Two near-billion-yuan rounds in five months show investors backing specialist inference chipmakers; Moffett's year-end SP200 launch will test whether sparse computing wins buyers beyond telecom carriers.
Moffett AI, a Shenzhen developer of chips for running AI models, said on November 3 it had closed a Series C+ round of nearly RMB1 billion. The round came about five months after a previous raise of the same size, Pedaily reported, and lifts the company’s fundraising across the two deals to close to RMB2 billion. The company, known in Chinese as 墨芯人工智能, did not disclose a valuation.
Industrial investors Renai Capital (仁爱资本), Haitang Fund (海棠基金) and Dongheng Petroleum (东恒石油) took part, along with unnamed industry players, according to Jiemian. State-owned and financial investors Biqiu Capital (比丘资本), Paila Capital (派拉资本) and Zhongcai Capital (中财资本) also took part. Moffett said it would spend the proceeds on a full upgrade of its inference chips, early research on its next-generation chip and an industry-research alliance around its technology.
What sets the round apart is its timing. Moffett AI has now raised close to RMB2 billion in about five months, two rounds of nearly RMB1 billion each. The money is going into inference, the work of running already-trained models for users. The company has not published financial statements, and its claims about sales growth come only from its own announcements.
The new money funds a second-generation card and a software stack
Moffett plans to launch a next-generation accelerator board, the SP200, before the end of the year, TechNode reported. It is also preparing an upgraded software layer called SECA, which it describes as a sparse-native architecture for deploying large models and running computing clusters, according to TechNode’s Chinese edition.
The company said its third-quarter revenue grew several times over, without giving figures. It runs inference clusters built on its own chips in four regions of China, the northwest, southwest, east and north, and a leading telecom operator has added Moffett’s sparse inference products to its compute services, Pedaily reported.
The previous round, a Series C announced on May 28, brought in Shenzhen Capital Group, Yanshan Technology and a Greater Bay Area fund among its new investors, according to Pedaily’s May report and the 21st Century Business Herald. At the time, Moffett said that money would pay for its second-generation card, supply-chain upgrades and hiring. The company’s earlier backers include Ant Group, which took part in its 2024 Series B, according to 36Kr‘s company database.
Moffett is betting that sparse computing lowers the cost per token
Moffett’s chips are built around sparse computing (稀疏计算): exploiting the sparsity in a model’s data and operations so that fewer calculations are needed to answer each query, TechNode explained. Founder and chief executive Wang Wei, whose company was set up in Silicon Valley in 2018 and is now headquartered in Shenzhen, has argued that the technique is central to supplying the next generation of inference capacity, Science and Technology Daily reported in August. The newspaper was covering the launch of a sparse-computing alliance backed by Moffett, at which Fudan University professor Zhang Qi spoke.
The company says its S30 card ranked first in three consecutive rounds of the MLPerf inference benchmark, according to the 21st Century Business Herald. In June it described its second-generation card as delivering two to eight times more compute at the same accuracy, with support for mixture-of-experts models and long inputs, and positioned it against inference cards from Cambricon and Huawei’s Ascend line. The newspaper said Moffett has filed more than 100 patents worldwide.
The commercial pitch rests on cost. China’s three big telecom carriers have begun selling compute in token packages, and IDC expects inference to make up 73% of AI compute demand by 2028, the 21st Century Business Herald reported. A chip that does less arithmetic per token can, in principle, be sold at a lower price per token. The company also says its second-generation card works with Nvidia’s CUDA software ecosystem, according to the newspaper, which would lower the cost for customers of moving models across.
The SP200 launch will show whether buyers follow the investors
The near-term marker is whether the SP200 ships on its year-end schedule and whether Moffett names customers beyond the unnamed telecom operator. A disclosed order or a public data-center deployment would give outside investors their first independent read on demand.
Two near-billion-yuan rounds in five months leave Moffett with a large capital base for a company that still publishes no revenue figures. The first disclosed revenue number would be the clearest check on its claim that third-quarter sales grew several times over.