As AI Heats Up, High Tech Turns to Spiders

The Cooling Frontier

The hunt for a way to keep the AI age cool — from the Moon to a silk thread

AI chips generate enormous heat, and keeping them cool is becoming one of the biggest challenges in technology. In search of solutions, engineers are exploring everything from lunar data centers and space-based cooling systems to an unlikely source closer to home: the golden orb-weaver spider, whose silk can conduct heat almost as efficiently as copper.

A blended analysis · engineering · economics · geopolitics · the long view

↑ The computing frontier is climbing — from terrestrial halls to subsea pods, orbital clusters, and a lunar archive.

01 ——The Invisible Infrastructure of Modern Civilisation

Ask most people where "the cloud" lives and they will point vaguely upward. The truth is far more terrestrial and far heavier. The cloud is a planetary archipelago of windowless warehouses — data centres — packed with servers, storage arrays, and networking gear, humming twenty-four hours a day. Every email opened, every film streamed, every payment cleared, every map route calculated, and now every AI prompt answered, resolves inside one of these buildings before the result races back across glass fibre to a screen.

For two decades this infrastructure stayed politely out of sight. It is now arguably the most strategically important physical asset class of the age. The reason is artificial intelligence. A traditional web request asks a server to fetch a stored page; a large-language-model query asks thousands of specialised chips to perform trillions of floating-point operations in concert.

02 ——The Demand Curve Bends Vertical

The numbers are sobering. In 2024, data centres worldwide consumed roughly 415 terawatt-hours (TWh) of electricity — about 1.5% of global supply. The International Energy Agency's Energy and AI analysis projects that figure will more than double to around 945 TWh by 2030, a quantity slightly larger than the entire electricity consumption of Japan today, and climb toward 1,200 TWh by 2035. Data-centre electricity is growing roughly 15% a year — more than four times faster than total electricity demand from every other sector combined.

↑ The IEA base case: a doubling by 2030, with AI-optimised "accelerated" servers expanding around 30% a year.

The growth is wildly uneven. The United States and China together account for nearly 80% of the projected increase. American data-centre demand is set to rise about 130% by 2030; in the IEA's reckoning, the US economy will soon consume more electricity processing data than producing all of its energy-intensive manufactured goods — aluminium, steel, cement, and chemicals — combined. China's data-centre load is projected to grow roughly 170%. Europe adds about 70%, Japan more than 80% from a smaller base. The US Department of Energy's Lawrence Berkeley National Laboratory found American data centres already drew 4.4% of national electricity in 2023 and could reach between 6.7% and 12% by 2028.

Driving all of it is the rise of the hyperscale data centre — campuses measured not in racks but in hundreds of megawatts, soon gigawatts, built by a handful of cloud giants. These are the cathedrals of the AI age, and they are the reason the cooling problem has gone from a facilities footnote to a board-level constraint.

03 ——Every Watt Becomes Heat

Here is the inconvenient law of physics at the centre of the whole story: a chip is not a perfectly efficient machine that turns electricity into computation. Almost all of the energy a processor consumes is ultimately dissipated as heat. A rack drawing 50 kilowatts of power is, for practical purposes, a 50-kilowatt heater. Push thousands of those racks together and you have created a thermal emergency that must be managed every second, or the silicon throttles, then fails.

For decades the answer was simply moving air: enormous fans, raised floors, and chilled-air aisles. That works up to roughly 20 kilowatts per rack. But a single rack of the latest AI accelerators can demand five to ten times that. Air, it turns out, is a feeble carrier of heat. Liquids carry it far better, which is why the industry is sprinting toward water and engineered fluids touching the chips directly. Cooling has become the binding engineering challenge of the era precisely because compute density is outrunning the physics of air.

04 ——The Cooling Ladder

The progression of cooling technology reads like a ladder, each rung handling more heat than the last.

↑ As power density climbs, liquids replace air — and in vacuum, radiation becomes the only exit for heat.

Air cooling remains the workhorse of legacy facilities: cheap, simple, familiar, and increasingly inadequate. Liquid cooling covers a family of approaches. Direct-to-chip cooling clamps cold plates onto the hottest components and pipes coolant through them, capturing heat at its source and handling 50 to 120 kilowatts per rack. Immersion cooling goes further, submerging entire servers in a non-conductive dielectric fluid so that every surface sheds heat at once, comfortably exceeding 100 kilowatts per rack and slashing the energy spent on fans.

Beyond the chip lie the macro-strategies. Seawater and lake cooling tap naturally cold bodies of water as a free heat sink. Heat-reuse systems capture the warm output and pump it into district heating networks, greenhouses, or even — as European researchers have proposed — water-purification and carbon-capture plants, turning a waste stream into a product. And then there are the emerging frontiers: two-phase immersion that exploits the latent heat of boiling, microfluidic channels etched directly into the silicon, and the radiative cooling that becomes mandatory once you leave the atmosphere entirely.

05 ——Power, Water, and the Strain on the Grid

If heat is the chip-level problem, power and water are the planetary one. A large data centre can consume up to 5 million gallons of water per day — the equivalent of a town of 50,000 people — largely for evaporative cooling. American data centres' indirect water footprint, through the power plants that feed them, was estimated at around 211 billion gallons in 2023. In Texas alone, researchers expect data-centre water use to leap from roughly 49 billion gallons in 2025 to as much as 399 billion by 2030. Troublingly, about two-thirds of new data centres built or planned since 2022 sit in already water-stressed regions.

The exact amount of water used by an AI query is still debated. Researchers at UC Riverside estimated that a 100-word ChatGPT response uses about 519 millilitres of water — roughly one standard water bottle. Technology companies report much lower figures, with Google saying a Gemini query uses around 0.26 millilitres and OpenAI estimating about 0.3 millilitres per query.

06 ——Into the Sea: The Underwater Data Centre

If the ocean is a vast, free, cold heat sink, why not put the data centre in it? Microsoft asked exactly that with Project Natick. Beginning in 2013, the company sealed servers inside a steel cylinder filled with inert nitrogen and sank it. In the landmark Phase 2 trial, a pod containing twelve racks and 864 servers — up to 27.6 petabytes of storage — was lowered off Scotland's Orkney Islands in 2018 and left untouched for two years.

↑ Sealed, dark, and human-free: Natick's pod failed servers at roughly one-eighth the land rate — yet could not be serviced.

The results startled even the engineers. Of 855 submerged servers, just six failed — a 0.7% failure rate — against 5.9% for an identical batch on land, roughly one-eighth as many failures. The credit went to a stable, cool, vibration-free environment, an oxygen-free nitrogen atmosphere that slowed corrosion, and the simple absence of humans bumping into things. Microsoft also noted that about half the world's population lives within 200 kilometres of a coast, so subsea pods could sit close to users while sipping local renewable power.

And yet, in 2024, Microsoft quietly confirmed it was no longer building subsea data centres anywhere. The reason was not failure but inflexibility. A sealed pod cannot be opened to swap a faulty server or — fatally, in the AI era — to upgrade to the next generation of chips. To service anything, you must raise the entire unit. In a market where hardware obsoletes every couple of years, an un-upgradeable box is a liability, however reliable. The idea is not dead, though: China's Highlander has deployed a commercial underwater module weighing over 1,400 tonnes off Hainan, and Microsoft says it will fold Natick's lessons into liquid-immersion designs on land.

07 ——Into Orbit: Computing Above the Clouds

The most audacious answer to Earth's power-and-water squeeze is to leave the planet. In orbit, sunlight is constant and unfiltered, so a solar array generates far more energy than the same panel on the ground — with no nights, clouds, or atmosphere to lose. There is no land to buy, no neighbours to anger, no aquifer to drain. To enthusiasts, it is the ultimate clean data centre.

The momentum is suddenly real. In November 2025, the startup Starcloud launched Starcloud-1 aboard a SpaceX Falcon 9 — a refrigerator-sized, 60-kilogram satellite carrying the first Nvidia H100 GPU ever flown, a chip roughly 100 times more powerful than anything previously operated in space. Within weeks it had trained an AI model and run Google's Gemma language model in orbit, both firsts. Backed by Nvidia and a $170 million Series A that valued it at $1.1 billion, Starcloud is pitching an eventual 5-gigawatt orbital data centre spread across a four-square-kilometre solar array, and a near-term Starcloud-3 craft designed to deploy from SpaceX's Starship.

↑ In space the easy part is power. The hard part is getting rid of heat — which can only leave as infrared radiation.

Google has entered with its own moonshot, Project Suncatcher, partnering with Planet Labs to launch two prototype satellites carrying its TPU chips by early 2027. The design imagines an 81-satellite cluster flying in a tight one-kilometre formation in a dawn-dusk, sun-synchronous orbit for near-continuous sunlight, lashed together by free-space optical links that have hit 1.6 terabits per second in bench tests. SpaceX, meanwhile, has sought regulatory room for as many as a million satellites for distributed computing, and a clutch of startups — Aetherflux, Aethero, and others — are crowding in.

But here is the counter-intuitive truth that orbital boosters tend to underplay: cooling in space is harder, not easier. In a vacuum there is no air or water to carry heat away by convection. The only way to remove heat is through thermal radiation, forcing operators to rely on large infrared-emitting radiator panels. For a data centre consuming hundreds of megawatts, those radiators could become enormous structures, adding significant mass, engineering complexity, and launch costs.

Space also introduces risks that terrestrial facilities rarely face: radiation damage to electronics, orbital debris, micrometeoroid impacts, and the challenge of repairing failed hardware. Google's tests found that its Tensor Processing Units (TPUs) — specialised AI accelerators designed for machine-learning workloads — could withstand roughly three times their expected five-year radiation exposure, although high-bandwidth memory remained the most vulnerable component. The results are promising, but they remain far removed from the reliability and easy maintenance of an Earth-based data hall.

The economics turn on a single number: launch cost. Starcloud argues an orbital facility could reach about five cents per kilowatt-hour if launch prices fall toward $500 per kilogram; Google models parity with terrestrial facilities if costs drop near $200 per kilogram by the mid-2030s. Both depend on Starship-class reusable rockets that are not yet flying commercially, with access expected around 2028–2029.

08 ——To the Moon: The Ultimate Off-Site Backup

If orbit is being promoted as a place to generate intelligence, the Moon is increasingly being seen as a place to protect it. The idea is not to run everyday computing tasks there, but to use it as a secure backup for humanity's most important data. Far from Earth, a lunar data vault could help safeguard critical information from natural disasters, wars, political instability, cyberattacks, and other threats that could damage data stored on our planet.

The pioneer is Florida's Lonestar Data Holdings, founded in 2021 to sell "Resiliency-as-a-Service" from what it calls the ultimate off-site location. On Intuitive Machines' first lunar landing in February 2024, Lonestar ran a software-defined payload that famously beamed a copy of the US Declaration of Independence to the lander and received the Constitution and Bill of Rights in return. A year later, its physical Freedom payload — an 8-terabyte solid-state drive paired with a radiation-hardened RISC-V processor running Linux — rode to the surface, though that lander toppled and ended the mission early. Lonestar has since signed a $120 million deal with Sidus Space for a fleet of six lunar-orbit storage spacecraft.

↑ Each rung trades readiness for ambition. The Moon is the most distant — and most patient — bet.

The appeal of long-term lunar archives is genuine: cold, stable temperatures in permanently shadowed craters, isolation from terrestrial catastrophe, and — visionaries suggest — eventual placement inside protected lunar lava tubes that shield hardware from radiation and micrometeorites. The idea dates back at least to a 2012 conference proposal for a lunar supercomputer.

The obstacles are equally genuine. Launch and landing remain expensive and unreliable; a "data centre" on the Moon today is a single drive and a chip, not a hall of racks. The 1.3-second one-way light delay rules out anything latency-sensitive, confining the use case to archival storage and disaster recovery. Repair is impossible. For now, the Moon is a backup tape in the sky — strategically intriguing, commercially embryonic.

09 ——Nature's Cooling Secrets: How Spiders Could Help Power AI

While engineers reroute heat across oceans and orbits, materials scientists are attacking the problem at the scale of molecules — and increasingly borrowing nature's blueprints. The poster child is an unlikely one: the silk of the golden orb-weaver spider.

↑ Iowa State researchers measured golden orb-weaver silk conducting heat like copper — and roughly 800× better than any other organic material.

In a 2012 study at Iowa State University, the engineer Xinwei Wang found that silk spun by Nephila clavipes conducts heat at around 416 watts per metre-kelvin — on a par with copper, better than silicon, aluminium, or iron, and roughly 800 times better than any other organic material ever tested. Strangest of all, stretching the silk increases its conductivity, the opposite of most materials, thanks to a near-defect-free molecular structure of protein nanocrystals linked by spring-like chains. The promise is biomimetic: flexible, lightweight, heat-dissipating fibres that could one day move warmth out of tightly packed electronics where rigid metal cannot reach.

Scientists are now exploring how to replicate and scale this behaviour in engineered materials. The idea is not to use natural spider silk directly, but to design synthetic fibres and composite films that mimic its molecular structure. If successful, these materials could be woven into ultra-thin thermal "networks" inside chips, wearables, and aerospace systems, quietly spreading heat away from hotspots and improving cooling in places where traditional metal heat sinks cannot fit.

10 ——The Winners and Losers of the AI Infrastructure Race

Every infrastructure super-cycle redistributes advantage. This one is no exception.

Winning countries will be those endowed with cheap, clean, firm power and the political will to build it fast: the United States, with its capital depth and gas-and-nuclear push; nations with abundant hydro, geothermal, or nuclear baseload such as the Nordics, Canada, and France; and Gulf states converting cheap energy and sovereign wealth into compute. Losing are regions where grids are saturated, water is scarce, or permitting is glacial — where the factories of intelligence simply cannot plug in.

11 ——The Long View

Strip away the chatbots and the valuations and the future of artificial intelligence resolves into something older and more physical: a story about energy, heat, and the human habit of going wherever the resources are. We mined coal, dammed rivers, and split atoms to power earlier industrial ages. We are now doing the same to power the age of thinking machines — and discovering that the limiting reagent is not intelligence but a place cool enough to host it.

The digital factories of the AI age are the most consequential infrastructure humanity has ever built, and the race to keep them cool is, in the end, a race to keep them running at all. It will be decided by engineers and physicists as much as by founders and financiers — and it is pulling the boundary of human industry off the planet for the first time. From the seabed to the stars, the search for somewhere cool enough has become the search for the next frontier itself.