Author | Chen Junda Editor | Moying
Xindongxi reported on November 3 that Moffett Artificial Intelligence (hereinafter referred to as "Moffett AI"), a domestic sparse computing chip company, announced it has recently completed a funding round of nearly CNY 1 billion. This round follows its previous round of nearly CNY 1 billion just five months ago, bringing the cumulative total of the two rounds to nearly CNY 2 billion.
This funding round for Moffett AI was participated in by industrial investors such as Ren'ai Capital, Haitang Fund, and Dongheng Petroleum, as well as prominent state-owned capital and financial investment institutions including Qiu Capital, Paramount 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 next-generation chips, while simultaneously advancing the construction of an industry-academia-research alliance for sparse computing.
Why are investors doubling down at this specific time? What changes have occurred in Moffett AI's commercialization? How does sparse computing translate into visible numbers on customers' ledgers? With these questions in mind, Xindongxi engaged in an in-depth dialogue with Guo Weijun, Vice President of Marketing at Moffett AI.
01. Capturing the Inflection Point of Inference Demand, Accelerating Commercialization
To understand Moffett AI's recent funding, it must be viewed within the broader industry context. Over the past year, large models have truly entered production environments, driving inference call volumes upward and causing costs to rise accordingly. Token consumption is no longer project-based but has become a fixed monthly expense akin to utility bills—a "computing power bill."
Against this backdrop, customers have placed higher demands on next-generation inference computing power. Guo Weijun summarized customer needs into three points: first, lowering comprehensive inference costs, which include not only token bills but also long-term expenditures on electricity, deployment, and compatibility; second, ensuring elastic supply and certain delivery of computing power; and third, controlling adaptation costs, enabling existing models and business systems to migrate with low barriers to entry.
Currently, there are many ways to improve overall system inference efficiency, with sparse computing considered one of the most promising directions. Sparse computing addresses a fundamental issue: while large models often contain tens or even hundreds of billions of parameters, only a small fraction is activated and participates in computation during each inference process.
Traditional chips do not distinguish between active and inactive parts, a method known in the industry as "dense computing." Sparse computing takes the opposite approach, calculating only the effective parts and skipping the rest. For the same inference task, reduced computational load leads to lower costs. Moffett AI has deeply cultivated technical validation in sparse computing for years, and recently, these accumulations have begun to yield concentrated results. These two funding rounds mark a critical watershed for Moffett AI, transitioning from years of single-point technical validation to scalable, commercialized delivery.
On the technical front, Moffett AI's collaborative design philosophy of "Algorithm—Software—Hardware" has been validated in real products, and an end-to-end software closed loop is gradually forming. Moffett AI's S30 and S40 computing cards have secured first place in MLPerf Inference tests for three consecutive editions, providing public test-backed evidence of engineering capability. On the delivery front, increasingly more intelligent computing centers and industry customers are incorporating sparse computing power into their selection criteria.
Moffett AI's sparse-dedicated inference chips have been deployed across four major regions—Northwest, Southwest, East China, and North China—with thousand-card clusters established. The company has also built partnerships with leading telecommunications operators, integrating sparse computing inference solutions into their computing service systems. Moffett AI's scalable delivery capabilities have been verified through these projects.
In terms of market trends, Guo Weijun observed another phenomenon: not every industry requires large general-purpose models. As AI applications enrich vertical industries, sectors such as finance, healthcare, exploration, and meteorology possess their own industry-specific models. These models can almost all benefit from sparse computing, and their implementation scenarios are beginning to show potential for cross-industry replication.
02. Sparse Computing Enters Real Production Environments, Helping Customers Calculate Cost Efficiency
Over the past few months, Moffett AI's list of implemented projects has grown rapidly, ranging from university research platforms to national computing power initiatives. We discussed several real customer cases. In scientific research scenarios, Moffett AI's sparse computing power has achieved quantifiable empirical results.
Focusing on spatiotemporal AI and big data frontiers, Moffett AI collaborated with the Institute of Earth and Space Technology at Peking University's Shanghai Lingang International Science and Technology Innovation Center. Based on proprietary sparsification technology, they completed the adaptation and empirical testing of two internationally mainstream remote sensing image data intelligent analysis foundation models on Moffett AI's computing cards.
Data shows that under conditions of virtually no loss in business accuracy and without requiring retraining for downstream tasks, model inference throughput increased by up to 2.88 times, and model compression reached up to 7.64 times, significantly reducing computing power consumption and processing time for large-scale remote sensing image interpretation.
If the Peking University project validated effectiveness in a single scenario, the next computing power project tested the ability of sparse computing to integrate with existing systems. Moffett AI participated in China's first offshore computing power Token factory, created by a Shanghai state-owned capital platform company. This project uses green electricity from offshore wind and seawater cooling to lower PUE and electricity costs. However, savings from green electricity could potentially be offset by the inference costs of general-purpose GPUs.
Moffett AI's approach involves lowering the single-token inference cost from the chip side, combined with a mixed heterogeneous computing mode, allowing "green electricity cost reduction" and "architecture cost reduction" to stack, further amplifying cost-saving effects. Notably, when communicating with customers, Moffett AI does not deliberately emphasize the technical concept of sparse computing. "What customers care about most is usability; they do not overly focus on whether the route taken is sparsification during the initial stages of the project," said Guo Weijun.
In actual commercial cooperation, customers pay more attention to overall input-output ratios: how many servers need to be deployed in the data center, what are the power consumption and O&M costs, and whether the final result meets business accuracy and ROI requirements. To address this, Moffett AI translates product value into specific cost calculations and is establishing a standardized pre-purchase simulation testing system, allowing customers to obtain corresponding simulation data before procurement to evaluate the actual benefits of the product in their own business scenarios.
Compared to performance metrics, migration costs are often a more insurmountable hurdle in customer decision-making. Existing evaluation scripts and scheduling systems are mostly built around dense computing architectures. Replacing chips requires not only redoing regression tests but also involves engineering modifications and business gray releases, bringing additional time and financial investments.
To tackle this issue, Moffett AI , on one hand, helps customers establish cross-architecture evaluation and supplier comparison systems through sparse evaluation standards and simulation tools. On the other hand, they provide heterogeneous adaptation toolchains, promoting collaborative work between sparse cards and customers' existing dense cards. For customers who have already deployed a large number of legacy chips, this approach reduces system migration and redundant construction costs, providing another path for the implementation of sparse computing.
03. From "Usable" to "Easy to Use", Software and Ecosystem Provide Momentum for Commercialization
Lowering the usage threshold for customers is only the first step toward commercializing sparse computing. Achieving wider implementation requires continuous product iteration, refinement of software toolchains, and joint support from the industrial ecosystem. Moffett AI is advancing its layout in three directions.
In academia, a mature collaborative system among algorithms, mathematical theories, hardware architectures, and software toolchains for sparse computing has not yet formed, and many key issues remain to be studied. To address this, in August of this year, Moffett AI initiated the establishment of China's first industry-academia-research alliance for sparse computing. It has already launched multiple joint projects with universities such as Fudan University, Tsinghua University, Xi'an Jiaotong University, Nankai University, and Shanghai Jiao Tong University.
Guo Weijun revealed that joint projects within the alliance are laid out according to the disciplinary characteristics of each school: Computer Science colleges focus on how combining "algorithms + hardware architecture" improves sparse computing performance; Mathematics colleges concentrate on matrix multiplication algorithm efficiency and the localization of sparse parameter patterns; Software disciplines focus on the development of sparse toolchains, compilers, and high-performance operators.
The task undertaken by the industry-academia-research alliance is not just scientific collaboration but also filling in the foundational elements required for the industrialization of sparse computing.
In Guo Weijun's view, the industry currently faces shortcomings in three aspects:
First is the shortage of compound talents; individuals who understand both algorithms/sparse mathematics and chip architectures are relatively scarce, and the training cycle for such talents typically requires 3 to 5 years;
Second is the lack of unified evaluation and acceptance standards; existing computing power delivery standards are primarily formulated around dense computing, making it difficult to accurately measure the actual capabilities of sparse computing. Guo Weijun stated that Moffett AI may launch acceptance standards for sparse computing power in the future;
Finally, foundational resources such as toolchains and datasets for sparse computing still need improvement. Guo Weijun believes that although this foundational work receives less publicity and attention, it is an indispensable link for the scaled industrialization of sparse computing.
Beyond the ecosystem, the pace of Moffett AI's product line iteration has not slowed. Its next-generation board SP200 is scheduled to be released by the end of this year. Guo Weijun revealed that real pain points exposed during customer use have been incorporated into the design of the next-generation architecture.
Regarding the software stack, Moffett AI's upgraded software stack SECA (Sparse-native Efficient Computing Architecture) is about to be released. From operator libraries and compilation to software-hardware collaboration, the entire software stack is redesigned around the sparse architecture. Supporting MaaS platforms and model framework adaptation management tools will also be launched sequentially. From ecosystem construction and product iteration to software stack upgrades, Moffett AI's momentum for commercialization is worth watching.
04. Conclusion: Is the Turning Point for Sparse Computing Here?
The application space for sparse computing is accelerating its extension from large models to more industries. In Guo Weijun's view, vertical domain models are customized for specific businesses and do not utilize the full capabilities of general-purpose large models, making them naturally suitable for sparse computing. For example, in industrial quality inspection scenarios, visual 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 holds the promise of reducing unnecessary computational overhead.
Embodied intelligence and world models share 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 based on current tasks, a pattern that aligns perfectly with sparse computing.
For Moffett AI, they have passed the critical node of proving the effectiveness of sparse computing. Now, the company is demonstrating through one real commercial case after another that sparse computing can be deployed at scale, continuously reduce costs, and bring quantifiable business benefits to more industries.