A San Jose chip startup that spent seven years working in near-total secrecy has come forward with $468 million in late-stage venture funding and a claim that it can help solve one of artificial intelligence's most expensive problems: getting data in and out of memory fast enough to keep pace with AI computation. Kepler Computing, founded in 2018, says its proprietary ferroelectric material and 3D stacking approach can match the density of 2-nanometer or 3-nanometer chips without relying on the ultra-expensive lithography tools that dominate cutting-edge chipmaking.
According to Dealroom, the funding round ranked in the 99th percentile of late-stage venture deals in its country, all-time, and drew backing from GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford, and Gates Frontier. GlobalFoundries alone put in $50 million and is also serving as Kepler's manufacturing partner, testing the startup's technology at its Vermont plant as well as in Singapore. Kepler's CTO and co-founder, Sasi Manipatruni, previously served as founding director of Intel's FEINMAN research center and co-invented MESO spin-orbit logic, while CEO Debo Olaosebikan holds a physics PhD from Cornell University and sits on the U.S. Department of Commerce's Industrial Advisory Committee, according to XenoSpectrum.
Seven Years of Quiet Materials Science
Before going public with its plans this month, Kepler's engineering team tested 35 distinct iterations of its proprietary ferroelectric composite material before settling on the formulation now headed for commercial production, according to Endroid News Network. That patience is central to the pitch: rather than shrinking transistors through Extreme Ultraviolet lithography, the standard path for cutting-edge chips, Kepler is layering new materials onto mature, widely available 28-nanometer fabrication lines.
That distinction matters because EUV machines, supplied by ASML, cost more than $150 million apiece and face multi-year delivery backlogs, per Endroid's reporting. By working within existing fabs instead, Kepler says it converted a mature 28-nanometer line into a working memory plant in eight months, versus the two to three years typically needed to build a new fab from scratch, according to TechPowerUp. Dealroom reports that the 3D stacking approach can boost chip density without EUV lithography at all.
Chasing a Market Locked Up by Three Giants
Kepler is applying its approach to SRAM, HBM, and other high-speed memory technologies, per Dealroom, aiming to build a new high-bandwidth memory architecture that could help ease a global memory-chip shortage. The company says the resulting architecture could deliver energy efficiency and bandwidth per watt comparable to high-speed SRAM while offering up to 10 times the memory capacity of standard SRAM and HBM, according to OC3D. Data movement between memory and processors is a central challenge in AI data centers, and memory remains a major bottleneck for AI machines, as reported by Qore.
Srini Ananth, managing director at Intel Capital, said market demand is driving interest in both Kepler's technology and existing DRAM or SRAM, per Dealroom's reporting. Kepler reportedly shifted its development plan to pursue SRAM and HBM in parallel after ChatGPT launched in 2022 and demand for an HBM alternative surged. That demand has piled up against a market that Counterpoint Research describes as tightly held by SK Hynix, Samsung, and Micron, with the leading suppliers sold out of production capacity through late 2026. Both Micron and SK Hynix are pouring billions of dollars into new HBM fabrication plants, per Dealroom, while HBM itself relies on stacks of DRAM.
Federal Backing Comes With a Stake in the Company
Kepler's private raise is paired with federal support. The U.S. Department of Commerce issued a July 2026 letter of intent under the CHIPS Act to provide up to $245 million to help Kepler develop AI memory technology, and the agreement required Kepler to grant the federal government a minority, non-controlling equity stake, per the Cato Institute. The federal support reflects a broader policy shift toward government equity stakes in chip companies.
Kepler says it plans to ship its first AI memory samples in late 2026, scale volume production at GlobalFoundries' Singapore plant in 2027, and begin U.S. manufacturing in 2028, according to OC3D's reporting. But the company's 2027 Singapore production allocation has been finalized, and the company still must prove it can produce its technology at scale, per Dealroom. Broader memory market inflation underscores the stakes: AI server demand pushed Q2 2026 conventional DRAM contract prices up 58% to 63% quarter-over-quarter, while NAND flash contract prices jumped 70% to 75%, according to TrendForce data cited by Endroid News Network. Surging prices and shortages of high-bandwidth memory for AI data centers have rattled budgets across the industry, and Kepler is betting its materials-science shortcut can loosen that squeeze before the incumbents lock in their lead with next-generation HBM4.
