America’s AI Revolution Is Going to Need a Lot More Space

America’s AI Revolution Is Going to Need a Lot More Space

You’re reading Dispatch Markets, a weekly dive into the forces driving economic growth—and those holding it back—featuring Scott Lincicome, Kyla Scanlon, Karl Smith, Marian Tupy, and Adam Ozimek. On November 3, voters sent a resounding message to their elected officials. Even though Republicans had a bad night, the data center industry might have envied them. Voters backed limits on data centers or rejected proposals that would have made them easier to build in 52 of the 53 local contests Ballotpedia tracked. Worse, the sole win comes with an asterisk: In Harwood, North Dakota, voters rejected an amendment that would have required extra review for manufacturing, warehousing, and energy projects, as well as data centers. Measures solely focused on data centers went worse for AI proponents. In Sunbury, Ohio, 79 percent of voters backed a ban on data centers above 25 megawatts in capacity. In Edgewater, Florida, 74 percent voted for a ban on new data centers altogether. According to an estimate from Lawrence Berkeley National Laboratory, data centers accounted for 4.7 percent of U.S. electricity use in 2024. By 2030, they could account for nearly 12 percent. It’s reasonable to ask who will pay for all that power. In 2024, Virginia’s legislative auditors found that data centers were paying the utility costs allocated to them, but that providing the infrastructure for their growing demand could still raise everyone else’s bills. Chips in Space Illustration by Noah Hickey/The Dispatch (Photos via Getty Images). Equally salient were water-use concerns. Despite industry assurances that newer cooling systems are more efficient, voters remained suspicious about the effects on the local drinking water supply. Yet suppose data center developers could answer every question about the electricity bill and reassure voters about the water supply. Would that settle it? It’s doubtful. University of Virginia researchers asked 809 people to consider building one of four types of facilities within 10 miles of home. The prospect of a data center drew more opposition than a nuclear plant. Among people who thought the data center developer would take advantage of the community, 90 percent withheld support. That seems understandable. More disquieting for those who care about America’s AI future, the survey found that even 41 percent of respondents who gave the developer the benefit of the doubt still opposed the project. That leaves open the question of which came first: distrust or opposition. It also gives us reason to seriously contemplate voter sentiments that someone else is making the decisions. In a recent survey, 62 percent said they had little or no confidence that officials would consider residents’ views when deciding whether and where to build data centers in their area. All this comes at a time when workers are being told that AI is about to transform the kind of work they’ve spent their lives doing. They have very little say over how or when that will happen, and they do not know how fundamentally their lives and their livelihoods will be disrupted. Then a measure appears on their local ballot about whether an AI company can build a data center in their town. This might be their only chance to fight back against the tide. It’s little wonder voters reacted the way they did, and marginal improvements to water-cooling systems or behind-the-meter power generation by data centers are unlikely to change their minds. For America to maintain its lead in building data centers, we’ll have to find places to put data centers. Fortunately, one final frontier remains: space. Google is testing its AI chips in low Earth orbit, and SpaceX has bet the farm on orbital data centers. To most people, this sounds a bit far-fetched. I, too, was skeptical at first, but repeated exercises in estimating costs and physical feasibility turned me into a convert. With a cheap enough ride and constant sunlight, the math behind putting chips on satellites in space rather than in giant buildings closer to home starts to pencil out. Outer space is closer than it used to be. Space is a big place. The closest region, known as low Earth orbit, extends from 60 to 1,240 miles above Earth’s surface. Yet for decades, satellite operators have favored a more distant location known as geostationary orbit. In geostationary orbit, a satellite circles 22,236 miles above the equator in the same direction Earth turns. One circuit takes 23 hours and 56 minutes, roughly as long as Earth’s rotation, and the satellite appears fixed in the sky. Geostationary orbit is handy for a satellite beaming television signals to customers in one region. The dish on the roof can stay pointed in one direction. It also works well for weather satellites that monitor a single slice of Earth’s surface. However, the higher the orbit, the harder it is to reach. The space shuttle Challenger launched NASA’s first tracking and data relay satellite in 1983. But Challenger could only fly up to about 200-300 miles above Earth’s surface, the lower end of low Earth orbit. From there, the crew released a separate rocket carrying the satellite. That rocket was supposed to carry the satellite up to geostationary orbit. It malfunctioned, and the satellite spent months climbing into place using its own tiny thrusters. The space shuttle process was clunky, but the alternative was nearly as bad. To reach geostationary orbit in one shot, you’d need to build a rocket two-thirds as big as Saturn V—the behemoth that took astronauts to the moon—and not a single bit of it could be reused. Because building a giant rocket for each launch was so expensive, satellites became a niche product, and space became the province of scientific and government agencies. The space shuttle was supposed to be reusable, but it was built by committee and ended up with limited range and little reusability. The giant red fuel tank was lost on every flight. The two solid rocket boosters on each side cost more to recover and refurbish than to build new, and the airplane-like orbiter’s high-tech heat shield consisted of thousands of expensive bespoke tiles that had to be hand-inspected after each flight for signs of damage. Fortunately, capitalism came to our rescue. SpaceX’s largest commercial rocket, the Falcon Heavy, can reach geostationary orbit and return two of three boosters back to Earth. Even more importantly, as we’ll see below, SpaceX’s workhorse rocket, the Falcon 9, can deliver a payload to low Earth orbit and return to Earth completely intact. The internet is full of videos showing the rocket returning to Earth using its engines in reverse. In a 2018 analysis, a NASA researcher put the shuttle program’s average launch to low Earth orbit at a cost of about $61,700 per kilogram vs. $1,410 per kilogram for the Falcon Heavy. That 40-fold drop has revolutionized the economics of putting things in space. Introduction to orbital economics. Midcentury America was rife with fantasies about a space-based economy, where we would live and work high above the clouds, possibly all the way to the surface of the moon. Needless to say, little of that has panned out as technology has developed more slowly than earlier generations imagined. But more fundamentally, escaping Earth’s gravity is hard, and for most projects the benefits of space are scant. There are a few exceptions. For those who want to observe large portions of the Earth’s surface at once, whether for meteorological or military purposes, or for those who want to communicate with isolated locations, the effort is worth it, even if it means going all the way out to geostationary orbit. Low Earth orbit is much closer. However, sending rockets into low Earth orbit has always presented a twofold problem. First, you need to launch more satellites to get continuous coverage, and second, you have to replace them much sooner than in geostationary orbit. See, the lower the orbit, the faster the satellite circles the Earth: The circumference is smaller, and gravity pulls on the satellite harder. So, while a satellite in geostationary orbit circles the Earth once a day, a satellite in low Earth orbit zips all the way around the Earth about every 90 minutes or so. Since the geostationary satellite circles the Earth at the same rate the Earth itself turns, it appears to stay fixed in the sky. By contrast, a satellite in low Earth orbit keeps a given spot in its field of view for only about 10 minutes. From the ground, it looks as if the satellite rises from the horizon, crosses the sky, and then dips below the opposite horizon over those 10 minutes. If you always want a satellite overhead, providing continuous coverage to a given spot, then you need to launch nine low Earth orbit satellites for every one geostationary satellite. In addition, while the atmosphere in low Earth orbit is incredibly thin, it’s still thick enough to matter. Given that friction, low Earth orbit satellites will fall out of orbit in about five years. Geostationary satellites can stay up indefinitely, though in practice their internal systems tend to fail after about 30 years or so. Still, that’s six times as long as a low Earth satellite, and continuous coverage takes nine times as many low Earth satellites the whole time. That means in total you have to launch 54 low Earth orbit satellites for every one geostationary. Before reusable rockets drove down launch costs, launching so many satellites—and replacing them so frequently—made continuous coverage from low Earth orbit too expensive for most commercial uses. It’s always sunny in space. Of course, once the satellite is up, the question of fuel remains. Over the past two decades, the price of solar power has fallen by 90 percent, from 41 cents to 4 cents per kwh. By comparison, electricity from natural gas generators costs about 9 cents. That makes solar the cheapest way to produce electricity. The problem, however, is that sunshine is fickle. On Earth, the sun is only out during the day, and even then its rays are too often diluted by clouds and precipitation.A data center needs to run 24 hours a day. So a solar-powered data center would require either battery storage or a backup power source, both of which eat into solar’s cost-effectiveness. But in space, the sun is always shining—if you pick the right path. If you look at a map showing where on Earth it’s day vs. night, you’ll notice the boundary between day and night follows a U-shaped path across the planet. If a satellite is set to follow that path, it'll never be nighttime. No backup batteries needed. On top of that, sunlight in space is a bit brighter, and the absence of rain means you don’t need tempered glass shielding to protect sensitive photo-electronics. All told, those factors drop the cost of power by about 90 percent. Hot chip, no lie. A key part of running a data center is cooling the computer chips. AI chips produce a lot of heat. They are also sensitive to temperature. As the temperature rises beyond 100 Celsius (212 Fahrenheit), their performance starts to degrade. Above 120 C, they can stop working altogether. On Earth, cooling typically involves running cold water through tiny pipes connected to the servers. Heat flows from the hot chips to the cold water, cooling them down. The water heats up. You still have to do something with the hot water. In the past, data centers would let it evaporate. This process, however, consumes an enormous amount of water and has fueled many an environmentalist’s nightmares. Newer systems are more efficient, but also more expensive. In space, cooling requires infrared radiators. Liquid coolant carries heat from the chips into broad, flat panels. Like all warm objects, the panels emit infrared light. That light carries energy away into space, cooling the panels and the liquid inside them. The thing is, the efficiency of infrared radiators is determined by what’s called the Stefan-Boltzmann equation. It says that when you increase an object’s temperature by X degrees Kelvin, the amount of infrared radiation it gives up every second goes up by X to the fourth power. That’s a massive increase in heat dissipated. Thus, ironically, the hotter the chips, the cheaper it is to keep them cool in space. Pencils out. Let’s put a price tag on all this. Elon Musk’s ultimate target is a launch cost of $10 to $15 per kilogram, but his estimates have been known to be … overly optimistic. Cooler heads put SpaceX’s ultimate launch costs at $200 per kilogram. Sending a megawatt’s worth of computers, solar panels, radiators, and supporting equipment into orbit would cost about $7 million. It sounds expensive until you look at the bill on Earth. Constructing a data center building with solar panels, battery backup, and low-water-use cooling costs just under $29 million per megawatt. And that is before property taxes. By contrast, orbital solar panels, infrared radiators, and other supporting hardware add about $4 million per megawatt. The engineering is well on its way to being established, but the economies of scale are still unproven. Nonetheless, at $200 per kilogram, the math works. And judging by this month’s election results, getting a data center into orbit may eventually be easier than getting one through a local zoning board. Markets FTW We’ve all seen the impressive feats accomplished by humanoid robots created by firms like Boston Dynamics. But it should come as no surprise to students of capitalism that one of the first commercial applications of so-called physical AI is in agriculture. Old-school commercial robots, like those used in automobile factories, do one precise task in exactly the same way over and over. They’re efficient but not flexible. Physical AI brings the flexibility and knowledge of LLMs like ChatGPT to automated machines. Carbon Robotics created a system that can recognize weeds and precisely shoot a laser at the vital meristematic tissue the weed needs to survive and grow. Laserweeding reduces, and potentially eliminates, the need for herbicides. Herbicides are a health concern for many consumers, but they also cut into yield for farmers, since it is virtually impossible to create a chemical that kills weeds reliably without harming the crop. The interesting thing is that the robot in this case is designed as a standard tractor attachment, making the conversion to laser weeding almost effortless for farmers. Next up, Carbon Robotics plans to market a kit that makes your tractor self-driving. The quiet serenity of farm life has never looked sweeter. Chart of the Week Cost of sending an object into space using the most sophisticated rockets over time. (Source: Visual Capitalist) Disclaimer: The opinions expressed above do not necessarily reflect those of the presenting sponsor.

Original Source

Read the full article at Thedispatch →

KhanList aggregates and links to publicly available news content. We do not host full articles from third-party sources. Always verify important information with original sources.