Kepler Computing has emerged from more than seven years of quiet development with a new approach to high-bandwidth memory. The startup says three-dimensional stacking and a proprietary ferroelectric material can increase memory density using existing semiconductor facilities, avoiding the need for the most advanced lithography equipment.
The San Jose company has raised $468 million from backers including GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford and Bill Gates through Gates Frontier. In July, the US Department of Commerce committed up to $245 million to support development of its high-performance AI memory technology in the United States.
New Materials Replace a Lithography Race
Leading chipmakers often use extreme ultraviolet lithography to fit smaller transistors into a fixed area. Kepler claims its stacking process can reach SRAM density comparable with 2-nanometer or 3-nanometer chips without relying on EUV. SRAM is the high-speed cache placed close to CPU, GPU and accelerator cores, while HBM stacks separate DRAM components to provide greater capacity.
Kepler’s HBM design aims to place compute closer to memory, reducing the energy required to move data. Its long-term target is to combine HBM capacity with data-movement energy comparable to SRAM. For SRAM, the company uses a low-voltage composite material with ferroelectric behavior. Chief technology officer Sasi Manipatruni says the team tested 35 composite iterations before choosing a material class.
These are company claims from development work, not independently verified production specifications. WIRED does not provide benchmark results comparing speed, power, yield or cost with commercial HBM and SRAM. The proposed density is meaningful only if the material can be manufactured consistently across large numbers of devices.
Existing Fabs Could Shorten the Buildout
Much of Kepler’s testing takes place at a GlobalFoundries facility in Singapore, where it combines its process with 28-nanometer chips. It has also run tests in Burlington, Vermont, and built “mini fabs” for development. GlobalFoundries is both a manufacturing partner and a $50 million investor.
Kepler says an early conversion created a next-generation fab in eight months, compared with a typical 24-month timetable. Its pitch is that modifying existing plants costs less and moves faster than constructing new memory factories that can require $20 billion to $40 billion, plus expensive equipment.
The approach still needs dedicated controls. A GlobalFoundries executive noted that Kepler’s composite contains iron, a difficult contaminant inside a production facility. The material must run on dedicated equipment or remain fully encapsulated. That requirement may limit how easily an existing fab can adopt the process and will be part of any cost comparison.
Two Thousand Wafers Are Only the Start
Kepler has run its technology on about 2,000 wafers. It plans to ship initial HBM samples later this year, increase production in Singapore next year and begin US chip production in 2028. Those milestones describe a route to scale, not completed capacity.
The timing is attractive because AI data centers have driven intense demand for HBM. Established suppliers are investing billions in new plants, betting that demand will remain high when construction finishes. Kepler offers a different bet: squeeze more capability from existing facilities before those plants arrive.
Its investors say the fundamental physics breakthrough has occurred. The remaining challenge is producing good results across thousands of wafers and millions of devices. That distinction is decisive. Kepler is developing SRAM and HBM in parallel after AI demand pushed the company to accelerate its HBM road map. It has substantial funding, government support and a manufacturing partner, but the memory bottleneck will only ease if its process delivers high yields and reliable parts at volume. A promising stack in a mini-fab is the beginning of that proof, not the end.
