Kepler Computing Will Start Memory Production in 2027

The startup plans to mass-produce 3D and ferroelectric memory to address AI-driven bandwidth limitations.

Updated on Oct. 5, 2026 in Semiconductors

Bold flat-color editorial illustration of a vertical stack of square layers, representing monolithic chip architecture for high-performance computing. — Kepler Computing announced it will begin mass-production of its proprietary 3D and ferroelectric memory chips in 2027 to address AI hardware bandwidth limitations. AI Illustration. Upload story photo >

Live Poll

Will new memory technologies successfully reduce the energy required for AI data centers by 2027?

Kepler Computing announced plans to begin production of its proprietary 3D and ferroelectric memory technology in 2027. The company, which has partnered with GlobalFoundries for manufacturing, has already allocated its entire initial production capacity.

Why it matters

AI workloads are straining current memory architectures, as SRAM scaling has largely stalled at the 5 nm node. This new architecture aims to reduce energy consumption and costs for high-performance computing tasks.

Kepler claims its memory delivers 5 to 10 times higher bandwidth per watt than standard HBM and 10 times the capacity of SRAM. The architecture relies on monolithic stacking and a backend-of-line process to bypass thermal issues typically associated with dense memory chips.

The players

Kepler Computing

A semiconductor startup founded in 2018 that focuses on proprietary 3D and ferroelectric memory technologies.

GlobalFoundries

A prominent semiconductor foundry providing the manufacturing infrastructure for Kepler Computing's planned memory production.

The details

The design utilizes ferroelectric materials—compounds that retain electrical polarization—and a backend-of-line process, which refers to the fabrication steps performed after transistors are created to establish metal interconnects. By using monolithic stacking—an approach where device layers are integrated within a single chip structure—the company aims to eliminate the thermal bottlenecks that often limit traditional memory density. This design is intended to solve the scaling wall encountered at the 5 nm node, where SRAM density improvements effectively ended.

2018: Kepler Computing was founded.
2027: Kepler Computing plans to start production.

The Tech Race

The industry is currently struggling to advance memory density beyond the 5 nm SRAM scaling wall. Kepler's effort aims to bypass this limit and directly compete with the performance gains traditionally sought through High Bandwidth Memory (HBM) architectures.

As all production capacity for 2027 is already allocated, hardware manufacturers will not have immediate access to this technology. The impact will first be felt by enterprise-grade AI infrastructure providers looking to replace existing SRAM or HBM setups with more efficient, high-density alternatives.

The takeaway

Kepler Computing is attempting to shift the memory paradigm by moving beyond the scaling limits of 5 nm SRAM. Observers should track the 2027 production ramp-up to see if the technology meets its performance benchmarks in a large-scale manufacturing environment.