Illustration: Dall-E
Every memory cell in a chip needs a gatekeeper: a device that stays firmly shut until it is called on, then opens instantly. Finding one that does both jobs reliably enough, at the scale chips demand, has stumped the semiconductor industry for decades. A team at the University of Southern California has built one.
In a study published in Nature Electronics earlier this month, researchers led by Professor J. Joshua Yang from the Ming Hsieh Department of Electrical and Computer Engineering at USC Viterbi School of Engineering and USC Stevens Schools of Computing and AI report a new kind of gatekeeper for memory cell, called selector or access device, the device that controls access to each memory cell, that outperforms the previous best by a factor of 100 and could enable memory arrays, the grids of cells that make up a chip, orders of magnitude larger than what was previously possible. This is because the new selector allows memory cells not only to become smaller but also to be stacked vertically, like floors in a high-rise building.
“Memory is the bottleneck for AI hardware,” Yang said. “And the bottleneck for memory is not the memory device itself. It is its gatekeeper.”
The implications reach across the semiconductor industry. It could help shrink the energy consumption of AI data centers, make it possible to run AI tools directly on a phone without relying on an internet connection, and eventually power computing systems in places where connectivity is impossible, including space.
The gatekeeper problem
In current chips, the selector’s role is played by a transistor, a three-terminal device that is simply too large to pack densely and cannot be stacked in three dimensions because it can only be built on a silicon substrate.
The new selector is built from five atomically thin layers of van der Waals materials, a family of two-dimensional crystals that can be stacked like sheets of paper one atom thick, the same family that produced Yang’s heat-resistant memory device reported in Science earlier this year. A human hair is roughly 10,000 times thicker.
The key innovation is the shape of the barrier brought in by the gatekeeper that controls whether current passes through. Previous selectors used a flat barrier that electricity had to climb over. Yang’s team used a graded, triangular shape instead.
“When you apply voltage, not only does the barrier become thinner, but the height also becomes lower, simultaneously,” Yang said. “Both increase the current. That is why it performs so much better.”
What the numbers mean
The device is 100 times better at distinguishing between its open and closed states than the previous best selector. It also survived more than one trillion switching cycles before the team stopped counting, not because the device failed, but because measuring further would have taken months. In principle, Yang said, these devices have no endurance limit, because only electrons move during operation, with no atoms displaced and no physical wear on the material.
The device also barely changes behavior with temperature, an important practical requirement for chips that run hot. And unlike previous selectors, it performs nearly identically from one memory cell to the next because of its atomic precision in fabrication control, which matters in a chip with billions of cells.
From the data center to your pocket, and beyond
AI tools like ChatGPT currently cannot run on a phone. They require too much memory and too much power, so they live in data centers and communicate over the internet. That means every question sent to an AI assistant travels to a remote server, gets processed there, and comes back, which costs energy, costs money, and requires sharing personal data with a third party.
With dramatically denser memory chips, that changes.
“You may be able to put the entire large language model on your phone someday,” Yang said. “You don’t have to share all your things to the cloud.”
The implications reach further still. Data centers consume enormous amounts of energy, and the industry has been searching for ways to move computing to places where power is cheaper and cooling is easier, including space.
“Elon Musk said we’re going to build a data center in space, and now it becomes more and more real,” Yang said. “But only if you have high temperature electronics can you talk about a data center in space. Otherwise, you cannot dissipate your heat efficiently, because space is a vacuum, where heat can be rejected only through radiation—and radiated power increases with the fourth power of the radiator (i.e., the chip) temperature.”
Two puzzle pieces falling into place
The high-temperature memory chip Yang’s lab reported in Science earlier this year was one piece of that puzzle. This selector, made from the same family of materials, is expected to withstand similar temperatures, though that has not yet been measured systematically.
“Previously, we had the memory cell but not its gatekeeper,” Yang said. “Now we have both.”
The devices are still assembled by hand in the lab. The materials used are relatively new to semiconductor manufacturing, and scaling to the 12-inch wafers foundries require will take time. Yang is candid about what the paper represents: a direction proven to work, not yet a product.
“A selector is the key for memory,” Yang said. “Memory is the key for electronics, for AI. This is one of the biggest challenges in electronics, and this research provides a promising avenue forward.”
The research was supported by the Army Research Office and the Air Force Office of Scientific Research. Collaborators included researchers from the University of Florida, the Air Force Research Laboratory, and the National Institute for Materials Science in Japan.
Published on September 21st, 2026
Last updated on September 21st, 2026