Share
The AI chip race has a memory problem, and a new startup thinks the fix belongs on the roof. Nagare Labs, a Georgia Tech spinout with former Intel CEO Pat Gelsinger on its board, says its Z2RAM technology can deliver SRAM-like speed at DRAM-like density, stacked directly above AI chips.
The pitch targets one of the toughest bottlenecks in modern computing. The ambitious part is the timeline, which stretches toward the end of the decade and beyond.
What Nagare Labs is actually building
Nagare Labs unveiled Z2RAM on October 26, 2026. The same day, the company announced a $5 million seed round led by Playground Global.
The startup was founded in 2026 by Georgia Tech professors Suman Datta and Shimeng Yu. Gelsinger serves as a board director and has praised Z2RAM for combining the strengths of DRAM, SRAM, and flash memory.
Those three memory types normally force engineers into trade-offs. SRAM is fast but space-hungry, DRAM packs more bits into less space but runs slower, and flash keeps data without power but is slower still.
Z2RAM tries to sidestep that choice with a two-transistor memory cell built on oxide semiconductors. Crucially, the design has no capacitor, the tiny charge-storing component that conventional DRAM depends on.
Removing the capacitor matters because it opens the door to fabrication using standard back-end processes. Back-end steps are the layers built on top of a chip after the transistors are formed, which is what lets Z2RAM sit directly above an AI accelerator’s logic die.
The numbers, and what they mean
Nagare Labs has laid out projections for its first-generation product. With a six-layer stack, Gen 1 is projected to reach more than 5x the bit density of 2 nm high-density SRAM.
Access speed is anticipated to be 14x faster than current DRAM. Read energy is designed to land at roughly one-third of standard SRAM.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
Those are projections, not shipping specs. The founders’ own published papers show a cell that has been improving quickly, and demonstrations have confirmed data retention exceeding 300 seconds at temperatures up to 85 degrees Celsius.
That retention figure deserves some attention. Conventional DRAM has to refresh its stored charge constantly, which costs energy and time, so a cell that holds data for minutes at elevated temperatures hints at a lighter refresh burden.
The 85-degree ceiling also matters because AI accelerators run hot. Memory that sits directly on top of one has to survive the heat coming up from below.
The memory wall, explained
Nagare Labs frames Z2RAM as an answer to the so-called memory wall in AI systems. The term describes a widening gap between how fast processors can compute and how fast memory can feed them data.
AI accelerators have grown dramatically more powerful at raw math. Moving model weights and data in and out of memory has become a defining constraint on how efficiently those chips actually run.
Stacking memory in 3D directly over the logic is one way to shorten the fetch. Nagare’s bet is that its oxide-semiconductor approach can do this using processes chipmakers already know, rather than requiring exotic new factory equipment.
A long road from lab to fab
Nagare Labs is targeting working memory-tile samples within four years, putting that milestone around 2030. Mass production is aimed for within a decade.
Gelsinger’s involvement adds credibility and connections. He spent decades in the chip industry before leading Intel, and his presence on the board signals that seasoned insiders see the concept as worth backing.
Playground Global leading the round is also notable. The firm is known for backing deep-tech hardware, a category where investors expect long timelines before revenue.
What this means for the AI hardware market
The risks are equally clear. Projected multiples against SRAM and DRAM are not the same as validated silicon, and incumbent memory makers are not standing still while a startup spends four years getting to samples.
Readers tracking the space should watch three things. The first is whether Gen 1 metrics hold up in multi-layer prototypes rather than single cells, and the second is whether Nagare announces a manufacturing partner.
The third is retention and reliability over time. Holding data for over 300 seconds at 85 degrees Celsius is a strong lab result, but production memory must perform across billions of cells, for years, under real workloads.
