Data Centers: what are the biggest unsolved problems?
In our data center market deck, you will find everything you need to understand the market
SUMMARY
Data centers’ biggest unsolved problems are securing enough power quickly, planning around uncertain AI demand, handling extreme rack density, integrating giant loads safely with the grid, and scaling the industrial infrastructure needed to support all of that growth.
The core mismatch is speed. AI computing capacity can be ordered and data centers can be built within a few years, while transmission lines, power plants, substations and major electrical equipment often take much longer.
Power availability has therefore become more important than the building itself. A developer may have land, financing and servers ready, yet still be unable to build because hundreds of megawatts cannot be delivered to that exact site.
The uncertainty around future demand makes the grid problem harder. Berkeley Lab’s U.S. scenarios range from roughly 521 TWh to 843 TWh of data-center electricity consumption in 2030, so utilities are being asked to make decades-long infrastructure decisions around an unusually wide range of outcomes.
Building dedicated power beside data centers will become more common, but it does not remove the bottlenecks. Gas turbines, transformers, switchgear and other equipment are themselves becoming constrained, while nuclear and geothermal usually cannot arrive quickly enough for projects that need electricity within the next few years.
Flexible AI demand could become one of the most valuable tools for speeding grid connections. Training jobs can sometimes move or pause, but the unanswered commercial question is how often operators will accept idle GPUs after spending billions of dollars on them.
Cooling looks more solvable than electricity. AI racks are moving toward power densities above 100 kW and, at the extreme, above 200 kW, forcing a shift toward liquid cooling, yet the basic thermal technology already works and operating practices are beginning to converge.
Water is increasingly a geographic constraint rather than a universal technical limit. Closed-loop, dry and adiabatic designs can sharply reduce direct water consumption, although water-stressed regions can still turn a technically viable project into a politically difficult one.
Huge computational loads are also becoming part of grid reliability itself. If clusters of data centers suddenly reduce consumption during the same voltage disturbance, grid operators can lose hundreds or thousands of megawatts of demand almost instantly, which is forcing new reliability rules for loads that were once treated as ordinary customers.
The broader pattern is that the hardest data-center problems now sit between industries. Utilities, generators, regulators, equipment manufacturers, cooling suppliers, local governments and data-center developers all need to expand in roughly the right order, and one slow layer can strand billions of dollars of equipment elsewhere.
Power availability remains the clearest number-one problem today. High-density cooling, water use and conventional facility engineering increasingly have visible technical paths forward; grid capacity, demand uncertainty, cost allocation and firm clean electricity depend on much larger systems that individual operators cannot fix on their own.
This market map, featured in our data center market deck, highlights top companies and startups in the data center market
Why have data centers suddenly become so hard to build?
Data centers are becoming much harder to build because AI demand is growing on software timelines while electricity, grid equipment and cooling infrastructure still expand on industrial timelines.
The scale change is unusually fast. The IEA's latest outlook puts global data-center electricity use at roughly 485 TWh in 2025 and around 950 TWh by 2030. AI-focused facilities account for much of that increase, with their electricity consumption expected to roughly triple over the period.
A data center can also move from planning to operation much faster than the infrastructure around it. The IEA estimates that a facility can be ready in two to three years, while new transmission lines and major generation projects often take much longer.
That mismatch explains why today's hardest data-center problems increasingly appear outside the server itself. Developers can buy computing hardware, design buildings and pour concrete quickly. Getting enough electricity to the site, cooling extremely dense racks and convincing utilities and communities to support the project are proving much harder.
Is getting enough power now the biggest data-center problem?
Yes. Getting enough electricity at the right location is currently the biggest unsolved problem facing large data centers.
Berkeley Lab's latest bottom-up estimate suggests U.S. data centers could consume about 649 TWh in 2030, or 11.8% of national electricity use. Its plausible range stretches from 9.5% to 15.3%. A few years ago, data-center demand was still commonly treated as a fairly small part of the electricity system.
The national percentages also hide the harder local problem. A large AI campus can request hundreds of megawatts at one site. The local utility therefore needs that capacity on one section of the grid, rather than somewhere else in the country.
Uptime Institute's 2026 survey confirms that operators themselves are becoming more worried about power availability. FERC has reached the same conclusion from the grid side: in June 2026, it ordered all six U.S. regional grid operators under its jurisdiction to justify or change the rules used to connect very large new loads.
Power has become unusually unforgiving as a constraint. More expensive cooling can still be installed. More servers can eventually be manufactured. A site without a credible route to hundreds of megawatts can simply become impossible to develop.
If you want more recent data on this point, please see our latest data center market report.
As this chart shows, and as featured in our data center market deck, search interest in data centers has increased significantly
Why are data-center grid connections taking so long?
Data-center grid connections are taking so long because utilities are being asked to approve enormous loads before the generation and transmission needed to serve them exist.
A request for 500 MW can trigger far more work than connecting another commercial building. The utility may need a larger substation, transmission upgrades, new generation, stability studies and agreements over who pays if the customer never uses all the capacity it requested.
The rules were also built for a slower world. FERC's recent interventions show how unusual the situation has become. After initially focusing on PJM, the largest U.S. power market, the regulator expanded its scrutiny across all six regional grid operators under its jurisdiction. FERC said existing rules may no longer be adequate for the rapid growth of data centers and other large loads.
Some developers are trying to shorten the wait by building beside power plants and connecting directly to generation. Even that approach has forced regulators to answer awkward questions. A data center beside a power plant may still depend on the wider grid during certain conditions, so regulators have to decide how much transmission service it needs and how much it should pay.
Physically connecting a large electrical load is well understood. The harder problem today is getting utilities, generators, developers and regulators to make decisions at anything close to the speed at which AI capacity is being ordered.
Do utilities actually know how much data-center power to build for?
No. Utilities still face a huge range of possible data-center electricity demand, which makes deciding how much infrastructure to build surprisingly risky.
Berkeley Lab currently estimates U.S. data-center electricity demand could reach anywhere from about 521 TWh to 843 TWh in 2030. The 322 TWh gap between those scenarios is enormous. It comes from different assumptions about AI-chip shipments, utilization, hardware efficiency and how quickly servers are replaced.
Project announcements create another layer of uncertainty. Developers often request power years before a campus is fully built, and several projects may effectively compete for the same future customers. A utility cannot safely assume that every announced gigawatt will arrive exactly as planned.
Building too little leaves real data centers waiting years for electricity. Building for every optimistic proposal could leave households and other businesses paying for grid infrastructure that ends up underused.
This forecasting problem is now tied directly to the AI business itself. Utilities effectively have to make long-lived infrastructure decisions while nobody yet knows how quickly AI inference will grow, how intensively GPUs will be used or how much computing efficiency will improve.
| U.S. data-center electricity in 2030 | Berkeley Lab estimate |
|---|---|
| Lower scenario | 521 TWh |
| Reference case | 649 TWh |
| Higher scenario | 843 TWh |
| Share of U.S. electricity in reference case | 11.8% |
This chart, featured in our data center market deck, illustrates yearly venture capital funding for data center startups
Can't data centers just build their own power plants?
Some data centers can bring their own power, and more are trying, but doing so creates a new set of bottlenecks rather than making the electricity problem disappear.
Developers are looking at onsite gas turbines, fuel cells, batteries, geothermal projects and nuclear power because waiting for the grid can delay a multi-billion-dollar AI campus. Ireland has already pushed this idea much further: new data centers seeking grid connections must provide generation or storage broadly matching their requested maximum demand.
Gas looks like one of the fastest options, yet even gas equipment is becoming constrained. The IEA recorded a roughly 70% jump in global gas-turbine orders in 2025 as utilities and large power users rushed to secure equipment. Turbine manufacturing capacity cannot expand as quickly as another AI cluster can be announced.
Nuclear and geothermal could provide attractive round-the-clock power over the longer term, but they generally cannot solve a campus that needs electricity within the next couple of years. Batteries help with peaks and grid support but cannot economically power a giant facility for long periods without another energy source behind them.
Bringing power directly to the campus will therefore become much more common. For now, it works best as part of the solution rather than a universal shortcut around the grid.
If you want more recent data on this point, please see our latest data center market report.
Can AI data centers simply turn down their computing when the grid is stressed?
Some AI data centers probably can reduce or shift electricity demand, but we still do not know how much computing operators will genuinely be willing to interrupt.
The idea is attractive. AI training jobs do not all need to run at a particular second. Some workloads can potentially move to another region, batteries can carry part of the load for a while, and lower-priority jobs can wait until electricity is easier to supply.
That flexibility could make a huge difference to grid connections. Instead of promising a new campus 500 MW during every hour of the year, a utility might connect it sooner if the operator agrees to cut consumption during a small number of stressed periods.
The harder question is how AI economics change the calculation. An operator that has spent billions of dollars on GPUs wants those GPUs working. Inference serving paying customers can also be much harder to interrupt than a flexible training run.
Utilities and operators are now testing these arrangements much more seriously, including through demand-response and flexible-interconnection programs. There is real potential here, but the amount of dependable flexibility available from a commercial AI data center remains one of the most important unanswered questions in the power debate.
This chart, featured in our data center market deck, shows how Equinix is capturing share in data centers
Are AI racks getting too hot for normal data-center cooling?
Yes. The newest AI racks are moving beyond the power densities that conventional air-cooled data centers were built around.
Uptime Institute's latest survey shows more operators now reporting peak rack densities above 30 kW. At the extreme end, the hardware roadmap is moving far faster: Uptime's recent research says upcoming AI systems are pushing toward rack power above 200 kW.
That jump changes the engineering problem. Moving heat away from a 5 or 10 kW rack with air is very different from continuously removing well over 100 kW from almost the same footprint. Fans need more energy, airflow becomes harder to control, and temperature differences inside the rack become more difficult to manage.
Liquid can carry much more heat than air, which is why direct-to-chip cooling is spreading so quickly through AI facilities. The difficult part now is keeping the rest of the facility architecture moving at the same pace as the chips.
As Uptime Institute's 2026 survey showed earlier, overall rack density is still rising more slowly across the entire installed base. That creates an awkward transition period in which operators may need to support ordinary air-cooled servers and extremely dense liquid-cooled AI hardware inside the same estate.
Has liquid cooling solved the AI data-center heat problem?
Liquid cooling can handle the heat, but data-center operators are still working out how to deploy and maintain it consistently at large scale.
Uptime Institute describes the current period as the first real wave of large-scale direct liquid cooling outside specialized supercomputing. Deployment grew fast enough that standards and operating practices struggled to keep up, especially around coolant chemistry, redundancy, maintenance and the dividing line between IT equipment and facility infrastructure.
Those details can become painfully practical. Someone has to decide who owns the coolant distribution unit, who responds to a leak, which coolant different server vendors accept, how pumps are made redundant and whether technicians can service one rack without disrupting neighboring equipment.
The good news is that some patterns are starting to emerge. Uptime's recent field research found operators converging around water-based systems, larger coolant distribution units and more clearly defined maintenance responsibilities.
Cooling is therefore one of the areas where the path forward already looks fairly clear. The core thermal technology works. The remaining challenge is turning liquid cooling into boring, standardized infrastructure that can be installed and maintained across thousands of facilities.
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 funding for data center startups
Is water still one of the biggest data-center problems?
Water is still a serious data-center problem in dry or water-stressed locations, but it no longer looks like an unavoidable technical limit on the whole industry.
Cooling designs can reduce direct water consumption dramatically. Microsoft, for example, has introduced a closed-loop data-center design that it says avoids more than 125 million liters of cooling water per facility each year.
Recent operating data also weakens the old assumption that operators always have to choose between low electricity use and low water use. Uptime Institute's 2026 analysis found that well-designed dry and adiabatic cooling systems can match the energy performance of evaporative systems across several climate zones while consuming little or no water.
Location still changes everything. Saving water in a wet, cool region has a different value from saving the same amount in Arizona or another water-stressed market. Electricity generation can also consume water away from the facility, so simply reporting onsite water use gives an incomplete picture.
We therefore rank water below power and high-density cooling today. The technology increasingly offers ways around heavy water consumption, while local water availability and the economics of choosing those alternatives remain unresolved.
| Data-center water problem | Where we stand |
|---|---|
| Water evaporated for cooling | Can already be reduced dramatically |
| Using drinking water for cooling | Alternatives increasingly exist |
| Energy penalty from saving water | Smaller than often assumed, but design-dependent |
| Water scarcity around specific campuses | Still a serious local constraint |
| Indirect water used to produce electricity | Much harder for the data center alone to control |
Can huge AI data centers actually destabilize the power grid?
Yes. Very large AI data centers are creating grid-stability problems that power-system rules are only beginning to catch up with.
NERC has been investigating cases where large groups of computational equipment suddenly reduce electricity consumption during grid disturbances. Servers and their power electronics can react to voltage problems very differently from older industrial loads.
That can create a strange event for grid operators: instead of suddenly losing a large power plant, they can suddenly lose a huge block of demand. If several nearby data centers respond to the same voltage disturbance in the same way, electricity consumption can drop by hundreds or even thousands of megawatts almost at once.
NERC's work accelerated further in 2026. It produced new guidance for computational-load operators, assessed gaps in existing reliability standards and began developing disturbance-performance requirements specifically for these loads.
The industry can engineer better ride-through behavior, battery support and controls. Common requirements are the part still catching up, and they need to arrive before data-center clusters get much larger. Today's biggest AI campuses are becoming important pieces of power-system infrastructure whether their owners planned to play that role or not.
This chart, featured in our data center market deck, compares the main business model options for hyperscale data center operators
Are transformers, turbines and skilled workers becoming the next data-center bottleneck?
Yes. Once developers secure land and electricity, shortages in physical equipment and skilled people can still slow the data-center buildout.
The turbine shortage is one example, but the pressure extends into transformers, switchgear, power electronics, cooling equipment, backup systems and advanced semiconductor components. Several of these industries historically planned capacity years ahead because demand was relatively predictable. AI has forced them to respond to a much faster investment cycle.
People are becoming scarce too. Uptime Institute's latest global survey found that more than half of data-center operators had difficulty finding qualified candidates for open positions. Staff turnover remains a problem because experienced technicians can move between rapidly expanding operators.
As pointed out above, that same Uptime survey also found operators increasingly worried about supply chains and power availability. Taken together, the findings show that "data-center capacity" is much broader than the number of buildings developers can finance.
A new AI campus requires the grid connection, transformers, switchgear, cooling plant, generators or batteries, GPUs, networking gear and trained people to arrive in roughly the right sequence. A shortage in any one of those layers can leave billions of dollars of other equipment waiting.
If you want more recent data on this point, please see our latest data center market report.
Who should pay for the grid upgrades that new data centers need?
The industry still has no settled answer on who should pay for major data-center grid upgrades, and governments are now becoming much less willing to leave that risk with ordinary electricity customers.
The problem starts with uncertainty. A utility may spend heavily on substations or transmission for a proposed data center that later opens more slowly than expected, uses less electricity or never gets built. The physical infrastructure still has to be paid for.
Utilities are responding with larger upfront payments, minimum-demand commitments, longer contracts and exit fees. FERC's current large-load proceedings are also heavily focused on whether existing tariffs protect customers while still allowing data centers to connect quickly.
Recent political decisions show how far the issue has moved. Virginia introduced a dedicated $0.011-per-kWh electricity consumption tax on qualifying data centers. New York went further and temporarily paused new hyperscale data-center permitting while it develops rules aimed at protecting ratepayers, the grid and local communities.
Ireland offers another version of the same answer. Data centers there already consumed 22% of national electricity in 2024, up from only 5% in 2015. New projects now face requirements covering matching generation or storage and additional renewable power.
These policies point in the same direction: data-center developers are increasingly being asked to carry more of the infrastructure risk created by their projects. How far governments can push that without making new capacity uneconomic is still being worked out.
This chart, featured in our data center market deck, shows the revenue mix across customer segments in the data center market
Can data centers keep growing this fast without making clean-power targets much harder?
Data centers can keep growing while electricity gets cleaner, but today's buildout is moving faster than round-the-clock clean power can be added in many markets.
Buying enough renewable electricity over a year sounds simple until we look hour by hour. AI servers need power at night, during windless periods and during grid emergencies too. Solar and wind can supply a large share of annual consumption, but a giant computing campus still needs firm electricity whenever renewable production falls.
That is why hyperscalers are signing nuclear, geothermal, storage and other firm-energy deals alongside conventional wind and solar contracts. The interest in these technologies has moved well beyond sustainability marketing; reliable electricity is becoming part of the competitive race to secure future computing capacity.
The timing remains difficult. Data centers can be built within a few years, while nuclear plants, transmission networks and many geothermal developments take longer. Gas therefore continues to fill part of the gap because it can provide controllable power more quickly.
This problem looks solvable over time, but the transition period will be messy. AI demand is currently arriving before enough firm low-carbon electricity exists in many of the places developers most want to build.
So what are the biggest unsolved problems in data centers today?
Power availability is clearly the biggest unsolved data-center problem today, followed by planning the grid around uncertain AI demand and safely operating much denser computing infrastructure.
Several problems that looked fundamental a few years ago are becoming easier to solve technically. Liquid cooling is maturing quickly. Dry cooling can sharply reduce water consumption. Conventional data-center outage rates have generally improved. Those issues still require work, but practical engineering paths through them are already visible.
Electricity remains harder because one company cannot solve it alone. Developers depend on utilities, power generators, equipment manufacturers, regulators and local communities, all operating on different timelines.
The next group of problems comes directly from the size of AI infrastructure. Rack power is heading toward levels that force liquid cooling, large computational loads are beginning to affect grid stability, and the supply chain has to scale everything from turbines and transformers to technicians at the same time.
For now, this is how we would rank the problems that still lack a clean, widely deployable answer:
| Rank | Biggest unsolved problem | Why it is still difficult |
|---|---|---|
| 1 | Getting enough power quickly | Data centers can be built much faster than new grid capacity |
| 2 | Knowing how much power to build | AI demand, utilization and project pipelines remain unusually uncertain |
| 3 | Handling extreme rack density | AI hardware is changing faster than cooling and facility standards |
| 4 | Connecting giant loads without destabilizing the grid | Grid rules for computational loads are still being written |
| 5 | Scaling the industrial supply chain | Transformers, turbines, electrical gear, cooling systems and people must all expand together |
| 6 | Deciding who pays for grid expansion | Utilities cannot safely leave speculative infrastructure costs with existing customers |
| 7 | Finding enough firm clean electricity | Data centers need reliable power every hour, while many clean projects take longer to build |
| 8 | Winning local permission to build | Electricity prices, water, noise and infrastructure costs are creating stronger political resistance |
| 9 | Managing water in stressed regions | Technical alternatives exist, but geography can still make water a hard constraint |
| 10 | Making flexible AI demand commercially real | Workloads can theoretically move or pause, but operators still need strong incentives to do it |
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, shows how hyperscale AI-ready campus technology has evolved over time
OUR METHODOLOGY
This analysis looks at the biggest unsolved problems in data centers by separating the question into the main constraints that can actually delay, reshape or prevent new capacity: power availability, grid connections, demand forecasting, cooling, water, grid stability, industrial supply chains, staffing, clean electricity and the cost of expanding the grid.
We investigated each area separately and then brought the findings back together. The ranking gives more weight to problems that affect a large share of the industry, can materially stop projects from being built, depend on infrastructure outside a data-center operator's direct control, and still lack a solution that can be deployed quickly and repeatedly at scale.
Fresh evidence matters particularly for data centers because AI infrastructure is changing quickly. Older assumptions about electricity demand, rack density, cooling systems and grid connections can become outdated fast, so we prioritized recent primary data, regulatory actions, technical work and operator evidence.
For electricity demand, we relied heavily on the International Energy Agency and Lawrence Berkeley National Laboratory. The IEA provides the global view of data-center and AI electricity growth, while Berkeley Lab's bottom-up U.S. scenarios are useful because they preserve the unusually wide uncertainty around future demand instead of reducing it to one forecast.
For grid connections, cost allocation and large-load regulation, we used material from the Federal Energy Regulatory Commission and the U.S. Department of Energy. FERC's recent proceedings are particularly useful because they show that large computational loads are already forcing changes to the rules used by regional power markets.
For grid-stability risks, we used North American Electric Reliability Corporation work on large and computational loads. NERC's guidance and standards work helps distinguish ordinary concerns about high electricity consumption from the newer reliability problem created when very large blocks of computing load react simultaneously to grid disturbances.
For cooling, rack density, staffing, supply chains and operating practices, we used Uptime Institute research and its latest global survey. We also used Microsoft's published zero-water cooling design as a concrete example of how direct cooling-water consumption can already be reduced through facility design.
We treated dedicated generation and clean-power options separately from ordinary grid supply. The IEA's work on power supply and gas-turbine orders, together with Ireland's Commission for Regulation of Utilities and Central Statistics Office data, helps show both the potential and the limits of asking large data centers to bring more of their own electricity infrastructure.
Key sources used for this analysis include: the International Energy Agency on global data-center electricity demand and AI growth, the IEA on data-center demand, geographic concentration and infrastructure lead times, the IEA on electricity supply for AI, Lawrence Berkeley National Laboratory on U.S. data-center electricity scenarios, Berkeley Lab and the U.S. Department of Energy on data-center load flexibility, the U.S. Department of Energy on large-load electricity tariffs and infrastructure risk, FERC on large-load integration, FERC's large-load interconnection proceeding, NERC's Large Loads Action Plan, Uptime Institute's Global Data Center Survey, Uptime Institute on direct liquid cooling maintenance, Uptime Institute on AI rack power and liquid cooling, Uptime Institute on dry and adiabatic cooling, Microsoft on its zero-water cooling design, the IEA on natural-gas turbine orders, Ireland's Commission for Regulation of Utilities on data-center connection policy, and Ireland's Central Statistics Office on data-center electricity consumption.
In our data center market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Data Center Tech: what are startups building now?
- Data Centers: what’s changing now?
- 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?
