The timeless thirst for smarter (traditionally, which means bigger) AI models and better adoption of those we have already got has led to an explosion in data-center construction projects, unparalleled each in quantity and scale. Chief amongst them is Meta’s deliberate 5-gigawatt knowledge middle in Louisiana, referred to as Hyperion, introduced in June of 2025. Meta CEO Mark Zuckerberg mentioned Hyperion will “cowl a big a part of the footprint of Manhattan,” and the primary part—a 2-GW model—will likely be accomplished by 2030.
Although the undertaking’s acknowledged 5-GW scale is the biggest amongst its friends, it’s simply one in every of a number of dozen related initiatives now underway. Based on Michael Guckes, chief economist at construction-software firm ConstructConnect, spending on data centers topped US $27 billion by July of 2025 and, as soon as the full-year figures are tallied, will simply exceed $40 billion. Hyperion alone accounts for a couple of quarter of that.
For the engineers assigned to deliver these initiatives to life, the combo of challenges concerned signify a novel second. The world’s largest tech corporations are opening their wallets to pay for brand spanking new improvements in compute, cooling, and network expertise designed to function at a scale that may’ve appeared absurd 5 years in the past.
On the identical time, the breakneck tempo of constructing comes paired with severe issues. Fashionable data-center development ceaselessly requires an inflow of non permanent staff and sharply will increase noise, visitors, pollution, and infrequently native electricity prices. And the environmental toll stays a priority lengthy after amenities are constructed because of the unprecedented 24/7 power calls for of AI knowledge facilities which, in keeping with one latest examine, could emit the equivalent of tens of millions of tonnes of CO2 annually within the United States alone.
No matter these points, massive AI corporations, and the engineers they rent, are going full steam forward on large data-center development. So, what does it actually take to construct an unprecedentedly massive knowledge middle?
AI Rewrites Constructing Design
The stereotypical data-center constructing rests on a bolstered concrete slab basis. That’s paired with a metal skeleton and poured concrete wall panels. The completed constructing is named a “shell,” a time period that suggests the construction itself is a secondary concern. Meta has even used gigantic tents to throw up non permanent knowledge facilities.
Nonetheless, the dimensions of the biggest AI knowledge facilities brings distinctive challenges. “The largest problem is usually what’s below the floor. Unstable, corrosive, or expansive soils can result in delays and require severe intervention,” says Robert Haley, vice chairman at development consulting agency Jacobs. Amanda Carter, a senior technical lead at Stantec, mentioned a soil’s thermal conductivity can be vital, as {most electrical} infrastructure is positioned underground. “If the soil has excessive thermal resistivity, it’s going to be troublesome to dissipate [heat].” Engineers could take tons of or hundreds of soil samples earlier than development can start.
There’s apparently no scarcity of eligible websites, nonetheless, as each the variety of knowledge facilities below development, and the cash spent on them, has skyrocketed. The spending has allowed corporations constructing knowledge facilities to throw out the rule e book. Previous to the AI growth, most knowledge facilities relied on tried-and-true designs that prioritized cheap and environment friendly development. Large tech’s willingness to spend has shifted the main focus to hurry and scale.
The unfastened purse strings open the door to bigger and extra strong prefabricated concrete wall and flooring panels. Doug Bevier, director of improvement at Clark Pacific, says some concrete flooring panels could now span as much as 23 meters and have to deal with flooring hundreds as much as 3,000 kilograms per sq. meter, which is more than twice the load international building codes normally define for manufacturing and industry. In some circumstances, the concrete panels have to be custom-made for a undertaking, an costly step that the economics of pre-AI knowledge facilities not often justified.
Concurrently, the time scale for initiatives can be compressed: Jamie McGrath, senior vice chairman of data-center operations at Crusoe, says the corporate is delivering initiatives in “about 12 months,” in comparison with 30 to 36 months earlier than. Not all initiatives are continuing at that tempo, however pace is universally a precedence.
That makes it troublesome to coordinate the labor and supplies required. Meta’s Hyperion website, situated in rural Richland Parish, Louisiana, is emblematic of this problem. As reported by NOLA.com, at the very least 5,000 non permanent staff have flocked to the world, which has solely about 20,000 everlasting residents. These workers earn above-average wages and produce a short-term enhance for some native companies, equivalent to eating places and comfort shops. Nevertheless, they’ve additionally spurred complaints from residents about visitors and development noise and air pollution.
This friction with residents consists of not solely these apparent impacts, however also things you might not immediately suspect, equivalent to mild air pollution attributable to around-the-clock schedules. Additionally vital are adjustments to native water tables and runoff, which may cut back water high quality for neighbors who depend on nicely water. These points have motivated just a few U.S. cities to enact data-center bans.
Knowledge Facilities Usually Go BYOP (deliver your personal energy)
Meta’s Richland Parish website additionally highlights an issue that’s precedence No. 1 for each AI knowledge facilities and their critics: energy.
Knowledge facilities have all the time drawn massive quantities of energy, which nudged data-center development to cluster in hubs the place native utilities have been attentive to their calls for. Virginia’s electric utility, Dominion Power, met demand with agreements to construct new infrastructure, often with a focus on renewable energy.
The facility calls for of the biggest AI knowledge facilities, although, have caught even probably the most responsive utilities off guard. A report from the Lawrence Berkeley National Laboratory, in California, estimated your complete U.S. data-center business consumed an average load of roughly 8 GW of power in 2014. Right this moment, the biggest AI data-center campuses are constructed to deal with as much as a gigawatt every, and Meta’s Hyperion is projected to require 5 GW.
“Knowledge facilities are exasperating points for lots of utilities,” says Abbe Ramanan, undertaking director on the Clean Energy Group, a Vermont-based nonprofit.
Ramanan explains that utilities typically use “peaker vegetation” to deal with additional demand. They’re often older, much less environment friendly fossil-fuel vegetation which, due to their excessive price to function and carbon output, have been due for retirement. However Ramanan says elevated electrical energy demand has kept them in service.
Meta secured energy for Hyperion by negotiating with Entergy, Louisiana’s electrical utility, for development of three new gas-turbine power plants. Two will likely be situated close to the Richland Parish website, whereas a 3rd will likely be situated in southeast Louisiana.
Entergy frames the brand new vegetation as a win for the state. “A core pillar of Entergy and Meta’s settlement is that Meta pays for the complete price of the utility infrastructure,” says Daniel Kline, director of power-delivery planning and coverage at Entergy. The utility expects that “buyer payments will likely be decrease than they in any other case would have been.” That will show an exception, as a recent report from Bloomberg found electrical energy charges in areas with knowledge facilities usually tend to improve than in areas with out.
The vegetation, which can generate a mixed 2.26 GW, will use combined-cycle gas turbines that recapture waste heat from exhaust. This boosts thermal efficiency to 60 percent and beyond, that means extra gas is transformed to helpful power. Easy-cycle generators, against this, vent the exhaust, which lowers effectivity to round 40 p.c.
Even so, whole life-cycle emissions for the Hyperion vegetation might vary from 4 million to over 10 million tonnes of CO2 every year, relying on how ceaselessly the vegetation are put in use and the ultimate effectivity benchmarks as soon as constructed. On the excessive finish, that’s as a lot CO2 as produced by over 2 million passenger automobiles. Luckily, not all of Meta’s knowledge facilities take the identical strategy to energy. The corporate has introduced a plan to energy Prometheus, a big data-center undertaking in Ohio scheduled to come back on-line earlier than the tip of 2026, with nuclear energy.
However different big tech corporations, spurred by the necessity to construct knowledge facilities rapidly, are taking a much less environment friendly strategy.
xAI’s Colossus 2, situated in Memphis, is probably the most excessive instance. The company trucked dozens of temporary gas-turbine generators to power the site situated in a suburban neighborhood. OpenAI, in the meantime, has fuel generators able to producing as much as 300 megawatts at its new Stargate data center in Abilene, Texas, slated to open later in 2026. Each use simple-cycle generators with a a lot decrease effectivity score than the combined-cycle vegetation Entergy will construct to energy Hyperion.
Demand for fuel generators is so intense, the truth is, that wait times for new turbines are up to seven years. Some knowledge facilities are turning toward refurbished jet engines to acquire the generators they want.
AI Racks Tip the Scales
The demand for brand spanking new, dependable energy is pushed by the power-hungry GPUs inside fashionable AI knowledge facilities.
In January of 2025, Mark Zuckerberg introduced in a submit on Facebook that Meta deliberate to finish 2025 with at least 1.3 million GPUs in service. OpenAI’s Stargate knowledge middle plans to use over 450,000 Nvidia GB200 GPUs, and xAI’s Colossus 2, an growth of Colossus, is built to accommodate over 550,000 GPUs.
GPUs, which stay by far the most well-liked for AI workloads, are bundled into human-scale monoliths of metal and silicon which, very similar to the info facilities constructed to deal with them, are quickly rising in weight, complexity, and energy consumption.
Nvidia’s GB200 NVL72—a rack-scale system—is presently a number one selection for AI knowledge facilities. A single GB200 rack accommodates 72 GPUs, 36 CPUs, and as much as 17 terabytes of reminiscence. It measures 2.2 meters tall, tips the scales at up to 1,553 kilograms, and consumes about 120 kilowatts—as a lot as round 100 U.S. properties. And this, in keeping with Nvidia, is only the start. The corporate anticipates future racks might consume up to a megawatt each.
Viktor Petik, senior vice chairman of infrastructure options at Vertiv, says the fast change in rack-scale AI methods has compelled knowledge facilities to adapt. “AI racks devour way more energy and weigh greater than their predecessors,” says Petik. He provides that knowledge facilities should provide racks with a number of energy feeds, with out taking over additional area.
The brand new energy calls for from rack-scale methods have penalties which are mirrored within the design of the info middle—even its footprint.
In 2022 Meta broke floor on a brand new knowledge middle at a campus in Temple, Texas. Based on SemiAnalysis, which research AI knowledge facilities, development started with the intent to build the data center in an H-shaped configuration common to other Meta data centers.
Building was paused halfway in December of 2022, nonetheless, as part of a company-wide review of its data-center infrastructure. Meta determined to knock down the construction it had constructed and begin from scratch. The explanations for this determination have been by no means made public, however analysts consider it was because of the outdated design’s incapability to ship adequate electrical energy to new, power-hungry AI racks. Building resumed in 2023.
Meta’s substitute ditches the H-shaped constructing for easy, lengthy, rectangular buildings, every flanked by rows of gas-turbine turbines. Whereas Meta’s plans are topic to alter, Hyperion is presently anticipated to comprise 11 rectangular knowledge facilities, every filled with tons of of hundreds of GPUs, unfold throughout the 13.6-square-kilometer Richland Parish campus.
Cooling, and Connecting, at Scale
Nvidia’s ultradense AI GPU racks are altering knowledge facilities not solely with their weight, and energy draw, but in addition with their intense cooling and bandwidth necessities.
Knowledge facilities historically use air cooling, however that strategy has reached its limits. “Air as a cooling medium is inherently inferior,” says Poh Seng Lee, head of CoolestLAB, a cooling analysis group on the Nationwide College of Singapore.
As a substitute, going ahead, GPUs will depend on liquid cooling. Nevertheless, that provides a brand new layer of complexity. “It’s all the best way to the amenities stage,” says Lee. “You want pumps, which we name a coolant distribution unit. The CDU will likely be related to racks utilizing an elaborate piping community. And it must be designed for redundancy.” On the rack, pipes hook up with chilly plates mounted atop each GPU; outdoors the data-center shell, pipes route by evaporation cooling models. Lee says retrofitting an air-cooled knowledge middle is feasible however costly.
The networking utilized by AI knowledge facilities can be altering to deal with new necessities. Conventional knowledge facilities have been positioned close to community hubs for simple entry to the worldwide internet. AI knowledge facilities, although, are extra involved with networks of GPUs.
These connections should maintain excessive bandwidth with impeccable reliability. Mark Bieberich, a vice chairman at community infrastructure firm Ciena, says its newest fiber-optic transceiver expertise, WaveLogic 6, can present as much as 1.6 terabytes per second of bandwidth per wavelength. A single fiber can help 48 wavelengths in whole, and Ciena’s largest prospects have tons of of fiber pairs, inserting whole bandwidth within the hundreds of terabits per second.
Meta’s Hyperion knowledge middle is below development in Richland Parish, La., on a sprawling website a couple of quarter the world of Manhattan.
Meta
This can be a level the place the dimensions of Meta’s Hyperion, and different massive AI knowledge facilities, could be misleading. It appears to suggest the bodily dimension of a single knowledge middle is what issues. However fairly than being a single constructing, Hyperion is actually a set of buildings related by high-speed fiber-optics.
“Interconnecting knowledge facilities is completely important,” says Bieberich. “You possibly can give it some thought as one logical AI coaching facility, however with geographically distributed amenities.” Nvidia has taken to calling this “scale throughout,” to distinction it with the concept that knowledge facilities should “scale up” to bigger singular buildings.
The Large however Hazy Future
The total scale of the challenges that face Hyperion, and different future AI knowledge facilities of comparable scale, stay hazy. Nvidia has but to introduce the rack-scale AI GPU methods it’ll host. How a lot energy will it demand? What kind of cooling will it require? How a lot bandwidth have to be offered? These can solely be estimated.
Within the absence of particulars, the gravity of AI data-center design is pulled towards one certainty: It have to be huge. New data-center designers are rewriting their rule e book to deal with energy, cooling, and network infrastructure at a scale that may’ve appeared ridiculous 5 years in the past.
This innovation is fueled by huge tech’s fats pockets, which shelled out tens of billions of {dollars} in 2025 alone, resulting in questions about whether the spending is sustainable. For the engineers within the trenches of data-center design, although, it’s considered as a chance to make the unattainable doable.
“I inform my engineers, that is peak. We’re being engineers. We’re being requested difficult questions,” says Stantec’s Carter. “We haven’t acquired to try this in a very long time.”
This text seems within the April 2026 print problem.
From Your Web site Articles
Associated Articles Across the Net
