Power and cooling are now the hard limits on data centres, on Earth and in orbit

Two stories broke within a day of each other this week that look unrelated until you read the numbers. One is about a refrigerator-sized satellite. The other is about 1,500 miles of transmission lines. Both are about the same problem: data centres have outgrown the power and cooling systems that were supposed to serve them.

Start with the satellite. Google's first Project Suncatcher test craft, named MVP, is scheduled to launch on October 1 aboard a SpaceX Falcon 9 as part of the Transporter-18 rideshare, according to Ars Technica. Inside are four of Google's custom TPU AI accelerators. The solar panels supply about one kilowatt, which Ars notes is roughly enough to run a microwave or a hair dryer.

That is not a data centre. It is a single server, in orbit, with a physics problem attached.

The cooling problem, in space and on the ground

AI accelerators generate a lot of heat, and heat is much harder to get rid of in a vacuum. There is no air to blow across a heatsink. Radiator systems do exist in space, but as Ars Technica reports, they are built to remove relatively small amounts of heat compared with what AI silicon produces. Google's answer is a layer of malleable thermal interface material connecting the chips to aluminum and copper heat pipes. Those pipes conduct heat into a radiator that projects it into space.

Even with that setup, the four chips can only run for spurts of about 15 minutes before they have to shut down and let the radiators catch up. Tom's Hardware, citing The New York Times, reports the same figure and adds that the chips will need periodic restarts to reset bit flips caused by radiation. The satellite is expected to handle AI requests for up to a year and stay in orbit for up to six years before re-entering the atmosphere, though Ars says this first test will operate for just a few months.

Google is not pretending this is a product. James Manyika, the company's senior vice president of research, told the Times: "We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years." He compared the effort to Google's driverless car work, which the company tested "like 15 years before anything showed up."

Note what is missing from this picture: a cost argument. Launching a single server into orbit to save on terrestrial cooling is not, on the numbers available, obviously cheaper than building the cooling on the ground. The point of the experiment is to find out what breaks.

On the ground, the grid is the bottleneck

The terrestrial version of the same problem showed up in a US Department of Energy announcement this week. The department said it would put $5.25 billion into grid improvements, split between $1.9 billion of federal funds and $3.35 billion in cost-share funding, across 31 projects in 26 states. The Register reported on 25 September that the program, run under the SPARK initiative with money from the GRIP program, is expected to unlock at least 23 gigawatts of additional capacity.

The mechanism matters. The projects are not adding generation. They are reconductoring or rebuilding more than 1,500 miles of transmission lines and deploying grid-enhancing technologies such as sensors and power flow control devices across 21,000 miles of infrastructure. In other words, squeezing more out of wires that already exist. US Secretary of Energy Chris Wright said in a statement that the investments "will get more out of the infrastructure we already have, move more electricity across the grid, and help deliver affordable, reliable, and secure power." The department claims the program will improve reliability and lower electricity costs for roughly 100 million Americans.

The Register is blunt that it may not be enough. It cites a warning from Moody's earlier in September that the pace of data centre construction exceeds the speed at which the US grid can add the required capacity. Moody's estimates US data centre electricity consumption will reach 426 TWh by 2030, nearly double the 2025 figure. That gap is why the industry has been spending on small modular reactors, geothermal generation and the recommissioning of ageing nuclear plants. It is also why chipmakers such as Nvidia are pushing for data centres that can hand capacity back to the grid when needed, according to The Register.

Why renewables alone do not close it

A common assumption is that clean energy solves the supply side. Data Center Knowledge laid out the limits of that argument in a piece that still reads as the clearest explanation of the constraint. Wind and solar output varies with weather, so a solar-powered facility needs an alternative source on cloudy days and a wind farm may not generate enough megawatts when it is calm. Geothermal runs independent of weather but needs specialised plants and a larger upfront investment; less than 0.5 percent of US electricity is generated that way.

Storage is the other half of the problem. Unless a site has massive battery arrays, excess electricity from wind, solar or geothermal cannot be retained. You use it or lose it. The practical consequence, as Data Center Knowledge puts it, is that powering a data centre purely with clean energy typically requires building sourcing facilities with a much higher total output capacity than the site needs at most times, or else supplementing with non-clean energy at peak.

None of that makes clean energy pointless. Sourcing even 20 percent of a facility's energy from renewables is a 20 percent cut in the carbon emissions tied to energy production. Combining that with efficiency measures such as liquid cooling and smart power management makes a measurable difference. It just does not make the grid optional.

The enforcement angle

When grid power is short, operators find workarounds, and sometimes those workarounds break the law. New Jersey this week ordered the operator of one of the East Coast's largest planned data centres to pay a $1.1 million fine for secretly installing and operating gas generators in violation of the state's Air Pollution Control Act, Ars Technica reported on 24 September.

DataOne was caught after an investigation by The Guardian and Floodlight News in August shared thermal drone footage showing 45 of 62 gas generators running. None had the permits required for generators with 37-kilowatt capacity or higher, despite running at 1,982-kw capacities, more than 50 times the state limit. New Jersey's Department of Environmental Protection noted that such generators emit carbon dioxide, nitrogen oxides, carbon monoxide and other combustion-related pollutants, which can worsen asthma, trigger heart attacks and cause early deaths.

DEP Commissioner Ed Potosnak called the action "by far the largest ever taken against a data center in New Jersey" and said data centres would not be "constructed or operated with impunity in this state." The company has 45 days to apply for permits or cease operations, but it can keep running the engines while it seeks them. A DataOne spokesperson told The New York Times the company would apply for the permits and "disagreed" with the fine, adding that it is transitioning to low-emission fuel cells.

Residents near the Vineland site are not satisfied. Nichole Gardner of the local nonprofit Sustain SJ told Ars the penalty is "insufficient, both in dollar amount and enforcement." Tiffany Leone-Vespa, who has lived in the neighbourhood for 18 years, told the Times: "What have we been breathing in since they've been running these generators without permits?"

What connects the two

The orbital experiment and the New Jersey fine sit at opposite ends of the same constraint. Google is testing whether it can escape terrestrial power and cooling limits by leaving the planet, and finding that heat is still heat and a kilowatt is still a kilowatt. DataOne is what happens when an operator decides the grid cannot wait and the permits can. The Energy Department is trying to buy time with reconductored wire. Manyika's estimate, as reported by Tom's Hardware, is that Google has about 15 years to go from a four-chip server to a full data centre. Moody's gives the US grid five years to nearly double the electricity it supplies to data centres. Those two clocks are not running at the same speed.