Reexen Technology raises $150 million for compute-in-memory AI chips
What's the deal? Reexen Technology, a Chinese compute-in-memory AI chip maker, has raised a cumulative US$137.5M (roughly $150 million). Strategic backers Meituan, XPeng Motors (through Starlight CapitalDealroom has a profile for this one. Try Dealroom →), Changcun Industry Fund, Yuanhe PuhuaDealroom has a profile for this one. Try Dealroom →, and OmniVision GroupDealroom has a profile for this one. Try Dealroom → led earlier tranches; Zhiwei CapitalDealroom has a profile for this one. Try Dealroom → and miHoYoDealroom has a profile for this one. Try Dealroom → joined recently, with existing investors Yinghao Capital and Weihao ChuangxinDealroom has a profile for this one. Try Dealroom → adding more.
What's the endgame? Founded in Shenzhen in June 2018, Reexen builds chips that fuse memory and computing across multiple layers. The money will fund research and tape-out of HBF and SRAM compute-in-memory chips, its Model-Processor-Memory (MPM) software stack, and products for cloud, edge, and device markets.
Why now? As large models shift to Mixture-of-Experts (MoE) architectures with trillions of parameters but sparse activation, systems need high-capacity, high-bandwidth, read-heavy storage. That is where high-bandwidth flash (HBF) fits, Reexen argues. At FMS 2026, SK HynixDealroom has a profile for this one. Try Dealroom → and SanDisk published the first HBF OCP specification, a sign the format is moving toward an industry standard.
What's the endgame? Reexen's product in development vertically stacks SRAM compute-in-memory with HBF into a 3D fused chip. On its engineering platform, DeepSeek-V4 reached about 20 tokens per second in a "single card + 4 SSD" setup; the company aims to raise single-card bandwidth and throughput more than tenfold.
What could go wrong? Reexen's compute-in-memory products are already in volume production, with device-side deals reportedly at vivoDealroom has a profile for this one. Try Dealroom →, Honor, and Lenovo. But the round itself proves little about mass-production progress, repeat orders, or real-world efficiency — the measures that will decide whether its bet pays off.
The signal: The round tracks a shift in domestic AI chip investment, from chasing raw compute toward solving the "memory wall" in large-model inference. Rival startups are drawing similar funding on bets around 3D-stacked storage and computing, a sign the sector sees memory bandwidth, not just processing power, as the next bottleneck.
Read more: sohu.com
Image credit: Generated with Gemini
