TetraMem memristor survives 700 degrees Celsius as startup moves AI chips to production wafers
The TL;DR
USC researchers have created a memristor that operates at 700 degrees Celsius, hotter than molten lava and above the temperatures of Venus, while TetraMem, a startup that sells the technology, is already moving room-temperature AI chips to 300mm production wafers with support from SK hynix and the CHIPS Act.
All human probes sent to Venus are dead. The occupants of the Soviet Venera survived between 23 minutes and two hours in an environment where the temperature exceeded 460 degrees Celsius. Their electronics, designed to withstand heat that can melt lead, still fail. The longest-lived mission in Venus exploration history lasted 127 minutes. Then the chips stopped working and the data stopped moving.
A team at the University of Southern California has created a memory chip that works reliably at 700 degrees Celsius, hotter than molten lava, and more than 200 degrees hotter than anything Venus can throw at it. The device, published in Science on March 26, 2026, held data for more than 50 hours at that temperature without refreshing, survived more than a billion switching cycles, and ran at 1.5 volts with a switching speed measured in tens of nanoseconds. Seven hundred degrees was not the limit of the device. It was the limit of the experimental equipment.
TNW City Coworking Space – Where your best work happens
A workplace designed for growth, collaboration, and endless networking opportunities at the heart of technology.
The device is a memristor, and the company that sells it, TetraMem, is already building smart AI chips that do machine learning with a speed and efficiency that conventional hardware can’t match. The extreme temperature version was dangerous. Its effects are non-existent.
Device
A memristor is a nanoscale component that stores information and performs calculations at the same time. The device developed by Joshua Yang’s team at USC consists of three layers: tungsten on top, hafnium oxide ceramic in the middle, and a single-thick graphene sheet on the bottom. Tungsten has the highest melting point of any metal. Hafnium oxide is a common insulator in semiconductor manufacturing. Graphene is a form of carbon, like diamond, that can withstand high temperatures without degrading.
In conventional memory devices, heat causes metal atoms from the top electrode to migrate through the ceramic layer to the bottom electrode, creating a permanent short circuit that kills the device. Graphene prevents this. Its surface chemistry with tungsten, as Yang said, is almost like oil and water. The tungsten atoms can’t find anything to focus them on and they just leave. No anchor, no short circuit, no failure.
The team didn’t just watch the result. Using electron microscopy, spectroscopy, and quantum-level computer simulations, they mapped the atomic interactions between graphene and tungsten to better understand why it works. That mechanistic understanding means that other materials with similar surface chemistry can now be identified, potentially making the device easier to manufacture on an industrial scale. Two of the three materials, tungsten and hafnium oxide, are already standard in semiconductor foundries around the world. Graphene is on the development roadmaps for both TSMC and Samsung.
AI communication
The effect of extreme temperature is the main one, but the commercial importance of the memristor lies elsewhere. More than 92 percent of the computing in AI systems is matrix multiplication, the core mathematical operation behind everything from image recognition to large language models. Today’s digital processors do it sequentially, step by step, consuming enormous amounts of energy. The memristor does it physically. When electricity flows through a device, Ohm’s Law, voltage multiplied by conductance, produces a response such as current. Repetition occurs instantaneously when the current is passed. There are no clock cycles. There is no bus to remember. No power is wasted shuffling data between the processor and storage.
This is in-memory computing: the data resides where the computation takes place, removing the von Neumann bottleneck that hinders all conventional processor architectures. The result is a calculation that is orders of magnitude faster and more energy efficient than GPU-based systems performing the same calculations.
The International Energy Agency projects that energy consumption in data centers will double by 2026, driven strongly by computing demands for AI training and interpretation. The AI industry’s response has been to build massive data centers, secure more power, and negotiate nuclear power contracts. The memristor-based architecture attacks the problem at a different level: not by providing more power on the same type of chip, but by creating a chip that requires orders of magnitude less power to perform the same calculations.
The demand for AI has driven a 90 percent increase in memory prices and a global DRAM shortage, forcing manufacturers to redirect capacity to high-bandwidth memory for AI accelerators. The memristor represents a very different approach. Instead of splitting the memory for processing and closing data between them at a huge cost of energy, it combines them. The architecture does not compete with DRAM in capacity. It competes with GPUs for the AI load of self-targeting.
Company
Yang co-founded TetraMem with three other original memristor research authors: Qiangfei Xia, Miao Hu, and Ning Ge. The company, based in San Jose, has built memory-based computer chips that students in Yang’s lab use every day to perform machine learning tasks. TetraMem has a partnership with SK hynix, the world’s second largest memory manufacturer, in a joint research project to develop in-memory computing for AI; by Andes Technology to combine its memristor architecture with the RISC-V vector processor; and NY CREATES at the Albany NanoTech Complex, where the company has successfully scaled its technology from 200mm to 300mm wafers, a standard industrial platform for mass production.
The NY CREATES partnership is very important. It was founded under the mission of CHIPS and the Science Act to strengthen the domestic semiconductor ecosystem, and featured what NY CREATES calls the split-fab model: companies develop and test chips in Albany before handing over the processes to a partner who is the founder of mass production. TetraMem memristors are no longer a laboratory curiosity. They are on 300mm wafers.
The US government’s investment in the CHIPS Act has reshaped the domestic semiconductor scene, with billions flowing into logic chip manufacturing. TetraMem’s approach through NY CREATES shows that the Act’s ambitions go beyond logic: infrastructure built to reproduce chips also enables fundamentally new computing architectures to reach production scale.
The market
The global memristor market was estimated at 420 billion dollars in 2025 and is expected to reach 4.5 billion dollars in 2030 and 21.7 billion dollars in 2035, growing at a compound annual rate of more than 48 percent. The broad analog AI chip market is expected to grow from 251 million dollars in 2025 to 2.5 billion dollars by 2035. The numbers pale in comparison to the 600 billion dollars Nvidia alone has generated in market capital from AI chip demand. But they represent the first stage of an architectural revolution.
Competitors include Mythic AI, Rain Neuromorphics, and a growing number of research labs at TSMC, Samsung, and KAIST that are building memristor crossbar arrays for the edge. TSMC’s mixed-precision processor scored 91.2 percent yield and 85 percent accuracy in image classification benchmarks. Asia-Pacific handset manufacturers are committed to embedding analog compute chips in flagship devices by 2026. Technology moves from paper to products.
The border
The high-temperature version of the memristor opens a category of computing that does not exist yet: the concept of the AI site in places where ordinary electronics cannot survive. A Venus lander equipped with memristor-based processors can analyze air samples, classify the composition of the earth, and make autonomous decisions without transmitting raw data to Earth and waiting for instructions. A geothermal drilling system can process sensor data at depths where the surrounding rock glows red. A nuclear reactor can use AI for diagnostics inside its containment vessel.
Researchers have proposed placing data centers in space to address AI’s energy needs, using space in orbit for cooling and solar energy for power. The memristor reverses the problem. Instead of moving data centers into space, it takes computing to the place where the data originates, whether that place is the surface of Venus, the inside of a jet engine, or the core of a fusion reactor.
NASA’s High Performance Spaceflight Computing processor, developed by Microchip Technology, delivers 500 times the performance of radiation-hardened spaceflight chips. But it was designed to cover the cold space of interplanetary travel, not a planetary furnace. The memristor lives both. A device rated at 700 degrees is virtually invulnerable to the 125-degree peaks that automotive computers regularly experience, the highly radioactive environment of deep space, or the hot rotation of low-Earth orbit.
Europe’s semiconductor industry has called for a fast-track Chips Act 2.0 to fund next-generation manufacturing capabilities beyond conventional logic and memory. Memristor-based in-memory computing is its own type of architecture that cannot be supported by such investment: European fabric technology that does not depend on access to Nvidia’s GPU supply chain or TSMC’s most advanced logic.
Yang was careful not to guard the timeline. Memory alone does not make a complete computer. High-temperature logic circuits must be developed and integrated around them. Current devices are built by hand in a small laboratory. But the missing part is done. The chip that survived the hotter temperatures than the mud became dangerous. There was no company to sell it.
