On 12 October, the primary market unveiled two major funding announcements on the same day.
ChinaVenture exclusively disclosed that AI Infra company Infinigence AI has completed a new round of financing of CNY 2.3 billion, with cumulative financing exceeding CNY 4.3 billion, and has secretly submitted a listing application to the Hong Kong Stock Exchange, with the earliest listing in the first half of 2027; almost at the same time, Pedata reported that RISC-V server CPU company Lanxin Computing has completed a nearly CNY 2 billion Series B financing, with old and new shareholders such as Shenchuangtou, Zhaoyi Chuangxin, and Sugoutou collectively increasing their investments.
One soft, one hard. One is a software platform that "schedules" distributed computing power, and the other is a hardware architecture that "integrates" general-purpose computing and AI inference into the same chip. There is almost no intersection in business, and the track labels are completely different: one grows in the large model ecosystem, and the other grows in semiconductor design.
But if these two financings are only regarded as good news belonging to two different industries, their true commonality will be missed—they are answering the same question: how to make money from computing power in the future.
This question is rewriting the competitive logic of the entire domestic computing power industry.
01. The Smartest Money Is Not in the Hottest Places
Let's first clarify the positioning of the two companies.
Infinigence AI was founded in May 2023, initiated by Wang Yu, a professor at the Department of Electronic Engineering at Tsinghua University, with Xia Lixue, a Ph.D. from the Department of Electronic Engineering at Tsinghua University, serving as the co-founder and CEO. According to Jensen Huang's "five-layer AI architecture," it neither makes chips nor models, but is stuck in the infrastructure layer between the model layer and the chip layer—in the company's own words, it is an "M×N" technical paradigm connecting "M models" and "N chips," organizing fragmented model capabilities and heterogeneous computing power to improve Token utilization efficiency.
Lanxin Computing made a similar choice. Founded in 2023, it abandoned the comfort zones familiar to RISC-V players, such as MCUs, edge computing, and small SoCs, on the very first day of its startup, and directly plunged into the globally blank field of high-performance RISC-V server large chips. Its customers are not consumer electronics manufacturers running small models on mobile phones, but operators, cloud providers, financial institutions, and data centers.
One does software, the other does hardware, but the positioning of both is consistent: instead of stepping into the most crowded spot, both choose to be in the middle layer that is "closest to the hustle and bustle, but does not enter the arena itself."
This is precisely where the capital in this round is most clear-headed. Over the past two years, large model companies and AI application companies have consumed a lot of attention, but the certainty of returns that actually lands in the pocket has repeatedly appeared on the "water sellers." The demand for computing power infrastructure comes from all models and all applications. Not betting on a single player and not bearing the risk of model route replacement—this is the most simple and rational choice for capital at a time when AI applications have yet to see stable returns for a long time.
The financing rhythm of Infinigence AI itself is a footnote: in the first round in August 2023, Shanhong, Beijiguang, and Jinshajiang entered; in December 2023, industrial parties such as Baidu, Tencent, and Zhipu were introduced; in September 2024, the Series A round introduced the "national team" such as the Shebao Zhongguancun Fund; since then, one round per year, until this round of CNY 2.3 billion once again refreshed the domestic AI Infra financing record. Capital has voted for the logic of the "water seller" with real money.
02. What They Are Actually Selling Is the Same Thing: Lower Per-Token Cost
If "stuck in the middle layer" is the commonality in positioning, then "how to bring down the cost of computing power" is the true product commonality of the two.
This year at WAIC, Infinigence AI disclosed the cross-cluster heterogeneous inference architecture PDD. According to the test results published by the company, this architecture can reduce the per-Token cost by 37.5%. Its approach is to unify the scheduling of heterogeneous resources at the system level—the chips can be NVIDIA, AMD, or Shengteng, Biren, Haiguang, Moer Xiancheng, Muxi, Suiyuan, and Pingtouge; the models can be DeepSeek, Kimi, Zhipu, MiniMax, and Jieyue Xingchen. The platform is only responsible for one thing: optimizing their efficiency to the maximum.
The logic of Lanxin Computing, on the other hand, occurs at a deeper level. The LX500 adopts a 48-core self-developed heterogeneous architecture—32 high-performance cores are responsible for high-throughput general-purpose computing, and 16 high-energy-efficiency cores focus on lightweight loads and AI computing power offloading. The chip natively integrates a 75 TOPS INT8 dedicated AI computing engine. What it aims to solve is the inherent cost of the traditional "CPU + AI accelerator card" architecture: high hardware procurement costs, high power consumption, architectural redundancy, and cumbersome adaptation. When inference tasks no longer require the additional purchase of a GPU card, what is saved is not just money, but also the loss of data transportation across chips.
One decouples at the system level, and the other integrates at the chip level; one schedules heterogeneity, and the other eliminates heterogeneity. The paths are different, but the indicators they point to are completely consistent: the cost curve per unit Token.
This is the most easily overlooked yet most critical industrial shift in 2026. For a considerable period, the language to measure computing power competitiveness was peak indicators—TFLOPS, computing power scale, number of cards, and the size of training clusters. But when Agentic AI shifts the focus of computing power demand from training to scheduling and inference, the subject of the industry evaluation system changes: from "how fast it can compute" to "how cheap it is to compute once."
Infinigence AI has reduced the per-Token cost by 37.5%, and Lanxin Computing saves customers from having to buy accelerator cards—essentially, they are doing the same thing: pushing "computing power economics" from the corner of financial reports to the center of the industrial stage.
03. They Are Both Tearing Down the Same Wall: Data Transportation
What is more interesting is that the technical paths of the two seem opposite, but the underlying logic is homologous.
Infinigence AI does "addition": abstracting heterogeneous chips and heterogeneous models into a unified resource pool. The more complex, the better; the value of the platform lies in digesting complexity. Lanxin Computing does "subtraction": integrating general-purpose cores and AI cores into one chip. The more integrated, the better; the value of the chip lies in eliminating intermediate links.
One expands outward, and the other converges inward. On the surface, they go in opposite directions, but in fact, they are tearing down the same wall—the classic paradigm of "CPU + independent accelerator card" that has dominated AI computing power for ten years.
The fundamental contradiction of this paradigm is called the "memory wall" in the industry: the faster the computing power chip, the more difficult it is to ignore the cost of transporting data among memory, cache, and accelerator cards. The traditional solution is to rely on process progress and bandwidth upgrades to tough it out, but when the dividends of advanced processes slow down and costs soar, the path of toughing it out becomes narrower and narrower.
Thus, the dividends of architectural innovation begin to take over the dividends of process innovation. Infinigence AI eliminates transportation at the software abstraction layer, and Lanxin Computing eliminates transportation inside the chip—these are two answers to the same industrial proposition on the software side and the hardware side. They jointly mark the arrival of a node: the narrative of China's semiconductor competition is switching from "process catch-up" to "architectural breakthrough."
Behind the two financings, there is also the same macro variable: autonomous and controllable.
But on closer inspection, the division of labor of the two in the domestic computing power map is completely different.
Infinigence AI plays the role of an "ecosystem aggregator." It simultaneously supports mainstream chips such as NVIDIA, AMD, Shengteng, Biren, Haiguang, Moer Xiancheng, Muxi, Suiyuan, and Pingtouge, essentially providing a unified upper-level interface for fragmented domestic computing power. The real dilemma of domestic chips has never been "inability to manufacture," but rather fragmented software ecosystems, high customer migration costs, and difficulty in being truly utilized. The value of Infinigence AI is to turn "usable" into "easy to use."
Lanxin Computing, on the other hand, plays the role of an "architecture definer." By choosing the open-source instruction set route of RISC-V, it essentially participates in the competition in the server chip field, which has long been monopolized by x86 and ARM, for the first time with the posture of "architectural autonomy" rather than "architectural follow-up." In September this year, RISC-V was officially written into the key tasks of the national five-year plan for the first time, and the "15th Five-Year Plan for the Development of the Electronic Information Manufacturing Industry" explicitly proposed promoting R&D and industrialization of RISC-V—the policy window and the industrial window opened at the same time.
Global industrial forces are also simultaneously increasing their bets on this track: NVIDIA publicly announced the hardware and software standards for RISC-V CPU adaptation to CUDA at Hot Chips 2026, opening the CUDA system, which has monopolized the AI computing power ecosystem for many years, to open-source architectures for the first time; Meta acquired the RISC-V team Rivos in 2025, and its computing power infrastructure built on RISC-V has achieved a deployment scale of hundreds of thousands. These moves indicate that the penetration of RISC-V into data centers is no longer a question of "whether it can or not."
One "strings together" domestic chips at the software layer, and the other "establishes" the autonomous route at the architectural layer. The bottleneck of domestic computing power has long shifted from "whether there are chips or not" to "whether chips can be used systematically." These two companies happen to be stuck at both ends of this closed loop.
04. Conclusion
The two financings on the same day are seemingly good news for two tracks respectively, but essentially they are two facets of the same industrial transformation.
They jointly declare: the focus of competition in China's computing power has shifted from "stacking chips" to "calculating costs"; what determines victory or defeat is no longer how fast a single chip runs, but who can press the per-Token cost curve lower and longer than others.
Their story is still in its early stages. But the direction is no longer blurred.
The information in this article is sourced from public corporate information. If you have any questions, please contact the editor's WeChat: xsychief.