Is power replacing chips as AI’s bottleneck?

Is Power Replacing Chips as AI’s Bottleneck?

Why is power suddenly being called AI’s biggest bottleneck?

Power is now being called AI’s biggest bottleneck because data-center construction is moving much faster than grids, power plants and electrical-equipment factories can follow.

The pace changed sharply during the past 18 months. The International Energy Agency’s latest review found that the operating capacity of advanced “AI factories” more than tripled over that period. Electricity use by AI-focused data centers then jumped about 50% in 2025, compared with 17% growth across data centers as a whole.

Spending has accelerated even faster. The IEA estimates that five large technology companies invested more than $400 billion in 2025 and could increase their combined capital expenditure by another 75% in 2026. If that estimate holds, their annual spending would approach $700 billion.

Power systems were never designed to absorb investment cycles like this. Utilities normally plan demand several years ahead and build assets expected to last for decades. AI developers can change their capacity plans after a new model, funding round or chip generation appears.

A large project also arrives as one concentrated load. Adding one gigawatt of demand in a particular county is much harder than adding the same amount gradually across millions of homes and businesses.

That mismatch has pushed electricity into conversations that used to focus almost entirely on GPUs. Developers now ask about available megawatts before they ask about land, tax incentives or construction costs.

What would it mean for power to replace chips as AI’s bottleneck?

Power replaces chips as the main AI bottleneck when electricity infrastructure determines the opening date of new computing capacity more often than processor deliveries do.

The bottleneck does not have to be the most expensive input. GPUs can account for most of a cluster’s equipment budget while a delayed substation keeps every purchased server switched off.

Three tests settle it.

First, can the company obtain the input at a price it can afford? Chips remain difficult here, especially for smaller buyers.

Second, how long does the input take to secure? Power connections, transmission upgrades and generation projects are increasingly losing this race.

Third, what happens when the project becomes ten times larger? A bigger chip order can improve a customer’s bargaining position. A bigger electricity request usually requires more engineering, more approvals and more grid construction.

Power currently wins the timing and scaling tests for many hyperscale projects. Chips still win the access and performance tests.

Test	Advanced chips	Usable power
What is actually scarce?	Latest GPUs, HBM and complete systems	Electricity at the required site, voltage and date
Common workaround	Older chips, custom silicon, cloud rentals or smaller models	Relocation, phased construction or onsite generation
Main consequence	Higher cost or weaker performance	The facility cannot operate
What happens at larger scale?	Large buyers gain leverage	Grid studies and construction become harder
Who feels it most?	Startups and smaller cloud customers	Hyperscalers building new campuses

If you want more recent data on this point, please see our latest AI infrastructure market report.

Has AI electricity demand really broken away from its old trend?

AI electricity demand has clearly broken away from the slower data-center growth seen during most of the previous decade.

US data centers consumed around 58 terawatt-hours of electricity in 2014 and 176 terawatt-hours in 2023, according to Berkeley Lab. That already represented a tripling in nine years.

Its latest reference case now reaches 649 terawatt-hours in 2030. Converted into a continuous average load, that is roughly 74 gigawatts, compared with about 20 gigawatts in 2023. Data centers would account for approximately 11.8% of total US electricity consumption.

The global figures are less concentrated but move in the same direction. The IEA expects worldwide data-center electricity use to rise from about 485 terawatt-hours in 2025 to roughly 950 terawatt-hours in 2030. AI-focused facilities are expected to triple their consumption during that period.

The forecasts remain uncertain. Model efficiency, GPU shipments, utilisation rates and the commercial success of AI products could move the final figure considerably.

Still, the direction no longer looks doubtful. Electricity demand keeps rising despite rapid improvements in computing efficiency. Companies are using each efficiency gain to run more queries, longer reasoning chains, video models and autonomous agents.

Measure	Earlier level	Latest central projection	Change
US data-center electricity use	176 TWh in 2023	649 TWh in 2030	About 3.7 times higher
Average US data-center load	About 20 GW	About 74 GW	Around 54 GW added
US electricity share	Around 4%	11.8%	Nearly triples
Global data-center electricity use	485 TWh in 2025	950 TWh in 2030	Nearly doubles
AI-focused data-center demand	2025 baseline	Three times higher by 2030	Grows faster than other data centers

Are AI chips still scarce today?

AI chips are still scarce today, particularly high-bandwidth memory and complete systems, but chip production is expanding much faster than the power system.

NVIDIA’s latest results make both sides visible. Its quarterly data-center revenue reached $75.2 billion, up 92% from a year earlier. Growth at that speed would be impossible without a rapid increase in the number and value of systems being delivered.

Demand is still running ahead of supply. NVIDIA said cloud GPUs were sold out during the Blackwell ramp, while Microsoft recently warned that it expects to remain capacity-constrained through at least the end of 2026 despite bringing more GPUs, CPUs and storage online.

Memory is now one of the tightest parts of the chip supply chain. Micron had already agreed the price and volume of its entire 2026 high-bandwidth memory supply before the year began. In its latest earnings remarks, it described the memory market as being in a period of significant shortage and said customers were committing to multi-year supply agreements.

The shortage goes beyond the processor die. An AI server requires HBM, advanced packaging, networking, power components and cooling equipment. One missing part can delay the finished rack.

Yet the semiconductor industry has a strong reason to expand quickly. AI components generate high margins, buyers commit billions of dollars in advance, and production can serve customers around the world.

A utility connection works differently. The infrastructure has to reach one specific site, pass local approvals and remain reliable for decades. Money helps, but it cannot compress every stage into a semiconductor-style product cycle.

Chips remain a serious bottleneck. Power is overtaking them mainly on delivery time, geographic flexibility and the difficulty of scaling from hundreds to thousands of megawatts.

Are power shortages already delaying AI data centers?

Power shortages are already delaying, phasing and relocating major AI data centers.

Dominion Energy offers the clearest public example. The utility has assigned connection dates through 2031 to around 25 gigawatts of new data centers in Virginia. Another 45 gigawatts of proposed projects have no offered connection date.

Roughly 64% of the capacity in that combined pipeline therefore lacks a schedule. Some projects are duplicated or speculative, but even serious developers cannot receive firm dates until Dominion knows which grid upgrades can be built and paid for.

A data center above 50 megawatts will often require a new substation and an extension of transmission lines, according to Dominion’s own connection guidance. A hyperscale campus can need several times that amount.

The problem has grown large enough to change federal policy. FERC recently ordered all six major US regional grid operators to justify or reform the rules used to connect data centers and other large electricity users. Its language focused directly on “speed-to-power,” a phrase that has become central to AI site development.

Companies are reacting before regulators finish rewriting the rules. They are splitting campuses into phases, building in less crowded markets, accepting temporary power arrangements and considering onsite generation.

A delayed GPU order can leave part of a data hall empty. A delayed power connection can leave the entire campus empty, including the servers that arrived on time.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Is the world actually running out of electricity for AI?

The world has enough electricity for AI, but many popular data-center regions cannot deliver enough power to the right site quickly enough.

The IEA expects data centers to consume around 3% of global electricity by 2030. That is a large load, but it hardly means AI will consume the world’s entire power supply.

The shortage is local. Northern Virginia cannot operate a Texas data center using unused electricity in West Texas unless the relevant generation and transmission links exist. Annual electricity production also tells us little about whether enough power will be available every hour.

AI campuses need generation, transmission, substations, backup systems, cooling infrastructure and fibre in the same place. A region can produce surplus electricity over a full year and still lack the equipment needed to serve another 500-megawatt customer.

Concentration makes the issue harder. The United States and China are expected to account for close to four-fifths of global data-center electricity-demand growth through 2030. Within those countries, developers continue to cluster around a limited number of established cloud and fibre hubs.

The real bottleneck is deliverable power: the correct amount, at the correct voltage, with high reliability and a credible opening date.

Regions with abundant generation but weak transmission may struggle. Regions with excellent fibre but congested grids face the opposite problem. Only a smaller number offer both.

Is power now deciding where AI data centers get built?

Power availability is now one of the first filters used to choose AI data-center sites, often ahead of land prices and tax incentives.

Google has openly described its approach as “energy first.” Its energy team now works with utilities and developers before final site plans are fixed, rather than selecting a campus and asking for power afterward.

The company has gone further than signing electricity contracts. Alphabet agreed to acquire Intersect, an energy and data-center infrastructure developer, for $4.75 billion plus assumed debt. Intersect brought several gigawatts of projects in development or construction, including a Texas site where electricity generation and data-center capacity are being built together.

That acquisition is the tell. Google decided that owning more energy-development capability could help it bring computing capacity online faster.

Other hyperscalers are following different versions of the same strategy. Meta announced nuclear agreements that could unlock as much as 6.6 gigawatts. Amazon recently unveiled 700 megawatts of new geothermal and solar-plus-storage capacity for future operations around Reno. Microsoft is supporting the restart of an 835-megawatt nuclear plant in Pennsylvania.

Many of these projects will take years. Advanced nuclear projects planned for the 2030s will do little for a campus waiting for power today.

The direction is already clear. Leading AI companies are becoming energy developers, financiers and anchor customers because ordinary utility procurement no longer fits their expansion plans.

Moving to a power-rich region still involves trade-offs. Training clusters can operate far from users, while inference services may need low latency, local data handling and direct links to existing cloud regions. Talent, fibre, water and permitting remain part of the decision.

Power does not choose the site alone, but it can now eliminate a location before the other factors are even discussed.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Are AI racks becoming too power-dense for existing data centers?

Many existing data centers cannot handle today’s AI racks without major electrical and cooling upgrades.

The strain inside the building can be as serious as the shortage outside it. The IEA found that AI-server power density increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027.

An advanced rack occupies roughly the space of a large refrigerator. By 2027, its peak electricity demand could equal that of 65 households, while the heat it produces may resemble the output of about 30 domestic gas boilers.

Older facilities were often built for racks using less than 20 kilowatts. Current AI installations commonly pass 100 kilowatts, and the newest designs can go much higher.

Air cooling becomes impractical at those densities. Operators need direct-to-chip liquid cooling, larger pumps, stronger floors, different power distribution and more sophisticated backup systems.

AI workloads also move quickly. The IEA has measured repeated server-load swings greater than 50% of rated capacity within one second. Batteries and power electronics must smooth those changes before they disrupt the facility or the surrounding grid.

A building can therefore have an available grid connection and still be unsuitable for frontier AI. Upgrading it may require rebuilding much of the equipment between the utility connection and the processor.

The bottleneck has reached the rack itself. Securing 200 megawatts for a campus does not guarantee that the operator can safely deliver hundreds of kilowatts into every cabinet.

Will more efficient AI chips solve the power bottleneck?

More efficient AI chips will slow the growth of electricity use per task, but they are unlikely to reduce total AI power demand anytime soon.

Efficiency is improving remarkably fast. The IEA estimates that energy consumption per basic AI task has recently fallen by at least a factor of ten each year. A simple text query now uses relatively little electricity.

AI companies are spending those savings on more demanding tasks. Reasoning models can generate and evaluate several possible answers before responding. Agents may plan, search, call tools, inspect their work and repeat failed steps. Video generation adds another large jump.

The difference can reach hundreds or thousands of times the electricity required for basic text generation, according to the IEA.

Usage is also growing. Major model providers reported roughly three times as many active users and five times as much revenue over the latest year covered by the agency. Lower prices bring more users, and better capabilities create entirely new categories of demand.

This is why AI-focused electricity consumption rose so quickly even while the underlying hardware became more efficient.

Better chips still help enormously. Without them, the same amount of useful AI work would require far more data centers. Custom accelerators from Google, Amazon, Microsoft and Meta also reduce dependence on general-purpose GPUs for certain workloads.

Efficiency slows the problem. It does not remove it, especially while companies keep using AI more often and for heavier tasks.

Is AI inference making the power bottleneck worse?

AI inference is turning power demand into a permanent operating requirement rather than an occasional cost of training new models.

Frontier training runs attract attention because they use thousands of accelerators for weeks or months. Once the run finishes, those processors can move to another job.

Inference continues whenever people and companies use the model. Search assistants, coding tools, customer-service systems and autonomous agents need capacity across several regions, every hour of the day.

Reasoning models increase that load. One visible answer may hide dozens of internal steps or tool calls. A user experiences one task while the infrastructure processes a much larger workload.

Inference also spreads geographically. A company may train a model in one giant, remote cluster, but customers often expect fast responses and local handling of sensitive data. Serving them can require smaller clusters across North America, Europe, Asia and other regions.

Training creates large peaks around a smaller number of facilities. Inference builds a broader and more persistent base load.

Specialised chips can reduce the electricity used for each response. Companies can also route simple requests to smaller models and schedule less urgent workloads when grids are under less pressure.

Those savings will help, although the volume of requests currently grows faster. As AI becomes a default feature inside software, search and communication products, inference will probably account for a growing share of the industry’s power requirements.

Can onsite power generation bypass the grid bottleneck?

Onsite generation can rescue some AI projects, but it rarely provides the easy shortcut developers hope for.

US developers are now considering natural-gas plants, fuel cells, batteries and other behind-the-meter systems because utility connections take too long. The IEA’s satellite tracking found that roughly one-fifth of proposed onsite gas projects had begun land clearing or construction.

Operating a reliable isolated power system is harder than matching a data center’s average consumption. Equipment has to cover maintenance, failures and sudden changes in server load.

The IEA calculates that an onsite gas system may need 30% to 70% more generation capacity than the data center’s normal requirement. A 500-megawatt campus could therefore need as much as 850 megawatts of generation equipment to maintain dependable service.

The equipment itself has become scarce. Global gas-turbine orders increased 70% in 2025. Turbine manufacturers, electrical-equipment suppliers and engineering firms now face the same wave of demand that overwhelmed GPU suppliers earlier in the AI boom.

Developers must also secure pipelines, air permits, water, backup equipment and community approval. Batteries help manage rapid changes in load but cannot replace a long-term energy source.

Nuclear projects offer reliable low-carbon power, although new reactors usually arrive too late for near-term construction schedules. Extending or restarting an existing plant can move faster, which explains the recent interest from Microsoft, Amazon and Meta.

The IEA expects between 15 and 27 gigawatts of onsite gas generation to power data centers by 2030, mostly in the United States. That would be meaningful without becoming the industry’s default model.

Onsite generation gives developers another route. It also replaces one infrastructure project with several others.

Do startups and hyperscalers face the same AI bottleneck?

Startups still experience a chip bottleneck, while hyperscalers increasingly experience the power problem behind it.

A startup usually does not request a 500-megawatt utility connection. It rents GPUs from Microsoft, Amazon, Google, Oracle or a specialised cloud provider.

Its immediate problems involve GPU prices, reservation limits and access to the newest systems. From that viewpoint, chips have clearly not been replaced.

The cloud provider sees the full chain. It must order processors, memory and networking equipment, place them inside a compatible data center and energise the facility. Any missing element reduces the capacity offered to customers.

Hyperscalers have ways to reduce chip risk. They place multi-year orders, design custom accelerators and spread purchases among several suppliers. Their scale gives them priority when supply is tight.

Power presents a different challenge. Even a company worth several trillion dollars cannot manufacture a transmission corridor by itself. It must work with utilities, regulators, local governments and equipment suppliers.

Large buyers also create their own problem. A startup may need a few megawatts distributed across existing cloud regions. A hyperscaler can request more electricity than an entire city uses.

The industry therefore has two versions of the bottleneck. Startups see a shortage of affordable computing capacity. Hyperscalers see the electricity, construction and equipment constraints that limit how much capacity they can sell.

Are AI data-center power forecasts exaggerated?

Many headline power forecasts are exaggerated, but the strongest projects still exceed what several grids can deliver.

Texas shows why raw project queues must be treated carefully. ERCOT was recently tracking 438 gigawatts of large-load connection requests, with data centers representing the vast majority.

Most of those requests had no completed studies. Some developers apply through several utilities or submit alternative sites before choosing one. Others may never secure funding, equipment or customers.

The gap between proposals and operation is enormous. ERCOT had approved around 9 gigawatts of large loads for energisation, while observed peak consumption from that group was only about 4 gigawatts. The operating load was therefore below 1% of the 438-gigawatt headline queue.

That comparison does not mean 99% of the proposed projects will fail. Future projects have had little time to progress. It does show why the queue cannot be used as a realistic demand forecast.

ERCOT has now moved toward a batch process that evaluates projects together and connects only the amount the grid can support in suitable locations. PJM has also tightened its approach by imposing minimum development periods, standard utilisation assumptions and clearer distinctions between firm and non-firm projects.

Financing will remove another layer of proposals. The largest data-center investments have become too big for some companies to fund entirely from their own balance sheets. Projects now depend more heavily on outside investors, debt markets and confidence that AI revenues will justify the cost.

The queue is inflated. The bottleneck affecting credible projects is still real, as Dominion’s scheduled and unscheduled pipeline and FERC’s intervention already show.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Which AI bottleneck will matter most through 2030?

Power will probably limit how quickly total AI capacity expands through 2030, while chips will continue to limit who gets the fastest systems.

The difference comes down to supply-chain speed.

Semiconductor companies can respond to committed demand by building fabrication, memory and packaging capacity. The process is expensive and technically difficult, but the finished products can be shipped worldwide.

Power infrastructure follows a slower route. Utilities need demand forecasts, regulatory approval, financing and cost-recovery plans. Transmission projects can cross several jurisdictions and face local opposition. Generation projects need fuel, equipment, land and permits.

The assets are also expected to last much longer. A chip may lose its performance lead after two or three years. A utility plans a substation or transmission line around decades of use. It cannot confidently spend billions whenever an AI developer submits an early proposal.

Current semiconductor shortages may still disrupt expansion. Micron’s supply commitments show that HBM remains tight, and geopolitical restrictions could limit access to advanced processors in some countries.

Still, the chip industry is already delivering rapid volume growth. NVIDIA’s data-center business almost doubled in a year. Electricity infrastructure has no comparable global production curve because every connection depends on local conditions.

More chips may even deepen the power problem. Faster deliveries allow developers to fill buildings sooner, raise rack density and request the next campus earlier.

For the rest of this decade, power looks like the slower and less flexible constraint. That gives it the stronger claim to being AI’s next major bottleneck.

Is power replacing chips as AI’s bottleneck?

Yes, power is replacing chips as the main bottleneck for building new hyperscale AI capacity, although chips remain the main bottleneck for accessing frontier performance.

The distinction is visible in company behaviour. The largest technology groups can reserve GPUs years ahead, fund custom processors and sign multi-billion-dollar memory contracts. They are still constrained, but those actions steadily increase their chip supply.

Their response to power has become more radical. They are acquiring energy developers, financing generation, supporting nuclear restarts, signing gigawatt-scale agreements and redesigning campuses around whichever electricity can arrive first.

Grid operators are changing their connection rules because ordinary processes can no longer handle the size and volume of requests. Utilities are giving dates to some projects while leaving tens of gigawatts unscheduled. Inside the facilities, rising rack density creates another electrical bottleneck after the grid connection is secured.

Chips have not become easy to obtain. HBM remains tight, complete systems are sold well in advance, and smaller customers still struggle to rent enough GPUs.

Chip scarcity increasingly changes the price, performance or allocation of a project. Power scarcity can decide whether that project opens at all.

Our judgment is mostly true. Power has taken the lead as the deployment bottleneck for the biggest AI infrastructure projects. Chips still control the technological frontier and remain the more visible constraint for startups, but the industry can manufacture additional processors faster than it can create additional powered sites.

Part of the AI market	Main bottleneck today
Frontier-model performance	GPUs, HBM and networking
Startup access to compute	Cloud GPU prices and allocations
Existing data-center expansion	Internal power and cooling density
New hyperscale campuses	Grid connections, generation and electrical equipment
National AI infrastructure	Deliverable power in suitable locations
Overall AI capacity through 2030	Increasingly power

If you want more recent data on this point, please see our latest AI infrastructure market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.