A memristor built at the University of Southern California kept storing data at 700°C for more than 50 hours and endured over one billion switching cycles without failure. The study, published in Science on March 26, 2026, describes a memory device that operates at roughly three and a half times the temperature at which conventional electronics begin to break down.

Joshua Yang, Arthur B. Freeman Chair Professor at USC's Ming Hsieh Department of Electrical and Computer Engineering, led the team. "You may call it a revolution," Yang said. "It is the best high-temperature memory ever demonstrated." Notably, 700 degrees was simply the ceiling of the team's test equipment. The device showed no sign of failure at that limit.

What is the device made of?

First author Jian Zhao built the component as a microscopic layered structure: a tungsten top electrode, a hafnium oxide ceramic layer in the middle, and graphene as the bottom layer. Tungsten carries the highest melting point of any element; graphene contributes exceptional heat resistance at a single atom's thickness.

The performance figures are concrete. The device retained data for more than 50 hours at 700 degrees without refresh, endured over one billion switching cycles at that temperature, and operated at 1.5 volts with switching speeds measured in tens of nanoseconds. Modern electronics typically start to fail above roughly 200°C.

Why does it survive the heat?

The breakthrough arrived partly by accident. The team originally attempted a different graphene-based device, which did not work. "To be honest, it was by accident, as most discoveries are," Yang said. "If you can predict it, it's usually not surprising, and probably not significant enough."

The failure mechanism in conventional devices is atomic migration. Heat drives metal atoms from the top electrode through the ceramic until they reach the bottom electrode, forming a permanent conductive bridge that short-circuits the device and leaves it stuck on. Graphene blocks this. Yang compared the tungsten–graphene interaction to oil and water: tungsten atoms approaching the graphene surface cannot attach, so they drift away rather than settle into a bridge.

The researchers confirmed the mechanism with electron microscopy, spectroscopy, and quantum-level simulations. Yang noted that other materials with similar surface properties could be identified, which may help scale the technology for industrial production.

Where could it be used?

Electronics rated above 500°C have long been a goal for space exploration. Venus has a surface temperature near that level, and every lander sent there has failed in part because of heat. Geothermal energy systems, nuclear and fusion installations, and deep industrial sensing present similar demands. Yang pointed to a mundane case as well: a device rated for 700 degrees would be extremely robust at the roughly 125°C found inside automotive electronics.

"We are now above 700 degrees, and we suspect it will go higher," Yang said.

What does it mean for AI?

Memristors can compute where they store. Instead of performing matrix multiplication step by step, the device exploits Ohm's Law — voltage times conductance equals current — so the result appears instantly as the measured current. The energy savings are the point.

"Over 92 percent of the computing in AI systems like ChatGPT is nothing but matrix multiplication," Yang said. "This type of device can perform that in the most efficient way, orders of magnitude faster and at lower energy."

Commercialization has already begun at room temperature. Yang and three co-authors — Qiangfei Xia, Miao Hu, and Ning Ge — co-founded TetraMem, which builds memristor-based AI chips. Yang's lab is already running machine learning tasks on working TetraMem silicon. The high-temperature variant could extend that capability to spacecraft and industrial sensors that process data on site.

How far from production?

Yang is candid about the gap between laboratory proof and product. Memory is one component of a complete system; high-temperature logic circuits still need to be developed and integrated. The current devices were assembled manually at very small scales, so volume manufacturing will take time.

"This is the first step," Yang said. "It's still a long way to go. But logically, you can see: now it makes it possible. The missing component has been made."

The materials picture is favorable. Tungsten and hafnium oxide are already standard in semiconductor production. Graphene is newer, but TSMC and Samsung are actively developing it, and it has been produced at wafer scale in research settings.

Who funded the work?

The research ran through the CONCRETE Center — Center of Neuromorphic Computing under Extreme Environments — a multi-university Center of Excellence led by USC and supported by the Air Force Office of Scientific Research and the Air Force Research Laboratory. Key experiments took place with Dr. Sabyasachi Ganguli's team at the AFRL Materials Lab in Dayton, Ohio; theoretical analysis involved USC researchers and collaborators at Kumamoto University in Japan.

Yang framed the Science publication as one step in a larger push. "Space exploration has never been so real, so close, and at such a large scale," he said. "This paper represents a critical leap into a much larger, more exciting frontier." With the 700°C ceiling set by test hardware rather than the device itself, the next round of measurements will likely define how far this materials stack can actually go.