Kepler emerges from 7-year stealth with AI memory claiming SRAM-like bandwidth, 10x capacity
morganlai · x · 2026-09-10
Chip startup Kepler has come out of stealth after 7 years, targeting AI's memory bottleneck with a new memory product it claims approaches the bandwidth-per-watt of SRAM while offering up to 10x the capacity of SRAM and HBM, aimed at high-throughput, high-interactivity inference. Sampling is planned for 2026 and production for 2027, and the company has been selected for up to $245M in proposed U.S. Department of Commerce support.
- Team pedigree: Co-led one of the industry's first 3D chip-stacking technologies, scaled seven generations of DRAM and the first 232-layer NAND to market, and built the world's first commercial physical synthesis tool, with work on ferroelectrics and beyond-CMOS devices.
- Thesis: Leadership chips without EUV dependence, and orders of magnitude more intelligence per watt by evolving from 2D silicon to 3D and beyond-silicon designs.
- Investor morganlai notes Kepler was among the first bets from his fund, backing founder Debo's vision to push every semiconductor's performance.
Related event: Stealthy chip startup Kepler emerges with new AI memory beyond silicon(8 posts)→
More from Infra
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- Gary Marcus: With Q3 data added, Anthropic's compute-efficiency growth shows deceleration — GaryMarcus · 2026-10-27
- Inference is just Kubernetes with a multi-trillion-dollar market, and the communities overlap — matt_slotnick · 2026-10-27
- vLLM eyes Rust migration as inference engines race to rewrite the stack — AccBalanced · 2026-10-27
- Resin lets an agent build standalone LLM inference engines; vLLM and SGLang optimize DeepSeek kernels differently — AccBalanced · 2026-10-27