MemSilicon now has two bets to fund, RRAM memory and a new inference chip, and its first LPU will not use RRAM, so the memory business must prove itself separately.
MemSilicon (铭芯启睿), a Beijing chip startup built on resistive memory, has closed a Series A of more than RMB400 million, according to the company’s announcement as reported by Pedaily and Jiemian on October 19. Boyuan Capital led the round. New investors were Bosch Ventures, Chuangling Capital, miHoYo, Zhongguancun Science City Company and GigaDevice. Seven existing shareholders added money, among them Xiaomi Strategic Investment, Jinqiu Fund, CDB Kechuang (国开科创) and Zhongke Chuangxing (中科创星), a list confirmed by a 36Kr newsflash the same afternoon.
The money goes to two things: bringing the company’s RRAM memory to market and developing a new inference chip it calls an LPU. Alongside the round, MemSilicon named Lu Tao, a former Graphcore president and Greater China general manager, as chief executive and co-founder. Pedaily describes the funding and the new strategy as the company’s step from an RRAM technology firm to a chip company that pairs memory and computing.
For investors, the shift changes what MemSilicon is selling. Until now it has pitched memory chips and memory designs licensed to other chipmakers. It is now also aiming at the market for running large AI models, where speed and cost per answer decide who wins. No valuation, customer or delivery date for the new chip has been disclosed.
MemSilicon’s rounds have grown fourfold in size since its angel financing
The company was founded in May 2024, and its technology traces back to a Chinese Academy of Sciences team led by academician Liu Ming. In March 2025 it raised an angel round of nearly RMB100 million led by Jinqiu Fund, with Lenovo Capital and Xiaomi Strategic Investment also investing, Pedaily reported at the time. That report said the company held more than 200 invention patents. In January 2026, a Pre-A round of more than RMB100 million followed, led by CDB Kechuang and Lenovo Capital, according to TMTPost. Liu Qi serves as chairman.
MemSilicon has raised three rounds in about 19 months, from an angel round of nearly RMB100 million to a Series A of more than RMB400 million. Lenovo Capital, a lead investor in the Pre-A, is not named among this round’s participants.
The product side moved days before the funding news. On October 12 the company taped out its first high-density RRAM chip, MountTai 2026 (泰山一号), meaning the design was sent to a factory for manufacture, the chip outlet Zhidx (芯东西) reported in a post on Ifeng. The 28-nanometer chip holds 1Gb, with access times under 20 nanoseconds, and is now being tested. The same report said MemSilicon’s embedded RRAM designs, licensed for use in display drivers, microcontrollers, SSD controllers and security chips, are already in customer products.
RRAM and the LPU address the same bottleneck in different ways
RRAM, short for resistive random-access memory, stores data as changes in a material’s electrical resistance, which is what its Chinese name, 阻变存储器, describes. MemSilicon’s MountTai 2026 is built on a standard 28-nanometer CMOS platform and places one transistor in charge of four memory cells to pack in more capacity, according to Zhidx. Its wider pitch has been compute-in-memory: processing data where it sits, which attacks the "memory wall," the delay and energy cost of shuttling data between memory and processors.
The LPU, which Jiemian spells out as Language Processing Unit, is a chip designed specifically to run large language models after they are trained, which the industry calls inference. In an interview with Leiphone, reposted by Sina Tech, Lu said the first-generation LPU targets complete models with trillions of parameters, without compression shortcuts, at 1,000 tokens per second for each user. The design keeps the speed of SRAM, the fast memory built onto processors, and adds a second memory layer for capacity so that fewer chips are needed per model.
Lu also said the first LPU will use a more mature memory medium than RRAM, which still needs work on capacity. Lu said MemSilicon’s position is clear: it will not fight GPUs head-on, and will focus on fast output from full models at better cost. Boyuan Capital chairman Jiang Hongquan said his firm’s trust in the company rests on a relationship with Lu that goes back to Graphcore ten years ago. EastFrontier recently covered how SUNMMIO’s new funding backs another alternative route to AI inference hardware in China.
Customer commitments will show whether the LPU plan holds
Lu has built an LPU core team of a dozen-plus people covering hardware, compilers and system software, and the company is in talks with leading cloud providers and large-model developers, according to the Leiphone interview. None has been named. Over the next two to three years, Lu said, he wants compute products in customers’ hands while large-capacity RRAM development continues.
Two milestones will test the plan. The first is test results from MountTai 2026, which would show whether the RRAM business can stand on its own. The second is a named cloud or model customer for the first-generation LPU, along with a tape-out date for that chip.