Is power replacing chips as AI’s bottleneck?

Why is electricity suddenly becoming a bigger problem for AI than GPUs?

Electricity is now holding up major AI data center projects even when companies have the money and the Nvidia GPUs ready to go.

The International Energy Agency's April 2026 report, Key Questions on Energy and AI, estimated that global data center electricity consumption increased 17% in 2025, while demand from AI-focused facilities jumped 50%. The growth is putting pressure on an electricity system that was never designed to expand at the pace of the semiconductor industry.

The problem became particularly visible when Microsoft pulled out of Nscale's planned AI data center in Loughton, Essex. The project was expected to support approximately 23,000 Nvidia GB300 GPUs, but reporting by The Guardian revealed that sufficient grid capacity might not arrive until the early to mid-2030s. The facility had originally been expected to open much sooner.

Amazon is facing a different version of the same problem. Recent Business Insider reporting revealed that AWS had committed approximately $5 billion to Nvidia GB300 server racks through an early-delivery program, with some equipment sitting in warehouses before the data centers needed to operate it were ready.

Has electricity already replaced AI chips as the main bottleneck?

Electricity has already overtaken GPUs as the immediate obstacle for some AI projects, but chip shortages remain serious enough that we cannot call power the industry's single biggest bottleneck.

The answer depends on where a project stands. A developer with 20,000 GPUs ordered and no grid connection has a power problem. Another developer with an operational facility but insufficient high-bandwidth memory or advanced processors still has a semiconductor problem.

The IEA's 2026 assessment found shortages across electricity infrastructure, chip manufacturing and financing. It specifically identified a shortage of high-bandwidth memory, an essential component of advanced AI accelerators, that could persist through at least the end of 2027.

Nvidia's transition from Blackwell to Vera Rubin adds another complication. Advanced GPUs depend on multiple supply chains working together, including memory, packaging, networking and power delivery.

There is also no comprehensive dataset comparing the number of AI projects delayed by electricity with those delayed by chips. Declaring power the industry's biggest bottleneck everywhere would go further than the evidence supports.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

How much electricity will AI data centers actually need?

AI data centers are on track to nearly double their global electricity consumption by 2030, and demand from AI-focused facilities is growing much faster than the industry average.

The IEA estimates that global data centers consumed approximately 485 terawatt-hours in 2025, up from around 415 TWh in 2024. Its latest central projection reaches 950 TWh by 2030.

That represents roughly 96% growth over five years, equivalent to approximately 14.4% annually.

The extra 465 TWh is equivalent to around 53 gigawatts of continuous electricity consumption throughout a year. In practice, the electricity system would need more installed capacity because of maintenance, fluctuations in generation, peak demand and reliability requirements.

The historical comparison is also revealing. Global data center electricity consumption grew roughly 12% annually over the preceding five years, according to the IEA's earlier analysis. Growth accelerated to approximately 17% in 2025, while AI-focused facilities grew 50%.

We also need to keep the global figures in perspective. Data centers represented approximately 1.5% of worldwide electricity demand in 2024, and the IEA expects that share to reach around 3% by 2030. The pressure is much greater in regions where facilities are heavily concentrated.

Metric	Earlier figure	Latest estimate or projection	Interpretation
Global data center electricity demand	415 TWh in 2024	485 TWh in 2025	Approximately 17% annual growth
Global demand by 2030	485 TWh in 2025	950 TWh	Approximately 96% growth
Additional annual consumption	2025 baseline	465 TWh by 2030	Equivalent to 53 GW of continuous load
AI-focused data center demand	2025 baseline	Approximately 3× by 2030	Much faster growth than conventional workloads
Share of global electricity	Approximately 1.5% in 2024	Approximately 3% by 2030	Modest global share
U.S. data center electricity	176 TWh in 2023	325–580 TWh by 2028	Berkeley Lab scenario range
U.S. electricity share	4.4% in 2023	6.7–12% by 2028	Much larger domestic exposure
AI server power density	2020 baseline	11× higher by 2025	Rapidly rising infrastructure requirements

Is the AI power shortage real, or are companies exaggerating future demand?

The AI electricity shortage is already disrupting real projects, but treating every proposed data center as though it will be built can substantially exaggerate future demand.

The IEA estimated in its 2025 Energy and AI report that grid constraints could delay approximately 20% of global data center capacity planned through 2030. That is a substantial share of the pipeline, even before considering construction costs and financing.

The delays also appear in utility planning. Lawrence Berkeley National Laboratory's June 2026 research documented how large electricity connections in the United States are being slowed by planning procedures, infrastructure requirements and uncertainty over future demand.

However, electricity utilities face another problem: developers often pursue several potential locations before committing to one. A company might submit connection requests for three sites while intending to build only one facility.

Some projects also lack confirmed financing or customers. Their inclusion in electricity-demand projections can make the future shortage look larger than it will ultimately be.

Microsoft's withdrawal from the Nscale project is much stronger evidence of a real constraint than another announcement of a planned AI campus.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Are AI data centers now getting delayed more by power connections than GPU deliveries?

Electricity connections are becoming the harder deadline for some AI data centers, although there is still no reliable industry-wide count showing whether power delays have overtaken GPU delivery problems.

A semiconductor manufacturer can distribute available GPUs across different customers and locations. A data center developer cannot move an unfinished transmission connection to another site.

According to the IEA, new transmission lines can take four to eight years in advanced economies. Meanwhile, specialized electricity equipment can take years to arrive.

Wood Mackenzie's transformer research found that average lead times increased from approximately 50 weeks in 2021 to 120 weeks in 2024. Large power transformers faced lead times ranging from 80 to 210 weeks.

That leaves AI developers exposed to two schedules that do not necessarily align.

A company might secure a GPU allocation twelve months before deployment but discover that a substation upgrade will take three years. The reverse is also possible: a finished data center might have electricity available but lack the advanced accelerators needed to fill it.

Could unused Nvidia GPUs become a serious financial problem?

Unused Nvidia GPUs are becoming a financial risk because companies can spend billions on equipment before earning anything from it.

Amazon provides a particularly useful example. Business Insider recently reported that AWS had changed its purchasing process to accept Nvidia equipment before the relevant data centers were ready.

The initial commitment involved approximately $5 billion in GB300 systems, with further purchases potentially reaching several billion dollars per month.

Amazon can afford to take such risks more easily than most competitors, but holding expensive processors creates several costs. The company must finance the purchase, store the hardware and manage the possibility that newer generations will arrive before all the equipment is deployed.

Consider a hypothetical $1 billion GPU installation with a four-year useful economic life. If deployment takes an additional twelve months, 25% of that assumed lifetime passes before the installation begins generating revenue. That does not automatically mean a 25% accounting loss, but the economic cost of waiting can be substantial.

The financial pressure is greater for companies relying heavily on outside capital. CoreWeave and other specialized AI infrastructure providers must coordinate debt financing, construction, equipment purchases and customer contracts.

Recent results from Nscale show how sensitive this business model can become. Financial Times reporting described a first-half 2026 net loss of approximately $1 billion against $141 million in revenue, while the company continued developing an infrastructure portfolio backed by major long-term contracts.

Those figures alone do not establish that electricity delays caused the loss. They do show how much capital can be committed before a young infrastructure business reaches meaningful revenue scale.

Are Nvidia's AI chips consuming electricity faster than data centers can adapt?

Nvidia's newest AI systems are pushing rack power requirements far beyond the levels traditional data centers were built to handle.

The company's GB200 NVL72 system combines 72 Blackwell GPUs in a liquid-cooled rack requiring approximately 120 kilowatts. Nvidia's GB300 NVL72 reference specifications allow for up to 142 kW per rack.

At that maximum rating, 1,000 GB300 racks would require 142 MW of IT power. With an illustrative power usage effectiveness of 1.2, the facility would need approximately 170 MW.

That is already the scale of a substantial industrial electricity connection.

The IEA estimates that AI server power density increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027.

Operators must deliver much more electricity into a relatively small space while controlling heat and rapid changes in electrical load. Existing buildings and electrical systems cannot always accommodate that shift.

There is also growing interest in battery systems that smooth those fluctuations. Nvidia has tested energy-storage-enhanced power equipment that reduced peak grid demand by approximately 30% during a specific AI training workload. The improvement came from smoothing power spikes, rather than cutting total electricity consumption by 30%.

Is AI's electricity problem global, or concentrated in a few regions?

AI electricity shortages are much worse in certain regions, especially established American data center hubs, than global electricity-demand figures suggest.

The IEA estimated that the United States accounted for approximately 45% of worldwide data center electricity consumption in 2024, compared with around 25% for China and 15% for Europe.

Within the United States, a large share of development is concentrated around Northern Virginia, Texas and other established cloud infrastructure markets.

Northern Virginia is a particularly important example. Its concentration of data centers has created major challenges around transmission availability and the ability to connect new large loads.

Texas has attracted developers through its land availability, electricity market and potential energy resources. However, large proposed data center loads are also forcing grid planners to assess whether future electricity supply and network infrastructure can keep up.

The geographical differences extend beyond the United States. The United Kingdom faces connection delays and ambitions to expand sovereign AI computing capacity. Recent reporting on Fuse Energy described plans for 142 MW of data center projects across 15 UK sites, with the company using its energy-development expertise to enter the computing market.

China faces a different situation. Its extensive electricity infrastructure and coordinated industrial planning can support large-scale development, while restrictions on access to certain advanced American processors create a separate semiconductor constraint.

Is nuclear power becoming essential to AI data centers?

Nuclear power is becoming an important part of Big Tech's long-term electricity strategy, but most AI data centers being built today cannot depend on new reactors arriving soon enough.

Google's October 2024 agreement with Kairos Power targets up to 500 MW of advanced nuclear capacity, with the first deployment planned around 2030 and additional projects extending through 2035.

In August 2025, Google, Kairos and the Tennessee Valley Authority announced a first 50 MW deployment associated with the Hermes 2 project in Tennessee.

Microsoft took another route by signing an agreement supporting the restart of a reactor at the Three Mile Island site in Pennsylvania. Reviving existing nuclear infrastructure offers a different development path from building entirely new reactor designs.

These commitments show that major technology companies are prepared to secure electricity years before they need it. Nuclear power offers continuous generation without the operational emissions associated with gas-fired electricity.

However, nuclear construction requires regulatory approval, specialized equipment, financing and years of work. The IEA expects natural gas and renewable energy to provide much of the near-term increase in data center electricity supply.

Nuclear could become a major competitive advantage for companies with access to reliable generation later in the decade and beyond. Its impact on projects facing connection problems today will be much smaller.

Can AI companies solve electricity shortages by building their own power plants?

Building private power plants can help AI companies bypass grid constraints, but even the largest developers are discovering that generating dependable electricity is expensive and complicated.

Natural gas offers a relatively established route because turbines can provide continuous electricity without relying on favorable weather conditions.

Yet the equipment is increasingly difficult to obtain, and projects require fuel connections, environmental permits and redundancy.

The IEA's 2026 research estimated that dependable onsite natural gas generation can require 30% to 70% more installed capacity than the critical electricity load being served.

For a data center requiring 500 MW of reliable electricity, that implies approximately 650–850 MW of installed generating capacity.

The agency also estimated that onsite gas installations could provide between 15 and 27 GW of worldwide data center generation capacity by 2030, with much of the development concentrated in the United States.

However, only around one-fifth of tracked onsite gas projects had begun clearing land or construction at the time of the assessment.

Could more efficient AI chips eliminate the electricity bottleneck?

More efficient AI chips are making individual tasks dramatically cheaper to run, but total electricity consumption keeps rising because companies are running more AI and building applications that demand much more computing power.

The IEA's April 2026 research estimated that energy consumption per AI task has fallen by at least an order of magnitude annually in recent years through hardware and software improvements.

The agency calculated that replacing every conventional internet search with a simple AI text query would require fewer than 4 TWh annually. That represents less than 1% of current worldwide data center electricity consumption.

The expensive applications are increasingly elsewhere.

Reasoning models can perform substantial computation before producing an answer. Video generation uses far more computing power than simple text generation. AI agents can make repeated model calls while completing a single task.

According to the IEA, these more demanding applications can use hundreds or thousands of times more energy per query than simple text generation.

This creates a difficult forecasting problem. A model might become ten times more efficient, but the resulting cost reduction could encourage developers to run it far more often or introduce applications that were previously too expensive.

The shift toward inference makes this especially relevant. Training frontier models requires enormous clusters, but deployed models can serve continuous demand from millions of users and automated workflows. Nvidia's newer systems increasingly target reasoning and inference performance, reflecting how much computing is needed after a model has been trained.

The industry still lacks standardized reporting that would let us establish a reliable global electricity split between training and inference. And efficiency gains alone are unlikely to make the capacity problem disappear.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Are AI companies now competing for electricity as aggressively as GPUs?

Big Tech companies are increasingly securing future electricity supplies years in advance, much as they have been reserving advanced semiconductor capacity.

Google's nuclear agreements with Kairos Power, Microsoft's support for restarting the Three Mile Island reactor and Amazon's involvement in nuclear energy projects all point in that direction.

The companies have different approaches, but their commitments reveal a shared concern about future electricity availability.

Google is supporting advanced reactor development. Microsoft has backed the reopening of existing nuclear infrastructure. Amazon has explored nuclear-linked data center development and partnerships alongside its broader renewable energy purchases.

There is also a growing distinction between buying electricity and securing the infrastructure that can actually deliver it.

A conventional power purchase agreement can provide financial exposure to electricity from a generating facility. It does not necessarily guarantee that hundreds of megawatts can reach a particular data center without additional network upgrades.

Companies with access to land, substations and reliable generating capacity may therefore have an advantage when choosing where to install the next generation of GPUs.

This is also why energy businesses are beginning to move into the data center market themselves. Fuse Energy's UK developments are an unusual example: a company originally focused on energy infrastructure is now pursuing computing facilities because controlling electricity development can help it secure attractive locations.

Could electricity costs become more important than the price of AI chips?

Electricity is becoming a larger part of AI operating costs, but high-end GPUs and the surrounding computing hardware still represent enormous upfront expenses.

The effect of electricity prices becomes clearer when we calculate the operating cost of Nvidia's GB300 NVL72 system.

A rack operating continuously at its maximum specified 142 kW would consume approximately 1.24 million kWh annually before accounting for cooling and other facility equipment.

At $0.06 per kWh, that costs around $74,600 annually. At $0.12, the electricity bill approaches $149,300.

With an illustrative power usage effectiveness of 1.2, the total facility electricity cost rises to approximately $89,600 and $179,200 respectively.

Across 1,000 racks, even a modest price difference becomes meaningful. However, rack utilization, commercial electricity tariffs, facility efficiency and contractual pricing will all affect actual spending.

The bigger financial problem may be access rather than price. A developer can sometimes justify paying more for electricity if an earlier connection allows expensive GPUs to begin generating revenue sooner.

Hypothetical electricity price	Annual IT electricity cost per GB300 rack	Including facility overhead (PUE 1.2)	Annual facility cost for 1,000 racks
$0.04/kWh	$49,757	$59,708	$59.7M
$0.06/kWh	$74,635	$89,562	$89.6M
$0.08/kWh	$99,514	$119,416	$119.4M
$0.10/kWh	$124,392	$149,270	$149.3M
$0.12/kWh	$149,270	$179,124	$179.1M
$0.15/kWh	$186,588	$223,906	$223.9M
$0.20/kWh	$248,784	$298,541	$298.5M

Are Nvidia and other AI chipmakers still struggling with shortages?

Nvidia and other AI chipmakers are still facing supply constraints, especially around high-bandwidth memory, making it premature to say that electricity has taken over from semiconductors everywhere.

TrendForce's August 2026 research projected tight DRAM supply through 2027 and reported that Nvidia was evaluating different memory configurations for Rubin Ultra amid HBM availability and validation concerns.

Advanced accelerators require much more than a finished GPU chip. They also need high-bandwidth memory, advanced packaging, interconnects and integration into working server systems.

Production problems in any of these areas can slow deliveries.

TrendForce also revised its estimate of Nvidia's high-end GPU shipment mix, expecting Blackwell to represent approximately 71% of shipments in 2026, compared with an earlier expectation of 61%. The change reflected difficulties and delays associated with moving toward newer products.

These shortages can overlap with power constraints. A delayed GPU delivery can leave a completed data center underused, while a delayed electrical connection can leave purchased GPUs sitting in storage.

Is the AI electricity shortage creating winners among infrastructure companies?

Companies that already control usable electricity connections and operating data center infrastructure are in a stronger position than developers whose projects still depend on major grid upgrades.

Securing land is relatively straightforward compared with obtaining hundreds of megawatts of deliverable electrical capacity in an already congested market.

Established operators with access to substations, cooling systems and utility contracts can potentially deploy computing equipment much faster than a new developer starting from undeveloped land.

That advantage also affects financing. Investors and lenders need confidence that GPU purchases, construction schedules and customer contracts will eventually line up.

The recent difficulties around Nscale's Essex development illustrate what can happen when that coordination fails. The company had ambitious plans and major technology relationships, but electricity availability disrupted an important project.

At the supplier level, companies producing transformers, switchgear, power-management systems and generation equipment could benefit from higher demand. Yet the IEA cautioned in its 2026 analysis that AI electricity demand is too small relative to the entire energy sector to justify expecting every energy company to receive an AI-driven valuation premium.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Could power shortages actually slow down the AI boom?

Electricity shortages could slow AI infrastructure spending substantially, even if customers continue adopting AI services.

The scale of the spending makes relatively small delays expensive.

According to the IEA, capital expenditure among five leading technology companies exceeded $400 billion in 2025 and was projected to increase by approximately 75% in 2026. That would imply annual investment above $700 billion if the projection materializes.

At the same time, companies are facing increasingly complicated schedules involving GPUs, construction, transmission, cooling systems and financing.

Amazon's reported early GPU purchases show how those schedules can get out of alignment. The company is spending heavily to reserve computing equipment while also managing the electricity and infrastructure required for deployment.

That creates a problem for investors. Capital spending can increase long before the resulting facilities contribute meaningful revenue.

However, we should distinguish between delays in supplying AI computing power and a collapse in customer demand.

AWS continues to report strong commercial growth. Its second-quarter 2026 revenue reached approximately $42.2 billion, representing 36.7% year-over-year growth, according to figures discussed by its finance leadership.

Could AI's electricity bottleneck disappear before 2030?

AI's electricity bottleneck is unlikely to disappear before 2030, although new generation capacity, batteries and better hardware could ease the pressure in some regions.

The IEA still expects global data center electricity consumption to reach approximately 950 TWh by the end of the decade. Meeting that demand will require substantial investment in generation, transmission, transformers and cooling equipment.

Some technical improvements are already promising.

Batteries can help smooth rapid changes in AI electricity consumption and reduce peak stress on electrical connections. The IEA estimates that data centers worldwide could install approximately 20–25 GW of battery storage by 2030.

More efficient processors can deliver additional computing performance per watt. Better power management can reduce the amount of electrical infrastructure required for a given workload.

There are also potential improvements in where companies build facilities. AI developers can increasingly favor locations with available electricity capacity rather than trying to expand exclusively around established data center clusters.

But demand may absorb much of the improvement. The IEA expects AI-focused data center electricity consumption to triple between 2025 and 2030, while electricity infrastructure often takes several years to expand.

Is power replacing chips as AI's bottleneck?

Yes, partly. Electricity is already replacing GPU availability as the immediate obstacle for a growing number of AI data center projects, but semiconductor shortages remain important enough that power cannot yet be called the industry's universal bottleneck.

The evidence is unusually concrete. Global data center electricity consumption increased approximately 17% in 2025, and AI-focused facilities grew 50%. The IEA expects worldwide data center demand to nearly double by 2030.

Recent developments also show the commercial consequences. Microsoft withdrew from Nscale's Essex project amid long electricity connection delays. Amazon is reportedly purchasing billions of dollars of Nvidia equipment ahead of deployment readiness. Meanwhile, electrical infrastructure can take years to build, even as successive GPU generations increase the power required per rack.

However, high-bandwidth memory shortages and the complex rollout of next-generation accelerators show that the chip supply problem remains unresolved.

Counting GPUs ordered, purchased or installed tells us less than it used to. A company also needs sufficient electricity, working cooling systems and a functioning facility before those processors can produce valuable computing output.

For an AI developer with secured chips but no grid connection, electricity is already the bigger problem. For another with available power but no advanced accelerators, chips remain the constraint.

The next phase of AI infrastructure competition will increasingly depend on which companies can combine both resources at the lowest cost and bring them into operation quickly.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.