The rapid expansion of artificial intelligence workloads has exposed physical and financial bottlenecks in semiconductor manufacturing, particularly around high-bandwidth memory. Chipmakers have traditionally depended on extreme ultraviolet lithography (EUV) to shrink transistor geometries and force higher density—an approach that demands multibillion-dollar capital outlays and years of construction for every new fab. Against this backdrop, Kepler Computing, a San Jose–based startup founded in 2018 by physicists and computer scientists, has emerged from seven years of stealth to pitch a radically different path.
Density Without Advanced Lithography
Kepler claims its proprietary 3D stacking approach and novel ferroelectric material achieve dense memory without touching EUV lithography. Instead of waiting for greenfield facilities or competing for scarce advanced packaging lines, the company designs its architecture to integrate directly into mature semiconductor plants, targeting the core throughput bottleneck in AI training and inference.
Industrial Partnerships and Deployment Timelines
To underwrite the hardware shift, Kepler has secured $468 million in private funding from investors including GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford, and Gates Frontier. Backing extends into public industrial policy: the US Department of Commerce committed up to $245 million under the CHIPS Act to develop high-performance ferroelectric 3D memory domestically.
Commercial validation is already underway in Singapore, where GlobalFoundries acts as both a manufacturing partner and a $50 million investor, pairing Kepler's memory architecture with existing 28-nanometer process nodes.
Bypassing leading-edge lithography through material science and legacy tools offers a compelling escape hatch for AI infrastructure operators watching hardware CapEx spiral. However, the real operational test lies in scaling new ferroelectric materials at production volumes. Established HBM incumbents hold mature yield curves; whether 28-nanometer retrofits can match those yields at commercial wafer volumes remains the decisive risk for Kepler's balance sheet.