Seven hundred degrees are around the lavadeep geothermal wells, probes that arrive on hostile worlds and last very little. They’re also at the point where traditional electronics usually give up. This is why the result arrived from California weighs more than a laboratory curiosity: in a study published on Science on March 26, 2026, the team led by Joshua Yang described a nanometric memory capable of reliably working at 700°Cwell beyond the range around 200 °C where conventional devices begin to fail.
The component belongs to the dei family memristorstiny devices that hold memory and computation together in the same structure. Here lies a substantial difference compared to classical electronics: instead of relying on electric charges that suffer from heat, the memristor conserves information by persistently modifying its resistance. In practice, it remains in the set state even without power. The prototype created by the team has a three-layer structure: tungsten top electrode, hafnium oxide in the center, graphene at the bottom. Two of these materials, tungsten and hafnium oxide, are already part of standard semiconductor foundry processes; graphene adds the decisive piece.
The numbers alone explain why this work has attracted so much attention. At 700°C the device retained data for over 50 hours without refresh, it passed more than one billion switching cyclesworked on 1.5 volts and with response times in the order of tens of nanoseconds. There is also a detail that matters a lot: 700 degrees coincides with the limit of the test instrumentation, so the maximum threshold of the component still remains to be measured.
The trajectory of discovery has the rough flavor of certain real turning points. The group was trying to build another graphene-based device. From there something unexpected emerged, and from that fertile error the right direction started. This often happens in serious research: a door remains closed, then a window opens that no one was looking. Yang said it with a frankness that sounds almost liberating: it happened by chance.
The detail that changes everything lies in a very thin interface
The fault that throws traditional memristors off track is quite clear. With heat, the metal atoms from the top electrode migrate through the ceramic layer and reach the bottom electrode. At that point a permanent conductive bridge is created, the device remains stuck in the on state and the game ends there. In this architecture, however, graphene behaves like a surface on which tungsten struggles to anchor itself: the atoms arrive at the interface, slide away, disperse, and the stable bridge that would cause the short circuit does not form. The analysis with electron microscopy, spectroscopy and quantum simulations has strengthened this interpretation of the phenomenon.
This technical aspect has a value that goes beyond the single prototype. The work suggests a design principle: look for other interfaces with similar surface properties, so as to obtain high-temperature devices that are easier to replicate and, in the future, more industrializable. The best result, in these cases, comes when an experiment stops being a record number and becomes a useful rule for those who come after.
Then there’s the chapter artificial intelligencewhich really matters here. Memristors have been of interest to those working on neuromorphic computing and efficient hardware for years, because matrix multiplication, one of the most used operations in AI systems, can be performed in an analogue way by exploiting Ohm’s law: voltage and conductance directly generate the current, and the result arrives as the electricity passes through the lattice of the devices. Compared to digital processors, which perform these operations step by step while consuming a lot of energy, the theoretical advantage lies in one orders of magnitude higher energy efficiency. Said in a less academic way: less waste, fewer steps, more work done where it is needed.
From Venus to the reactors, passing through the car
The possible applications read almost like a catalog of places hostile to electronic life. The surface of Venus travels around the 465°C with pressures reaching approx 92 times terrestrial ones, and the landers that touched the planet resisted for very short times, in the order of hours. A device that works above 700 degrees immediately changes the mental picture: longer missions, sensors capable of remaining active, local data processing in environments where today everything shuts down quickly. The same goes for deep geothermal drilling, nuclear plants, fusion systems and even the automotive industry, where the control electronics operate at enormously lower temperatures but would benefit from much wider reliability margins.
On the industrial front, a first bridge with the market already exists. Yang, along with co-authors Qiangfei Xia, Miao Hu and Ning Ge, co-founded TetraMema company working on memristor AI chips for room-temperature applications, and the lab already uses working components made by the company for machine learning tasks. The high temperature version fits into that trajectory, with a very concrete idea: processing on site on probes, sensors and industrial systems, without having to send everything elsewhere for calculation.
Here, however, you need to keep your feet on the ground. We are faced with a laboratory prototypeassembled by hand on a sub-micrometer scale. Memory alone does not build a complete computer: it also requires logic circuits compatible with extreme temperatures, system integration, stable manufacturing processes and years of development. The encouraging side remains strong: tungsten and hafnium oxide are already familiar materials to the industry, wafer-scale graphene production has been demonstrated in research, and major industry roadmaps are already exploring low-dimensional materials and 2D architectures to extend future miniaturization.
The research was born inside the CONCRETE Centera multi-institutional center led by USC and supported by the Air Force Office of Scientific Research and the Air Force Research Laboratory. The main experimental measurements also took place at the Materials Lab in Dayton, Ohio, while the theoretical part involved researchers from USC and the University of Kumamoto. In short, the picture is already that of the technologies that matter: many groups, a long supply chain, a lot of money, potentially enormous applications. Three open questions remain, and they are the right ones: how far does the operational limit really go, what other combinations of materials can replicate this behavior, and how much is left until integration into complete systems. The answer will come in the next few years. The first brick, meanwhile, is already red-hot.
