Z2RAM: Nagare Labs memory aims to combine SRAM speed with DRAM density for AI

Nagare Labs, a startup spun out of research at Georgia Tech, emerged from stealth on October 26, 2026 with the announcement of Z2RAM, a new memory architecture designed to attack the so-called "memory wall" that limits artificial intelligence accelerators. On the same day, the company announced a $5 million seed round led by Playground Global.

Among its backers is Pat Gelsinger, former Intel CEO and a general partner at Playground Global, who sits on Nagare Labs' board. The startup was co-founded by Georgia Tech professors Suman Datta and Shimeng Yu.

What Z2RAM is

In AI accelerators, the processor usually computes faster than memory can deliver data — and fetching a value from HBM memory takes roughly 1,000 times more energy than operating on it, according to Nagare Labs. SRAM is fast but has little capacity; DRAM and HBM store more, but sit off-chip, with higher latency and energy costs.

Z2RAM tries to escape that dilemma with a two-transistor cell built from oxide semiconductors, with no capacitor — the component conventional DRAM uses to store charge and which is reaching its physical scaling limit. A write transistor places charge on the storage node, and a read transistor senses it without draining it, enabling fast, non-destructive reads. The memory layers are deposited directly on top of the accelerator, on the same die, instead of being stacked beside it.

The numbers are projections, not measured results

It is important to separate what was announced from what was demonstrated. For the first generation, with six layers, the company projects density more than 5 times greater than 2 nm high-density SRAM, access 14 times faster than DRAM and read energy 3 times lower than SRAM. These figures come from layout simulations, not measurements on a commercial product. Laboratory demonstrations have confirmed data retention of more than 300 seconds at temperatures up to 85 °C.

The technology is still at the laboratory-prototype stage. Nagare Labs is seeking a manufacturing partner and aims for working memory-tile samples in about four years, around 2030, with mass production within a decade. In other words: even if the promises hold, Z2RAM will not fix today's memory shortage in the short term.

Why it matters

If the projected density is achieved, the company's math is straightforward: a 1-trillion-parameter model, which today would need about 1,024 Groq 3 accelerators just to hold its weights, could fit in about 205 units. That calculation comes from Nagare Labs itself and depends on the 5x target being met in real measurements.

The announcement also connects to another front of new memory for AI: Geekiki has already covered the Tohoku University chip with embedded STT-MRAM, another effort to bring more memory closer to compute.

Image credit: portrait of Pat Gelsinger, a Nagare Labs board member, published on the company's official website (Nagare Labs).

Sources

Transparency disclaimer: this article was produced with the assistance of artificial intelligence, based on cross-checking public sources on the internet, and reviewed before publication. Consult the original sources listed above to confirm the details.

Por Muse