The Infrastructure Super Intelligence Will Need

The Change Everyone Expects

Nobody in the data center business is surprised by change anymore. Processors arrive faster than buildings can be designed around them. Models grow. Racks get denser.

What should surprise people is where the hardest change is happening. The public conversation about artificial intelligence is a conversation about chips: who makes them, who can buy them, how fast the next generation will be. Yet in November 2025 Microsoft’s chief executive described a different problem. “It’s not a supply issue of chips; it’s the fact that I don’t have warm shells to plug into,” he said, using the real estate term for buildings ready for tenants (TechCrunch). Weeks earlier, OpenAI’s chief executive had written that building a gigawatt of new AI infrastructure every week “will require innovation at every level of the stack, from chips to power to building to robotics” (Abundant Intelligence).

Power. Building. That is facility infrastructure: the power delivery and energy supply, the backup systems, the cooling and heat rejection, and the building and controls that carry the computers. It is the least glamorous part of a data center. Today it is also the most critical part of the industry’s innovation, and the least told. When the facility makes the news, it arrives as a problem: a strained grid, a rising bill, a contested hearing. It is rarely covered as the place where the decisive engineering of the AI era is being done.

The Trump administration now calls the technology Super Intelligence. Whatever name it carries, it will run in buildings, and those buildings will need infrastructure that does not yet exist at scale. To see what that infrastructure must become, start with how far it has already come.

2005 to 2025: The Building Learned to Disappear

In 2005 a data center spent about as much energy keeping its servers powered and cooled as it spent running them. The federal government’s 2007 report to Congress assumed a facility used two units of energy for every unit its computers used, and found that the power and cooling infrastructure accounted for 50 percent of total data center consumption. U.S. servers and data centers used about 61 billion kilowatt-hours in 2006, 1.5 percent of the nation’s electricity, and the report projected more than 100 billion by 2011 if trends held (U.S. Environmental Protection Agency).

Then the industry went to work on the building. It adopted a common yardstick, power usage effectiveness (PUE), which divides a facility’s total energy by the energy its computers use, and the largest operators turned that yardstick into a public contest. When Facebook opened its Prineville, Oregon data center in 2011, it reported an initial power usage effectiveness of 1.07, against an average of 1.5 for its existing facilities, using 480-volt distribution and no central uninterruptible power supply, and it released the designs to the industry as the Open Compute Project (Meta Engineering). The building had no chillers (The Register). Industry thermal guidelines widened the range of conditions servers were allowed to run in, which let operators cool with outside air. By 2016 machine learning was running the cooling plant, cutting the energy Google used for cooling by up to 40 percent (Google).

It worked. U.S. data center electricity use rose by only about 4 percent from 2010 to 2014, to about 70 billion kilowatt-hours, about 1.8 percent of the nation’s electricity (Lawrence Berkeley National Laboratory). Efficiency had beaten the forecast. At the best-run fleets, the overhead that once equaled the computing load fell below a tenth of it. Google’s company-wide power usage effectiveness reached 1.09 in 2024, “only 0.02 better than a decade ago” (TechCrunch). The building had learned to disappear. It became so efficient that almost no one outside the industry thought about it.

Then the curve turned. Average rack density doubled, from 6.1 kilowatts in AFCOM’s first State of the Data Center survey to 12 kilowatts in its 2024 report (Data Center Knowledge), while racks built for AI went far beyond either figure. U.S. data center electricity climbed from 1.9 percent of national consumption in 2018 to 4.4 percent in 2023, with 6.7 to 12 percent projected for 2028 (Lawrence Berkeley National Laboratory). In September 2024, Microsoft signed a 20-year power purchase agreement that cleared the way to restart Three Mile Island Unit 1 as the Crane Clean Energy Center (Constellation). A software company was contracting for an atomic plant’s output to keep its buildings running. The facility was back.

2025 to 2026: The Years Everything Changed

In about eighteen months the industry redesigned how power and cooling reach a rack of computers. Each change was a facility change.

The rack headed for a megawatt. In late 2024 a reference design for NVIDIA’s GB200 NVL72 platform supported up to 132 kilowatts per rack (Business Wire), more than ten times the industry’s average rack. In May 2025 NVIDIA said it was leading a transition to new power infrastructure “to support 1 MW IT racks and beyond, starting in 2027” (NVIDIA). Google wrote that machine learning “will require more than 500 kW per IT rack before 2030” (Google Cloud). By October 2026 the Open Compute Project reported “a clear path to 1MW racks in power and cooling” (Open Compute Project).
The power chain was declared obsolete. For a decade, servers drew power from a low-voltage bus inside the rack. At a megawatt that stops working: “The physics of using 54 VDC in a single 1 MW rack requires up to 200 kg of copper busbar” (NVIDIA). The answer is direct current at far higher voltage, 800 volts or plus and minus 400 volts, delivered into the rack. By August 2026 Google, Microsoft and NVIDIA reported that “More than 80 partners are now developing 800 VDC-compatible infrastructure” (Open Compute Project). Energy storage moved into the power path too. Thousands of processors training a model in lockstep create power swings the grid can feel, and storage built into the newest racks cuts the peak demand the grid sees by 30 percent in one published test (NVIDIA).
Liquid cooling became mandatory. The newest AI racks are cooled entirely by liquid, with no fans in the system, and their processors hold full performance “even with coolant entering the rack at 45 degrees Celsius” (NVIDIA). Water that warm can shed its heat to outside air through dry coolers, without chillers, in favorable climates, which NVIDIA says sharply reduces the water a facility uses for cooling. Microsoft went further, testing a microfluidic system that removed heat up to three times better than cold plates, and warning that “the current cooling technology will put a ceiling on progress in just a few years” (Microsoft).
The building got its own specification. In 2026 AMD, Google, Meta and NVIDIA published, through the Open Compute Project, design principles for the AI data center building itself. They set a design target of 8,000 pounds for the heaviest single rack and require an 18-foot minimum ceiling height for rows fed by two busways (Open Compute Project). The companies that make the chips and run the clouds are now writing floor loads and ceiling heights.
The facility became a grid actor. On July 10, 2024, a single transmission fault led about 1,500 megawatts of voltage-sensitive load to drop off the grid at once, and about 1,260 megawatts did not return for hours (North American Electric Reliability Corporation). In May 2026 the grid’s reliability organization issued an essential-action alert on the risks computational loads pose to the bulk power system (North American Electric Reliability Corporation), and in July 2026 the Federal Energy Regulatory Commission directed it to write reliability standards for those loads by the end of the year (Federal Energy Regulatory Commission). The largest operators answered with rules of their own: in September 2026 Google, Meta, Microsoft and NVIDIA published an open specification committing their AI computing load to ride through grid disturbances and to “restore to 90% of pre-fault load within 2 seconds” (Open Compute Project). The power electronics inside a data center now matter to the stability of a regional grid.

2027 to 2050: More Innovation Is Required

Forecasts of data center demand keep rising. EPRI projects that data centers will use 9 to 17 percent of U.S. electricity by 2030, up from 4 to 5 percent today (EPRI). The International Energy Agency sees global data center use roughly doubling, from 485 terawatt-hours in 2025 to 950 in 2030 (International Energy Agency). The frontier is moving to campuses measured in gigawatts: Anthropic has written that “the U.S. AI sector needs at least 50GW of electric capacity by 2028 to maintain global AI leadership” (Anthropic).

In September 2026 the Trump administration began calling the technology Super Intelligence (The White House). Its plan for the infrastructure was set out the year before, under a heading that could serve as this industry’s brief: “Develop a Grid to Match the Pace of AI Innovation” (The White House). Three questions decide whether the facilities can keep that pace.

How data centers will be powered

The industry’s own design forum now names the constraint: “Energy availability, not land, fiber, or hardware supply, is now the primary constraint on new infrastructure deployments,” and “on-site and standby generation is becoming core infrastructure, not just a backup measure” (Open Compute Project).

The answer is a layered supply. Most data centers still prefer a grid connection (International Energy Agency), but the grid cannot deliver everywhere in time. Where it cannot, on-site natural gas is the industry’s bridge. The International Energy Agency estimates that 15 to 27 gigawatts of on-site gas may power data centers by 2030, mostly in the United States, and that reliable on-site supply requires overbuilding generation by 30 to 70 percent relative to demand (International Energy Agency). That overbuild is the hidden cost of leaving the grid: a campus that supplies itself must carry the reserve a utility once carried for it.

The Open Compute Project, which launched an Energy Project for generation and storage at its October 2026 Summit, says that “understanding the best alternatives for on-site energy generation is still an unsolved problem” (Open Compute Project). Even backup power is being reworked: Microsoft is “developing open requirements for reliable backup systems that reduce reliance on diesel” (Microsoft).

Advanced atomic generation is the next layer. The International Energy Agency expects the first small modular reactors to come online around 2030 (International Energy Agency), and in September 2026 the Nuclear Regulatory Commission approved construction of what it called the nation’s first commercial small modular reactor, at the Tennessee Valley Authority’s Clinch River site (Nuclear Regulatory Commission).

How data centers will be cooled

Cooling is moving inward and getting warmer. Room air gave way to cold plates on the chip, and the next step puts coolant inside the chip itself, which Microsoft has already tested (Microsoft). Warmer coolant lets a facility reject heat to outside air across more of the year, cutting the energy and the water that chillers and cooling towers consume. How heat leaves the building, and how little water it takes to move it, will separate the facilities that scale from the ones that stall. Water is already one of the first questions communities ask.

What facility infrastructure must do for Super Intelligence

Every constraint on the next twenty-five years points to an innovation the facility must deliver.

Equipment scarcity points to new power electronics. When a transformer takes three years to arrive (Reuters), the industry needs conversion equipment that can be built faster and consolidated into fewer, smarter units.
Labor scarcity points to the factory. The construction industry needs an estimated 349,000 net new workers in 2026 alone, and about one-fifth of all electricians are over 55 (Associated Builders and Contractors). Power and cooling systems built as modules in factories, then set in place on site, stretch a scarce workforce further.
Grid reliability points to facilities that ride through faults. In April 2026 a fault much like the one of 2024 led to only 500 megawatts of sustained data center load loss, an improvement the reliability organization said “may be attributable to the changes made by the data centers.” The same report recorded a load reduction of nearly 4 gigawatts in July 2026 (North American Electric Reliability Corporation). The risk is growing.
Security points to hardened campuses. The national plan says the data centers where AI works with some of the government’s most sensitive data “must be resistant to attacks by the most determined and capable nation-state actors” (The White House). That is a facility requirement before it is anything else.

The ideas that draw the headlines are further out. Google plans to launch two prototype satellites “by early 2027” to explore AI computing in orbit (Google Research). Even there, the hard problems are facility problems: how to power the computers and how to reject their heat. On the ground, the decisive work goes on largely unnoticed: direct current to the rack, warm water at the chip, firm power at the fence, and buildings that hold the grid steady.

How Island Roadhouse Is Innovating

Island Roadhouse Data Centers designs the facility as the product. Our customers bring their own compute; we provide long-term, power-backed data center capacity designed for where AI is going.

Power that does not wait on the grid. The campus is designed so that campus generation carries the compute load.
Halls designed for direct current. Our GPU and HPC halls are 800 VDC-native at the customer’s rack: each of their electrical blocks distributes 800 VDC to the rack from two independent DC power blocks, with a centralized DC-coupled energy store for ride-through. Every hall is planned with 2N electrical distribution to the rack, fed by a 2N-capable generation block. The halls are planned on a slab, with overhead distribution and no raised floor.
Cooling at the chip, and low water. Our GPU and HPC halls are designed for closed-loop warm-water direct-to-chip liquid cooling, managed by our proprietary thermal management for optimal energy use, on a campus designed to be low-water.
Space ready for accreditation. Our initial Missouri campus plan includes a Secure Mission Hall: a GPU and HPC hall planned as an ICD 705 accreditation-ready shell, designed so that customer-sponsored accreditation can follow, and sharing the campus’s high-density design rather than standing apart as a one-off build.

None of this is a single invention. It is the facility designed as one system, with power, cooling and the building engineered together from the site up.

The Reality

Change in data center technology is expected. The change that will decide how fast Super Intelligence arrives is happening in the facility. It means power delivered as direct current to the rack. It means cooling that reaches the silicon. It means generation at the fence and buildings that hold the grid steady.

At Island Roadhouse Data Centers, we believe the infrastructure Super Intelligence will need is facility infrastructure, and that the companies willing to innovate there will set the pace for everyone else. That work rarely makes the news. It will decide the future anyway.