TechPowerUp reported that Kepler Computing has emerged from seven years of stealth mode to address the ongoing memory crisis with a custom alternative to High Bandwidth Memory (HBM). The startup claims its technology utilizes Ferroelectric RAM (FeRAM) and a novel 3D stacking technique to maintain standard HBM capacity while significantly reducing manufacturing costs. Crucially, this approach does not rely on cutting-edge semiconductor nodes or EUV lithography, a departure from the standard production methods used by major memory suppliers.
GlobalFoundries 28 nm conversion
Kepler’s strategy centers on leveraging existing mature nodes to manufacture HBM and SRAM. In early collaborations with GlobalFoundries, the company reported converting a standard 28 nm manufacturing facility into a memory fab in just eight months. This timeline is significantly faster than the typical 24-month lead time associated with traditional DRAM production. The company has processed approximately 2,000 wafers of this new DRAM technology, with the first HBM samples anticipated later this year.
$468 million from Intel and AMD
The startup has secured approximately $468 million in funding over the past few years from GlobalFoundries, Intel Capital, and AMD Ventures. These investors are likely to be future customers, suggesting a closed-loop ecosystem for the new memory technology. Kepler aims to begin full-scale volume manufacturing at GlobalFoundries’ Singapore facility next year, with manufacturing in the United States slated to start in 2028. The company claims to have overcome significant scaling challenges, undergoing 35 iterations of composites before settling on a single scalable design.
Huawei Ascend 950DT price surge
The emergence of Kepler comes against a backdrop of severe HBM supply constraints and price surges. TechNews reported that Chinese AI chip manufacturers are increasingly relying on the gray market to obtain advanced HBM, with purchase prices reaching several times that of buyers outside China. Huawei’s latest Ascend 950DT accelerator chip has seen its quoted price exceed 250,000 RMB, representing a 50% increase in just two months. This cost pressure is directly transferring to product selling prices, highlighting the urgent need for alternative memory solutions.
SanDisk HBF vs Kepler FeRAM
Kepler’s FeRAM approach is one of several emerging alternatives to traditional DRAM-based HBM. SanDisk has been promoting High Bandwidth Flash (HBF), which stacks NAND flash memory dies to achieve capacities 10 to 100 times larger than DRAM per module. While HBF focuses on massive capacity for AI inference, Kepler’s reported ability to use FeRAM on mature nodes to achieve HBM-like performance and capacity could offer a unique, cost-effective pathway for memory expansion. Both approaches aim to alleviate the bottlenecks in AI infrastructure by providing high-bandwidth memory without relying solely on advanced DRAM nodes.
Yield and viability skepticism
Despite the reported progress, skepticism remains regarding the viability of Kepler’s technology. Many memory startups have emerged in recent years but struggled to achieve yield and technology viability. If Kepler has truly overcome these technical challenges, the industry will be watching closely for the first HBM-like chips and their density, bandwidth, and other specifications. The success of this venture could significantly impact the memory market, particularly for buyers and capacity planners facing HBM shortages and price surges.