DeepSeek Sparks Sparse Computing, Domestic Chip Companies Accelerate Breakthrough: This Company Raises Nearly 1 Billion Yuan Again

DeepSeek Sparks Sparse Computing, Domestic Chip Companies Accelerate Breakthrough: This Company Raises Nearly 1 Billion Yuan Again

Q3 revenue achieved multi-fold growth

Yunzhong, reporting from Aofei Temple

QbitAI | WeChat Official Account QbitAI

Recently, Moffett AI (hereinafter referred to as "Moffett") completed a nearly 1 billion RMB C+ round of financing.

Just five months after the previous nearly 1 billion RMB round, Moffett has completed two consecutive financing rounds in a short period, reflecting the capital market's high recognition of its technological barriers and commercialization progress in the field of dedicated sparse inference chips.

Leveraging the first-mover advantage formed by technological accumulation and product deployment, as well as the momentum of accelerated commercialization, Moffett has taken the lead in positioning itself at the window of concentrated release of inference computing power demand, further highlighting the generational advantage of its sparse dedicated inference chips.

Betting on sparse computing, forging a differentiated chip path

Currently, the continuous evolution of large model technical architecture is driving sparse computing from an auxiliary optimization method to an important position in AI computing power design.

With the accelerated application of Mixture of Experts (MoE) architecture, large model inference exhibits obvious conditional computation characteristics; meanwhile, the sparse structures in weights and activations also provide space for reducing redundant computation.

Taking DeepSeek-V4-Pro as an example, it has approximately 1.6 trillion total parameters, but only about 49 billion parameters are activated when processing each token, accounting for about 3% of the total parameters.

How to convert model-level computational savings into actual inference efficiency improvements and cost reductions is becoming a key issue that AI computing power development needs to solve.

To efficiently handle such workloads, the choice of chip architecture is a key variable.

General-purpose GPUs need to accommodate multiple types of workloads such as training, inference, and scientific computing, making it difficult to perform deep collaborative optimization of algorithms and hardware for specific sparse patterns.

Dedicated chips for sparse inference, on the other hand, can reduce invalid computation and data movement overhead through co-design of algorithms, compilers, and hardware. The value of this architectural difference is being re-evaluated by the industry.

Moffett has long been deeply engaged in this technical direction. Since its establishment in 2018, Moffett has focused on the field of sparse computing, inventing the "weight sparsification + activation sparsification" dual-sparsity algorithm, and based on this, gradually formed a technical system of "algorithm-software-hardware" co-design.

To turn this technical system into mature chip products, it requires full-stack engineering capabilities spanning algorithms, compilers, chip architecture, and software toolchains.

Moffett's founding team members come from Carnegie Mellon University and top semiconductor companies such as Intel, Qualcomm, and Marvell, possessing both algorithm research and chip engineering backgrounds, with mass production experience of dozens of products and over 5 billion chips.

With this engineering capability, the performance efficiency of Moffett's products has been verified in internationally authoritative benchmark tests.

In the MLPerf™ Inference tests, the S30 computing card won first place in some of the test categories it participated in for three consecutive editions, maintaining industry-leading energy efficiency ratio and per-compute-power inference throughput in mainstream tasks such as computer vision and natural language processing, verifying the engineering feasibility of the dedicated sparse inference solution under real data center workloads.

Multi-scenario deployment and ecosystem co-building, sparse computing enters a new stage

Scaled deployment is further consolidating Moffett's leading position in the field of sparse inference chips.

The inference cluster built by Moffett based on its self-developed chips has been strategically deployed in four major regions: Northwest, Southwest, East China, and North China. On this basis, Moffett's commercialization is accelerating across the board:

In the third quarter of 2026, revenue achieved multi-fold growth, and sparse computing power is deeply integrating into China's computing power system construction, serving the development of the intelligent industry.

Specifically, in supporting the construction of new computing power infrastructure, Moffett participated in the country's first offshore computing power Token factory project built by a Shanghai state-owned platform company.

The project relies on offshore wind green electricity and seawater cooling to reduce PUE and electricity costs, but the inference cost of general-purpose GPUs may still offset the cost-reduction benefits brought by green electricity. To address this pain point, Moffett leverages its sparse architecture advantages, combined with a hybrid heterogeneous computing power model, to reduce the per-token inference cost from the chip side, achieving a superposition of "green electricity cost reduction" and "computing architecture cost reduction" advantages.

Focusing on the frontier fields of spatiotemporal AI and spatiotemporal big data, Moffett provides a cost-effective computing power foundation for applications such as spatiotemporal big data intelligent computing with sparse computing power.

Moffett has collaborated with the Institute of Earth and Space Technology at Peking University Shanghai Lin-gang International Science and Technology Innovation Center, and based on Moffett's self-developed sparsification technology, achieved adaptation and actual testing of two internationally mainstream remote sensing image intelligent analysis foundation models on Moffett's computing cards:

Under conditions of almost no loss of business accuracy and no need to retrain downstream tasks, inference throughput increased by up to 2.88 times, and model compression reached up to 7.64 times.

This means that the computation saved at the model level can be converted into actual processing efficiency improvements through software-hardware collaboration, allowing large-scale remote sensing image interpretation to be completed with less computing power and in less time, and also providing engineering verification for sparse computing power to undertake geospatial foundation model inference.

As real business needs are integrated into chip design and software development, Moffett also converts its technological accumulation and scaled deployment experience into continuous iteration capabilities, driving the evolution of sparse dedicated inference chips from "usable" to "easy to use".

In terms of product iteration, the new generation board product SP200 will be officially launched before the end of the year.

In terms of adaptation ecosystem, Moffett's upgraded software stack SECA (Sparse-native Efficient Computing Architecture) is about to be released, which will strengthen support for large model inference deployment and cluster operations, further lowering the threshold for model migration and deployment.

In addition, Moffett extends its technological advantages to the ecosystem level, further enhancing its industry influence in the field of sparse computing.

In August this year, Moffett initiated the establishment of the "Sparse Computing Industry-University-Research Alliance", focusing on frontier research, key technology breakthroughs, and achievement transformation, consolidating its technological accumulation and industrial practice as the foundation for industry collaborative innovation, and promoting the continuous development of sparse computing technology.

The competition in large models is shifting from "computing more" to "computing smarter".

As the continuous expansion of model scale is increasingly constrained by computing power costs and energy consumption, the core proposition of the industry has changed:

How to carry more intelligence with less computation.

Sparse computing is precisely one of the direct answers to this proposition.

Moffett's value lies in providing a differentiated technological path for Chinese AI chips to participate in the next stage of competition.

*This article is published with authorization from QbitAI, and the views are solely those of the original author.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.