The Notes Tesla Left

A thermomagnetic heat sink for cooling data without water

OBSERVE

The Thirst Behind the Screen

A data center is a building full of machines that turn electricity into answers and heat. The answers leave on fiber. The heat stays behind, and somebody has to carry it out.

The cheapest way to carry it has been water. Evaporative cooling works like sweat on skin. Water soaks up heat as it turns to vapor and drifts off, taking the heat along. It’s effective and cheap to run, and the water it uses is gone, out of the tower and into the sky. Industry estimates put a one-megawatt facility on evaporative cooling at 20,000 to 50,000 gallons a day. The new AI campuses are measured in hundreds of megawatts.

That’s how a town ends up arguing with a server farm over the same aquifer. Water use has climbed to the top of community concerns about new data center construction, and places like Phoenix, which drew the industry with cheap land and power, now count as water-scarce.

The water is a choice. Dry coolers move heat to outdoor air through closed loops and fans, the way a car radiator does, and evaporate none of it. They cost more electricity, particularly on hot afternoons when outdoor air is a poor place to dump heat. A 2026 Uptime Institute brief found that well-designed dry and adiabatic systems can match evaporative efficiency across much of the country’s climate range with zero or near-zero water use. A study of California sites found dry cooling saved millions of gallons a month for less than one percent added cost of ownership. The price shifts with climate and grid, but the choice is on the table.

We’ve worked this problem at building scale before, in “Passive Cooling for Data.” This piece goes down to the chip.

The Heat Problem Underneath

The electricity a server draws leaves the room as heat, watt for watt. A data center’s chillers, towers, pumps, and fans amount to a second building bolted onto the first, and the industry’s own efficiency score, PUE, is a measure of how big that second building is. Tech firms already pay for a data center that cools the data center.

Heat moves along a chain. Chip to heat sink, heat sink to air or fluid, fluid to the building loop, building loop to the outdoors. Each handoff takes energy, and at the end of the chain, the heat gets thrown away like trash at the curb. The whole system is built around getting rid of energy that arrived at the chip in perfectly usable form a moment earlier.

DESIGN

Energy Is Energy Is Energy

This idea started as a thought experiment. Heat is energy in disorder, molecules jostling in random directions like a crowd milling around a fairground. Electricity is energy in order, charge moving one direction down a wire like that same crowd filing through a gate. Same energy, different arrangement. Getting from one to the other is a mechanical problem: find the mechanism that lines up the crowd.

Power plants solve it with steam and turbines. Thermocouples solve it with semiconductors. We wanted something that could bolt onto a heat sink with no turbine and no steam: a coil wrapped around a core, sitting in the path the heat already travels.

A coil makes current when the magnetism inside it changes. That’s Faraday’s law, and it carries a catch. Steady heat flowing through a steady core settles into a steady magnetic state, and a steady state makes no current. The magnetism has to rise and fall.

Nature supplies the lever. Magnetic materials lose their magnetism above a temperature called the Curie point and get it back when they cool below it. Swing a material’s temperature back and forth across that line, and its magnetism swings with it. A coil nearby feels each swing as a pulse of voltage.

One material with one switch point has a narrow window, though. A chip’s temperature moves with its workload, cooler at idle and hotter under full load, and a single switch point only pulses when the heat happens to sit near its threshold. So use several. Give each magnetic element its own switch point, and different temperatures wake different elements. Light work sets the low-threshold elements pulsing, heavy work sets the high-threshold elements pulsing, and the stretch in between gets covered by the ones in the middle. Like a set of wind chimes tuned to different notes, whichever way the breeze blows, something rings.

What drives the swing is open. A thermal capacitor on the hot side that fills and releases, like a geyser. A gate on the cold side that opens and closes. A pulse of compressed air. A pumped liquid loop on a valve. Chemistry. Electricity. The requirement is temperatures that cross their thresholds on a repeating cycle. The method can be fitted to the job.

Flowers for Nikola

Then we found out Tesla had already drawn it.

U.S. Patent 428,057, “Pyromagneto-Electric Generator,” was filed in 1887 and granted May 13, 1890. Tesla built it on two laws he called well known: a conductor develops electricity when the magnetism around it varies, and iron loses its magnetic properties when heated and regains them when cooled. His claim describes a magnetized core, a conductor in its field, a source of heat applied to part of the core, and an artificial cooling device to bring that part back down. That’s our device, on paper, 136 years ahead of us. Edison filed a pyromagnetic generator of his own, also built on iron.

Neither machine went anywhere. Iron gives up its magnetism around 770°C, so the device needed a furnace on one side, and the trickle of current it made didn’t justify the fire. Tesla moved on to alternating current and changed the world with it. The pyromagnetic generator became a footnote.

Tesla had the principle right. What he lacked were materials that switch near room temperature and a task that needed the machine. He was trying to make electricity, and as a power plant his generator was weak. As a device that makes heat flow do work on its way out of a hot room, it was waiting on computer chips that wouldn’t exist for most of a century. This is a habit with Tesla. He invented things without knowing what they’d be used for, wrote them down carefully, and left the notes where someone could pick them up when the right job came along.

The job is here.

Turning Tesla Around

Tesla tuned his machine to make the most electricity. We tune it to move the most heat, and that changes the parts list from top to bottom. The research community has been rebuilding Tesla’s machine for two decades now, and a paper published October 20, 2026, by a team at the University of Science and Technology Beijing supplies several of the pieces.

Materials. Gadolinium and engineered alloys in the LaFeSi and MnFePSi families switch near chip temperatures, and the alloys can be tuned to switch where you want them. The Beijing team’s numbers show why the choice matters. Gadolinium switches around 17°C, and when they warmed its cold side from 5°C to room temperature, its output fell 86 percent. A hydrogenated LaFeSi composite switching around 62°C lost only 16 percent under the same change. Chip-side heat with room-temperature cooling puts the tuned alloys in the lead. Some of these materials also soak up extra heat as they cross their transition, the way ice soaks up heat while it melts without getting warmer, which turns each element into a heat sponge at its switch point.
From core to switch. Tesla wrapped his coil around the hot iron. The Beijing design puts the magnetic material to work as a switch instead. Permanent magnets and iron bars form a closed loop, with a single coil wound on a center bar. Thin Curie plates sit in two branches, one above the center bar and one below. Heat the top branch and cool the bottom, and the top plates go weak while the bottom plates carry the field, so the magnetism runs through the bottom loop. Swap the temperatures and it runs through the top. Either way it passes through the center bar, in opposite directions, so the coil sees the field flip back and forth like a railroad switchyard sending the same train through the station first eastbound, then westbound. They report the flip roughly quadruples output, and the plain coil-around-material layout manages about a twenty-fifth of their voltage per gram of material. Using small plates as switches also cuts the amount of costly magnetic material.
Many elements, many switch points. Tesla used one iron core. We use a family of switching units, arranged two ways. In series along the coolant path, the warmest-switching units sit nearest the chip and the coolest nearest the cold side, each working its own slice of the temperature drop and handing the heat down the line. Side by side, units tuned to different load conditions cover the chip from idle to full tilt. The Beijing paper lays out the series version for a cooling pipeline, high-switch-point units near the hot inlet and low-switch-point units downstream, and tested two materials across shifted temperature windows to show each holds up best in its own range. A 2025 study ran a series of thermomagnetic units with switch points spread from 300 to 420 K off a single heat source. Units sitting far from their threshold at a given moment still conduct heat, so idle units keep the path open instead of blocking it.
Staged in time. Temperature staging decides which units respond to which heat level. Time staging decides when each unit cycles. The paired-branch design already runs its two branches half a cycle apart, one heating while the other cools. Extend that across several units on offset schedules, and at any moment some branches are dumping heat while others take it in. The cold side keeps working through the whole cycle instead of opening and closing all at once, so heat leaves the chip at a steadier rate and the buffer between chip and stack can be smaller. The electrical pulses overlap too, filling in the gaps between them. The Beijing array ran its six units in lockstep. Offset timing is untested, and it’s where this design steps past the published work.
Thin plates, fluid contact. Tesla heated with flame and steam. The Beijing team used gadolinium plates one millimeter thick with one-millimeter gaps for water to flow through, and the plates heated or cooled almost completely in about half a second. Their best cycle ran six seconds. A chip is a steady, known heat source an inch away, and fluid moving through thin gaps is the fastest way to swing the plates.
A new job for the coils. The electrical output is small. Six Beijing units in series made about 0.8 volts and 159 microwatts, enough for a small sensor. Measured against the heat flowing through, conversion efficiency was a rounding error. The wires carry a sliver of the energy out of the room, and the cooling system carries the bulk. That’s fine here, because the goal is cooling. With units pulsing at different moments, their outputs get rectified and fed into a shared capacitor, the way several small streams fill one pond. That stored current helps power the device’s own controller, and each unit’s coil doubles as a sensor reporting where its plates sit relative to their switch point. Read together, the coils give a running temperature map of the stack.

The same physics also runs backward. Magnetize one of these materials and it warms; remove the field and it cools. Drive the coils instead of harvesting from them, and the plates become a heat pump with no refrigerant and no compressor. That’s magnetic refrigeration, an active research field of its own. It costs power and belongs to a later version, but the hardware is shared.

The Stack

From the chip outward:

Interface. Standard thermal paste or pad on the chip lid.
Spreader and buffer. A copper cold plate that holds the chip at a steady temperature and hands its heat to a coolant loop. This is the thermal capacitor.
Bypass path. A conventional cooling path from the cold plate that keeps working if the switching stack stalls. If a valve sticks or the controller fails, heat still escapes and the chip throttles instead of cooking.
Switching cartridge. A row of paired-branch units. Each unit has permanent magnets, iron bars forming a closed loop, a coil on the center bar, and two small chambers of thin Curie plates with fluid gaps. Along the row, switch points step down from warmest nearest the chip to coolest nearest the cold side. Samarium-cobalt magnets hold their strength near hot electronics better than common neodymium ones.
Valve gate. Two closed fluid loops feed the cartridge: a hot loop carrying heat off the chip’s cold plate, and a cold loop running out to a dry cooler. Solenoid valves alternate which branch of each unit gets hot fluid and which gets cold, with each unit’s timing offset from its neighbors. The water stays sealed in pipes and goes back around. None of it evaporates.
Collection and control. A rectifier on each coil feeding a shared storage capacitor. A temperature sensor on the cold plate and a small microcontroller running the valves, powered in part from that capacitor and reading the coils as a temperature map to time the switching.

INTERVENE

The Dare

A concept on paper asks readers to take it on faith. This one comes with a parts list, and a published blueprint to crib from. The Beijing team put their prototype’s components at about $31.

Step one: build the paired-branch unit. Two small permanent magnets, iron bar stock for the loop, a coil of magnet wire with a thousand-plus turns on the center bar, and thin gadolinium plates in two small printed chambers. Add two water reservoirs, one hot around 65°C and one ice-cold, two small pumps, four solenoid valves, a microcontroller, and an oscilloscope on the coil. Flip the hot and cold flows between the chambers every three seconds. The scope should show a pulse in one direction, then the other, each time the flows swap. That reproduces the published result at desk scale.

Step two: offset it. Build a second unit on the same loops and run the pair two ways. First in lockstep, both switching at once. Then with the second unit’s valves delayed by a quarter cycle. Compare the scope traces for smoothness, and track how fast the hot reservoir loses temperature as a stand-in for cooling. If the staggered pair pulls heat more steadily, the time-staging idea stands, and that’s a result nobody has published yet.

From there, the questions are engineering: how many units, how thin the plates, how far apart their switch points, how fast a cycle, how far to offset the timing, which magnet, and how the plates hold up. The Beijing team ran 83 cycles with no wear and said longer testing is needed. At a six-second cycle, a server would put a unit through about 14,000 cycles a day. Then there’s the question of when pump mode earns its power bill. Those questions belong to anyone who wants them.

We’re casting this one on the waters. Tesla left the notes in 1890. Build the test, break the design, improve it, and publish what you find.

Sources