Kepler Computing Will Start Memory Production in 2027
The firm intends to manufacture advanced memory chips using a partnership with GlobalFoundries.
Updated on Oct. 5, 2026 in Semiconductors
Live Poll
Will new memory technologies successfully reduce the energy required for AI data centers by 2027?
Kepler Computing plans to begin production of its proprietary 3D and ferroelectric memory in 2027. The startup has entered a manufacturing partnership with GlobalFoundries to produce these alternatives to traditional HBM and SRAM.
Why it matters
AI models currently stress existing memory architectures, driving up energy costs and performance bottlenecks. Kepler Computing aims to resolve these constraints by providing higher bandwidth and capacity compared to conventional hardware.
The proprietary 3D and ferroelectric technology uses monolithic stacking to mitigate thermal issues. The design leverages a backend of line process to increase memory density significantly beyond current industry standards.
The players
Kepler Computing
This startup focuses on developing high-performance, proprietary 3D and ferroelectric memory technologies.
GlobalFoundries
This major semiconductor manufacturer provides large-scale production facilities for various tech partners.
The details
Kepler Computing claims its memory technology offers 5 to 10 times higher bandwidth per watt than HBM and 10 times the capacity of SRAM. By using monolithic stacking of devices, the company asserts it can avoid the thermal problems associated with traditional memory architectures.
Timeline
Kepler Computing was founded in 2018.
Production of the new memory technology is scheduled to begin in 2027.
The Tech Race
Kepler Computing is attempting to solve the scaling crisis that stalled SRAM performance at the 5 nm node. This development mirrors the broader industry trend of shifting away from legacy memory architectures toward specialized, stackable hardware to sustain AI growth.
The company has already allocated its entire production capacity for 2027, signaling tight supply for early-adopting customers. This shift could eventually lead to more energy-efficient AI processing in consumer and enterprise hardware.
The takeaway
The move suggests a long-term transition toward specialized memory architectures to overcome power-to-bandwidth constraints in AI development. Industry observers should watch for how successfully the company integrates these chips with logic components in future hardware designs.
Further reading
For more on the current manufacturing landscape, explore our Semiconductors section.
Source note: This article includes information reported by EE Times.
Live Poll
Will new memory technologies successfully reduce the energy required for AI data centers by 2027?
