Moxin AI raises nearly RMB 1B in C+ round as sparse-inference chips target LLM costs

Moxin AI, known in Chinese as 墨芯人工智能, has raised close to 1 billion yuan in a C+ round, according to a report published by QbitAI on November 6. The article notes the round comes only five months after a previous raise of roughly the same size, and frames the pace as investor confidence in the company's specialised sparse-inference silicon.

The report is vendor-supplied: QbitAI states at the end of the piece that it is an authorised reprint and that the views belong to the original author. Funding details and operating figures therefore come from the company, with no investor list, valuation or breakdown of use of proceeds disclosed.

Moxin was founded in 2018 and has concentrated on sparse computation, developing what it describes as a dual sparsity approach covering both weights and activations. Around that algorithm it has built a co-design stack spanning compilers, chip architecture and software tooling, and the report says founding team members came from Carnegie Mellon and from Intel, Qualcomm and Marvell.

The bet rests on a structural shift in inference. The article cites DeepSeek-V4-Pro, which carries about 1.6 trillion total parameters but activates only around 49 billion per token, roughly 3 percent. As mixture-of-experts architectures spread, conditional computation becomes normal, and general-purpose GPUs built for training, inference and scientific computing alike struggle to specialise for particular sparsity patterns.

On deployment, the report says Moxin's inference clusters now span four regions of China and that third-quarter 2026 revenue grew several-fold, without giving absolute figures. It also points to MLPerf Inference results, where the company's S30 card has topped some of the categories it entered for three consecutive rounds; the qualifier matters, since it signals strength in selected workloads rather than across the board.

Concrete projects include an offshore compute token factory built with a Shanghai state-owned platform, which uses offshore wind power and seawater cooling to cut power usage effectiveness, with Moxin's chips targeting the remaining per-token cost. The company also adapted two mainstream remote-sensing image analysis models to its cards in work with a Peking University research institute, reporting up to 2.88x higher inference throughput and up to 7.64x model compression with no downstream retraining.

On product timing, a new board called SP200 is slated to ship before the end of the year, alongside an upgraded software stack named SECA aimed at large-model deployment and cluster operations. In August the company also set up an industry alliance for sparse computing research.

What matters next is less the architecture story than adoption: how many mainstream models run on these cards without friction, and whether the software stack can convert theoretical sparsity gains into stable throughput at customer scale. The published figures are company-reported and category-limited, so customer counts and real cost curves will be the test.