--- title: "Sparse Computing Startup Moxin AI Raises Nearly RMB 1 Billion C+ Round for Inference Chips" date: 2026-11-05 category: Capital & Markets site: NeuroAI canonical: https://neuroai.site/a/na-capital-moxin-sparse-inference-c-plus-ro language: en --- # Sparse Computing Startup Moxin AI Raises Nearly RMB 1 Billion C+ Round for Inference Chips > Moxin AI closes a nearly RMB 1 billion (≈ US$140M / HK$1.1B) C+ round, its second large raise in five months, as it pushes sparse inference silicon and a new SP200 card to year-end launch. The most expensive part of a deployed model is not training it. It is answering the same question, millions of times a day, forever. A Shenzhen chip company betting that most of that work can simply be skipped has now raised money twice in five months to prove it. Moxin AI (墨芯人工智能), officially Moxin Artificial Intelligence Technology (Shenzhen) Co., Ltd., announced a C+ round of nearly RMB 1 billion (≈ US$140M / HK$1.1B). The round brought together industrial investors including Ren'ai Capital, Haitang Fund, and Dongheng Petroleum, alongside state-owned and financial institutions such as Bigiu Capital, Paila Capital, and Zhongcai Capital. Proceeds go to a full-stack upgrade of the company's sparse inference chips, pre-research on the next generation of silicon, and an industry-academia alliance around sparse computing. The round lands roughly five months after the company's previous raise of a similar size, lifting cumulative funding to nearly RMB 2 billion (≈ US$280M / HK$2.2B). ## What sparse computing actually does Inference is the phase where an already-trained model handles real tasks — answering questions, recognizing images, processing video. Moxin's core idea is to cut the cost of that phase at the source: using both weight and activation sparsity, its hardware identifies computations that will not meaningfully change the output and skips them, so the same hardware investment carries more useful inference traffic. The approach matters more now for two reasons. Demand for inference is climbing sharply as agent-style applications run continuously rather than in bursts; figures attributed by company-linked coverage to China's National Data Administration put daily token consumption in the country past the hundred-trillion scale earlier this year, up more than 40% from the end of last year. And on the supply side, computing centers face tightening constraints on power, energy quotas, and capital budgets, which pushes buyers to measure success in cost per token, energy efficiency, and total cost of ownership rather than peak performance. Sparse computing is not the only answer to that squeeze, and the company itself acknowledges the hard part is engineering. Sparsity identified in algorithms only pays off if chip architecture, compilers, and the software stack can all execute it efficiently — a mismatch anywhere turns theoretical savings into nothing. Moxin's founding team is built around that full chain: CEO Wang Wei, a Carnegie Mellon ECE master's graduate who worked at Qualcomm and Intel on multiple CPU generations; co-founder and chief scientist Yan Enxu, also from Carnegie Mellon, specializing in machine learning and sparse algorithms; and co-founder Lu Yong, who did chip development at SK Hynix and Marvell. ## From S-series cards to a year-end SP200 The company has shipped commercial products built on its in-house Antoum chip. Its S4, S30, and S40 compute cards carry one, three, and four chips respectively, targeting different deployment scales, with software handling model deployment, device monitoring, and cluster management. Customers buy a package that fits into standard servers and runs specific models — not a chip on a datasheet. Per company statements, Moxin's inference clusters now span four regions of China — northwest, southwest, east, and north — and its sparse inference solutions have been adopted into the computing services of leading telecom operators. The company also reports revenue growing severalfold in the third quarter of 2026; that figure is a company claim, unaudited. Two launches are planned before year-end: the SP200, a next-generation compute card, and an upgraded software stack called SECA (Sparse-native Efficient Computing Architecture), which strengthens large-model deployment, software adaptation, and cluster operations. Both remain announcements rather than shipped products. ## Why the capital keeps arriving - **Inference economics is now its own investment thesis.** Investors are no longer funding model training speed; they are funding whoever can lower the recurring bill of running models at scale, and sparsity is one of a small number of credible levers. - **Two large raises in five months signal conviction, not fashion.** Moxin's May C round drew names including Shenzhen Capital Group, a Greater Bay Area fund, and corporate backers; the C+ adds industrial parties, suggesting customers rather than only financial investors now see value in the roadmap. - **MoE architectures lean sparse.** As mixture-of-experts models spread, inference loads increasingly look conditional and sparse by construction — structurally friendly terrain for hardware designed to skip work. The open question is whether sparsity's advantages survive contact with the specific models and accuracy tolerances each customer runs. That will show up in operator-scale deployments, not in funding announcements. ## Key takeaways - Moxin AI raised a nearly RMB 1 billion (≈ US$140M / HK$1.1B) C+ round from Ren'ai Capital, Haitang Fund, Dongheng Petroleum, Bigiu Capital, Paila Capital, Zhongcai Capital and others; cumulative funding now approaches RMB 2 billion (≈ US$280M / HK$2.2B). - The company, founded in 2018 in Shenzhen, builds sparse inference chips and S-series compute cards around its Antoum silicon, with clusters across four regions of China and telecom operator partnerships. - An SP200 card and the SECA software stack are slated for launch by year-end; proceeds also fund next-generation chip pre-research. ## Honest limitations - The revenue growth claim for the third quarter is a company statement without audited figures, customer counts, or pricing details; no independent financial data verifies it. - The National Data Administration token-consumption figure was relayed through company-linked coverage rather than read directly from the agency's filings, and single-sourced numbers should be treated as indicative. - The SP200 and SECA remain planned launches; benchmarks, pricing, and availability are not yet public, and comparisons with incumbent GPU-based inference stacks remain untested in public. ## What readers can do now - If you run inference workloads, benchmark sparsity against your own models and accuracy thresholds before assuming the savings translate — the discount is workload-dependent, and vendor demos are not your production mix. - If you track China's semiconductor funding cycle, watch whether the year-end SP200 launch comes with named customers and published benchmarks; that is the difference between a roadmap and a business. - If you procure computing capacity, add cost per completed task and energy per token to your evaluation sheet; peak performance alone increasingly misprices the actual bill. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-capital-moxin-sparse-inference-c-plus-ro Free to quote with attribution and a link to the original.