They Cooled 5,000 Watts per Square Centimeter in Space. Now Diamond Goes on AI Chips

Akash Systems is a Bay Area, venture-backed company that starts at the periodic table rather than in the middle of a supply chain. Felix Ejeckam and Ty Mitchell, both materials scientists, grow diamond in the lab and use it to build chips, boards, and systems. They began in space, where they say they have made and launched some of the fastest satellite radios humans have built. They are now applying the same stack to AI, claiming they can make a GPU run faster than before, or cut data-center energy, because of that ground-floor materials work.

Ejeckam's Cornell PhD in electrical engineering, with a minor in materials science and device physics, was about joining dissimilar materials so that one plus one equals ten — optics with silicon electronics, for example. He and Mitchell founded Akash in 2017 on the same idea. Artificial diamond is the most thermally conductive material that occurs in nature or can be grown; silicon, gallium nitride, or any other semiconductor does the computing. Bring them together, he says, and each does what it is good at. He likens it to people: different strengths, combined, exceed the sum.

Mitchell is the silicon-carbide person. After his PhD he went to Cree, later Wolfspeed, which had strong SiC materials for radar, EV power electronics, and LEDs. The expensive lesson: if you only sell a Schottky diode or a MOSFET, the carmaker already has a silicon IGBT and will switch when you match price. Promise 200 extra miles of range, a 40 percent gain, and you still get a polite nod. Make the car, or at least the power module, and the customer starts to believe. Akash's rule from day one was build the system and go to the customer. Space came first because it is harsher than AI: limited area, no airflow, no liquid cooling, brutal reliability. Power densities they cool there sit at 4,000 to 5,000 watts per square centimeter. A typical ground server chip is a full order of magnitude less, a couple of hundred watts per square centimeter. If the same technology, even backed off for a faster ramp, works in space, they expect a home run in a rack.

Ejeckam holds up a lab-grown diamond wafer with the Akash name on it and presses it into an ice cube. Heat from the room, and from his fingers, flows through the diamond so fast that the wafer sinks in "just like butter." The stone feels cold because it yanks heat away on contact. "You can cut ice with your fingers," he says. That conductivity is what they want sitting on a GPU.

The first commercial target is a 10°C drop, which he says is already worth millions of dollars for a data center with even a small number of servers. Over the next twelve months they want 20, 30, then 40 degrees. In space they have already taken temperatures down 80 to 90 degrees. Diamond cooling does not replace fans or liquid loops. It stacks on them. Today's cooling already lives at three layers: hot-aisle/cold-aisle containment in the hall; CDUs, manifolds, and fans at the rack; fan-out packaging and stacked HBM at the chip. Akash's current Nvidia H200 servers ship liquid-cooled and diamond-cooled at once. Extra thermal margin can be spent on a cooler chip or on running hotter to buy more transactions per watt.

Sarah Guo notes that anything with moving parts makes her nervous, and that even announced liquid-cooled GB200 NVL72 systems still have to not leak. Mechanical complexity is real. Ejeckam's harder claim is that if the industry does not add a physics-or-chemistry attack on heat, projected AI data-center demand will crash the grid, and if it does not, electricity will be exorbitant. A laptop on an ice pack does not speed up the CPU. Too many thermal barriers sit between the pack and the die. You have to go to the heat source.

Guo lists the complaints she hears from people who run large data centers: chip supply, GPU reliability, power, heat. Mitchell does not split them. "Everything you mentioned is heat." Space taught the same lesson. Treat symptoms and you play Whac-A-Mole; the fault pops up somewhere else unless you go to the heat source. One blade they ship has eight GPUs plus scores of other equally hot chips, and there are billions of such devices in the world. Akash, he says, is still scratching the surface.

The failure modes are already familiar on the floor. Servers die on arrival often enough that the industry has a name for it, infant mortality. When a die climbs to 80–90°C, the operating system backs off GPU workload — thermal throttling. Inference the customer paid for slows down. Performance that used to double or triple every few years starts to stall because power, water, and parts cannot keep up with the heat inside the box. Delays in device rollouts are the early form of getting stuck; Mitchell expects more of them.

He inventories the sandwich: an aluminum chassis, a copper heat sink, epoxy bonding, the silicon chip, gold-tin solder, an FR4 or other polymer board, gold bumps, another board. He is only listing some of the layers. Every interface is a thermal barrier. Almost nobody is attacking that stack from the inside. Samsung HBM had heating issues; Blackwell's rollout slipped for heat; on Nvidia's last earnings call an analyst asked about heating and Jensen Huang, Mitchell says, skillfully declined a direct answer. Those questions will multiply. Miss two or three quarters, as Intel just illustrated with Pat Gelsinger's departure, and you fall off the mountain. Guo adds that most people in machine learning treat GPUs as abstract software. Erratic, temperature-dependent behavior is a new domain they now have to budget for.

Akash does not make GPUs. It buys the hottest chips in the world, cools them, and hands the extra performance envelope to a system architect. The commercial entry is as a server maker for inference, training, data-center operators, and cloud providers — the people already stuck at the performance edge. Super Micro and other SIs are the neighborhood; Nvidia remains the chip source.

Ejeckam goes further. With diamond, he says, they can hyper-accelerate Moore's law: in two years, the performance that used to take six or seven. Moore's law is transistors packed closer and closer until thermal crosstalk stops the squeeze. That crosstalk, in his telling, is now the limiter on inference speed. Open it, raise density, and a 90-minute feature film that currently takes days under thermal limits could be generated in seconds, in production. The business is to sell that unlocked density, not a new GPU architecture.

Growing the diamond is less mystical than jewelry ads suggest. You start with a seed crystal, as with silicon, silicon carbide, gallium arsenide, or indium phosphide, and use plasma-enhanced chemical vapor deposition in a hot, high-pressure reactor. Diamond is carbon. The specialty is not only the growth. It is intimately coupling diamond atoms to semiconductor atoms with physics and chemistry. That bond is not trivial, which is why materials work is slow, and it is what they bring to both space and AI. Market pull, they say, is already strong.

The week of the taping, Akash announced NextGen, which they describe as India's largest sovereign cloud provider — national data that does not leave the country. Ejeckam has seen the same requirement in space: satellite data stays inside the owner's borders. The pattern, he thinks, will repeat country by country. NextGen is also a scale bet. The company has to ship everything within the next quarter, the step where small hardware firms usually trip. NextGen's leaders, in his telling, saw a materials attack on a problem that will not go away, and a vendor that might actually ramp. Larger 2025 orders are in discussion, and U.S. operators with the same heat problem are in the pipeline too.

On American manufacturing Ejeckam is blunt: the United States is not doing enough. Leadership has to run the whole value chain, not rest on Nvidia and AMD. AI is commercial and national-security technology, and it cannot be outsourced. Akash is one of the few small CHIPS Act recipients. Publicly, they work with Raytheon; he says the technology lets U.S. defense keep a military edge that has not existed before, and that it will create jobs in California and Texas. Vinod Khosla and Peter Thiel are among the investors. The pandemic, he adds, laid the supply chain bare: key chips had to be backfilled from overseas that can now be made at home.

Mitchell's version is colder. America will do this; the only question is whether it does it because it wants to or because it has to. Corporate culture chases earnings, so if Asia is cheaper, the factory goes to Asia. Chips, servers, software, frames, racks, and the halls themselves can all be built in the United States, and federal money should keep that chain here. AI, he thinks, is how the country competes without cutting wages: per-capita output of parts, chips, and equipment up a thousand times, even a hundred thousand. When he was in graduate school, Cornell had four or five machine shops and every freshman engineer learned a mill. Now almost every professor has a 3D printer. He expects AI to do that to the factory floor. His father-in-law was a World War II radar operator who chased German submarines off Florida; the boats could not understand how they kept being found. Radar won the Battle of Britain. New technology, RF then and AI now, always hits harder than the first users guess.

Guo asks whether transformers and diffusion models will matter in their field. Both founders light up. Mitchell already uses AI as a research assistant. The first move in a core-technology problem is finding what was already tried. The key to a solution might sit in a 1963 paper that went nowhere because the author could not run enough iterations. Pull that work forward, attach modern loops, and the pace of innovation changes. Inverse design, chemical-space search, faster DFT: he treats them all as question-and-iterate problems, the same as medicine or code. What used to block people was the cost of a run. A partner that can do a billion iterations a second changes which hypotheses you can even start.

Ejeckam pushes the thought experiment. Entrepreneurs, he says, imagine inside today's constraints. Remove the thermal envelope, run processors a billion times faster, collapse a hundred years of modeling into a second, and the hard part is asking the right question. On his list: drugs dialed to every form of cancer; a battery that takes a car from San Francisco to New York on one charge. "We're limited by our ability to frame the questions, and only by our own imagination." That is the note they leave on: heat is a physics limit, diamond is one way back into the die, and AI might be how materials science stops waiting decades between ideas.