On November 3, Moxin AI announced the completion of a nearly 1 billion yuan Series C+ financing round—exactly five months after it secured nearly 1 billion yuan in a Series C round in May of this year. Combined, the two rounds have injected nearly 2 billion yuan into the accounts of this company specializing in sparse dedicated inference chips.

Image source: Moxin official website

First, let's look at the list of investors in this round. Rena Capital, Haitang Fund, and Dongheng Petroleum brought industrial capital to the table, while Biqiu Capital, Paila Capital, and Zhongcai Capital represented state-owned and financial investment. Compared with the May round (Shenzhen Capital Group, Yanshan Technology, Greater Bay Area Common Home, etc.), Moxin has reinforced its "state-owned capital + industrial giants + financial capital" structure once again within just half a year.

An interesting detail is the entry of energy-backed investors such as Dongheng Petroleum. The computing power business and electricity costs are inherently intertwined, and what industrial investors likely value is the company's logic of lowering the cost per token.

In 2026, global AI chip investment and financing continues to climb, and dedicated inference chips have become the most concentrated direction for capital. Goldman Sachs predicts that dedicated AI chips will account for 40% of the market this year. Inference computing power is taking over the banner from training, and industry estimates suggest that inference computing power will account for more than 70% of global computing power this year.

The demand side is shifting, and power and energy consumption constraints on the supply side are also tightening. The old path of simply stacking hardware is becoming increasingly costly. Moxin's sparse approach happens to land right on this demand node of doing more work with less computation.

This fundraising is mainly directed to two areas: first, full-stack upgrades of sparse dedicated inference chips; second, preliminary research on next-generation chips. In addition, it will continue to advance the sparse computing industry-academia-research alliance. Moxin initiated this alliance in August this year and has already brought in Fudan University, Tsinghua University, Xi'an Jiaotong University, Nankai University, and Shanghai Jiao Tong University for joint research projects. It wants to turn the sparse route it bet on in 2018 from one company's technological choice into a track backed by academic and industrial alliances. A large amount of computation in large-model inference is invalid, and Moxin uses the "dual sparsity" of weights and activations to skip this portion, allowing the same hardware to run more effective inference.

Founder Wang Wei often cites a figure for reference: Nvidia's A100 supports only 2x sparsity, while Moxin can achieve 32x. This figure comes from the company's own statement and is the pillar of its technological narrative. On the hardware side, the solid product Moxin can present is the S30 compute card, which uses a 12nm process to achieve 1.2 times the performance of 4nm GPU products while consuming only one-third of the power. On the 176 billion-parameter BLOOM model, four S30 cards achieved a content generation speed of 25 tokens per second, surpassing eight A100s.

These are results from its MLPerf tests and internal tests. In nature, they are still the company's self-verification, but at least they provide proof that "less computation can translate into efficiency."

What is noteworthy is that Moxin's commercialization has been accelerating over the past six months. Moxin says revenue grew severalfold in the third quarter of 2026. This figure is somewhat vague, but the supporting deployment scenarios are specific. It participated in China's first offshore computing power Token factory built by Shanghai state-owned capital, using a sparse architecture to reduce per-token inference costs. In addition, it worked with the Peking University Shanghai Lingang team to adapt two mainstream remote sensing foundation models to its own compute cards, increasing throughput by up to 2.88 times and compressing models by up to 7.64 times, with almost no loss in accuracy and no need for retraining.

Image source: Moxin official website

It is reported that the inference clusters Moxin has built based on self-developed chips have now been deployed in Northwest, Southwest, East China, and North China. It has also established cooperation with leading telecom operators, embedding sparse inference solutions into their computing power service systems and building a nationwide sparse inference computing power network.

Moxin board secretary Wang Shuaiyu said during the Series C round: "To evaluate an AI chip company, one should not only look at single-card theoretical computing power, but also at the effective computing power and energy efficiency ratio for completing equivalent tasks in real clusters." This now looks more like a test question the company set for itself. After all, sparse computing is a technological route that directly diverges from the CUDA ecosystem, with all its advantages concentrated in inference cost.

Once the sparsification of large-model inference workloads does not unfold as expected, or general-purpose GPUs manage to drive costs down through volume and software, the window for this path will narrow.

Before the end of the year, Moxin's SP200 will be launched, and the SECA software stack will also be released. Capital has already placed its bet on the table with two rounds totaling nearly 2 billion yuan. What the next six months will show is whether Moxin can turn "severalfold growth" into harder scale figures, and whether that computing power network can truly deliver energy efficiency.

On the policy side, the Ministry of Industry and Information Technology and the National Development and Reform Commission recently made clear their intention to tackle dynamic sparse computing and lay out software-hardware coupled inference chips. In July, Moxin joined the MIIT Artificial Intelligence Standardization Technical Committee and also became one of the first members of the CAICT Token Special Working Group. The technological route and policy direction have converged—an external tailwind it may not have anticipated at the beginning of the year.