 888 miles, while annual demand for high-capacity transmission may be close to 5,000 miles. It is not that the United States cannot build transmission; it has built faster in the past. The problem is that today’s permitting, cost allocation, cross-state coordination, and community resistance have made transmission slower than generation.

Chapter 6: FERC and Large-Load Rules Will Decide Who Pays

Institutional rules are also being pushed by engineering pressure. In June 2026, FERC issued large-load-integration show cause orders to six RTOs/ISOs, asking them to explain or revise rules around large-load interconnection, network-upgrade costs, co-located generation, interruptible load, and resource adequacy. These documents are not just regulatory news. Who pays for data center interconnection, whether costs are shifted to ordinary customers, how much upgrade funding project owners must post in advance, and whether behind-the-meter generation can bypass part of the queue all affect site selection and utility capex recovery.

The United States also does not have one unified national grid. Physically, the chain is the classic sequence of generation, transmission, substation step-down, distribution, and end use. But assets and dispatch are split across investor-owned utilities, municipal power providers, rural cooperatives, independent power producers, transmission owners, and regional markets. PJM, MISO, SPP, CAISO, ISO-NE, NYISO, and ERCOT are RTOs/ISOs that function like dispatch centers, wholesale exchanges, and open transmission platforms at the same time. A nationally growing load ultimately has to enter many local rules and regional price systems.

PJM is one of the most important windows to watch. It covers a broad region across the eastern and mid-Atlantic United States, and Northern Virginia, the world’s densest data center cluster, sits inside it. Data center pressure will show up simultaneously in load forecasts, capacity prices, transmission planning, RTEP, tariffs, and large-load interconnection rules. PJM’s changes are not the whole United States, but they illustrate one point: when high-reliability load concentrates in a region, the electricity gap first turns into capacity prices, interconnection rules, equipment queues, and local politics.

Cost allocation will become increasingly sensitive. Data centers bring tax revenue, jobs, and investment, but the costs of transmission, substations, backup capacity, and grid hardening may eventually enter ordinary customer bills. State regulators and local communities will not accept this transfer unconditionally. Large-load-specific rates, take-or-pay contracts, minimum usage commitments, higher interconnection deposits, interruptible service, self-supply power, storage, and load-curve disclosure will gradually move from special clauses to standard negotiations. Project Kilby shows one path: large customers pay for part of the incremental supply and supporting infrastructure themselves in exchange for speed and certainty.

Chapter 7: Supply Chain, Water, and Communities Will Become Part of Permitting

Supply-chain constraints have not disappeared. The United States wants to keep data centers, grid equipment, and critical energy manufacturing as domestic as possible. But storage cells, inverters, solar modules, power-electronic components, copper and aluminum processing, some low- and medium-voltage equipment, cooling-system components, power-semiconductor packaging and testing, and some transformer materials still depend on global supply chains. The more urgent the expansion, the more obvious the tension becomes: policy wants more links to stay in the United States, while project sites still need global supply to deliver on time. Tariffs and industrial policy can change cost curves, but they cannot instantly create qualified capacity.

Water and communities will matter more and more. As AI rack density rises, heat rejection and cooling choices directly shape site selection. Water-scarce, high-temperature regions will be more resistant to evaporative cooling, and communities will care about noise, cooling towers, transmission lines, substations, and electricity bills. In its public materials for Pecos, Microsoft specifically emphasized closed-loop cooling, non-potable water, and not drawing power from the existing public grid. These details show that permitting for large data centers is not just an engineering issue. It is also a community issue.

Looking down the value chain, the closest medium- to long-term bottlenecks remain primary electrical equipment and the grid side. Large power transformers, GSUs, high-voltage switches, breakers, substations, E-houses, medium-voltage distribution, busbars, and cable have long lead times, low substitutability, and slow certification. Hitachi Energy, Siemens Energy, GE Vernova, Eaton, Schneider, ABB, Hubbell, nVent, Prysmian, and Nexans represent different businesses from grid equipment to distribution connectivity. They should not be put in one basket simply because they are all “electrical.”

Chapter 8: The Investment Map Should Be Layered by What Each Company Monetizes

The logic on the generation and power-source side is different. GE Vernova, Siemens Energy, and Mitsubishi map to gas turbines and large power equipment. Caterpillar and Cummins map to reciprocating gas engines, backup power, and on-site generation. Bloom Energy maps to fuel cells, with a value proposition around shortening the wait from project planning to actual power supply. Constellation and Vistra look more like platforms holding available power resources, capacity-price elasticity, and long-term contracts. Chevron’s entry shows that oil and gas companies may eventually sell not just natural-gas molecules, but a more certain power service that bundles gas, generation equipment, and long-term supply arrangements.

Cooling and thermal management can no longer be treated as back-end data center equipment. Coolant distribution units (CDUs), cold plates, dry coolers, chillers, pumps and valves, quick disconnects, and monitoring systems will become part of platform design as rack density rises. Vertiv, Modine, and nVent do not matter simply because they have liquid-cooling products. The more important question is whether they can integrate thermal management, power management, field integration, and aftermarket service. Storage, microgrids, and flexible resources provide peak shaving, backup, frequency regulation, and load shaping, with Fluence, Tesla, GE Vernova, Eaton, Powell, and Flex positioned in different places.

Utilities and power-resource platforms are closer to holders of “approvable grid capacity.” Dominion, NextEra, Southern, and similar regulated utilities have advantages in regional monopoly positions, regulated asset bases, and long-term capex recovery, while facing pressure from electricity-price politics, capex approval, and customer pushback. The attraction of power-resource platforms such as Constellation and Vistra comes from dispatchable generation, capacity prices, and large-customer long-term contracts, but they also face fuel, regulatory, decommissioning, and asset-reinvestment issues.

Chapter 9: Conclusion: The U.S. Power Market Is Moving from Selling Electricity to Delivering Usable Power

Taking a longer view, the near-term shortage is reliable capacity that can be connected quickly, so natural gas, on-site generation, backup power, storage, and interruptible-load arrangements will become important first. The medium-term shortage is equipment, substations, transmission lines, EPC capacity, and local permitting, so power equipment and grid construction will enter a longer development cycle. The long-term shortage is an institutional capability: power markets must be able to pay ahead of time for reliable capacity, flexibility, transmission, and load management. Otherwise, real-time prices alone cannot build enough power supply and grid capacity several years in advance.

In that sense, the U.S. power market is moving from a market that sells electricity to a market that delivers usable power. The central tension over the next several years is the gap between reliable electricity demand and the speed at which the power system can expand. In the short term, natural gas will fill part of the gap. In the medium to long term, power equipment, substations, transmission, distribution, behind-the-meter supply, and cooling systems will enter a longer construction cycle. The companies worth studying over the long run are not necessarily the ones with the loudest stories, but those that can deliver on time, continue to be chosen by customers, and keep their position in the next project cycle.

Looking back at Project Kilby, the most important point is not the surface label that “Microsoft used natural gas.” It is that large technology companies have begun to pay separately for speed, certainty, and reliable capacity. Whoever controls connectable power, long-lead equipment, field delivery capability, cooling systems, and backup power is closer to the real bottleneck of the AI factory era. When watching the U.S. power market from here, it is better to focus less on broad slogans and more on four concrete questions: whether the project has signed a real power-supply agreement, whether equipment has been ordered and scheduled, whether regulation allows cost recovery, and whether final revenue can turn into gross margin and cash.

---