Data centers topped headlines in 2025, serving as physical monuments of how artificial intelligence turned from abstract code into real, tangible infrastructure sprawled across the United States. Last year, tech giants spent roughly $580 billion turning empty fields, deserts and abandoned factories into “mini cities” packed with GPUs, cooling systems, fiber networks and dedicated substations.
Now, as 2026 kicks off, the race is shifting from bloated, multi-billion deals and rapid scaling toward a sharper focus on execution — building “inference factories” closer to users and within real-world constraints that finally kick the training wheels off. What happens next will shape where AI actually lives, how fast it responds and who ultimately benefits.
How Data Centers Defined 2025
First, let’s take a look back at how data centers became the main character of the AI boom, and how it all played out.
Data Centers Powered the the AI Boom
America’s AI boom is being built one multi-billion-dollar deal at a time. Giants like Amazon, Microsoft, Google and Meta are leading the charge, and collectively poured an estimated $320 billion into AI infrastructure in 2025 alone.
New sites like Amazon’s $23 billion Ohio expansion and Microsoft’s so-called “world’s most powerful” data center in Wisconsin are set to go live in 2026, while older facilities are retrofitted at similar breakneck speed — as seen with Elon Musk’s 100,000-GPU supercomputer Colossus, which was built in 122 days inside a shuttered Electrolux factory. Perhaps the most ambitious is the Stargate Project, a $500 billion, multi-year plan to sprawl across multiple states, intended to deliver up to 10 gigawatts of AI-ready power, equivalent to the total power consumption of New York City and San Diego.
Since 2022, spending on AI infrastructure has tripled, with Meta and OpenAI pledging hundreds of billions more over the next decade. This once-in-a-generation “investment supercycle” is projected to require $3 trillion in global capital by 2030, a shift so profound that U.S. spending on data center construction has officially surpassed that of traditional office buildings for the first time in history.
OpenAI’s Multi-Pronged Expansion
In 2025, OpenAI shifted from a single-cloud dependency to a multi-provider infrastructure strategy, striking large-scale deals with AWS, Oracle, Google, Microsoft and GPU-focused neoclouds like CoreWeave to secure scarce compute capacity. Its $38 billion AWS agreement and $300 billion Oracle deal lock in access to hundreds of thousands of GPUs and gigawatts of power, signaling that long-term capacity — not just price — is now the primary constraint in AI growth.
By renegotiating exclusivity with Microsoft and spreading workloads across providers, OpenAI reduces outage risk and bottlenecks while increasing leverage in an increasingly supply-constrained data center market. Meanwhile, the company’s joint data center projects with SoftBank, the UAE and major chipmakers inherently shape where, how and at what scale new facilities get built.
Circular Financing and the ‘AI Bubble’
As these deals unfolded over the past year, they revealed themselves to be part of a self-reinforcing web that connected the same handful of companies. This funding method, known as circular financing, works like a loop: investors pour capital into startups, which then use that money to buy the investors’ own chips, cloud services or general data-center infrastructure, effectively recycling the funds while locking in guaranteed revenue through these intermingled partnerships.
OpenAI, Microsoft, Nvidia, Oracle, AMD and SoftBank put the circular financing model in action throughout 2025. Microsoft, for instance, invested billions in OpenAI, which then funneled much of that money back into Microsoft’s Azure cloud, creating immediate demand for the very services funding the investment. Similarly, OpenAI contracted Oracle to build out the Stargate Project data centers, while also using AMD and Nvidia GPUs purchased with investor-backed capital. These arrangements compressed months — or even years — of fundraising and procurement into a matter of weeks, accelerating growth while guaranteeing revenue for the companies supplying the hardware, cloud capacity and infrastructure.
At the same time, circular financing’s inherent structure tends to blur real market demand. When suppliers invest heavily in startups that then spend that money on their own products, it becomes harder to tell genuine market growth from inflated, manufactured illusion. Analysts point out that when a company’s customers are also its investors, or when revenue rises faster than actual usage, these are clear indicators of a bubble forming. And so far, all the classic signs — steep valuations, overbuilt infrastructure, speculative investment and artificially engineered growth — are all paying out in the “AI bubble.”
AI Data Centers’ Energy Consumption Became a Top Issue
AI data centers pushed the country’s power grid to its limit in 2025. The average facility already uses as much electricity as 100,000 homes, with the largest new sites expected to consume 20 times that, according to the International Energy Agency.
In 2023, data centers accounted for more than four percent of total U.S. electricity use. That number is likely to triple by 2028, according to the Department of Energy. At the local level, residents near major data center hubs are already paying the price, with dried up taps and compromised wells as utility bills climb.
These facilities are massive water consumers, too. Cooling a single AI server can use up to two liters of water per kilowatt-hour, and large facilities may evaporate up to half a million gallons of drinking water daily. One chatbot query alone can consume ten times the electricity of a standard web search, while rapid hardware turnover — new models replacing old ones every few weeks — wastes the energy of prior training.
The environmental stakes are staggering. By 2030, Cornell researchers estimate that AI data centers could release anywhere from 24 to 44 million metric tons of CO₂ annually, equal to adding up to 10 million cars to U.S. roads, while consuming 6 to 10 million Americans’ households worth of water. Let’s not forget the additional environmental cost that comes with making and transporting GPUs and other AI infrastructure essentials to data center sites, including rare earth mining, toxic chemical use and e-waste containing mercury and lead.
At this rate, it’s unlikely that the major names in AI will meet their 2030 net-zero pledges. Even the best-case scenarios project massive water use and carbon emissions comparable to those of entire countries.
Experts Called to Diversify Energy Sources
Despite climate pledges, fossil fuels still power most modern data centers. In fact, more than 40 percent of U.S. data centers run on natural gas, compared with 24 percent on renewables, 20 percent on nuclear and 15 percent on coal. To secure reliable, 24/7 electricity, tech companies are increasingly turning to nuclear, including plans to reopen shuttered plants like Three Mile Island by 2028 following a $1.6 billion overhaul, and a chain of long-term deals that give Meta access to more than six gigawatts of nuclear-powered capacity.
At the same time, companies are layering in faster-to-build alternatives such as large-scale, 100-megawatt solar projects, as well as experimental ventures into geothermal, hydrogen and fusion, like Microsoft’s plan to build its fusion portfolio up to 50 megawatts by 2028. The result is a rapid, high-stakes reshaping of the U.S. energy landscape, driven less by climate idealism than by the sheer electricity demands of AI.
NeoClouds Carved Out Their Own Corner in the Market
Neocloud providers are a new breed of cloud companies built specifically to handle AI’s massive computational demands, specifically when it comes to training large models. Unlike legacy clouds like AWS, Azure or Google Cloud, these startups narrowly focus on raw compute, securing capacity faster and at lower cost.
Tech giants poured tens of billions into neocloud partnerships last year. Deals like Anthropic’s $50 billion Fluidstack agreement and Meta’s $14 billion CoreWeave contract highlight the massive demand for these neo-newcomers, specialized to help companies avoid supply bottlenecks, save money and efficiently scale GPU-heavy workloads. Currently, neoclouds are serving as a complement rather than a replacement for traditional hyperscalers, but are actively reshaping how AI infrastructure is built and deployed worldwide.
What’s to Come for Data Centers in 2026 and Beyond
This year will likely be a turning point for data centers, where speed, efficiency and proximity to users matter just as much as sheer scale. At large, the industry is grappling with new technical limits, supply chain pressures and shifting energy demands. How top contenders navigate these challenges will define who leads the next wave, and shape data center growth for years to come.
Below are some anticipated data-center trends expected to unfold in the upcoming year.
The Shift From Training Campuses to ‘Inference Factories’
2026 marks a pivot from the initial phase of building massive AI training clusters to the next, deploying infrastructure for real-time AI use, or inference. Inference needs to be close to users, creating a demand for local, modular “micro-data centers” that strictly fulfill edge computing needs. Specialized chips for mass AI compute are overwhelmingly replacing one-size-fits-all GPUs, prioritizing throughput and energy efficiency over raw training power. Companies like Nvidia and Qualcomm are rolling out inference-focused accelerators suited for these distributed networks. The next natural move for the industry is shifting toward a distributed network of inference nodes, putting intelligence closer to the devices and people that use it.
Data Centers Will Be Limited By How Much Power They Can Generate
Power scarcity has shifted from a temporary bottleneck to the primary constraint limiting AI and hyperscale growth. Utilities across the U.S. and Europe are revising load forecasts upward, with data centers accounting for a disproportionate share of projected demand growth. In response, grid interconnection timelines are stretching into the late 2020s in some mature hubs, effectively capping near-term expansion regardless of land availability or capital backing.
Operators are responding by taking matters into their own hands. New projects increasingly include on-site or nearby power generation — gas turbines, fuel cells, large-scale batteries and hybrid microgrids — to get around years-long delays. Even nuclear is moving from speculation to reality, with long-term agreements on existing plants and new small modular reactors explicitly pitched as “AI-ready” baseload.
Soon, the most competitive data center campuses will feel more like energy hubs than simple server farms. Dedicated generation, dispatch rights and carbon performance will be baked into financing and customer contracts. Access to literal power, not land or demand, will become the deciding factor in where and how these facilities are being built.
Policy Will Determine the Speed and Viability of Projects
Regulation has shifted from a yes-or-no hurdle to an operational variable that dictates timelines, cost and risk. In many markets, it’s common knowledge that permits, environmental reviews and grid access can stretch projects from months to years, forcing developers to plan around policy as much as land or demand. Utilities are now starting to shift costs back onto the centers themselves rather than dumping it on local communities.
For example, Ohio’s American Electric Power now requires large data centers to pay for significant portions of grid upgrades upfront, a move that could slow or discourage future builds. Simultaneously, the federal government is trying to shorten the time it takes regulators to review data center grid connections from years to roughly 60 days. As red tape catches up with private AI infrastructure expansion, a given project’s success will soon depend as much on policy alignment as it does on market demand.
Liquid Cooling as the Standard Design
We have officially reached the thermal limit of air cooling. In other words, air can no longer move heat away from a chip as fast as modern chips generate it. Rack densities have leapt from a legacy average of 15 kilowatts to more than 40 kilowatts. When training, that rate can even jump to 100 kilowatts, moving liquid cooling from a somewhat niche feature to a non-negotiable norm.
The dominant cooling architecture so far are direct-to-chip (D2C) systems, where coolant is piped directly onto the cold plates on GPUs. Another method called immersion cooling, which submerges servers in thermally conductive fluid, is becoming a favorite for high-density edge nodes where space is extremely tight and noise needs to be minimized.
Construction Delays Are Nigh
The data center supply chain has become a front‑and‑center bottleneck for new builds. Even when projects have their permits and financing lined up, long queues for critical infrastructure like transformers, switchgear, cabling and prefab electrical systems are now measured in years, not months, slowing construction well past planned deadlines.
Amazon has publicly noted transformer shortages delaying hyperscale data center builds in Virginia and Ohio, underscoring how grid equipment availability can hold up massive facilities. For reference, the average lead time for small-scale transformers hit roughly 120 weeks in 2024, with large power transformers taking as long as 210 weeks. Oracle has also reportedly delayed some of its data center developments by as much as a year due to labor and material shortages (not to mention the handful of Amazon’s sites have been disrupted by public backlash that has escalated into legal action).
This type of demand is now industrializing delivery models. To cope with unprecedented demand, top operators are treating data center construction like a factory line, standardizing designs into repeatable, plug-and-play units, locking in factory output years in advance and bringing energy and equipment procurement in-house to avoid getting stuck at the back of the queue. Given the rapid pace of the AI race, the ability to execute on schedule may be as important as having the land or capital to build at all in the upcoming year.
Frequently Asked Questions
How will data centers shift in 2026?
Data centers will shift from headline-grabbing scale to disciplined execution. In 2026, the focus moves away from who can build the biggest campus and toward who can actually get projects online, with reliable power, realistic timelines and workable designs. This may look like smaller, inference-focused facilities closer to users, built-in energy sites, supply-chain delays and regulation increasingly shape what gets built — and where.
What is a data center?
A data center is a physical facility that houses servers, storage systems and networking equipment used to store, process, and deliver digital data. It provides the power, cooling and security needed to keep online services and applications running reliably, particularly those involving artificial intelligence.
Can data centers ever become obsolete?
While the specific hardware and infrastructure within them may become outdated, data centers themselves are highly unlikely to become obsolete. In fact, they are growing in popularity and scale to support advancements in artificial intelligence.
