Here is an uncomfortable truth about the AI revolution: the next phase may not be decided by who designs the fastest chip.
It may be decided by who can deliver a gigawatt of electricity before the competition does.
Nvidia and AMD can design increasingly powerful accelerators. TSMC can manufacture them. Memory suppliers can surround them with HBM, and optical companies can connect them into enormous clusters.
But none of that infrastructure produces a single useful token until someone supplies the electricity required to turn it on.
The problem is no longer theoretical. The Department of Energy estimates that data centers could consume approximately 11.8% of total U.S. electricity by 2030, with scenarios ranging from 9.5% to 15.3%. These are not ordinary commercial buildings. Some of the largest proposed AI campuses require power on the scale of a small city. Department of Energy
This is where companies like Bloom Energy, Constellation, Vistra, Talen, GE Vernova, Eaton, and a growing collection of advanced-nuclear developers enter the AI supply chain.
But they are not interchangeable.
Bloom Energy is trying to solve the immediate time-to-power problem. Existing nuclear operators own scarce sources of dependable, carbon-free electricity. Gas turbines provide another near-term bridge. Small modular reactors may become part of the longer-term solution, but most remain years away from commercial operation.
Here is how I think about the AI energy stack, where the real opportunities are, and the proof points that determine whether this becomes a durable investment cycle or another infrastructure boom built ahead of economic demand.
The New Bottleneck: Time-to-Power
The AI industry does not merely need more electricity. It needs enormous amounts of electricity in very specific locations, on a very aggressive construction schedule.
The Opportunity and the Evidence:
A traditional data center might require tens of megawatts. The largest AI campuses are now being planned in the hundreds of megawatts, with some expected to reach multiple gigawatts as they expand.
The electricity may exist somewhere on the grid. That does not mean it can be delivered to the data center.
New generation must be built. Transmission lines may need to be expanded. Substations, transformers, switchgear, and power-distribution systems must be installed. Every one of those steps introduces permitting, engineering, equipment, and interconnection delays.
A hyperscaler can order servers faster than a utility can construct a transmission line.
What This Means in Plain English:
A company can own billions of dollars of GPUs and still have no functioning AI factory because it cannot plug them in.
The scarce resource is not simply the electron. It is the electron delivered to the correct piece of land on the correct timeline.
Why It Matters:
Every month that an AI campus sits unfinished represents idle capital, lost cloud revenue, and compute capacity that cannot be rented to customers.
That changes the purchasing decision. A data-center developer may willingly pay more for electricity if the alternative is waiting several years to begin generating revenue.
The winning power technology does not necessarily have to produce the cheapest electricity in the country. It has to produce dependable electricity where and when the customer needs it.
What to Watch Out For:
Power forecasts can become self-reinforcing. Utilities build infrastructure because developers promise enormous loads, while developers announce campuses assuming the infrastructure will arrive.
If AI demand disappoints, utilities and ratepayers could be left with underutilized generation and transmission assets. Regulators are already examining how the costs of serving large data-center loads should be allocated without shifting the risk onto ordinary customers. Department of Energy
Watch actual energized megawatts, not proposed campuses. A power agreement is not the same thing as an operating data center.
Bloom Energy: Selling Time, Not Just Electricity
Bloom Energy occupies one of the most interesting positions in this ecosystem because it can help developers bypass part of the traditional utility timeline.
The Opportunity and the Evidence:
Bloom manufactures modular solid-oxide fuel-cell systems that generate electricity directly at the customer’s location.
Instead of burning fuel inside a turbine, Bloom’s Energy Servers use an electrochemical process to convert natural gas, biogas, hydrogen, or a blend of fuels into electricity. The systems can operate alongside the grid or provide continuous onsite power through an independent microgrid.
The important word is onsite.
Bloom says its initial Oracle deployment became operational in 55 days. Oracle subsequently entered into a master agreement supporting the procurement of as much as 2.8 gigawatts of Bloom fuel-cell capacity, with an initial 1.2 GW contracted for deployment across U.S. projects. Bloom Energy–Oracle agreement
That is not a pilot project. It is utility-scale capacity being assembled from modular units directly beside the compute infrastructure that will consume the electricity.
What This Means in Plain English:
Bloom allows a data-center developer to bring part of its own power plant with it.
The company is not simply selling a cleaner generator. It is selling control over the construction schedule.
Why It Matters:
The economics of an AI factory are different from the economics of an ordinary commercial building.
If several billion dollars of GPUs are waiting for a grid connection, the developer is not optimizing for the lowest possible electricity rate. It is optimizing for the fastest reliable path to revenue.
Bloom’s advantage is not that it will always produce the cheapest electron. Its advantage is that it may produce a usable electron years earlier.
This also makes Bloom different from a traditional backup-power supplier. Diesel generators usually sit idle until the grid fails. Bloom’s fuel cells can operate continuously as the primary source of electricity.
What to Watch Out For:
Bloom is sometimes described as though it produces emissions-free electricity. That is not accurate for most installations today.
Most Bloom Energy Servers currently operate on natural gas. They avoid combustion and produce fewer local pollutants than many conventional generators, but they still emit carbon dioxide. Bloom can operate on biogas or hydrogen and can be paired with carbon capture, but those alternatives remain less available and often more expensive. Bloom explains these limitations in its own annual report.
Bloom also does not eliminate infrastructure risk completely. It replaces part of the electrical-grid dependency with a natural-gas dependency. A multi-hundred-megawatt installation still requires sufficient gas-pipeline capacity, permits, equipment, financing, and long-term maintenance.
Watch how quickly Bloom converts contracted gigawatts into operating systems, whether manufacturing capacity keeps pace, whether service margins improve as the installed base grows, and how heavily the business becomes concentrated around a small number of AI customers.
The contracts are evidence of demand. The installations, margins, and cash flow will determine whether that demand creates durable shareholder value.
The Uncomfortable Bridge: Natural Gas
The AI industry talks extensively about nuclear power and renewable energy. But much of the electricity needed over the next several years is likely to come from natural gas.
The Opportunity and the Evidence:
Natural-gas generation is dispatchable. It can operate regardless of whether the sun is shining or the wind is blowing, and it can be constructed more quickly than a traditional nuclear plant.
Bloom’s fuel cells largely depend on natural gas today. Conventional gas turbines provide another route to onsite or grid-connected generation. GE Vernova has explicitly identified data-center construction as a significant driver of future gas-turbine demand. GE Vernova 2025 annual report
The hyperscalers may prefer carbon-free electricity, but a carbon-free reactor that arrives in 2035 cannot energize a data center scheduled to open in 2027.
What This Means in Plain English:
Natural gas is the bridge between the AI infrastructure companies want to build now and the cleaner power system they hope to build later.
Bloom converts that gas into electricity through fuel cells. GE Vernova and other equipment manufacturers convert it through turbines. Different technologies are competing to solve the same immediate problem: how to produce dependable power before the grid catches up.
Why It Matters:
This makes the AI energy trade larger than fuel cells or nuclear power alone.
It reaches upstream into natural-gas production and pipelines, across into turbines and generators, and downstream into the switchgear, transformers, cooling systems, and electrical equipment required to deliver that power to the rack.
If the United States wants to build AI infrastructure at the proposed pace, natural-gas availability may become as important as semiconductor availability.
What to Watch Out For:
Gas is dependable, but it is neither price-stable nor emissions-free.
Pipeline constraints can delay projects. Higher gas prices can weaken project economics. Local opposition and emissions requirements can complicate permitting. Turbine production itself can become a bottleneck if orders grow faster than manufacturing capacity.
There is also a strategic risk. If developers build large amounts of gas infrastructure as a temporary solution, those assets may eventually compete against nuclear, geothermal, storage, or stricter carbon requirements.
Watch gas-pipeline availability, turbine lead times, fuel-cell efficiency, and whether hyperscalers are willing to sign long-term agreements for gas-powered electricity rather than treating it as a temporary bridge.
Existing Nuclear: The Scarce Asset Hiding in Plain Sight
Nuclear power solves almost the opposite problem from Bloom Energy.
Bloom offers modularity and speed. Nuclear offers enormous scale, long asset life, and continuous carbon-free generation.
The Opportunity and the Evidence:
The existing U.S. nuclear fleet operated at approximately a 92% capacity factor in 2024. These plants are designed to produce large amounts of electricity almost continuously, which closely matches the round-the-clock demand profile of an AI data center. U.S. Energy Information Administration
That helps explain why hyperscalers are signing long-term nuclear agreements.
Meta entered into a 20-year agreement supporting continued operation of Constellation’s Clinton Clean Energy Center. The agreement begins in 2027 and helps preserve more than one gigawatt of nuclear generation while supporting a small capacity increase. Meta–Constellation agreement
Microsoft signed a 20-year agreement supporting the restart of Constellation’s Crane Clean Energy Center, formerly Three Mile Island Unit 1. The project is expected to return approximately 835 MW of carbon-free capacity to the grid. Constellation–Microsoft agreement
Talen Energy expanded its relationship with Amazon through a power-purchase agreement that can supply as much as 1,920 MW from the Susquehanna nuclear facility at full quantity. Talen–Amazon agreement
What This Means in Plain English:
The fastest way to add nuclear power may not be to build a new reactor.
It may be to prevent an existing reactor from closing, increase the output of an operating plant, or restart a reactor that has already been built.
Why It Matters:
Existing nuclear plants have become strategically scarce assets.
Companies such as Constellation, Vistra, and Talen already own operating generation. They do not have to prove that nuclear fission works. Their opportunity is to extend plant lives, increase output, restart selected facilities, and sign long-term agreements with creditworthy customers willing to pay for dependable carbon-free power.
Those agreements can change the economics of an existing plant. Instead of selling all its electricity into a volatile wholesale market, the operator can secure long-duration revenue from a hyperscaler that values reliability and clean energy.
What to Watch Out For:
Existing nuclear plants still face maintenance, outage, regulatory, and political risks.
A signed agreement does not automatically create new electricity. Some contracts preserve capacity that might otherwise have retired. Others support a modest uprate. A restart must still complete repairs, obtain regulatory approvals, procure fuel, reconnect to the grid, and remain within budget.
Watch the number of incremental megawatts actually added through uprates and restarts, the length and pricing structure of hyperscaler contracts, licensing progress, outage performance, and the capital required to keep older plants operating safely.
The relevant metric is not the number of nuclear announcements. It is the number of additional dependable megawatts reaching the grid.
Advanced Nuclear: Enormous Promise on a Longer Clock
Small modular reactors could eventually bring nuclear power closer to the data center. But the distinction between an operating reactor and a proposed reactor matters enormously.
The Opportunity and the Evidence:
Google has agreed to purchase as much as 500 MW from a series of Kairos Power reactors. The first reactor is targeted for 2030, followed by additional deployments through 2035. Google–Kairos agreement
Amazon invested in X-energy and is supporting an Energy Northwest project expected to begin with 320 MW and potentially expand to 960 MW. Commercial operation is targeted for the 2030s. Amazon and X-energy are also working toward more than 5 GW of U.S. nuclear capacity by 2039. Amazon nuclear program
Meta has expanded the model further through agreements involving Vistra, TerraPower, Oklo, and Constellation that could support as much as 6.6 GW of existing and new nuclear power by 2035. Meta nuclear agreements
The hyperscalers are effectively helping advanced-reactor developers build an order book before commercial production has fully matured.
What This Means in Plain English:
Big technology companies are using their balance sheets and future electricity demand to help create a nuclear industry that does not yet exist at commercial scale.
They are acting as anchor customers for the first reactors, much as airlines place large orders to help support a new aircraft platform.
Why It Matters:
The customer problem that historically held back advanced nuclear is beginning to change.
A reactor developer now has potential customers with enormous power requirements, strong credit, long investment horizons, and a willingness to sign multi-decade agreements.
That demand could support factory construction, supply-chain development, regulatory work, and repeated deployments. Repetition is essential because the promise of small modular reactors depends on building standardized units rather than redesigning every plant from scratch.
What to Watch Out For:
An agreement is not a reactor.
Advanced-nuclear projects still face licensing, fuel availability, construction, financing, and cost risks. Several designs require specialized fuels that do not yet have mature commercial supply chains. The first unit may be expensive, and delays in the first project can affect every reactor behind it.
This is also where investors must distinguish between very different kinds of exposure.
Constellation, Vistra, and Talen own operating assets. Oklo and NuScale are development-stage bets on future deployment. Cameco, BWX Technologies, and Centrus Energy provide varying forms of uranium, enrichment, fuel, equipment, and nuclear-industry infrastructure.
These companies may all benefit from renewed nuclear investment, but they do not carry the same risk.
Watch regulatory milestones, actual construction starts, fuel-supply agreements, committed customer payments, estimated construction costs, and whether the first commercial plants remain on schedule.
A presentation showing dozens of future reactors is not enough. The industry must build the first one economically and then prove it can repeat the process.
The Other Firm-Power Candidate: Geothermal
Nuclear is not the only source of carbon-free electricity capable of operating around the clock.
The Opportunity and the Evidence:
Enhanced geothermal uses modern drilling and subsurface technology to reach heat sources that conventional geothermal projects could not economically access.
Google has already worked with Fervo Energy on an operating project in Nevada. A subsequent arrangement approved in 2025 will add 115 MW of new around-the-clock geothermal power to the Nevada grid supporting Google’s data centers. Google–Fervo agreement
The Department of Energy estimates that geothermal plants generally operate with capacity factors of approximately 90%, placing them much closer to nuclear than solar or wind in terms of continuous availability. Department of Energy
What This Means in Plain English:
Enhanced geothermal is trying to use oil-and-gas drilling techniques to build underground sources of dependable clean electricity.
It is less mature than conventional gas generation but could be deployed faster and in smaller increments than traditional nuclear plants.
Why It Matters:
AI data centers need power every hour, not merely an annual renewable-energy credit.
Solar, wind, and batteries will remain important because they can add low-cost clean generation relatively quickly. But variable renewable electricity still needs transmission, storage, demand flexibility, or firm generation when weather conditions change.
Geothermal could eventually occupy the space between renewables and nuclear: carbon-free, relatively compact, and capable of producing continuously.
What to Watch Out For:
Geothermal remains dependent on geology, drilling performance, financing, and project-specific execution.
Fervo is privately held. Ormat Technologies provides public-market exposure to conventional geothermal, but it is not a direct substitute for investing in next-generation geothermal development.
Watch drilling costs, output per well, long-term production performance, new power-purchase agreements, and whether enhanced geothermal can expand beyond a handful of favorable locations.
The Delivery Layer: Generation Is Not Enough
Producing electricity is only half the problem.
The electricity must still be transformed, switched, protected, cooled, and delivered through the facility before it reaches the GPU.
The Opportunity and the Evidence:
This brings in companies such as Eaton, GE Vernova, Quanta Services, and Hubbell.
Eaton supplies switchgear, uninterruptible power systems, transformers, busways, power-management software, and other equipment connecting the grid to the data-center rack. Its Electrical Americas orders accelerated sharply in early 2026, with management identifying data-center demand as a major driver. Eaton results
GE Vernova participates in generation and grid equipment. Quanta helps construct transmission, substations, and other power infrastructure. Hubbell provides components used across utility transmission and distribution systems.
These companies do not produce the electricity consumed by AI. They provide the physical infrastructure that makes the electricity usable.
What This Means in Plain English:
A gigawatt promised in a power contract is not the same thing as a gigawatt reaching the GPU.
Between the generator and the accelerator sits an enormous chain of transformers, substations, switchgear, cables, backup systems, voltage conversions, and cooling equipment.
Why It Matters:
The bottleneck can migrate.
If generation becomes available but transformer deliveries take too long, the transformer becomes the constraint. If the substation is completed but the transmission line is delayed, the transmission line becomes the constraint.
This is the same lesson we saw inside the semiconductor supply chain. The system can move only as fast as its slowest physical component.
What to Watch Out For:
High demand does not guarantee unlimited profits.
Manufacturers must expand capacity without sacrificing quality or margins. Utilities can delay projects. Customers can postpone campuses. Equipment backlogs can turn into cancellations if expected AI demand fails to materialize.
Watch electrical-equipment backlogs, book-to-bill ratios, transformer and turbine lead times, manufacturing expansion, utility capital-expenditure guidance, and the pace at which backlog converts into revenue.
The Ultimate Test: Electricity Must Produce Valuable Tokens
Every layer of this energy buildout ultimately depends on the same economic question that supports Nvidia and AMD.
Can the AI applications built on top of the infrastructure produce enough economic value to justify what is being constructed beneath them?
The Opportunity and the Evidence:
AI developers are no longer paying only for GPUs.
They are paying for land, buildings, networking, cooling, generation, grid connections, backup systems, and long-term electricity contracts. As clusters become larger, energy becomes a more visible part of the cost of producing each token.
Cheaper and more dependable power can lower inference costs, increase utilization, and improve the economics of AI applications.
What This Means in Plain English:
The energy companies do not get paid because AI is exciting.
They get paid because someone believes the electricity can be converted into tokens that customers will eventually purchase for more than the full cost of producing them.
Why It Matters:
The power buildout can continue only as long as hyperscalers believe additional compute will earn an acceptable return.
If AI agents become useful, reliable, and economically productive, demand for compute and electricity could remain enormous. If monetization stalls, the hyperscalers may slow construction before many of these proposed power projects are completed.
What to Watch Out For:
The most dangerous assumption is that every announced gigawatt will be required simply because it has been announced.
Framework agreements are not always firm purchase orders. Backlog is not revenue. A proposed reactor is not operating capacity. A reserved grid connection is not an energized data center.
Watch:
- Data-center electricity consumption against the DOE’s forecasts.
- Bloom’s contracted capacity versus operating deployments.
- Nuclear uprates, restarts, and license extensions that add or preserve actual megawatts.
- Small modular reactor licensing, fuel procurement, construction, and commissioning.
- Gas-turbine, transformer, and switchgear delivery times.
- Hyperscaler capital-expenditure and power-purchase guidance.
- The all-in electricity cost per unit of useful AI compute.
- Whether AI applications generate enough revenue and cash flow to support the infrastructure beneath them.
The Bottom Line
The AI revolution is no longer only a semiconductor story.
It is becoming an electricity story.
Bloom Energy sits at the front of that transition because it can provide modular onsite power without waiting for the traditional grid-development cycle. Its advantage is speed, but its dependence on natural gas means it should be viewed as a lower-emission bridge—not an automatically carbon-free solution.
Existing nuclear operators occupy a different position. Constellation, Vistra, and Talen own scarce assets capable of producing dependable carbon-free power today. Extending, uprating, and restarting those plants may be the fastest practical route to additional nuclear generation.
Advanced reactors from companies such as Oklo, NuScale, Kairos, TerraPower, and X-energy represent the longer-term possibility. Their opportunity is enormous, but their commercial, regulatory, fuel, and construction risks remain equally real.
GE Vernova, Eaton, Quanta, Hubbell, and the rest of the electrical supply chain provide the infrastructure connecting all of this generation to the rack.
The semiconductor industry determines how much compute can theoretically be produced.
The energy system determines how much can actually be turned on.
Conviction in the AI buildout therefore requires more than watching GPU shipments. It requires watching fuel-cell deployments, natural-gas availability, nuclear restarts, reactor licensing, grid construction, electrical-equipment backlogs, and the economics of the tokens produced at the end of the chain.
The chips may be ready.
The grid is not.
Educational only. Nothing here is financial advice.