How AI Data Centers Are Rebuilding the U.S. Power Stack

August 18, 2026

Overview

Artificial intelligence is beginning to change the architecture of electricity supply, not only the quantity of electricity that data centers consume. The first phase of the AI infrastructure debate focused on whether utilities could deliver enough power. That remains the central institutional constraint. But once the grid becomes slow, congested, politically contested, or unable to offer a credible energization date, a second question appears: what kind of power system is operationally viable around an AI campus?

This article begins where the Hi K Robot Analysis America's Power System Was Not Built for AI Civilization leaves off. That analysis examined the U.S. grid as an institutional system: regulated utilities, regional grid operators, interconnection queues, transformer shortages, local politics, ratepayer protection, and the widening gap between the speed of digital investment and the speed of physical infrastructure. The conclusion was that the scarce resource is not electricity in the abstract. It is deliverable power at a known place and time.

The technological response is now becoming visible. AI developers are not choosing one replacement for the grid. They are assembling multiple power pathways at once. Existing nuclear plants are being restarted, uprated, relicensed, and placed under long-duration contracts. Competitive generators are turning merchant plants into contracted infrastructure for hyperscalers. Natural gas is being used where speed and dispatchability matter more than a perfect carbon profile. Fuel cells are moving behind the meter to energize campuses before transmission upgrades arrive. Co-location is being redesigned around hybrid arrangements that preserve access to the public grid. Advanced nuclear developers are moving from concept demonstrations toward commercial programs backed directly by technology companies. Battery storage is becoming the fast coordination layer that links all of these systems together.

The result is not a single winning technology. It is the emergence of a new AI data center power architecture: a stack in which generation, grid connection, onsite power, storage, contracts, controls, and compute scheduling are designed together. If this architecture continues to develop, the data center of the 2030s may look less like a building that purchases electricity and more like a digitally controlled industrial energy system that happens to produce machine cognition.

Scope Note: This is an analytical, educational, and non-commercial assessment of emerging technical pathways and infrastructure architectures over a 5–15-year horizon. Company examples and financial figures are used only to show operating scale, deployment choices, and structural incentives; they are not endorsements, valuation judgments, investment recommendations, or predictions of company performance. Company-reported targets, backlogs, contracted capacity, customer commitments, and project timelines are treated as statements as of their publication dates and remain subject to financing, permitting, construction, supply chains, fuel availability, customer performance, and execution. Timelines for advanced nuclear and other emerging systems remain conditional on licensing and project delivery. The U.S. cases examined here do not imply that one national energy system is superior or universally transferable.

The Power Problem Is Becoming an Architecture Problem

A conventional data center could treat electricity as a utility service. The developer selected a site, negotiated a connection, installed backup generators and uninterruptible power systems, and expected the public grid to remain the primary energy source. That model works when the requested load is small relative to the surrounding network and when new capacity can be delivered on a predictable schedule.

AI changes both assumptions. Large training campuses can request hundreds of megawatts or more, and the economic value of the installed compute can make a multi-year power delay more costly than a higher electricity price. At the same time, high-density accelerator clusters place demanding requirements on cooling, power conversion, redundancy, voltage quality, and the ability to respond to rapid changes in load. At AI-campus scale, the power system increasingly cannot be treated as a design step that follows the servers. Electrical supply, cooling, redundancy, and compute deployment are becoming more tightly coupled in project planning.

This distinction matters because no generation technology solves every requirement. Existing nuclear plants can provide large blocks of firm, low-carbon energy but cannot be copied rapidly from one site to another. Natural-gas generation can be dispatchable and comparatively fast to develop, but fuel pipelines, turbine supply, emissions permits, and carbon goals create constraints. Fuel cells can be modular and located beside the load, yet they still require fuel and long-term service. Batteries respond almost instantly but store energy rather than create it. Advanced nuclear could eventually combine firm output with smaller modular units, but most commercial deployments remain years away.

The emerging solution is therefore layered rather than singular. A campus might use firm grid service for one portion of its load, onsite generation for another, batteries for fast response, a long-term nuclear or gas contract for economic certainty, and workload management to reduce demand during constrained periods. The architecture becomes valuable because the weaknesses of one component can be partially covered by another.

Power pathway	Primary role in an AI campus	Structural advantage	Primary constraint
Existing nuclear	Large blocks of firm, low-carbon energy	Operating sites, workforce, licenses, transmission access	Finite fleet, outages, uprate limits, relicensing and restart work
Natural-gas generation	Dispatchable near- and medium-term supply	Flexible operation and mature generation technology	Fuel infrastructure, turbine availability, emissions and carbon exposure
Onsite fuel cells	Behind-the-meter primary or supplemental power	Modular deployment close to the load	Fuel cost, servicing, emissions when using natural gas
Advanced nuclear	Potential modular firm power in the 2030s	New designs, possible factory replication, high energy density	Licensing, first-of-a-kind construction, fuel and manufacturing scale
Battery storage	Fast balancing, ride-through, peak management and coordination	Millisecond response and modular deployment	Duration, cycling economics and dependence on an external energy source

Deployment maturity matters as much as technical promise. Existing nuclear plants, conventional natural-gas generation, fuel cells, and battery storage already operate commercially and can be contracted or deployed within known industrial systems, even though each remains constrained by site, equipment, fuel, permitting, or duration. Nuclear restarts, uprates, and customer-backed new gas plants sit in a second category: they use established technologies but still depend on project-specific construction, licensing, procurement, and commissioning. Advanced modular reactors occupy a third category. Their potential significance is real, but broader system impact remains conditional on successful licensing, first-of-a-kind execution, qualified fuel, repeatable manufacturing, and fleet deployment. The pathways therefore do not carry equal present-day deployment weight.

This is the key shift. AI electricity demand is not simply creating a market for more generation. It is forcing customers to think about the complete path from fuel or reactor to generator, from generator to grid or campus, from power electronics to GPU rack, and from instantaneous load behavior back to the control software that orchestrates the compute.

Power Autonomy Is Moving from Option to Design Requirement

Two August 2026 cases show why power autonomy is becoming a design requirement rather than an optional resilience feature. In Vineland, New Jersey, Nebius's expansion moved from an unresolved August 5 hearing into a special August 17 planning-board process for an amended two-phase campus. The City of Vineland's official agenda described both phases as already under construction and listed a much larger internal energy system around the data halls: multiple ground-mounted and multi-story Bloom power areas, LNG infrastructure, a chiller building, water treatment, and modular megabay structures. Subsequent reporting on August 18 said the board approved the amended Phase 2 plan. The sequence is important because it shows that behind-the-meter power can shorten dependence on transmission without making local permission irrelevant. In Texas, Governor Greg Abbott on August 3 ordered a comprehensive audit of the state's rapidly expanding data-center pipeline as ERCOT faced about 474 GW of proposed new load requests—more than five times the state's historical peak demand—with officials estimating that roughly 90 percent of the queue was associated with data centers. These cases, also examined from the institutional side in the K Robot Analysis of the U.S. power system, show that even after a developer secures land, compute equipment, and capital, the power pathway can still be delayed by grid verification, local permitting, water, emissions, or community consent.

That pressure changes technical design. Developers have stronger incentives to bring generation closer to the load, phase energization, add storage, secure powered land before the servers arrive, and finance customer-backed power so that a campus is not dependent on one future grid date. Power autonomy is therefore moving from a resilience feature toward a design objective: not complete isolation from the grid, but enough control over generation and delivery to keep compute deployment from being synchronized to the slowest external infrastructure process.

Nebius and CoreWeave: Energized Megawatts Become an Operating Constraint

The latest public filings and quarterly disclosures from Nebius and CoreWeave provide a direct operating view of the U.S. AI power constraint. Neither company is a regulated utility; both sell AI compute. Yet both now report power capacity, infrastructure deployment, customer commitments, and capital intensity alongside cloud revenue. Their disclosures show that growth depends not only on access to GPUs, but on how rapidly contracted land, electrical capacity, data halls, cooling, networking, and compute equipment can be converted into active capacity.

Metric	Nebius (NBIS), Q2 2026	CoreWeave (CRWV), Q2 2026	Why it matters for U.S. power
Quarterly revenue	Company-reported: about $582 million group revenue; about $575 million from Nebius AI cloud	Company-reported: about $2.58 billion	Revenue scale shows why power availability has become part of the operating buildout rather than a peripheral utility input.
Capacity signal	Company target: 5 GW of contracted power by year-end 2026; planned deployment of more than 1 GW per year beginning in 2027	Company-reported: 1.5 GW active power and approximately 3.7 GW contracted power at June 30	The operating constraint is the conversion from contracted megawatts into energized megawatts.
Capital intensity	Company-reported: about $5.7 billion of Q2 capital expenditures	Company-reported: about $14.1 billion of cash purchases of property and equipment in the first half of 2026	AI cloud expansion is pulling financing into GPUs, data-center infrastructure, electrical systems, and powered sites at the same time.
Demand financing	Company expectation: more than $9 billion of customer prepayments in 2026; roughly 70% of Q2 deals included prepayments	Company-reported: approximately $104 billion of revenue backlog at June 30, excluding more than $25 billion of net new customer commitments added in early Q3	Long contracts, backlog, and prepayments can support infrastructure financing before all contracted capacity is online, while leaving execution and customer-concentration risk in place.

Nebius reported Q2 2026 group revenue of about $582 million, including about $575 million from Nebius AI cloud, and said growth was driven in part by capacity scaling alongside pricing and utilization. The company reported about $5.7 billion of capital expenditures during the quarter, primarily for GPUs, GPU-related hardware, and data-center expansion. It also raised its year-end contracted-power target to 5 GW and stated that it plans to deploy more than 1 GW per year beginning in 2027. Those capacity figures are company targets rather than energized output, so their structural value lies in showing the scale of infrastructure the company is attempting to convert into operating compute.

CoreWeave reported Q2 2026 revenue of about $2.58 billion and approximately $104 billion of revenue backlog at June 30, with the company noting that the figure excluded more than $25 billion of net new customer commitments added in early Q3. During the quarter, CoreWeave said it brought nearly 500 MW of additional power into service, taking active power to 1.5 GW, while contracted power reached approximately 3.7 GW. Its data-center page subsequently listed 51 facilities, 1.5 GW of active power, and 4.2 GW of contracted power capacity. The difference between active and contracted capacity is important: contracted power indicates a future supply path, while active power is closer to the capacity that can support operating infrastructure.

Taken together, these disclosures indicate a partial change in the sequence by which large U.S. AI loads are being built. Utility planning and interconnection remain essential, but some cloud operators are securing land and power ahead of final energization, using customer commitments, prepayments, debt, leases, and infrastructure partnerships to finance capacity in parallel with grid and facility development. The public grid remains a central institution in the supply path, while private contracts and onsite systems are taking on a larger role in determining when a specific campus can operate.

One observable U.S. response to the AI power constraint is therefore a more distributed, contractual, customer-financed, and hybrid buildout around the existing grid. New utility generation and transmission remain necessary for many sites, but they are being supplemented by behind-the-meter generation, dedicated power projects, colocated infrastructure, long-term PPAs, batteries, flexible workloads, powered-land portfolios, and project finance tied directly to AI customers. This can shorten some transmission or financing dependencies, but it transfers more execution risk to cloud operators, generators, equipment suppliers, capital providers, fuel systems, and local permitting processes. The emerging architecture changes where the bottleneck sits; it does not remove the need to satisfy physical and institutional constraints.

NRG: Customer-Backed Generation Exposes Conversion Risk

NRG Energy's August 2026 disclosure provides a cleaner operating example of conversion risk. The company said it had aligned on principal commercial terms with a global cloud and AI hyperscaler for a 1.2 GW combined-cycle gas project in Texas, while stating that the project remained subject to final documentation and approvals. NRG also reported that its 415 MW T.H. Wharton project had reached commercial operations and that two additional Texas Energy Fund projects remained in development. The contrast is useful: one project had crossed into operating generation, while the 1.2 GW hyperscaler project remained a proposed customer-backed pathway.

A gigawatt of proposed AI demand is therefore not the same as a gigawatt of operating electricity demand or generator revenue. Conversion depends on customer commitment, definitive agreements, financing, permits, turbines and other equipment, fuel infrastructure, grid studies, construction, commissioning, and final energization. Customer-backed generation can reduce some demand-risk and financing uncertainty, but it does not remove project-execution, fuel, permitting, or equipment constraints. The distinction between announced, contracted, construction-ready, and energized megawatts remains central to the power architecture described in this article.

Independent Power Producers Move Closer to the Compute Layer

The institutional split between regulated utilities and independent power producers becomes technologically important in this second stage. A regulated utility is primarily responsible for serving customers within a territory under approved tariffs and reliability obligations. A competitive generator earns money by operating power assets, selling electricity and capacity, hedging market exposure, and negotiating bilateral contracts. These two business models create different responses to AI demand.

Before the current data center cycle, many competitive generators lived with substantial merchant exposure. Their plants sold into wholesale markets whose prices changed with weather, natural-gas prices, outages, transmission constraints, capacity markets, and regional supply-demand balances. A nuclear or gas plant could be physically valuable and still experience volatile economics if market prices were weak.

AI creates a new counterparty with a different objective. A hyperscaler does not only want inexpensive energy. It may value a long contract, a defined quantity of capacity, a credible operating life, an adjacent development site, and a generator willing to coordinate new transmission or data center infrastructure. This can convert part of a merchant fleet into something closer to contracted digital infrastructure.

The change is visible in agreements across the industry. Constellation Energy is restarting an existing Pennsylvania nuclear unit under a 20-year agreement with Microsoft. Talen Energy expanded its Susquehanna relationship with Amazon to a long-term arrangement that can reach 1,920 MW. Vistra signed 20-year agreements with Meta supporting more than 2,600 MW of nuclear energy and planned uprates across three PJM plants. These are not identical transactions, but they all demonstrate the same structural movement: the generator is being pulled closer to the planning horizon of the compute customer.

This does not eliminate merchant markets. It changes the mix. A generator can keep some output exposed to wholesale prices while placing another portion under longer contracts. The result may be a portfolio that combines contracted cash-flow visibility with merchant upside and operational flexibility. For the AI customer, the important point is different: a bilateral relationship can reserve industrial capability that might otherwise be exposed to a market the customer does not control.

Nuclear Power Returns Because Existing Infrastructure Is Hard to Recreate

The renewed interest in nuclear power is often described as a change in public attitudes toward nuclear energy. That is part of the story, but it is not the most important structural explanation. Existing nuclear plants combine several assets that are extraordinarily difficult to reproduce on an AI product cycle: a licensed site, a trained workforce, a high-capacity generator, fuel logistics, cooling systems, grid interconnection, transmission access, and decades of operating knowledge.

For much of the 2010s, some merchant nuclear plants struggled against low wholesale power prices and inexpensive natural gas. In that environment, a large plant could be technically reliable yet economically vulnerable. AI reverses the scarcity. The value of firm output rises when new large loads are searching for continuous capacity, while the time required to construct new generation makes an existing plant more valuable than its raw megawatt rating suggests.

Crane: Restarting an Existing Nuclear Asset

Constellation's Crane Clean Energy Center illustrates the logic. The plant, formerly Three Mile Island Unit 1, shut down in 2019 for economic reasons. Microsoft later signed a 20-year agreement that supports the restart. Constellation says the project will restore about 835 MW of carbon-free generation to the grid and now describes the unit as potentially returning as early as 2027.

The important point is not that restarting a nuclear plant is easy. It requires equipment restoration, regulatory review, workforce rebuilding, inspections, licensing work, and grid coordination. The point is that the project begins with a physical system that already exists. The site does not need to invent a nuclear workforce, acquire an entirely new location, or establish a transmission corridor from nothing.

For AI infrastructure, that difference can be decisive. A restarted plant can be viewed as a way to recover a dormant power node rather than create a new one. Similar logic applies to life extensions and uprates. Extending the operating life of a reactor preserves an existing energy platform; an uprate attempts to extract additional megawatts by replacing or improving equipment inside a plant that already has a grid position.

Vistra and Meta: Keeping Existing Reactors Alive and Adding Output

Vistra's agreements with Meta add another layer. The 20-year contracts cover operating generation from Perry and Davis-Besse and planned uprates at Perry, Davis-Besse, and Beaver Valley. Vistra said the full package totals 2,609 MW, including 433 MW of incremental output from uprates, with capacity entering over time through 2034.

This is technically different from building a new reactor fleet. The customer is helping support the long operating life of existing plants while also financing a pathway to increase output. The underlying infrastructure already includes reactors, turbines, cooling systems, transmission connections, operating licenses, and staff. The capital is therefore being directed toward preserving and expanding an installed energy machine.

If similar contracts spread, one likely consequence is that the boundary between corporate energy procurement and utility-scale generation planning will become less distinct. A technology company may not own the plant, but its long-term demand can determine whether a generator pursues license extension, equipment replacement, uprates, or additional transmission work.

Talen and Amazon: Nuclear Power Meets Data Center Geography

Talen's Susquehanna site is important because it connects generation to geography. Amazon originally acquired the adjacent Cumulus data center campus, creating one of the most visible experiments in co-locating a major computing load next to a nuclear plant. The concept was intuitively attractive: instead of transporting power across a constrained network to a distant data center, place the data center beside the generator.

The later agreement became larger and more integrated with the public grid. Talen's 2025 PPA can supply Amazon with up to 1,920 MW of nuclear power through 2042, with the full quantity expected no later than 2032 and the possibility of acceleration. The existing co-located arrangement is transitioning to a front-of-the-meter structure in which Susquehanna supplies the PJM grid, Talen acts as Amazon's retail supplier, and PPL provides transmission and delivery.

That evolution is crucial. It shows that co-location does not necessarily mean isolation. The strongest architecture may be physical proximity combined with continued grid participation. The data center gains a nearby generation anchor and an existing energy site, while the broader system retains a defined interface for transmission, reliability, and cost allocation.

Natural Gas Becomes the Bridge Between AI Time and Nuclear Time

Nuclear power solves one set of problems but creates another: time. Existing reactors can be extended or uprated, and selected closed plants can potentially restart, but the number of such opportunities is finite. New conventional nuclear units remain major construction projects. Advanced reactors may become more modular, but their commercial fleet buildout is still emerging.

AI developers cannot suspend infrastructure expansion until every long-term clean-energy technology is mature. This creates a bridge market for dispatchable generation that can be developed with familiar equipment, financed under existing commercial models, and operated around the clock when fuel is available. In the United States, that bridge is increasingly natural gas.

Combined-cycle gas turbines convert fuel into electricity through both a gas turbine and a steam cycle, improving efficiency relative to simple-cycle generation. Simple-cycle turbines and reciprocating engines can offer faster response and modularity for some applications. In both cases, gas generation has a property AI developers value: it can be dispatched when needed rather than waiting for weather conditions.

The trade-off is that a gas plant is not self-contained. It depends on turbine manufacturing, gas pipelines, compression infrastructure, fuel contracts, air permits, cooling design, and the price and availability of natural gas. A region can have abundant gas production and still face local pipeline constraints. A hyperscaler can secure a power plant and still discover that the fuel system, not the electrical system, is the next bottleneck.

NRG and the Customer-Backed CCGT Model

NRG Energy provided a concrete example in August 2026 when it announced that it was advancing a Bring Your Own Power (BYOP) strategy with a hyperscaler around a 1.2 GW combined-cycle gas turbine project in Texas. Here, BYOP is NRG's own industry-facing term for a specific customer-backed generation model. It is distinct from the Build / Bring / Buy analytical taxonomy used in the companion K Robot Analysis article: the earlier taxonomy classifies three broad ways hyperscalers can secure power, while NRG's BYOP names one particular commercial structure for bringing customer-supported generation to a large load.

The timing of the announcement was instructive because it came during the same week that Texas tightened scrutiny of the data-center pipeline. NRG's second-quarter Adjusted EBITDA was $1.217 billion and Adjusted EPS was $1.49, while Texas segment Adjusted EBITDA fell to $381 million from $512 million a year earlier. Those figures do not invalidate the long-term AI power thesis; they show why a generator can believe in that thesis while still wanting a customer-backed structure that protects a specific project from the uncertainty between forecast load and an energized campus.

This structure is important because it changes who carries the demand risk. Instead of a utility building a large power plant based primarily on a forecast that future data centers will arrive, the large customer can support the generation project directly. The generator receives a more credible path to financing and construction; the data center receives a more specific supply pathway; and regulators can more clearly ask whether the customer rather than ordinary ratepayers is carrying the incremental cost.

Natural gas is therefore more than a fuel choice in this architecture. In the AI era it can function as a time-conversion technology: it converts a customer's willingness to accept fuel and carbon exposure into earlier dispatchable capacity. That can be rational when the economic value of operating compute earlier is greater than the cost difference between gas generation and a slower alternative.

The Carbon Constraint Does Not Disappear

The fact that gas can arrive sooner does not make its environmental trade-offs disappear. Hyperscalers have corporate carbon goals, communities may oppose new combustion sources, and future policy can change the economics of emissions. Carbon capture, lower-carbon fuels, renewable gas, hydrogen blending, and contractual clean-energy matching may reduce parts of the problem, but none automatically removes the physical emissions associated with fossil natural gas.

This creates a likely portfolio logic. Gas can fill near- and medium-term gaps while nuclear, renewables, storage, transmission, geothermal, or other clean firm resources expand. Whether that bridge becomes temporary or entrenched will depend on how quickly alternative firm capacity scales, how carbon rules evolve, and whether AI demand grows fast enough to keep dispatchable generation scarce.

For this reason, the most resilient power developers may be those that can operate across more than one generation pathway. A portfolio containing nuclear, gas, geothermal, storage, and retail supply can respond differently to a customer's timeline than a company dependent on a single technology.

Co-Location Is Evolving from Bypass Strategy to Hybrid Architecture

When grid connection becomes the bottleneck, placing a data center beside a generator appears to offer a shortcut. Electricity can travel a short physical distance. The project can reuse land, substations, switchyards, fuel infrastructure, water rights, or transmission corridors associated with an existing power site. In some configurations, part of the load can sit behind the meter and reduce the amount of electricity that must be transported through the public network.

But a data center rarely becomes an electrical island simply because it is adjacent to a generator. Large computing campuses still value backup supply, black-start planning, maintenance support, reserve capacity, and the ability to import or export power under abnormal conditions. Nuclear plants refuel. Gas units trip. Fuel cells require maintenance. Batteries eventually discharge. A public grid remains valuable even when the campus is designed to minimize its dependence on that grid.

This is why the Talen-Amazon evolution matters beyond one transaction. The model is moving away from the simplistic idea that a hyperscaler can place a private wire beside a power plant and ignore the surrounding system. The emerging architecture is hybrid: physical proximity, contractual control, dedicated infrastructure, and continued participation in a larger network.

That hybrid model may be more durable because it assigns explicit roles. The generator provides energy and capacity. The grid provides network redundancy, balancing, and regional reliability. The data center contributes a large creditworthy load. Storage and onsite systems can reduce peaks or ride through disturbances. Software can coordinate how and when each layer responds.

The PPA Is Becoming an Infrastructure Design Tool

A power purchase agreement is often described as a financial contract for buying electricity. In the AI buildout, the most important PPAs are becoming closer to infrastructure design documents. They can determine how long a plant remains open, whether a restart is economic, whether an uprate proceeds, how much new capacity is financed, when contracted volumes ramp, and which party carries market risk.

This is a significant change from the earlier corporate renewable PPA model. A financial or virtual PPA could help a company match annual electricity consumption with renewable generation elsewhere on the grid. That structure remains useful for carbon accounting and project finance. But an AI campus that needs a credible physical power pathway cares about additional variables: deliverability, firm capacity, location, transmission rights, operating life, outage behavior, ramp schedule, fuel exposure, and the relationship between contracted energy and actual grid service.

The Talen-Amazon agreement illustrates the ramp concept. The final 1,920 MW quantity is not assumed to appear on day one; contracted delivery increases over time. Vistra's Meta agreements similarly combine existing output with future uprates. Constellation's Microsoft agreement supports a plant restart whose output does not exist until the physical facility returns to service.

In other words, a PPA can reserve future industrial capability. It can coordinate capital before the megawatts exist. This is one reason technology companies are becoming more deeply involved in energy projects: the contract is no longer only a way to purchase electricity after infrastructure is built. It can become part of the mechanism that causes the infrastructure to be built, restarted, extended, or expanded.

Onsite Fuel Cells Push Generation Inside the Data Center Boundary

The first article used onsite power only as evidence that grid constraints are changing development behavior. This section deliberately resumes those examples at the technical layer, where the differences in scale, modularity, fuel use, deployment time, and campus design become the subject rather than the policy backdrop.

Large power plants and grid-connected PPAs solve energy at the regional level. Fuel cells attack the problem from the opposite direction. Instead of moving the data center toward a distant source of electricity, they move generation directly beside the data center.

Bloom Energy's solid oxide fuel cells electrochemically convert fuel into electricity without the combustion process used by turbines or reciprocating engines. When supplied with natural gas, the system still produces carbon dioxide, so it is not zero-carbon generation. Its strategic attraction lies elsewhere: modularity, high power density, onsite siting, comparatively low local air pollutants, low water use, and a deployment pathway that can reduce dependence on new long-distance transmission.

CoreWeave: From One Fuel-Cell Site to a Multi-Gigawatt Power Portfolio

Bloom's 2024 partnership with CoreWeave placed fuel cells at a high-performance data center in Volo, Illinois, with commissioning originally targeted for the third quarter of 2025. The importance of that project was not its absolute scale. It demonstrated that an AI cloud operator was willing to treat onsite generation as part of the computing platform rather than as emergency backup equipment.

By the second quarter of 2026, the more important CoreWeave story was the scale of the entire power portfolio. The company reported 1.5 GW of active power after bringing nearly 500 MW online during the quarter and approximately 3.7 GW of contracted power at June 30. Its current infrastructure page now lists 51 data centers, the same 1.5 GW of active power, and 4.2 GW of contracted capacity. The difference between active and contracted power is not an accounting footnote. It is the physical backlog the company must convert into operating AI capacity.

The financial statements make that conversion visible. CoreWeave generated $2.575 billion of Q2 revenue, but it also carried approximately $104 billion of revenue backlog at quarter end and disclosed more than $25 billion of additional customer commitments in early Q3. In the first half of 2026 it paid $14.1 billion for property and equipment, with the company describing the spending as infrastructure investment across its GPU fleet, networking equipment, servers, switches, and related assets. This is why access to power has become inseparable from access to capital: a customer commitment is economically useful only if CoreWeave can finance and energize the corresponding physical stack.

CoreWeave's model also shows why the future solution is unlikely to be a single power technology. The company leases and operates capacity across many facilities rather than waiting to own one vertically integrated utility system. It can combine utility service, colocation, powered land, onsite generation, liquid cooling, high-density electrical design, and cross-cloud networking across different markets. Its earlier Bloom project is therefore one component of a broader strategy that diversifies the routes by which a GPU cluster can reach an energized state.

This is partial electrical self-sufficiency at portfolio scale rather than a literal off-grid cloud. The public grid still supplies and balances much of the system, but CoreWeave is increasingly responsible for assembling the surrounding infrastructure fast enough that power availability does not become the pacing item for customer contracts. In that sense, the AI cloud operator is moving one layer downward into energy development.

Oracle: Onsite Power Reaches Gigawatt Scale

The scale changed dramatically in 2026. Oracle and Bloom expanded their relationship under a master agreement supporting up to 2.8 GW of fuel cell capacity, with an initial 1.2 GW contracted and deployment underway across U.S. projects.

Gigawatt-scale procurement changes the category. Fuel cells are no longer being considered only for a small resilience layer around a conventional grid connection. They are being evaluated as utility-scale blocks assembled from modular equipment at or near the load.

This does not imply that every data center can economically or operationally become fuel-cell powered. The economics depend on fuel prices, gas access, equipment cost, maintenance, emissions policy, and the availability of utility power. But when the alternative is waiting years for transmission reinforcement, the relevant comparison is not simply levelized electricity cost. It is the cost of power plus the value of time.

Nebius: Power Autonomy Becomes an Explicit Design Objective

The 2026 Nebius-Bloom agreement makes this logic even clearer. Nebius selected Bloom for a first project targeting 328 MW of installed behind-the-meter capacity, expected to be operational during 2026. The company explicitly cited fast time-to-power, reduced dependence on new transmission build, minimal water use, and a lighter permitting burden relative to combustion-based alternatives.

Nebius's second-quarter results now show why that 328 MW project matters financially. Group revenue reached $582.3 million, Nebius AI cloud revenue reached $574.9 million, and AI cloud adjusted EBITDA margin was 49.7 percent. Management said capacity scaling was the principal driver of revenue growth. At the same time, Q2 capital expenditures were approximately $5.7 billion, and the company expects more than $9 billion of customer prepayments during 2026. Roughly 70 percent of Q2 deals included prepayments, with those funds covering an estimated 50–60 percent of associated capital expenditure. This is a direct mechanism through which future AI demand can finance present-day electrical and compute infrastructure.

The deployment pipeline is expanding beyond one New Jersey site. Nebius told shareholders that it brought additional capacity online in the United States and Europe during the second half of Q2, had delivered all capacity tranches then due under its Microsoft contract, was progressing construction at owned U.S. AI factories, and was building capacity for a second Meta agreement targeted for early 2027. The company raised its 2026 year-end target to 5 GW of contracted power and said it plans to deploy more than 1 GW of capacity per year starting in 2027. In Missouri, construction has begun on its first gigawatt-scale U.S. AI factory in Independence, while the Vineland project represents the complementary strategy of placing a substantial power system directly behind the meter.

The August 17 Vineland special-meeting agenda shows how far that concept extends. The amended campus plan included multiple Bloom power areas, LNG infrastructure, cooling and water-treatment systems alongside the data-center buildings. Four days before that meeting, WHYY reported that the city had issued stop-work orders covering LNG-tank work and fuel-cell units that officials said were proceeding before the required approvals. Subsequent reporting said the planning board approved the Phase 2 amendment on August 17. The earlier hearing is therefore not only a story about delay. It is also evidence that the object being permitted is no longer merely a data center. It is a coupled compute-and-energy campus whose internal infrastructure now resembles a small industrial power system.

The important concept is therefore not full energy independence. A large campus can still use the grid, batteries, backup generation, and other resources. The structural change is power autonomy: the data center tries to control enough of its own energy stack that the entire project schedule is no longer determined by one external interconnection date. In practice, the objective is partial self-sufficiency combined with grid optionality.

Nebius is also adding a second response to the constraint: an asset-light infrastructure-partner model in which outside owners finance and operate facilities while Nebius supplies its systems architecture, hardware design, software stack, and customers. That structure attacks two bottlenecks simultaneously—capital and capacity—and broadens the number of parties that can transform land plus power into saleable AI cloud capacity. If it works at scale, the solution to the power shortage will not come only from utilities building faster. It will also come from new contractual forms that let infrastructure capital, power developers, cloud operators, and end customers divide the buildout among themselves.

If that model spreads, the campus boundary and the commercial boundary will both expand. Data center developers may increasingly secure gas connections, generation equipment, switchyards, batteries, controls, water systems, and service contracts at the same time they secure GPUs and cooling systems, while customers and financing partners fund parts of the capacity before it is energized. Power procurement becomes part of product deployment, and local permitting becomes part of power engineering.

Advanced Nuclear Is Not One Technology

The phrase "small modular reactor" can hide more than it explains. Advanced nuclear developers are pursuing different reactor coolants, fuels, power outputs, heat-storage systems, manufacturing strategies, and commercial models. Their common promise is that nuclear power might become more repeatable and modular than the very large light-water projects that defined the previous construction era. Their common challenge is that first-of-a-kind systems must still prove licensing, construction, fuel supply, manufacturing, cost, and operating performance.

For AI infrastructure, advanced nuclear is attractive because the long-term demand profile is unusually compatible with firm generation. A large campus can anchor a multi-decade power contract and may have enough credit quality to support early development. Technology companies are therefore doing more than waiting for utilities to decide which reactors to build. They are becoming early commercial partners.

Oklo: A Scalable Power Campus Backed by Meta

Oklo and Meta announced an agreement in January 2026 supporting development of a nuclear power campus in southern Ohio that could scale to 1.2 GW. The structure allows Meta to provide prepayment and development funding intended to improve project certainty. Oklo says pre-construction and site characterization begin in 2026, with a first phase targeted as early as 2030 and potential expansion toward the full 1.2 GW by 2034.

The architecture is notable because it links modular deployment with customer financing. Rather than begin with one enormous plant that must achieve full scale before the customer receives power, multiple units could theoretically be added over time. If the licensing and construction model proves repeatable, generation could expand in increments closer to the growth pattern of a data center campus.

The uncertainty remains substantial. Commercial timelines depend on Nuclear Regulatory Commission approvals, fuel availability, component manufacturing, construction performance, and the ability to replicate early units economically. The important signal is therefore not that 1.2 GW is guaranteed. It is that a hyperscaler is willing to fund the development pathway years before the electricity is available.

Kairos Power: Demonstration Before Fleet Scale

Kairos Power is following a deliberately iterative path. Its technology is a fluoride salt-cooled high-temperature reactor using TRISO fuel. In April 2026, the company broke ground on Hermes 2 in Oak Ridge, Tennessee. Kairos describes Hermes 2 as its first commercial-scale reactor and the first power-producing Generation IV reactor to receive an NRC construction permit. The plant is designed to provide up to 50 MW to the Tennessee Valley Authority grid and is the first deployment under Kairos Power's agreement with Google.

The significance is methodological. Nuclear projects have historically suffered when developers attempted to move from design to very large commercial construction before enough manufacturing, licensing, and operating knowledge had been accumulated. Kairos is attempting to learn through a sequence of hardware programs before expanding toward a fleet.

If this approach works, the key innovation may be less the reactor physics than the industrial process: standardized modules, factory fabrication, repeated regulatory engagement, and a supply chain that learns from each unit. For AI customers, that could eventually create a nuclear product that behaves more like repeatable infrastructure and less like a one-off national megaproject.

TerraPower: Nuclear Generation with Built-In Energy Storage

TerraPower's Natrium architecture addresses a different problem. The design combines a 345 MW sodium-cooled fast reactor with molten-salt energy storage that can increase output to 500 MW for more than five hours. In January 2026, TerraPower and Meta announced an agreement supporting up to eight Natrium plants, representing up to 2.8 GW of baseload generation and as much as 4 GW of output when the storage systems discharge at higher power.

This coupling of reactor and storage is especially relevant to AI because it separates steady reactor operation from variable electrical output. A reactor can operate near a stable thermal condition while the storage system absorbs heat and later increases electricity production during periods of higher demand. In principle, the plant behaves less like a fixed block of baseload and more like a firm generator with built-in flexibility.

TerraPower reached an important regulatory milestone in March 2026 when the NRC approved the construction permit for the first Natrium plant at Kemmerer, Wyoming. The company expects that project to be completed around 2030, while the Meta-supported units are targeted to begin arriving later, with initial units as early as 2032.

Those dates are project targets rather than guarantees. But Natrium demonstrates why the future power stack may blur the line between generation and storage. Instead of bolting a battery onto every generation technology after the fact, some future plants may be designed around flexible output from the beginning.

X-Energy: Modular High-Temperature Reactors and Industrial Scale

X-energy and Amazon are pursuing another modular pathway. X-energy's Xe-100 is designed as an 80 MW high-temperature gas-cooled reactor using TRISO-X fuel, with multi-unit plants that can scale from hundreds of megawatts upward. Amazon and X-energy have described an initial 320 MW project with Energy Northwest in Washington and a broader ambition to support more than 5 GW of U.S. deployments by 2039.

The Xe-100 concept matters because modularity is not only about smaller reactors. It is about building a standardized unit that can be repeated. A 320 MW plant can use four modules; larger sites can add more units if economics, licenses, fuel supply, and customer demand support expansion.

Again, the industrial bottleneck is the important question. Advanced nuclear requires a manufacturing ecosystem for specialized fuel, reactor vessels, heat exchangers, control systems, construction labor, quality assurance, and licensed operators. A design becomes civilization-scale infrastructure only when those supply chains can reproduce it reliably.

The Real Competition Is Between Deployment Systems, Not Reactor Brands

It is tempting to rank advanced reactors as if they were competing consumer products. That misses the larger structural test. Oklo, Kairos, TerraPower, and X-energy are not only competing on reactor design. They are competing on the ability to convert design into repeated physical deployment.

The decisive variables will include licensing cadence, fuel qualification, manufacturing yield, construction labor, cost control, customer financing, site replication, grid integration, and operational learning. A technically elegant reactor that cannot secure fuel or manufacture components at volume may lose to a less novel system that can be built repeatedly. A reactor with a high projected output is not strategically valuable until it can move through permits, construction, commissioning, and commercial operation.

The meaningful metric is therefore not theoretical reactor performance alone, but the conversion rate from technical design into licensed, financed, fueled, constructed, commissioned, and continuously operating megawatts. That conversion rate determines whether a promising technology remains a demonstration node or becomes repeatable infrastructure.

This is the same principle that appears throughout the AI Civilization Map: capability matters only when it can be converted across layers. A reactor design is a technical node. Commercial nuclear capacity requires capital, institutions, supply chains, labor, fuel, licensing, land, transmission, and customers to align with it.

Technology-company participation may accelerate this conversion because hyperscalers bring large balance sheets and long-duration demand. But their money does not repeal physics. If a reactor fuel cannot be produced at scale, if a component factory is late, or if a project cannot pass safety review, a cloud contract cannot create missing megawatts by itself.

BESS Becomes the Fast Coordination Layer

Generation technologies operate on different time scales. Nuclear plants excel at sustained output but do not exist to chase millisecond load spikes. Gas turbines can be dispatchable, yet thermal equipment still has ramp constraints. Fuel cells can follow load but operate within equipment limits. The grid itself can experience voltage disturbances or local congestion. AI accelerators, by contrast, can shift power consumption much faster than traditional industrial systems.

This is why battery energy storage remains important even in a campus supplied by firm generation. The role is not to replace a nuclear plant or a gas pipeline. It is to coordinate the electrical system at a faster time scale.

As discussed in AI Data Centers Are Breaking the Grid, BESS can provide rapid frequency and voltage support, peak shaving, ride-through, and renewable shifting. In a more complex AI microgrid, the battery can also manage transitions among the public grid, onsite generation, backup systems, and compute load. A battery may absorb a sudden reduction in server demand before a thermal generator can ramp down, or discharge during a rapid load increase before other assets respond.

The deeper point is that storage acts as an interface. It converts slow energy assets into a system that can behave more quickly. That makes BESS less like a separate power source and more like the power stack's coordination fabric.

This also connects to the mineral and manufacturing constraints explored in From Batteries to Minerals. If storage becomes standard inside gigawatt-scale data center campuses, the electrical architecture propagates backward into cell manufacturing, power electronics, lithium and other mineral supply chains, thermal management, and fire-safety systems. The power stack is therefore linked to multiple layers of industrial capacity.

Compute Scheduling May Become Part of the Power Plant

The architecture becomes more interesting when demand itself becomes controllable. AI workloads are not all identical. Some inference must respond immediately to users. Some training jobs can checkpoint. Data preprocessing, evaluation, synthetic-data generation, and batch workloads may tolerate scheduling changes. Distributed cloud systems can also move some work geographically when network, model, data, and regulatory conditions allow.

If data center operators expose even a small amount of this flexibility to their energy control systems, the boundary between compute orchestration and power orchestration begins to disappear. A future scheduler may consider not only GPU availability and network topology but also battery state of charge, generator output, electricity prices, grid constraints, and carbon intensity.

The most sophisticated AI campus could therefore optimize two flows simultaneously: electrons and tokens. When power is abundant, flexible workloads accelerate. When the grid is stressed, batteries discharge, onsite generation increases, or non-urgent jobs pause. When a reactor refuels or a gas unit is offline, workloads migrate or contractual supply covers the gap.

This does not mean computing becomes infinitely flexible. Large distributed training runs can be sensitive to interruption. Real-time inference has service-level requirements. Moving workloads may create data-sovereignty, latency, and networking constraints. But even partial flexibility can change the amount of firm infrastructure required to serve the same annual compute output.

In that sense, software becomes part of energy infrastructure. A megawatt avoided at the system peak can be as valuable as a megawatt newly constructed, provided the workload can shift without destroying the economic value of the computation.

From Data Center to Power Campus

The combined effect of these trends is a change in what a data center campus contains. The old stack began with a utility interconnection and then added electrical distribution inside the fence. The emerging stack may begin earlier, with dedicated generation and energy contracts outside or inside the fence.

Traditional data center	Emerging AI power campus
Utility connection is the primary power source	Utility connection is one layer among grid, onsite generation, PPAs and storage
Backup generation is mostly emergency equipment	Onsite generation can become primary, bridge, or supplemental power
Energy procurement follows facility development	Energy strategy can precede and determine facility development
Carbon accounting can rely heavily on annual matching	Physical deliverability and hourly reliability become more important
IT scheduling and energy operations are largely separate	Compute orchestration may respond directly to power conditions
Power plant and data center are separate industries	Generation, storage and compute are increasingly co-designed

This transition can create new categories of infrastructure company. Competitive generators can become powered-land developers. Data center operators can become energy project sponsors. Fuel-cell manufacturers can become campus power suppliers. Nuclear developers can structure deployments around hyperscaler demand. Battery companies become part of compute reliability. Software providers can optimize the entire energy-compute system.

The important phrase is co-design. A 1 GW campus cannot be treated as a conventional building with a very large electric bill. It is closer to a heavy industrial district whose product is computation. The power plant, cooling plant, network, storage system, and accelerator fleet all become parts of the same production machine.

What Could Break the New Power Architecture

The emergence of a new architecture does not mean every announced project will be built. In fact, the more layers a project combines, the more execution points it creates. The most important uncertainties are physical and institutional rather than rhetorical.

AI Demand Could Arrive More Slowly Than the Power Pipeline

Electricity projects are long-lived. AI hardware cycles are fast. If future compute demand grows more slowly than expected, or if accelerator efficiency improves faster than electricity consumption, power developers could be left with infrastructure built around load forecasts that do not fully materialize.

Long-term contracts, customer contributions, phased construction, and modular designs reduce this risk but cannot eliminate it. The distinction between requested megawatts, contracted megawatts, construction-ready megawatts, and energized megawatts remains essential.

Natural Gas Could Move the Bottleneck Upstream

A gas-fired campus can reduce dependence on transmission only to become dependent on pipelines, turbines, compressors, fuel contracts, and air permits. Rapid growth in gas generation can create competition with residential heating, industry, LNG exports, and conventional utility demand. In some regions, pipeline expansion may be politically as difficult as transmission construction.

The result would be a relocation of the bottleneck rather than its elimination. AI infrastructure repeatedly converts one constraint into another as it moves deeper into the physical economy.

Advanced Nuclear Depends on Becoming a Manufacturing Industry

The most important advanced-nuclear risk is not whether one demonstration reactor works. It is whether the technology can be reproduced. A commercial fleet needs qualified fuel, repeated licensing, factories, suppliers, construction crews, financing, maintenance systems, and operating talent.

That is why the current projects are better understood as industrial learning systems. A first reactor can validate physics and licensing. A second can expose manufacturing problems. A fleet reveals whether the technology can actually scale.

Onsite Power Still Requires Community Legitimacy

Moving generation behind the meter can shorten transmission dependencies, but it also moves more energy infrastructure into the local community. Natural-gas plants, fuel cells, substations, batteries, LNG storage, pipelines, and cooling systems create questions about emissions, noise, safety, land, water, and emergency response. Vineland's multi-hearing approval process and Texas's decision to audit a roughly 474 GW large-load pipeline show how quickly those questions can become part of the project schedule even when the underlying demand for AI infrastructure remains strong. The fact that Vineland advanced after additional review does not remove the legitimacy constraint; it demonstrates that local review has become one of the engineering gates through which private power infrastructure must pass.

The technical implication is that equipment selection is increasingly shaped by permitting characteristics as well as electrical performance. Lower local emissions, lower water requirements, quieter operation, modular construction, smaller footprints, or easier fuel logistics can matter as much as nominal efficiency when a project needs political permission to operate. The architecture is therefore being optimized not only for electrons and cost, but for permission to build.

The Grid Remains the System of Last Resort

The most aggressive onsite architecture still benefits from a wider network. A public grid pools generators and loads across geography. It provides reserves, balancing, transmission diversity, emergency support, and economic dispatch. Replacing all of those functions inside every private campus would require enormous redundancy.

The likely future is therefore not grid versus off-grid. It is a renegotiation of the boundary. Some generation moves closer to the customer. Some storage moves inside the campus. Some power remains contracted remotely. The grid becomes one component in a hybrid system rather than the only primary source.

The AI Civilization Map: Foundation Becomes an Active Technology Layer

Within the AI Civilization Map, this transition sits first in Energy & Physical Infrastructure, the Foundation layer on which the rest of the AI system depends. But the consequences do not stop there.

The Foundation now interacts directly with Semiconductors, Compute & Packaging. Higher rack density changes the shape of the power system. Accelerator deployment schedules change the value of early energization. Liquid cooling changes the electricity and water balance. Chip efficiency changes how many tokens a megawatt can produce.

It also reaches Capital, Institutions & Operating Layers. Long-duration PPAs, hyperscaler prepayments, customer-backed generation, nuclear licensing, project finance, fuel procurement, and grid-service arrangements determine which technical systems leave the drawing board.

And it reaches Geopolitics, Sovereignty & Constraints. Nuclear fuel supply, gas infrastructure, turbine manufacturing, transformer capacity, domestic reactor supply chains, enriched fuel, critical minerals, and local permitting can all become strategic constraints. An AI civilization cannot scale merely by possessing better models if it cannot reproduce the physical energy systems that keep those models running.

The correct unit of analysis is therefore not one reactor, one gas plant, one battery, or one fuel-cell company. It is the conversion chain:

fuel and generation technology -> physical plant -> grid or onsite connection -> storage and power electronics -> data hall -> accelerators -> usable compute.

Each arrow can become the bottleneck. The power architecture matters because it is an attempt to make those arrows more controllable.

Current Deployment Patterns Favor a Portfolio Architecture

The AI power debate is often framed as a contest: nuclear versus gas, grid versus microgrid, batteries versus generation, conventional reactors versus SMRs. Current deployment patterns point instead to combinations of technologies selected around local constraints. The evidence does not establish that one portfolio will persist everywhere; it shows that no single pathway presently satisfies the full set of speed, scale, reliability, fuel, permitting, carbon, and network requirements across all sites.

Existing nuclear can anchor large low-carbon loads where plants already operate. Uprates and life extensions can preserve capacity faster than entirely new construction. Natural gas can supply dispatchable bridge power where fuel and permits are available. Fuel cells can move generation beside the data center and reduce dependence on transmission. Batteries can absorb fast electrical changes and integrate multiple sources. Advanced nuclear may add new modular firm power in the 2030s if licensing and manufacturing scale successfully. The public grid remains the network that connects and stabilizes the wider system.

Different geographies may assemble different portfolios. A Pennsylvania campus may anchor itself around nuclear generation and PJM. A Texas campus may combine gas, batteries, grid service, and private generation. A constrained urban site may value modular onsite fuel cells. A future industrial campus may be built around several advanced reactors and thermal storage. The architecture depends on the local bottleneck.

This makes the evolution a Perspectives question rather than a forecast of one winning energy technology. The future possibility space is being created by combinations. If the power system remains slow, more energy infrastructure may move toward the compute campus. If advanced nuclear proves repeatable, nuclear could become more modular. If natural-gas constraints tighten, storage and clean firm alternatives gain value. If workload flexibility improves, software can reduce the amount of physical capacity needed at the peak.

The architecture could continue changing as the binding constraint moves from one layer to another.

Conclusion

The first AI electricity problem was whether the grid could deliver enough power. The second problem is what happens when the answer is not fast enough.

The response now forming across the United States can be read as a partial reconstruction of the energy stack around compute. Competitive generators are entering longer-duration relationships with data centers. Existing nuclear plants are being restarted, extended, or uprated where project economics and approvals support them. Natural gas is being used as a near- and medium-term dispatchable pathway. Co-location is evolving toward hybrid arrangements that combine local generation with network services. Fuel cells are being deployed behind the meter, advanced-nuclear developers are using technology-company demand to support development programs, and batteries are being used as a fast-response coordination layer. Each pathway carries different cost, fuel, permitting, execution, and carbon constraints.

None of these pathways is guaranteed to scale as announced. Nuclear projects can be delayed. Gas can encounter pipeline and emissions constraints. Fuel cells remain dependent on fuel economics and maintenance. Batteries do not create primary energy. Data center demand itself can disappoint. The point is not that one technology has solved AI power.

The deeper change is that large AI data-center developers are taking a more active role in the design, financing, location, and control of the energy systems that serve them. Nebius and CoreWeave make that shift measurable: their latest public disclosures describe power capacity, capacity activation, infrastructure spending, customer prepayments or backlog, and contracted demand alongside cloud revenue. Energization is therefore becoming an operating metric because contracted compute cannot be delivered until the associated facilities and power capacity are available.

If this direction persists, AI infrastructure could increasingly be built as a coupled system: generation produces electrons, storage shapes time, networks move power and data, accelerators convert electricity into computation, and software coordinates the interfaces. The server hall remains the visible compute asset, while deployment performance may increasingly depend on whether the underlying power architecture can be assembled and energized on the same schedule as the compute it is intended to support.

Under current constraints, the AI power shortage is creating incentives to redesign the relationship between generation, grids, contracts, onsite systems, storage, and compute. In the AI Civilization Map, that makes the Foundation layer an increasingly active engineering constraint on how quickly higher layers can be deployed, rather than a passive background assumption.

Sources

• Constellation Energy - Crane Clean Energy Center
• Talen Energy - Expanded Nuclear Energy Relationship with Amazon
• Vistra - Second Quarter 2026 Results
• NRG Energy - Second Quarter 2026 Results and 1.2 GW Bring Your Own Power CCGT Project
• Reuters via KSL - Texas Data Center Audit and Large-Load Pipeline
• Nebius - Q2 2026 Shareholder Letter Filed with the SEC; Nebius - 328 MW Bloom Energy Behind-the-Meter Deployment; City of Vineland - August 17 DataOne Special-Meeting Notice; WHYY - Vineland Stop-Work Orders; NBC10 Philadelphia - Phase 2 Planning-Board Approval
• CoreWeave - Second Quarter 2026 Results; CoreWeave - Q2 2026 Form 10-Q; CoreWeave - Current AI Data Center Footprint and Power Capacity
• Bloom Energy - CoreWeave Onsite Power Partnership; Bloom Energy - Oracle Multi-Gigawatt Fuel Cell Framework
• Kairos Power - Hermes 2 Groundbreaking and Google Deployment Path; Oklo - Meta-Supported Ohio Nuclear Campus
• TerraPower - Meta Agreement for Natrium Advanced Nuclear Plants; X-energy - Amazon-Backed Xe-100 Deployment Program

Reproduction is permitted with attribution to Hi K Robot (https://www.hikrobot.com).