Moffett AI Raises Nearly 1 Billion Yuan Again in 5 Months, Sparse Computing Inference Chips Accelerate Commercialization

Chip Inside reported on November 3 that today, domestic sparse computing chip company Moxin AI (hereinafter referred to as "Moxin") announced it has recently completed nearly 1 billion yuan in financing. This round comes only 5 months after its previous round of nearly 1 billion yuan, bringing the two rounds to a cumulative total of nearly 2 billion yuan.

This round of financing for Moxin was jointly participated by industrial parties such as Ren'ai Capital, Haitang Fund, and Dongheng Petroleum, as well as well-known state-owned and financial investment institutions such as Biqiu Capital, Paila Capital, and Zhongcai Capital. The raised funds will be invested in the full-stack upgrade of sparse dedicated inference chips, with a focus on pre-research for the next-generation chip, while also advancing the construction of a sparse computing industry-academia-research alliance.

Why is capital increasing its bet at this point in time, what changes have occurred in Moxin's commercialization, and how does sparse computing become visible numbers on customers' books? With these questions in mind, Chip Inside held an in-depth conversation with Guo Weijun, Vice President of Moxin's Marketing Department.

1. Seizing the node of explosive inference demand, commercialization is fully accelerating

To understand this financing round for Moxin, it must be viewed against the broader industry backdrop.

Over the past year, large models have truly entered production environments, inference call volumes have continued to rise, and bills have soared. Today's token consumption is no longer project-based; it is a fixed monthly "computing bill" like water and electricity.

Against this backdrop, customers have raised higher requirements for next-generation inference computing power. Guo Weijun summarized customer demands into three points: first, lower comprehensive inference costs, including not only token bills but also long-term expenses such as electricity, deployment, and compatibility; second, computing power supply must be elastic and delivery must be certain; third, control adaptation costs so that existing models and business systems can migrate with a low threshold.

At present, there are many ways to improve overall system inference efficiency, and sparse computing is considered one of the most noteworthy directions. What sparse computing addresses is actually a simple problem: large models often have hundreds of billions or even trillions of parameters, but in each inference, only a small portion is actually activated and participates in computation. Traditional chips make no distinction here, which the industry calls "dense computing"; sparse computing does the opposite, computing only the parts that actually matter and skipping the rest. For the same inference task, the amount of computation is reduced, and costs fall accordingly.

Moxin has been deeply engaged in technical validation of sparse computing for many years, and recently these accumulations have begun to pay off in concentrated fashion. These two financing rounds are an important watershed as Moxin moves from years of single-point technical validation toward large-scale, commercial delivery.

At the technical level, Moxin's "algorithm-software-hardware" co-design approach has been validated in real products, and an end-to-end software closed loop is gradually taking shape. Moxin's S30 and S40 compute cards have won first place for three consecutive editions in the MLPerf Inference tests, and its engineering capability has public benchmark backing.

At the delivery level, more and more intelligent computing centers and industry customers are also bringing sparse computing power into consideration for selection. Moxin's sparse dedicated inference chips have been rolled out across four major regions—Northwest, Southwest, East China, and North China—with thousand-card clusters deployed, and it has established cooperation with leading telecom operators to incorporate sparse computing inference solutions into their computing power service systems. Moxin's large-scale delivery capability has also been validated in these projects.

At the market level, Guo Weijun also observed a trend: not every industry needs large models. As AI applications in vertical industries continue to enrich, fields such as finance, healthcare, exploration, and meteorology all have their own industry models, and almost all of these models can benefit from sparse computing, with deployment scenarios beginning to show the possibility of cross-industry replication.

2. Sparse computing moves toward real production environments, Moxin helps customers calculate costs clearly

Over the past few months, from university research platforms to national computing projects, Moxin's list of deployment projects has been rapidly lengthening. We also discussed several real customer cases.

In research scenarios, Moxin's sparse computing power has already produced a set of quantifiable measured results. Targeting the frontier fields of spatiotemporal AI and spatiotemporal big data, Moxin worked with the Institute of Earth and Space Technology of the Peking University Shanghai Lingang International Science and Technology Innovation Center to complete adaptation and testing on Moxin compute cards of two internationally mainstream remote sensing image intelligent analysis foundation models based on its self-developed sparsification technology.

Data shows that, under conditions of almost no loss in business accuracy and without the need to retrain downstream tasks, model inference throughput increased by up to 2.88 times and model compression reached up to 7.64 times, significantly reducing the computing consumption and processing time of large-scale remote sensing image interpretation.

If the Peking University project validated the effect in a single scenario, then the next computing project tests the ability of sparse computing power to integrate with existing systems. Moxin participated in the first domestic offshore computing token factory built by a Shanghai state-owned platform company: this project uses green power from offshore wind and seawater cooling to lower PUE and electricity costs, but the money saved on green power may be offset by the inference costs of general-purpose GPUs.

Moxin's approach is to reduce the per-token inference cost from the chip side, and then combine it with a hybrid heterogeneous computing model so that "green power cost reduction" and "architecture cost reduction" reinforce each other, further amplifying the cost-saving effect.

It is worth noting that when communicating with customers, Moxin does not deliberately emphasize the technical concept of sparse computing. "What customers care about most is whether it is easy to use; they will not focus too much at the early stage of a project on whether the route is sparsification."

In actual commercial cooperation, customers pay more attention to overall input-output: how many servers need to be deployed in the data center, what the power consumption and O&M costs are, and ultimately whether business accuracy and ROI requirements can be met.

To this end, Moxin translates product value into concrete cost calculations and is working to establish a standardized pre-purchase simulation testing system, so that customers can obtain the corresponding simulation data before procurement and evaluate the actual benefits of the product in their own business scenarios.

Compared with performance metrics, migration cost is often a harder threshold for customers to overcome in decision-making. Most of customers' existing evaluation scripts and scheduling systems are built around dense computing architectures. Changing chips not only requires rerunning regression tests, but also involves engineering modifications and business gray releases, bringing additional time and capital investment.

To address this issue, Moxin, on the one hand, helps customers establish a cross-architecture evaluation and vendor comparison system through sparse evaluation standards and simulation tools; on the other hand, it also provides a heterogeneous adaptation toolchain to promote collaboration between sparse cards and customers' existing dense cards. For customers who have already deployed a large number of existing chips, this approach can reduce the cost of system migration and duplicate construction, and also provides another path for the implementation of sparse computing.

III. From "usable" to "easy to use," software and ecosystem provide momentum for commercialization

Lowering the customer usage threshold is only the first step toward commercializing sparse computing. To achieve broader implementation, continuous product iteration, improved software toolchains, and joint support from the industrial ecosystem are also needed. Moxin is advancing its layout in three directions.

In academia, a mature collaborative system has not yet formed among sparse computing algorithms, mathematical theory, hardware architecture, and software toolchains, and many key issues still require further research. To solve this problem, in August this year, Moxin initiated the establishment of China's first sparse computing industry-academia-research alliance, and has currently launched multiple joint projects with universities including Fudan University, Tsinghua University, Xi'an Jiaotong University, Nankai University, and Shanghai Jiao Tong University.

Guo Weijun revealed that the joint projects within the alliance are laid out separately according to each school's disciplinary strengths: the computer science schools focus on how the combination of "algorithms + hardware architecture" can improve sparse computing performance; the mathematics schools focus on the computational efficiency of matrix multiplication algorithms and the identification of sparse parameter patterns; and the software discipline focuses on the development of sparse toolchains, compilers, and high-performance operators.

▲Plan of the Sparse Computing Industry-Academia-Research Alliance

The task undertaken by the industry-academia-research alliance is not only scientific research cooperation, but also to fill in the foundational elements required for the industrialization of sparse computing. In Guo Weijun's view, the industry currently still faces shortcomings in three aspects:

First, there is a shortage of interdisciplinary talent. People who understand both algorithms and sparse mathematics and are also familiar with chip architecture are relatively scarce, and the training cycle for such talent usually takes 3 to 5 years.

Second, evaluation and acceptance standards have not yet been unified. Existing computing power delivery standards are mainly formulated around dense computing, making it difficult to accurately measure the actual capability of sparse computing. Guo Weijun said that Moxin may also launch acceptance standards for sparse computing power in the future.

Finally, foundational resources such as toolchains and datasets for sparse computing still need to be improved.

Guo Weijun believes that although these foundational efforts have a lower profile and receive less attention, they are indispensable links for the large-scale development of the sparse computing industry.

Beyond the ecosystem, Moxin's product line iteration has also not stopped. Its new-generation board card SP200 is planned to be launched before the end of this year. Guo Weijun revealed that the real pain points previously exposed in customer use have been incorporated into the design of the next-generation architecture.

On the software stack side, Moxin's upgraded software stack SECA (Sparse-native Efficient Computing Architecture) is about to be released. From operator libraries and compilation to software-hardware co-design, the entire software stack has been redesigned around the sparse architecture, and the accompanying MaaS platform and model framework adaptation management tools will also be launched successively.

From ecosystem building to product iteration and then to software stack upgrades, Moxin's commercialization momentum is worth watching.

Conclusion: Has the turning point for sparse computing arrived?

The application space for sparse computing is accelerating its expansion from large models to more industries. In Guo Weijun's view, vertical-domain models are customized for specific businesses and do not need all the capabilities of general large models, so they are naturally suited to sparse computing.

For example, in industrial quality inspection scenarios, vision models often only need to identify local defect areas in images, while a large amount of background information does not directly participate in judgment. Sparse computing is expected to reduce unnecessary computational overhead. Embodied intelligence and world models also have similar needs. Guo Weijun analyzed that the human brain does not process all information in the surrounding environment to the same degree, but filters and focuses on key information according to the current task, which coincides with the model of sparse computing.

For Moxin, it has already passed the key milestone of proving that sparse computing is effective. Now, through one real commercialization case after another, the company is proving that sparse computing can be deployed at scale, continue to reduce costs, and bring quantifiable business benefits to more industries.