The Chip That Laughs at Lava

In 1982, the Soviet Venera 13 lander touched down on the surface of Venus, snapped a few photos of cracked orange rock, and died after two hours and seven minutes. Its electronics, cocooned inside a pressurized vessel filled with melting lithium nitrate to soak up heat, simply could not survive the planet's 460°C surface for any longer. That two-hour record has stood for more than four decades. Every Venus lander since has been a study in how aggressively you can refrigerate a chip while it slowly cooks anyway.

A team at the University of Southern California just made that entire approach look quaint.

In a paper published in Science on March 26, 2026, USC engineers led by Joshua Yang and first author Jian Zhao demonstrated a memory device that runs at 700°C without breaking a sweat. No pressure vessel. No cooling. No clever heat-sink trickery. The chip just sits there at a temperature that would melt aluminum, lead, zinc, and a long list of other metals you would normally find inside electronics, and it keeps storing and switching data. The team did not even find the device's failure point. They stopped at 700°C because that was the upper limit of their test equipment.

A sandwich, a hostile surface, and a happy accident

The device is a memristor, a two-terminal component that stores information as a resistance state rather than as charge. Think of it as a switch that remembers what position it was last in, even with the power off, and can also do certain kinds of math directly through Ohm's law rather than by shuffling bits through a CPU. Memristors have been a darling of AI hardware research for years because they collapse the wall between memory and compute, the wall that makes large language models so absurdly hungry for power.

Physically, the USC chip is built like a microscopic club sandwich. The top slice is tungsten, which has the highest melting point of any element at 3422°C. The middle is hafnium oxide, a ceramic insulator already used throughout the semiconductor industry. The bottom is a single sheet of graphene, one carbon atom thick. The finished devices range from 200 nanometers by 1 micrometer up to 1 square micrometer, small enough that you could fit several million of them on a postage stamp.

The reason this stack works is one of those discoveries that makes physicists grin. In a normal high-temperature memristor, atoms from the top metal electrode get restless under heat. They drift down through the insulating layer and eventually form a conductive thread of metal that bridges both electrodes. This is called a filament, and once one forms permanently, the device is dead. Short-circuited. Useless.

Graphene blocks that migration in a way that nobody quite expected. Yang describes the relationship between tungsten and graphene as oil and water. The tungsten atoms reach the graphene surface, find no good place to anchor, and refuse to clump together into a filament. The hostile interface is the whole trick. No filament, no short circuit, no failure.

Yang is candid about how they got there. They were working on a different graphene design. It was not behaving the way they wanted. In trying to figure out why, they noticed something odd, ran more tests, and ended up with a device that broke every published record for high-temperature memory. He told reporters, more or less, that if you can predict a discovery in advance, it usually is a minor discovery.

What the numbers actually look like

The headline figure is 700°C, but the supporting data is what should make engineers sit up. At that temperature, individual devices held their stored state for over 50 hours without needing a refresh. When the team tested 30 separate devices, the average retention time worked out to roughly 145 hours, with the best units lasting 170 hours. They demonstrated more than one billion switching cycles at 700°C. Switching voltage stayed around 1.5 volts. Pulse widths landed in the tens of nanoseconds, with the team reporting around 30 nanoseconds. The ON/OFF current ratio held above three orders of magnitude across the entire range from room temperature to 700°C.

For anyone who has worked with memory hardware, those are not lab-curiosity numbers. Those are numbers that say, "We could build something with this."

Compare that to where the field was last week. NASA's Glenn Research Center has spent years developing silicon carbide integrated circuits that can survive Venus-like conditions. In 2016 they ran a SiC ring oscillator for 521 hours at 460°C inside their Glenn Extreme Environments Rig, a 20-ton chamber built specifically to simulate the Venusian surface. That was a genuine breakthrough at the time, more than a hundred times longer than any previous Venus electronics had managed. But silicon carbide topped out around the upper 400s. The USC memristor operates at 240°C above that and is also a memory device, not just a logic circuit. Different beast entirely.

There is one important footnote. A memristor is a memory and compute element, not a complete computer. To build something useful at 700°C, you also need logic circuits, interconnects, and packaging that can survive the same conditions. Those pieces do not exist yet. Silicon carbide and gallium nitride logic exists for the lower extreme-temperature range, but pairing it with this new memory at 700°C is its own engineering project. The current devices are also handcrafted in a research cleanroom one at a time. Mass production is a separate mountain to climb.

That said, two of the three materials in the stack are already at home in commercial fabs. Tungsten is everywhere in modern chips. Hafnium oxide has been a standard high-k dielectric for nearly two decades. Graphene is the new ingredient, but TSMC and Samsung have both been working on wafer-scale graphene production, and research labs have been pulling off square meters of the stuff for years now. The supply chain is closer than it looks.

Why AI people care more than space people

If the only application were Venus landers, this would still be a beautiful piece of material science. But Venus landers are not really why Yang's lab spun out a startup called TetraMem to commercialize the work.

The pitch is matrix multiplication. According to Yang, more than 92% of the computation in a system like ChatGPT consists of multiplying matrices together, the same operation, over and over, billions of times per query. On a conventional GPU, every multiplication means moving numbers from memory into a multiplier, doing the math, then writing the result back. The energy cost of moving the data dwarfs the energy cost of the actual arithmetic. This is the famous memory wall, and it is the main reason AI data centers consume small countries' worth of electricity.

A memristor crossbar array does the math differently. You apply voltages along the rows, the resistance values stored in each memristor multiply those voltages, and the resulting currents sum together along the columns. Ohm's law plus Kirchhoff's law, executed in the analog domain at the speed of electricity. No data movement. No clock cycles. The answer just appears as a current, all at once, for the entire matrix.

The performance numbers people throw around for in-memory computing of this kind are several orders of magnitude better in both speed and energy than what GPUs can do. Whether the USC team's specific device hits those numbers in production is unproven. But the architecture is real, the physics is real, and competitors like Mythic and Rain AI have been chasing the same prize for years.

The 700°C tolerance changes the conversation in a subtle but important way. It means you can stop spending half your power budget on cooling. Modern data centers burn roughly 40% of their electricity on chillers, fans, and air handlers just to keep silicon below its operating ceiling. A chip that does not care about temperature is a chip that can run in places where you could never put a server today. Inside an engine bay. On a turbine blade. Down a borehole. Pressed against the heat exchanger of a small modular reactor. Anywhere a sensor wants to be a computer instead of just shipping its data somewhere cooler.

A car engine bay hits about 125°C in normal operation. That is now well within the comfortable middle of this device's operating range, not anywhere near a stress condition. Geothermal drilling tools that currently have to be hauled up every few hours when their electronics give up could potentially stay down indefinitely. A nuclear reactor could carry instrumentation directly inside the containment vessel rather than running miles of shielded cable out to a control room.

For Venus, the implications are obvious. NASA has flirted with a long-duration Venus surface station for years, and the Russian Venera-D mission, currently targeted for no earlier than 2029, has been kicking around proposals for a 24-hour surface lander in collaboration with NASA. The bottleneck has always been electronics that can think for more than two hours without melting. A memristor that runs at 700°C with billion-cycle endurance is not the whole answer, but it removes one of the biggest blockers.

The work was done through USC's CONCRETE Center, the Center of Neuromorphic Computing under Extreme Environments, which is funded primarily by the Air Force Office of Scientific Research and the Air Force Research Laboratory. The experimental side ran through Sabyasachi Ganguli's team at the AFRL Materials Lab in Dayton, Ohio. The Air Force's interest in extreme-environment electronics is not a mystery. Hypersonic vehicles generate enormous amounts of heat from atmospheric friction. So do the hot sections of jet engines, where embedded sensors could enable far more efficient combustion if only the sensors could survive. There is a long unclassified literature on why DARPA and AFOSR have been quietly funding this research for two decades.

Yang's quote about the discovery being a leap toward a more exciting future is the kind of thing principal investigators say when their grant manager is in the room. The more interesting reaction came from one of his colleagues, who noted that fewer than ten labs in the world can even test devices at 700°C, and most of them are in materials science departments rather than electrical engineering. The USC group happened to be in the right department, with the right collaborators, when an experiment that was not working revealed something that worked spectacularly.

Sometimes that is how the technology that ends up running our lives gets built. A graduate student fiddles with a stack of three materials, looks at a result that should not be possible, and rewrites the temperature limit of computing by 500 degrees in a single afternoon.

The Venera 13 record might finally fall.