← All posts

Blog · On agents · November 12, 2026

A five-million-dollar memory bet

Nagare Labs raised $5 million to build Z2RAM, a memory that stacks on top of AI chips. Memory supply is a disaster right now, and I think that makes it a good time to try.

By Timothy O’Brien3 minute readFiled under On agents

Nagare Labs came out of stealth on October 26 with $5 million and a memory technology called Z2RAM. It is a Georgia Tech spinout founded by two professors there, Suman Datta and Shimeng Yu. Playground Global led the round, and Pat Gelsinger, the former Intel CEO and a general partner at Playground, joined the board.

What Nagare Labs is building

An AI chip has two kinds of memory to work with today. SRAM sits on the chip and is fast, but it uses a lot of chip area per bit. HBM holds far more, but it sits beside the chip, and moving data back and forth costs time and power. Z2RAM is meant to go between them. The cell is two transistors and a shield, with no capacitor. Nagare says it can deposit the memory in layers directly on top of the processor, as 3D NAND flash adds layers, using standard 300mm equipment and no EUV. The first version is planned at six layers, and the roadmap goes past 100.

None of that is a product yet. Tom’s Hardware went through the founders’ published papers and found that the measured results come from related research cells. The best of them held data for more than 300 seconds at 85°C. Nagare declined to share measured numbers for Z2RAM itself, and the first-generation figures it quotes, five times the density of SRAM and fourteen times the speed of DRAM, are projections. Datta told The Register the company is at “literally lab prototypes,” with samples targeted within four years and production within the next decade. It has no fab and is looking for a partner. Capacitorless memory has also been tried before. AMD licensed a design called Z-RAM about twenty years ago and did not pursue it.

The squeeze behind it

Memory supply is a disaster right now. Samsung, SK Hynix, and Micron have shifted production toward HBM and server memory for AI data centers, which leaves less of everything else. IDC described it last December as “a potentially permanent, strategic reallocation of the world’s silicon wafer capacity.” Micron closed Crucial, its 29-year-old consumer brand, to keep supply for its larger customers. A CSIS report in September cited forecasts that data centers will take about 70 percent of all memory produced this year, that manufacturers will meet only about 60 percent of demand, and that DRAM prices on new automotive contracts will run 70 to 100 percent higher.

Why this is when you bet

I think a shortage like this is a good time for a hardware bet, even one this early. When memory is cheap and easy to get, a chip designer has little reason to qualify an unproven memory from a startup with no fab. When suppliers can cover about 60 percent of demand, that designer has a reason to take the meeting. It also helps that the design is meant to run on equipment fabs already have. It doesn’t need EUV, and the layers are deposited on one die instead of bonded together from separate dies.

I wouldn’t put odds on this company. IDC expected the shortage to last well into 2027, and by the founders’ own estimate production is years past that. If IDC is right that the reallocation is permanent, the timing still works. If the market swings back to a glut, as memory markets have before, it gets much harder. Either way, people who know chips are funding new memory designs at the prototype stage in the middle of a shortage.

I’ve spent the fall writing about agent memory as software: who writes it and who can correct it. All of that software runs on physical memory, and this year there isn’t enough of it. Nagare is the only hardware entry in the registry so far.