Moffett AI closed roughly RMB 1 billion of C+ funding on 3 November, five months after landing roughly RMB 1 billion in a C round in May. The two rounds together put close to RMB 2 billion into a company that designs sparse inference chips.
The investor list is worth reading closely. Ren’ai Capital, Haitang Fund and Dongheng Petroleum bring industrial money. Biqiu Capital, Paila Capital and Zhongcai Capital represent state and financial capital. Set that against the May round, which included Shenzhen Capital Group, Yanshan Technology and Greater Bay Area Common Home, and Moffett has spent six months thickening the same three-part structure of state funds, industrial corporates and financial investors.
One detail is more interesting than the rest. Dongheng Petroleum is an energy company. Compute and electricity cost were already the same business, and an industrial investor from that side is most likely buying the logic of a lower cost per token.
Why sparsity is suddenly fundable
Global AI chip investment has climbed through 2026, with dedicated inference silicon the most concentrated destination. Goldman Sachs forecasts that dedicated AI chips will take 40 per cent of the market this year, and industry estimates put inference above 70 per cent of global compute.
Demand is shifting at the same time as the supply-side constraint tightens. Power and energy budgets are the binding limit now, and the old approach of stacking more hardware gets more expensive with every rack. Moffett’s sparse route sits precisely on the demand for doing less arithmetic and finishing more work.
The proceeds go to two places: a full-stack upgrade of the sparse inference chip line, and pre-research on the next generation. A third use is organisational. The company founded a sparse computing industry-academia alliance in August and has already pulled in Fudan, Tsinghua, Xi’an Jiaotong, Nankai and Shanghai Jiao Tong for joint projects. The ambition is to turn an eight-year-old technical bet made inside one company into a track with academic and industrial backing.
The pitch is straightforward. A large share of the arithmetic in model inference produces nothing useful. Moffett skips that share through dual sparsity on weights and activations, so the same silicon completes more effective inference.
Founder Wang Wei uses a comparison to anchor it. Nvidia’s A100 supports only 2-fold sparsity, he says, while Moffett reaches 32-fold. That number comes from the company and is the pivot of its technical story.
The hardware it can actually point to is the S30 compute card. On a 12-nanometre process the company claims 1.2 times the performance of a 4-nanometre GPU product at one third of the power consumption. On the 176-billion-parameter BLOOM model, four S30 cards are said to generate content at 25 tokens per second, ahead of eight A100s. Those results come from MLPerf and internal testing, so they remain the company’s own account, but they do give the claim of less compute for more output something to stand on.
Commercialisation is moving, imprecisely
Over the past half year deployment has accelerated. Moffett says third-quarter 2026 revenue grew several-fold, which is vague, but the use cases attached to it are specific.
It took part in the first offshore compute token factory built by Shanghai state capital, using its sparse architecture to compress the cost of a single token of inference. With a Peking University team in Shanghai’s Lin-gang zone it ported two mainstream remote-sensing foundation models onto its own cards, lifting throughput by up to 2.88-fold and compressing models by up to 7.64-fold with almost no accuracy loss and no retraining.
Inference clusters built on its own silicon are now spread across northwest, southwest, east and north China, and the company has partnerships with leading telecom operators, embedding sparse inference into their compute service offerings to build a national sparse inference network.
Board secretary Wang Shuaiyu set the test during the C round. Judging an AI chip company, he said, should not look only at single-card theoretical compute but at effective compute and energy efficiency in a real cluster completing equal tasks. That reads now as a question the company set for itself, because sparse computing is a route that deliberately diverges from the CUDA ecosystem and puts all its advantage into inference cost.
The risk is symmetric. If sparse inference workloads do not spread as expected, or if general-purpose GPUs cut costs through volume and software, the window narrows. The SP200 chip and the SECA software stack are both due before the end of the year.
Policy has moved towards the route. The Ministry of Industry and Information Technology and the National Development and Reform Commission have both named dynamic sparse computing and software-hardware coupled inference chips as targets. Moffett joined the ministry’s artificial intelligence standardisation technical committee in July and became a first-batch member of the CAICT token working group. A technical bet made in 2018 has run into a policy direction the company could not have counted on at the start.
Capital has now placed close to RMB 2 billion on the table across two rounds. The next six months decide whether several-fold growth converts into a harder scale number, and whether that national inference network delivers the energy efficiency it was built for.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.ofweek.com/ai/2026-11/ART-201714-8120-30710342.html.