The Georgia Tech spinout says its Z2RAM architecture could bring faster, denser memory closer to compute, reducing an increasingly expensive AI infrastructure bottleneck.
Nagare Labs has emerged from stealth with $5 million in early-stage financing led by Playground Global, as the Atlanta semiconductor startup develops a new form of memory for AI accelerators.
The company was founded by Georgia Tech researchers Suman Datta and Shimeng Yu. Former Intel CEO Pat Gelsinger, now a general partner at Playground Global, is involved as a backer and board member.
AI's Memory Wall
Modern AI processors can perform vast numbers of calculations, but getting data to those processors is increasingly expensive in energy and latency. SRAM is fast but occupies considerable space, while larger DRAM and high-bandwidth memory structures introduce different tradeoffs.
Nagare is developing Z2RAM, a capacitorless memory architecture based on oxide-semiconductor transistors. The company proposes stacking memory layers directly above logic, reducing the physical distance that data must travel.
Its goal is to combine the access speed associated with SRAM with higher density more comparable to DRAM, using processes compatible with established semiconductor manufacturing infrastructure.
The company says its first generation targets five times the density of SRAM and projects substantial improvements in energy use and access speed. These are design targets and company projections, not independently verified mass-production performance.
From Research to Fabrication
Datta has contributed to semiconductor transistor technology, while Yu is known for work on emerging memory and AI-hardware simulation.
Their research background gives Nagare technical credibility, but converting an architecture from experimental devices to commercially viable silicon remains difficult. Fabrication yields, thermal reliability, compatibility with foundry processes and total cost will determine whether Z2RAM can be adopted.
Nagare plans to hire memory-circuit and device-process engineers while working with foundries, AI accelerator designers and packaging partners.
The company is entering a market where demand for higher-bandwidth, lower-energy memory is expanding rapidly alongside AI compute spending. Its $5 million financing will support the earliest stages of proving that an alternative memory layer can be manufactured at scale.
If Nagare succeeds, it could improve how efficiently accelerators use expensive compute. If not, its architectural advantages may remain confined to research demonstrations.
Sources: Nagare Labs · Dealroom