钛媒体报道,资本正大举投入中国国产芯片领域,覆盖新架构、新产线及整条产业链。
Three days ago, Metax (沐曦) and UBTech (优必选) jointly launched Xixuan Chuangzhi (曦选创智), which announced it had completed a 500 million yuan angel round. Two days before that, smart network interface card maker Muchuang (沐创) and silicon carbide power module maker Lipsi (利普思) also disclosed new funding. Compute chips, data interconnects and power-supply devices, which span several links of the compute infrastructure stack, are all attracting capital.
A week earlier, compute-in-memory chip company Houmo Intelligence (后摩智能) reportedly completed a funding round of several billion yuan. Advanced packaging firm Yicheng Technology (奕成科技) and Lingchuan Technology (凌川科技) also disclosed funding of several billion yuan and 1 billion yuan, respectively. From October 31 to November 6, domestic integrated circuit companies disclosed roughly 11.1 billion yuan in funding, more than half of all disclosed funding that week. ChinaVenture Research Institute data show that semiconductor VC/PE investment in the third quarter of this year reached about 109.2 billion yuan, up 91% year on year.
Many companies have set mass-production targets around 2027, while some firms that have just closed large rounds do not plan to launch their first chip until 2028. Capital is supporting both the expanded delivery of existing products and upfront investment in new architectures and processes that have not yet generated sales. What will this wave of funding mean for competition among domestic chipmakers?
**01 Capital Flows Toward Different Compute Pathways**
ChinaVenture Research Institute data show that semiconductor VC/PE investment in the third quarter reached about 109.2 billion yuan, up 91% year on year. That is well above the 49.6% growth rate of the overall VC/PE market. The number of investment deals also rose by more than 50%. Since October, at least 10 AI chip companies have publicly disclosed funding. They cover compute-in-memory fusion, server CPUs, intelligent driving and embodied intelligence.
On November 4, Houmo Intelligence (后摩智能), a compute-in-memory chip company, reportedly completed a B round worth several billion yuan. The funds will be used to expand M50 deliveries, develop the next-generation M6X and 3D compute-in-memory products, and continue investment in its software toolchain. At its ecosystem partner conference in October, the company disclosed that less than a year after the M50 entered mass production, it had won more than 150 product design wins and was serving more than 100 customers.
Houmo's compute-in-memory technology tightly integrates storage and compute units to reduce energy consumption from data movement. Its products are aimed mainly at local inference on PCs, smart terminals and edge devices. These devices have tight limits on power, memory and thermal headroom, so chips must run models under full-system conditions. Existing products require co-development with customers' complete systems, and next-generation products also require sustained investment.
Weina Hexin (微纳核芯) is also pursuing a compute-in-memory architecture, but at a different commercialization stage. On October 28, the company announced it had completed a C2 round of more than 1 billion yuan, bringing its cumulative funding over the past two months to more than 2 billion yuan. Its first-generation three chips have only recently all taped out, and it expects to deliver them to customers for testing in early 2027 and begin volume supply. The new funds will go not only to product iteration but also to stocking supply chains for volume production. The company is working with memory makers and advanced packaging partners to prepare for subsequent manufacturing.
Stacking storage and compute units vertically requires coordinating chip design, memory supply and packaging processes at the same time. Weina Hexin disclosed that edge-side customers have joined specification definition and algorithm adaptation, while cloud-side customers are evaluating solutions. The first-generation product has not yet achieved meaningful sales volume, so customer onboarding and manufacturing preparation are proceeding in parallel.
Others are starting from specific applications. Lingchuan Technology (凌川科技), which disclosed a 1 billion yuan B round on November 4, was formerly the heterogeneous computing and chip division of Kuaishou. According to investor Qiming Venture Partners, its SL200 intelligent video chip has sold nearly 100,000 units, with customers including Kuaishou, Alibaba Cloud and Baidu Cloud. Its next-generation product shifts to AI video generation and uses 3D stacking. First silicon has recently returned from the foundry, and the company expects large-scale shipments in 2027.
The customer base built through video processing gives Lingchuan a testing and collaboration channel. However, generative video involves different computing tasks, so the new product must still be revalidated. The company has launched its own software platform, which supports models such as Tongyi Wanxiang and Hunyuan Video. This round will also support software and market expansion.
In early November, Moffett AI (墨芯) disclosed a C+ round of nearly 1 billion yuan. The funds will continue to go toward sparse inference chips and research on next-generation products. Unlike compute-in-memory, which reduces data movement, sparse computing reduces unnecessary computation through hardware-software coordination. The company plans to launch its SP200 accelerator board by year-end and to update its software stack to support large-model deployment and cluster operations.
Xixuan Chuangzhi (曦选创智), which has just closed its funding round, is targeting robotics. The company plans to integrate the computing required for perception, decision-making and motion coordination into a single main control chip, with a general-purpose GPU as its core compute unit. It is also developing a software toolchain. The first chip is still in architecture design. The company plans to provide a development and validation platform to industry partners in 2027 and to launch its first SoC and supporting solution in 2028.
Metax contributes GPU architecture and software experience, while UBTech brings the power consumption, interface and deployment requirements of robotics tasks into product definition. From edge inference and video generation to robot control, different workloads create openings for new architectures to enter the market. They also require chip companies to develop software and validate systems around specific tasks.
**02 From Packaging Lines to Power Devices**
High-performance chips increasingly use advanced packaging to integrate processors, memory and other dies, which raises the investment needed for manufacturing support. Yole Group expects global advanced packaging revenue to exceed traditional packaging revenue for the first time in 2026, with related capital expenditure approaching US$28 billion. New chips seeking performance and cost advantages need matching packaging processes, equipment and materials. This extends funding needs from design into manufacturing.
On November 5, Chengdu Industrial Investment Group (成都产投集团) disclosed that advanced packaging company Yicheng Technology (奕成科技) had completed a C round of several billion yuan. This is Chengdu Industrial's fourth investment in the company. Yicheng processes chips on large square panel carriers, which can offer better area utilization than circular wafers.
Chairman Li Chaoliang (李超良) disclosed in late October that the total investment in the company's first plant is 5.5 billion yuan. Construction began at the end of 2021, and the production line was fully commissioned for mass production only in October 2024. Packaging for consumer chips is now in mass production. The company's 2.5D packaging for AI chips targets large-scale mass production in 2027, and a second plant is still in planning. This funding is meant to extend existing operations into more complex packaging processes, not simply to replicate existing capacity.
Chengdu Industrial Investment Group previously funded Yicheng's pilot line, process validation and mass-production stages. In 2025, it also converted convertible bonds into equity, helping the company reduce its debt-to-asset ratio. Packaging companies often invest well before revenue stabilizes, and yield fluctuations or process adjustments can delay deliveries. Continuous investment gives companies time to complete these adjustments, but it also ties capital recovery closely to yield ramp-up and customer onboarding.
According to data disclosed by Li, equipment localization at Yicheng's first plant has reached 70%, and materials localization has reached 60%. Domestic equipment and materials suppliers gain not only procurement opportunities but also the chance to improve processes alongside customers. This helps new packaging routes gradually build cost competitiveness.
Bangxin Semiconductor (邦芯半导体), which disclosed a B+ round of more than 600 million yuan in early November, has located its production base near its memory customers. The company supplies etch, photoresist removal and thin-film deposition equipment. According to its investors, it delivered more than 100 units in 2025. The first phase of its Wuhan Huazhong base began production on October 28, handling equipment assembly, testing, process validation and delivery. Lead investor Jiangxia Science & Technology Investment (江夏科投) not only provided capital but also supplied standardized factory space and coordinated other investors to participate.
Lipsi (利普思), which disclosed a B round of nearly 100 million yuan on November 11, will direct its funds mainly toward R&D and global market promotion of AI data center products. Its existing silicon carbide power modules are used in electric heavy trucks, energy storage and industrial power supplies. The company now hopes to enter power conversion and protection in data centers. Several products have been sampled to customers, and some are in reliability validation. Related business will scale up only after validation is complete.
Lipsi is also building a plant in Yangzhou, expected to begin production in April 2027, with an annual capacity of 1.5 million modules at full output. Plant commissioning and new customer orders do not naturally happen at the same time. The funding must also cover commissioning, validation and inventory preparation in between.
**03 From Capacity Locking to Co-Investment**
In a late-October study, TrendForce noted that cloud providers' capital expenditure continues to drive growth in AI server shipments and output value, with ASICs and Chinese chips seeing increased use. At the same time, advanced packaging, HBM and substrate supply remain tight, which limits rack-level shipment growth. For new chip companies, market expansion creates opportunities to win customers. But after completing R&D, they must still organize supply and adapt to systems before their products reach customers.
On October 28, Korean inference chip company FuriosaAI announced it had completed a US$620 million Series D round. The funds will go to expanding production and deployment and to developing next-generation products. Its RNGD inference accelerator entered mass production in January this year. The company plans to produce 20,000 accelerator cards this year and says it has secured wafer manufacturing and memory capacity to support more than doubling output in 2027.
FuriosaAI is also working with partners including Broadcom to develop a third-generation product. This year it has released four major software development kit updates, adding model support and deployment features. After mass production, spending on software and next-generation R&D continues, so the new funding must support both delivery and subsequent product development.
In July, Samsung SDS added RNGD-based compute services to its cloud platform. Customers can subscribe on demand without first buying servers. The service runs alongside existing GPU services, allowing enterprises to assign suitable inference tasks to the new chip, test cost and performance in real business operations, and then decide how far to deploy. This is closer to ongoing procurement than announcing partnerships or publishing test results, but production targets are still not the same as orders received.
Compared with FuriosaAI's mass-produced RNGD, UK-based Fractile has not yet launched its first commercial chip but has still raised a large round. On November 11, Fractile announced a US$720 million Series C round. Investors include Google's CapitalG, the UK Sovereign AI Fund and British Business Bank. The company plans to launch its first commercial chip in 2027. The UK government hopes this will strengthen domestic AI hardware capabilities and reduce reliance on overseas technology supply. Support for this technology route is arriving before commercial delivery.
Chinese companies are also using investment to organize partnerships early. Muchuang (沐创), which disclosed a new funding round on November 10, received investment from Lihe Suanxin Fund and Baiao Software. The funds will go toward R&D on smart NICs and GPU interconnect chips. Lihe Suanxin Fund's sponsors include compute chip company Dongfang Suanxin, while Baiao Software provides the firmware needed for server startup, configuration and management.
Muchuang's N30-400G smart NIC has been adapted to several GPUs, and the company is co-developing high-speed interconnect solutions with domestic GPU makers. To enter compute clusters, chips must interface with GPUs and adapt to server management software. This round links several types of companies, and the cooperation directly involves interfaces, software and system validation. It is not simply an increase in one company's R&D budget.
The same kind of industrial collaboration extends into packaging. On November 6, Biwin Storage (佰维存储) announced that its subsidiary Guangdong Xincheng Hanqi (广东芯成汉奇) had received 600 million yuan in capital injections. Investors include Zhuhai Biren, a wholly owned subsidiary of Biren Technology (壁仞科技). The storage company retains control, while an AI chip company takes part in investing in an advanced packaging and testing platform. This helps connect processing and validation needs earlier. Whether the equity partnership leads to stable purchasing will still depend on the onboarding of specific projects.
Once design companies secure funding, demand for test tape-outs, pilot production, inventory and packaging validation may increase ahead of large-scale deployment in end markets. Manufacturers can also use these projects to build process know-how and customer relationships. However, R&D-stage procurement should not be taken as direct evidence of long-term capacity utilization. Subsequent capacity expansion will still depend on product validation and order progress. Many companies have set mass-production targets around 2027, and this round of capital is giving more domestic solutions the chance to enter customer supply chains. Whether these investments in chip R&D, manufacturing support and system integration ultimately translate into market share will depend on whether companies can deliver reliably and win sustained customer purchases.
*Originally published by Semiconductor Industry Observer (半导体产业纵横); republished by TMTPost with authorization.*