Data Centers: what’s changing now?
In our data center market deck, you will find everything you need to understand the market
SUMMARY
Data Centers: what’s changing now? AI is turning data centers into giant energy-and-infrastructure projects where power access, cooling, grid capacity and financing increasingly matter as much as the computing hardware itself.
The scale of the buildout is already physical, not just forecast. Global construction pipelines have jumped sharply, North American supply is still being absorbed almost immediately, and single campuses are now being planned in gigawatts rather than tens of megawatts.
The real scarce asset is often usable electricity rather than land. A cheap site has limited value if hundreds of megawatts cannot be connected for years, which is why “powered land” and realistic interconnection timelines are becoming central to site selection.
That power bottleneck is changing who a data-center developer has to become. Large projects increasingly involve gas generation, fuel cells, batteries, nuclear contracts, substations and transmission upgrades, so the boundary between a computing project and a power project is getting blurry.
AI hardware is also changing the building itself. Rack densities above 100 kW, and in some cases around 200 kW, push operators toward liquid cooling, heavier power distribution and mechanical systems designed around far more heat in the same physical footprint.
Efficiency is still improving, but it is losing the race against scale. A very efficient 500 MW facility still consumes vastly more electricity than an older 30 MW site, so PUE alone says less about the industry’s overall resource footprint than it once did.
Governments are reacting because the largest campuses can now affect local electricity prices, transmission plans, water systems and zoning. Ireland, Virginia, Singapore and Thailand are taking different approaches, but all are pushing large operators to carry more of the infrastructure burden they create.
Geography is therefore shifting. Northern Virginia remains dominant, but Atlanta, Texas and Johor can win new projects when they offer faster access to power, room for utility-scale infrastructure and fewer constraints on very large campuses.
Hyperscalers are shaping more of the market because one customer can now trigger hundreds of megawatts of construction. Their hardware roadmaps influence cooling, power procurement, rack design and even where entire data-center clusters get built.
The biggest risk is moving from empty-building risk toward return-on-capital risk. Today’s tight vacancy and strong preleasing argue against an immediate glut, but the industry is committing enormous sums to assets whose buildings may last decades while the GPUs inside them can become economically old within a few years.
This market map, featured in our data center market deck, highlights top companies and startups in the data center market
Why are data centers changing so much right now?
Data centers are currently turning into huge energy projects built around computing, and AI is forcing that change much faster than the old cloud boom did.
The jump in electricity demand gives us the clearest measure. The International Energy Agency expects global data-center electricity use to reach roughly 945 TWh by 2030, more than twice today’s level. AI is the biggest driver: electricity used by accelerated servers is expected to grow around 30% a year through 2030, compared with roughly 9% for conventional servers.
The projects themselves are also getting much bigger. Meta is expanding its Richland Parish complex in Louisiana to 5 GW of compute capacity and says investment there will exceed $50 billion. The accompanying energy plan includes seven new gas-fired generating plants, three grid-scale batteries, nuclear uprates and additional purchased power.
A 5 GW campus sits in a completely different category from the data centers the industry was building a decade ago. At that scale, securing the servers is only one part of the job. Developers also need enough electricity generation, substations, transmission equipment, cooling infrastructure, water, financing and local approval to make the campus possible.
That is the big change running through the industry today. Data centers used to consume infrastructure around them. Increasingly, they have to bring much of that infrastructure with them.
How big is the AI data-center boom right now?
The AI data-center boom is already huge by almost every physical measure we can track, and the latest construction numbers are still moving upward.
The IEA expects worldwide data-center electricity consumption to more than double by 2030. Its base case puts consumption near 945 TWh, slightly above Japan’s entire electricity use today. Data centers would still represent less than 3% of global electricity demand, but they are expected to grow more than four times faster than electricity consumption across the rest of the economy.
Construction tells the same story. Cushman & Wakefield's latest global study found roughly 31.7 GW of data-center capacity under construction, compared with 12.5 GW in the previous edition of its study.
North America alone had 7,481 MW under construction in CBRE's latest first-half market review, up 24.8% from a year earlier and above the previous construction record.
Entire established data-center hubs were once measured in hundreds of megawatts. Developers are now discussing single campuses measured in gigawatts.
Forecasts can still be wrong, especially when they depend on uncertain AI usage several years out. But tens of gigawatts are already being built.
| Measure | Earlier level | Latest level or forecast | What changed |
|---|---|---|---|
| Global data-center electricity use | Current base | ~945 TWh by 2030 | More than doubles |
| Global capacity under construction | 12.5 GW in prior Cushman study | ~31.7 GW | More than 2x |
| North American construction | Prior-year level | 7,481 MW | +24.8% |
| Meta Richland Parish campus | Originally smaller | 5 GW planned | Now utility-scale |
As this chart shows, and as featured in our data center market deck, search interest in data centers has increased significantly
Are companies actually filling all these new data centers?
Yes. Data-center demand is currently absorbing new capacity so quickly that record construction has barely loosened the market.
CBRE's latest review of North America's eight primary markets is unusually convincing here. Supply jumped 33.7% in one year to 10,903 MW, yet vacancy still sat at a record-low 1.4%.
More than 80% of the 7,481 MW being built had already been committed to customers. Only around 1,500 MW remained available for preleasing, which CBRE estimated was equivalent to roughly six months of demand at the current absorption rate.
Northern Virginia is even tighter. The world's biggest established data-center cluster had vacancy of just 0.2%.
Prices are moving with that scarcity. CBRE reported higher asking rents across every major deployment size, including increases of 8.3% for 3–10 MW requirements and 6.7% for deployments above 10 MW.
Developers managed to expand primary-market supply by roughly one-third in a year without creating meaningful spare capacity. For now, demand is arriving about as fast as operators can turn electricity and buildings into usable data-center space.
Is electricity now more valuable than land for data centers?
Yes. For large data centers today, access to usable electricity often determines a site's value more than the land itself.
Finding land for a giant industrial building is relatively easy. Finding the same land with several hundred megawatts that can actually be delivered within a useful timeframe is much harder. CBRE now describes power availability and infrastructure delivery times as the biggest factors shaping site selection and leasing.
That has pushed “powered land” to the center of the industry. The valuable asset is increasingly a parcel where the developer has already moved through difficult utility discussions, grid studies, permits and interconnection work.
A 1 GW computing campus running continuously near full load would consume roughly 8.8 TWh over a year before we adjust for utilization and other operational details. That puts a single campus into the electricity-demand range of major industrial operations.
AI hardware also ages quickly. Waiting several extra years for grid power can be especially painful when the chips intended for the site may have gone through multiple generations by the time electricity arrives.
These days, buying land without understanding its real power timeline can mean buying a data-center site that cannot become a data center. It sounds obvious, but plenty of the economics now hinge on exactly that.
If you want more recent data on this point, please see our latest data center market report.
This chart, featured in our data center market deck, illustrates yearly venture capital funding for data center startups
Are data centers starting to generate their own electricity with gas and fuel cells?
Yes. Onsite power, especially gas-backed generation and fuel cells, is moving from backup infrastructure toward a serious way of getting large AI data centers online faster.
Oracle's Project Jupiter in New Mexico is one of the clearest examples. Oracle, BorderPlex Digital Assets and Bloom Energy are developing an architecture using up to 2.45 GW of Bloom fuel cells for the AI campus.
That amount of onsite generation is comparable with the output of multiple large conventional power stations. Oracle's broader agreement with Bloom covers up to 2.8 GW of fuel cells, with 1.2 GW initially planned across U.S. projects.
Meta's Louisiana expansion is approaching the same problem differently: its arrangement with Entergy supports seven new gas plants, batteries, nuclear uprates and additional purchased electricity.
The IEA expects natural-gas generation serving data-center growth to increase by roughly 175 TWh through 2035. Renewables are expected to add much more electricity overall, but their output is variable, while new nuclear takes longer to build. That gives dispatchable power a near-term advantage.
This creates an uncomfortable reality for technology companies that have spent years promising lower-carbon operations. Their AI infrastructure needs firm power faster than many clean alternatives can currently be deployed.
Hybrid setups are therefore likely to spread. A large project can combine grid supply, onsite generation, batteries and contracted external power instead of depending on one source.
The fastest data-center developers today increasingly behave like power developers as well.
Can electricity grids actually handle the data-center boom?
Some can, but grid capacity is becoming one of the hardest limits on data-center growth in several important markets.
The United States shows the scale of the challenge. Berkeley Lab's reference scenario puts data centers at roughly 12% of U.S. electricity consumption by 2030, while other credible scenarios land above or below that depending on AI growth and efficiency.
The IEA expects data centers to account for nearly half of the increase in U.S. electricity demand through 2030. By then, it expects U.S. data centers to consume more electricity than the country's production of aluminium, steel, cement, chemicals and other energy-intensive goods combined.
Ireland already gives us a preview of what extreme concentration can look like. Data centers went from 5% of national electricity consumption in 2015 to 22% in 2024. Irish regulators now require new facilities to take much more responsibility for the generation, storage and renewable supply associated with their demand.
The engineering challenge also goes beyond annual electricity totals. Grid operators must be able to deliver huge quantities of electricity at the right place and time, cope with sudden changes in load and keep the network stable during disturbances.
A country may have plenty of generation in aggregate while a specific substation or transmission corridor has no room for another 500 MW campus.
This chart, featured in our data center market deck, shows how Equinix is capturing share in data centers
Is nuclear power really coming back because of data centers?
Yes. Data-center demand has become one of the strongest commercial forces behind nuclear power's revival, although most new reactors will arrive too late to solve the industry's immediate electricity shortage.
The list of technology buyers has become difficult to dismiss. Microsoft backed the planned restart of the former Three Mile Island Unit 1 through a long-term power agreement. Google has signed agreements supporting Kairos Power's advanced reactors. Amazon invested in X-energy and projects based on its small modular reactor technology. Meta has built a portfolio of nuclear agreements covering both existing generation and future capacity.
Google has also agreed to buy half of the output from Fortum's Loviisa nuclear plant in Finland for 22 years beginning in 2030 as it expands AI infrastructure in the country.
The attraction is obvious from a data-center operator's perspective. Nuclear plants can produce large quantities of electricity around the clock with very low operational carbon emissions. That combination is unusually useful for AI campuses that may need hundreds of megawatts continuously while their owners still have emissions targets.
Existing nuclear plants can help sooner through life extensions, uprates and restarts. New reactor designs are slower. The IEA expects the first small modular reactors contributing to this demand wave around 2030.
Nuclear is therefore becoming strategically important now, even though natural gas, renewables, storage and existing grid generation will carry much more of the near-term load.
If you want more recent data on this point, please see our latest data center market report.
Is liquid cooling becoming mandatory for AI data centers?
For the densest AI systems, liquid cooling is quickly becoming part of the standard design rather than an exotic upgrade.
Uptime Institute's latest global survey found that more operators are reaching peak rack densities of at least 30 kW. Its separate cooling study this year focused specifically on the continued adoption of direct liquid cooling, reflecting how central the technology has become to new high-density deployments.
The industry's ordinary racks are still much less extreme. Uptime says typical rack density has been moving toward roughly 10 kW, and much of the installed data-center base remains below that.
Frontier AI hardware is pulling away from that baseline. Configurations built around the newest accelerator generations can push individual rack requirements well above 100 kW, with some designs approaching or exceeding 200 kW.
A 200 kW AI rack can release around twenty times as much heat as a 10 kW conventional rack in roughly the same physical footprint.
That changes the building around the computers. Cooling loops, pumps, heat exchangers, power distribution, busways and mechanical layouts increasingly need to be designed for AI loads from the beginning.
Air cooling will remain common across conventional workloads, and much of the existing data-center stock will keep running useful workloads for years. But free floor space alone no longer tells you whether a building can take the hardware customers actually want to install.
This chart, featured in our data center market deck, illustrates yearly funding for data center startups
Are data centers getting more efficient while using more electricity?
Yes. Data centers are still getting more efficient, but AI demand is growing far faster than those efficiency gains can compensate for.
Uptime Institute's latest global survey found another gradual improvement in average power usage effectiveness, or PUE. The easy gains, however, are largely behind the industry. Moving an old facility from very poor efficiency toward a modern design can save a lot of electricity; squeezing another small improvement from an already efficient hyperscale site is harder.
Meanwhile, the computing load is exploding. The IEA expects electricity use by accelerated servers, driven mainly by AI, to grow around 30% a year through 2030.
A highly efficient 500 MW facility still consumes much more electricity than a mediocre 30 MW one. Better cooling and power systems reduce overhead, but they do not erase the energy used by the processors themselves.
This is why PUE tells us less about the industry's overall impact than it once did. Operators increasingly also care about the carbon intensity of electricity, server utilization, water consumption and how flexible loads can be around periods of grid stress.
Is water becoming a serious problem for data centers?
Yes, in water-stressed areas. Water is becoming a real constraint for some data-center projects, although the problem varies much more by location and cooling design than electricity scarcity does.
Evaporative cooling can cut the electricity required for cooling but consume substantial water. Other designs can dramatically reduce onsite water use while creating different energy, cost or engineering trade-offs.
The issue has become especially visible in places attracting a sudden wave of large projects. Thailand recently asked operators to suspend work on 49 planned data-center developments while the government prepares dedicated rules covering electricity, water, cooling and zoning. Among the issues under discussion are closed-loop cooling requirements and guarantees around water availability.
Major hyperscalers are also spending more effort demonstrating that their water use can be managed. Google's latest environmental reporting says its water-stewardship projects replenished around 7.7 billion gallons in 2025, equivalent to 78% of its reported freshwater consumption.
National statistics can make water demand look small, while the local picture can be very different. A project drawing from one constrained watershed competes with the people, farms and industries connected to that same system.
This chart, featured in our data center market deck, compares the main business model options for hyperscale data center operators
Why are governments suddenly getting tougher on data centers?
Governments are getting tougher because today's biggest data centers can reshape local electricity, water and infrastructure needs on a scale that ordinary commercial buildings never did.
North America is already seeing the change. CBRE says local opposition and zoning delays are now serious obstacles to data-center projects, alongside electricity availability.
Virginia has started changing how very large electricity users pay for infrastructure. New rules are designed to make large-load customers bear more of the cost created by transmission and distribution assets built for them, rather than leaving those costs with ordinary electricity customers.
Ireland has gone further. New data centers face requirements around matching generation or storage and additional renewable electricity because the sector already represents an exceptionally large share of national power demand.
Singapore is taking a controlled-growth approach. Its Green Data Centre Roadmap initially set out at least 300 MW of additional capacity, with projects expected to meet tighter resource-efficiency standards. The government has lately reiterated that further growth will remain measured, while new legislation raises baseline sustainability requirements for data centers.
Thailand is now dealing with the same issue from a different starting point. Its requested pause covering 49 projects came as authorities worked on data-center-specific rules after a rapid investment wave exposed gaps in existing zoning, electricity and water regulations.
Across very different markets, regulators are converging on the same basic idea: unusually large data centers should shoulder more of the infrastructure burden they create.
If you want more recent data on this point, please see our latest data center market report.
Are data centers leaving Northern Virginia for places like Atlanta, Texas and Johor?
Data centers are spreading much faster into new regions, but Northern Virginia is still far from losing its position as the industry's dominant established cluster.
Northern Virginia had roughly 4.5 GW of existing inventory in CBRE's latest North American review, more than any other primary market, and vacancy was only 0.2%. Customers clearly have not abandoned it.
The interesting change is happening in the development pipeline. Atlanta reached roughly 2.9 GW under construction, overtaking Northern Virginia's construction pipeline for the first time.
Texas is also attracting enormous AI campuses. Its appeal goes beyond cheap land: developers can pursue sites with room for generation, substations and gigawatt-scale electrical infrastructure that would be much harder to assemble inside mature urban clusters.
Asia shows the same pattern. Johor has rapidly become a major alternative for hyperscale expansion around Singapore, while Singapore itself deliberately controls the amount of additional capacity it approves.
Large AI training workloads make this geographic shift easier. Some computing jobs do not need to sit milliseconds away from the end user, so developers can trade a less central location for faster electricity access.
For the largest AI projects today, a place with available power can beat a famous data-center market where the next meaningful connection is years away.
| Market | What is happening now |
|---|---|
| Northern Virginia | Still the largest established U.S. cluster and almost full |
| Atlanta | Construction pipeline has overtaken Northern Virginia |
| Texas | Attracting very large power-led AI campuses |
| Johor | Emerging rapidly as a major Asian hyperscale location |
| Singapore | Allowing additional capacity selectively rather than freely |
This chart, featured in our data center market deck, shows the revenue mix across customer segments in the data center market
Are hyperscalers taking over the data-center industry?
Hyperscalers are currently gaining more control over where data-center capacity gets built, what it looks like and how quickly the industry invests.
Synergy Research counted 1,360 hyperscale data centers worldwide at the end of 2025. Hyperscale operators represented roughly 48% of global data-center capacity, and Synergy expects that share to reach 67% by 2031.
Traditional enterprise-owned capacity is moving in the opposite direction. Synergy expects its share to fall to around 19% by 2031.
Uptime Institute found another milestone this year: for the first time in its global survey, third-party data-center facilities and services accounted for a larger share of surveyed IT workloads than enterprise-owned facilities.
Colocation companies still have a large role because hyperscalers lease enormous amounts of space as well as building their own campuses. Around 40% of hyperscale capacity was in leased facilities in Synergy's latest estimate.
A handful of cloud and AI companies can now trigger hundreds of megawatts of construction with one contract. Their hardware roadmaps also influence rack density, cooling systems, electricity procurement and building design across the companies supplying them.
Why are companies spending so much money on data centers?
Companies are spending extraordinary amounts on data centers because AI infrastructure now combines expensive chips with buildings, power equipment and energy projects at unprecedented scale.
The biggest technology companies give us the clearest view. Amazon expects roughly $200 billion of company-wide capital expenditure in 2026, with AI infrastructure a major reason for the increase. Alphabet expects approximately $180–190 billion. Meta expects roughly $130–145 billion.
Taken together, those three companies are pointing to more than $500 billion of capital expenditure in a single year.
And they are only part of the buildout. Microsoft, Oracle, CoreWeave, OpenAI's infrastructure partners, specialist cloud operators and major colocation companies are all investing heavily at the same time.
JLL estimates that adding around 100 GW of global data-center capacity through 2030 could require roughly $3 trillion of total investment.
The money now extends well beyond the building. As seen above, Meta's 5 GW Louisiana project comes with new generating plants, batteries and nuclear uprates. Oracle is pursuing gigawatts of onsite fuel cells. Other developments require dedicated substations, transmission upgrades and cooling systems designed for extreme rack densities.
This chart, featured in our data center market deck, shows how hyperscale AI-ready campus technology has evolved over time
Is data-center financing becoming dangerous?
Yes. Financing has become one of the clearest risks in the data-center boom because capital commitments are growing faster than the industry's track record for monetizing AI at this scale.
The largest technology companies can fund a large share of their investment from cash flow, but even they are increasingly using debt and other structures alongside their own balance sheets. Smaller AI infrastructure companies depend much more heavily on external financing.
The latest example is Iren. The AI infrastructure company says it could invest up to $30 billion by mid-2027 and has raised roughly $19 billion through different financing channels while expanding around large Nvidia-based deployments.
Across the sector, financing now includes ordinary corporate bonds, project debt, private credit, asset-backed structures and other vehicles tied to long-term data-center contracts.
The uncomfortable part is the mismatch in lifespans. A data-center building can operate for decades. Transmission assets can last even longer. The GPUs that initially justify a project may be economically old after only a few years.
A long hyperscale lease with a strong customer can reduce that risk substantially. A heavily financed campus built around one fast-growing AI customer, optimistic utilization assumptions and short-lived hardware is much more exposed.
If you want more recent data on this point, please see our latest data center market report.
Could the data-center boom still become a bubble?
Yes, but today's evidence points more toward a future returns problem than a current glut of empty data centers.
If developers had already massively overbuilt, we would expect excess space, weakening rents and difficulty leasing new projects. The latest North American numbers still show the reverse: vacancy is around 1.4%, rents are rising and most construction has customers attached before completion.
As pointed out above, more than 80% of North America's primary-market construction pipeline was already committed in CBRE's latest review. That is not proof that every project being planned today will earn a good return, but it does show that the current physical market remains extremely tight.
The risky part sits further ahead. Companies are committing hundreds of billions of dollars on assumptions that AI usage will keep rising, customers will remain creditworthy, expensive chips will stay heavily utilized and future computing revenue will cover the enormous infrastructure built around them.
The larger projects get, the more painful forecasting errors become. A company being wrong about demand for a 30 MW facility is manageable. Being wrong about several gigawatts can hurt developers, lenders, utilities and local power systems simultaneously.
A data-center bubble could therefore emerge through poor returns on enormous amounts of capital even if most of the buildings themselves get leased.
In our data center market deck, we identify pain points entrepreneurs should prioritize
What is really changing in data centers right now?
Data centers are currently becoming a new class of industrial infrastructure where electricity access matters almost as much as computing technology.
The numbers across the industry line up unusually well. Electricity demand is rising rapidly. Construction is at record levels. Large markets remain almost full. AI racks are getting much denser. Developers are installing liquid cooling and pursuing their own power supply. Nuclear plants are signing technology-company contracts. New gas generation is being tied directly to AI campuses. Governments are rewriting connection and resource rules. Hundreds of billions of dollars of annual capital spending are flowing into the buildout.
The key question for a major data-center project has therefore changed. Finding customers and constructing the building are no longer enough. The difficult part is assembling hundreds of megawatts of electricity, getting that power delivered on time, cooling extremely dense hardware and persuading utilities, regulators, lenders and nearby communities that the project can work.
That shift explains why Atlanta, Texas and Johor can suddenly compete with older hubs, why technology companies are signing nuclear agreements, why Oracle is buying gigawatts of fuel cells and why Meta's latest campus comes with an energy system that looks more like infrastructure planning for a small region.
For now, demand still looks strong enough to justify an enormous amount of construction. The harder test comes next: whether the industry can build the power system quickly enough, and whether all the AI revenue expected from these campuses eventually earns a good return on the extraordinary amount of money being spent.
If you want more recent data on this point, please see our latest data center market report.
OUR METHODOLOGY
This analysis tests what is actually changing in data centers today by breaking the broad question into the main areas where change can be observed directly: demand and construction, electricity and grid access, power generation, cooling and water, regulation, geography, market structure, capital spending and financial risk.
We looked for recent, concrete measures rather than broad industry narratives. We gave more weight to electricity consumption, megawatts under construction, vacancy, preleasing, rack density, secured generation, regulatory decisions and committed capital because those measures show physical or financial activity more directly than general announcements or sentiment.
Large projects from Meta, Oracle and individual utilities were used as examples of what the extreme end of the market now looks like. We checked those examples against broader datasets from energy agencies, regulators and established data-center research groups so that no single project carried a conclusion by itself.
Forecasts and observed market conditions played different roles. Forecasts were used to understand the possible scale and direction of the buildout, while construction, leasing, electricity consumption and operating data were used to show how much of the change is already visible today. Where several independent measures pointed the same way, we treated the conclusion as stronger.
Some geographic comparisons were chosen because they isolate specific changes clearly. Ireland shows what unusually concentrated data-center electricity demand can do to grid policy. Northern Virginia and Atlanta separate the size of an established market from the direction of new construction. Singapore and Johor show how capacity can shift across neighboring markets when power, land and regulation create different development conditions.
We treated the bubble question separately from the demand question. Current vacancy, preleasing and rents help show whether the industry has already built more usable capacity than customers want. Capital commitments, financing structures, hardware replacement cycles and expected AI revenues help assess whether the money being invested will eventually earn adequate returns.
We prioritized first-hand information wherever possible: government statistics and regulatory decisions for electricity and policy, company disclosures for specific projects and spending commitments, and technical documentation for hardware. For broader market measurements that no single public authority collects, we used established industry research.
Key sources used for this analysis include: the International Energy Agency on data-center electricity demand, CBRE's North America Data Center Trends review, Cushman & Wakefield's Global Data Center Market Comparison, Meta on its Richland Parish expansion, Oracle on Project Jupiter, Lawrence Berkeley National Laboratory on U.S. data-center energy use, Ireland's Central Statistics Office on data-center electricity consumption, Ireland's Commission for Regulation of Utilities on connection policy, Uptime Institute's Global Data Center Survey, Synergy Research Group on hyperscale capacity, JLL's Global Data Center Market Outlook, Singapore IMDA's Green Data Centre Roadmap, and the Royal Thai Government on tighter data-center assessment covering power, water and environmental impacts.
This chart, featured in our data center market deck, shows the revenue mix by region across Europe, Asia, North America, Africa, and South America in the data center market
Related blog posts
- Data Center Tech: what are startups building now?
- Data Centers: what are the biggest unsolved problems?
- The latest funding news in the data center market
- What are the top startups in the data center market?
- What are the fundraising trends in the data center market?
- How strong is fundraising in the data center market right now?
