Signals Inbox·July 28, 2026·AI Infrastructure

Is energy more valuable than GPUs now?

Energy has not overtaken GPUs in economic value, but a ready-to-use megawatt can now decide whether billions of dollars of AI hardware earn anything at all.

Q1Why are people saying energy matters more than GPUs now?

Right now, energy can delay an AI data center even after its GPUs are secured, so the idea that energy is more valuable has become plausible.

The latest IEA review shows how fast the physical build-out has accelerated. Five large technology companies spent more than $400 billion on infrastructure in 2025, and their combined spending is expected to rise by roughly three-quarters in 2026. Over the same 18-month period, the capacity of cutting-edge “AI factories” more than tripled.

Companies are adding industrial-scale power demand on a software-company timetable. Chips can be ordered, financed and shipped through a global supply chain. A large electricity connection still depends on a local utility, new transformers, substations, transmission work and permits. Those pieces often move much more slowly.

That gap explains the argument. Power has become the immediate obstacle in many projects, even though GPUs continue to carry more revenue, profit and technical value across the wider AI market.

Q2What does “more valuable” mean for energy and GPUs?

GPUs still win on dollars and profits; ready-to-use power wins when the question is whether the next AI campus can open.

Three tests clear up most of the confusion. First, which input costs more? GPU servers and networking usually dominate the initial bill. Second, which layer captures more profit? Nvidia’s economics sit far above those of a normal electricity supplier. Third, which missing input can stop a project today? In several crowded data-center regions, power decides the answer.

A company may already own the chips and still be unable to switch them on. Another approved megawatt then becomes extremely valuable because it unlocks the servers, buildings and customer contracts behind it.

We judge the title in two ways: where the AI industry creates and captures the most economic value, and which resource currently controls the next step of expansion.

What “more valuable” means in practice

Q3Is the world actually running out of electricity for AI?

Globally, the answer is no; locally, some AI hubs are already short of usable power.

The IEA estimates that data centers used about 485 terawatt-hours in 2025. Its central projection reaches roughly 950 terawatt-hours in 2030, or around 3% of global electricity demand. That is a major new load, yet the world’s power system can absorb it with enough investment.

The strain looks very different when a one-gigawatt campus lands in a single county. One utility may suddenly need to serve demand comparable to a sizeable city within two or three years. Global percentages hide that local shock.

This is where the debate gets muddled. At the world level, AI can still be supplied without overwhelming the power system. Around dense data-center markets in the United States and Europe, available capacity can already be booked years ahead. It is a local infrastructure squeeze, not a worldwide energy shortage.

Q4Is power already delaying AI data centers?

Yes. Power is already delaying AI data centers, and grids are adapting much more slowly than demand for new compute.

AI-focused data-center electricity use jumped 50% in 2025, according to the IEA, compared with 17% growth for data centers overall. Utilities are receiving requests for campuses measured in hundreds of megawatts, while transformers, turbines and high-voltage equipment face their own backlogs. Global gas-turbine orders rose 70% in 2025.

Chip supply remains tight too, particularly for high-bandwidth memory. The difference comes from geography. Once a chip leaves the factory, it can travel to many markets. A power connection must be designed and approved for one specific site.

The density of the hardware makes the problem harder. AI server power density increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027. More compute now fits into each building, but every rack asks much more from the electrical and cooling systems.

CoreWeave shows how quickly those needs can grow. Its active power rose from about 70 megawatts at the end of 2023 to more than one gigawatt in its latest quarter, a greater than fourteenfold increase in just over two years. Contracted power also passed 3.5 gigawatts. The company reports megawatts alongside revenue and customer commitments because power now sets the amount of GPU capacity it can bring online.

Q5What kind of energy is actually scarce for AI?

The scarce resource today is a megawatt that can actually reach the data center on schedule.

Electricity may be abundant in one region while a project waits elsewhere. Wind and solar output can also be plentiful for several hours and then fall when a campus still needs power. Cheap generation hundreds of miles away helps only when transmission can carry it to the site.

So a developer needs more than an annual energy contract. It needs a grid connection, enough local equipment, power during tight hours and a backup plan for outages or sudden swings in demand. Batteries, onsite generation and flexible workloads can all help.

Google’s Meitner Energy Center in Texas offers a recent example. The company is building a data center beside new generation so both can come online together. That cuts the risk of completing an expensive campus and then waiting for the local grid to catch up.

Q6Does electricity now cost more than the GPUs?

No. Electricity still costs only a fraction of the GPU capacity it supports.

AWS currently lists an eight-GPU B200 instance at $98.84 per hour in several US regions. Continuous use would cost about $866,000 a year before any operating-system charge.

Nvidia rates its eight-GPU DGX B200 system at a maximum of 14.3 kilowatts. Allowing another 20% for cooling and other facility power brings the draw to about 17.2 kilowatts. At $100 to $120 per megawatt-hour, a full year of continuous operation would cost roughly $15,000 to $18,000 in electricity. Local tariffs and utilization change the figure, but not by enough to alter the conclusion.

The cloud price also pays for CPUs, networking, buildings, support, financing and unused capacity. We cannot treat the gap as pure GPU profit. It still tells us where most of the money goes: into the computing system and the service built around it.

Annual cost comparison for eight B200 GPUs

Q7Who captures more profit today: Nvidia or power suppliers?

Nvidia currently captures far more AI profit than ordinary power suppliers.

Its latest reported quarter produced $81.6 billion of revenue, including $75.2 billion from data centers, up 92% from a year earlier. Across fiscal 2026, Nvidia generated $215.9 billion of revenue with a 71.1% gross margin. Very few infrastructure companies reach that scale with those margins.

Most electricity suppliers work under regulated returns or competitive wholesale prices. A power plant in a tight market can earn strong returns, but the owner rarely has Nvidia’s combination of proprietary hardware, software lock-in and yearly product upgrades.

The gains are concentrated in selected parts of the energy chain. Gas-turbine makers, transformer suppliers, nuclear developers, battery companies and owners of powered land have gained the clearest advantage. The IEA’s recent market analysis found the same concentration.

Q8Does CoreWeave spend more on power or GPUs?

CoreWeave’s latest accounts say hardware and financing cost far more than power.

The company produced $2.08 billion of quarterly revenue and still lost $740 million. Technology and infrastructure expenses reached $1.27 billion, equal to 61% of revenue, while net interest expense came to $536 million.

The filing lets us compare what changed from a year earlier. Spending on data-center utilities and power increased by about $85 million, and depreciation tied to power installation and distribution rose by another $60 million. Depreciation for servers, switches and other technology infrastructure increased by roughly $642 million to $1.1 billion.

The gap is too large to dismiss. CoreWeave needs more than one gigawatt of active power to run its fleet, so electricity clearly controls deployment. The larger financial burden still comes from buying, installing and financing the compute itself.

Q9Are GPUs becoming easier to get now?

Older GPUs are easier to book now, while frontier capacity still carries a steep premium.

AWS prices an H100 at about $4.72 to $5.19 per hour through Capacity Blocks, depending on the region. A B200 costs about $12.36 per hour. The premium has simply shifted to the newer generation.

Buyers also have more choices than they did during the first ChatGPT rush. H100s, H200s, B200s, Google TPUs, Amazon Trainium and specialized inference chips now cover a wider range of budgets and workloads.

Frontier model builders still want the best memory, networking and performance per watt available. High-bandwidth memory is expected to remain constrained through at least the end of 2027, and Meta recently signed a long-term infrastructure partnership with Nvidia. Companies are still lining up for leading GPUs.

Q10Why does AI electricity demand keep rising despite better GPUs?

AI electricity demand keeps rising because people are using far more compute, even as each individual task gets cheaper.

The IEA found that energy use per AI task has fallen by at least tenfold each year in recent years. Better chips, smaller models and smarter serving software have made a simple text response remarkably efficient.

The products are becoming heavier at the same time. Reasoning, video generation and agentic workflows can use hundreds or thousands of times more energy than a basic text query. An agent may call a model repeatedly, search databases and run tools before returning one answer.

Usage is also expanding faster than efficiency. Data-center electricity consumption rose 17% in 2025, while AI-focused facilities used 50% more power. The IEA now expects total data-center demand to move from about 485 terawatt-hours in 2025 to around 950 terawatt-hours in 2030.

Cheaper computation creates more demand for computation. The cost of each useful AI task can keep falling while the industry’s total electricity use keeps climbing. Both can be true.

Q11Can AI companies move their GPUs to cheaper power?

AI companies can move training jobs toward cheap power, but live services have much less freedom.

A long model-training run can operate in a remote region when the site has enough chips, fiber and electricity. The finished model can then be copied elsewhere. Data processing, evaluations and some internal jobs can follow the same pattern.

Live inference has tighter limits. Users expect fast responses, companies keep applications near their existing cloud databases, and regulated sectors may require data to remain inside a country or region. Sending huge volumes of information between distant centers can create another cost and latency problem.

Companies can still shift work by the hour. Google has signed one gigawatt of demand-response capacity with US utilities, allowing selected machine-learning jobs to slow down or move during stressed periods. That can speed up grid connections and reduce the amount of backup infrastructure needed. It works well for flexible jobs; widely used assistants, hospitals and financial systems require more consistent service.

Q12Are hyperscalers becoming energy companies?

Yes. Hyperscalers are now behaving like energy developers on a utility scale.

Meta has assembled nuclear agreements covering up to 6.6 gigawatts. Amazon is backing small modular reactors with a longer-term goal of more than five gigawatts of new nuclear capacity by 2039. Google has started pairing new data centers with dedicated generation.

A single campus can require one or two gigawatts, and waiting for the normal utility cycle may leave billions of dollars of servers idle. Large technology companies are securing generation, storage, grid upgrades and flexible-load agreements much earlier.

A decade ago, electricity was mainly a bill to pay after the data center was built. These days, power planning starts before the site is chosen. Cloud and AI still produce the profits, but the companies increasingly finance the infrastructure that keeps those businesses growing.

Q13Does the nuclear rush solve AI’s power problem soon?

Nuclear can support AI later, but it will contribute too little, too late for most projects already in the queue.

Existing plants can help sooner through license extensions, higher output and restarts. They already have grid links, trained staff and large sites, which makes their output especially attractive to data-center developers.

New reactors run on a slower schedule. Meta’s TerraPower and Oklo projects are aimed mainly at the early 2030s. Amazon’s stated nuclear target stretches to 2039. Those programs may become important if AI demand stays high, but many campuses need power years before the reactors arrive.

The next few years will depend on renewables, gas, batteries, grid upgrades and flexible demand. Nuclear strengthens the longer-term plan. It does not clear the immediate backlog.

Q14What will power the next wave of AI data centers?

The next wave of AI data centers will use renewables, gas, batteries and existing nuclear together.

Wind and solar can be built quickly and often provide the cheapest energy. Their output changes with the weather, so operators combine them with grid purchases, storage and power sources that can run whenever needed.

Natural gas offers that control, which explains the growing pipeline of onsite projects in the United States. The IEA’s latest satellite tracking found that around one-fifth of planned onsite gas projects for US data centers had already begun land clearing or construction. Even then, a reliable onsite system may need 30% to 70% more generating capacity than the campus’s peak demand, and turbine supply is tight.

Batteries will become more common as AI racks create fast changes in load. The IEA estimates that data centers could install 20 to 25 gigawatts of battery storage worldwide by 2030. Existing nuclear plants add steady large-scale output, while newer reactors become more relevant later.

Likely power sources for the next AI build-out

Q15Which energy assets are becoming most valuable?

Powered sites and approved grid connections are becoming the most valuable energy assets in AI.

Cheap land has limited appeal when the utility cannot energize it for five years. A site with approved power, fiber and cooling can start earning revenue much sooner, so developers will pay a premium for it.

Transformer lead times, turbine orders and transmission queues have all tightened. A completed substation or a signed connection agreement may be worth more to a developer than a cheaper electricity contract that depends on future infrastructure.

The assets gaining the biggest premiums solve several problems at once. Co-located generation reduces dependence on grid expansion. Batteries handle sudden changes in AI demand. Existing nuclear sites combine land, transmission and steady output. That is where the energy premiums are building.

Q16Could power constraints weaken Nvidia?

Power limits may slow Nvidia’s sales while increasing the value of every watt its systems save.

A data-center operator with a fixed 100-megawatt allocation wants the most useful AI output from each megawatt. Faster chips, stronger networking and better software can raise revenue without another grid connection. Nvidia sells the full system, which helps when power becomes the ceiling.

Two pressures could change that. Customers may divert spending from servers toward substations, generation and cooling. They may also choose Google TPUs, Amazon Trainium or specialized inference chips when those options deliver a better result per dollar or watt.

So far, power has barely dented Nvidia’s growth. Its latest quarter set another data-center revenue record, and newer architectures remain in heavy demand. Customers care more about performance per watt now, but Nvidia’s lead still looks intact.

Q17So, is energy more valuable than GPUs now?

No. GPUs still hold more economic value overall, although ready-to-use power now controls many expansion decisions.

GPUs carry the larger purchase price, create the computing performance and capture the strongest profits. Electricity is sold in a much more competitive and regulated market, even when a particular power plant or connection becomes extremely valuable.

Power now determines how quickly many projects can open. A developer can often reserve chips and raise financing faster than it can secure a large grid connection. Powered land, substations and reliable generation have become some of the hottest assets around AI.

The broad claim is exaggerated. Energy has reached the same strategic importance as GPUs when a company chooses where and when to build its next cluster. Across the full AI value chain, GPUs still capture more economic value.