智东西（公众号：zhidxcom）
作者 | 李水青
编辑 | 漠影
智东西10月12日获悉，国内领先的AI原生基础设施企业无问芯穹近日已完成C轮融资，融资额23亿元，累计融资额达43亿元，单轮融资额与累计融资额均创国内AI Infra（人工智能基础设施）行业新高。
据了解，本轮融资已完成交割，公司正在筹备IPO。
从2023年成立至今，这家脱胎于清华大学电子工程系的创业公司，已经走到资本市场聚光灯下。背后既有国资机构，也有产业投资方、市场化财务投资机构等对AI基础设施赛道的押注。
放眼海外，AI基础设施公司的估值也在快速攀升。Fireworks AI最新估值达到175亿美元，Baseten达到130亿美元，Together AI达到83亿美元。英伟达对这三家公司的投资，更凸显了算力产业链对AI基础设施的重视。
当下，全球AI产业的竞争正沿着算力、模型向Token生产效率和应用生产力延伸。如何把多元算力转化为稳定、高效、低成本的Token服务，正在成为一门重要生意。
23亿元融资背后，资本究竟看中了国产Token工厂的哪些特质？无问芯穹的竞争力，是踩中了市场的风口，还是已经构建起足够深的技术壁垒？让我们一探究竟。
一、从清华电子工程系出发，啃下异构算力硬骨头
很多人知道，中国AI大模型领域有不少企业都脱胎于清华，智谱、月之暗面、生数科技等头部企业皆是其中的代表。
无问芯穹的创立，同样深深植根于清华大学电子工程系的学术脉络。
公司发起人汪玉是清华大学电子工程系长聘教授，时任系主任，IEEE Fellow；联合创始人兼CEO夏立雪的学士、博士均毕业于清华大学电子工程系，长期从事人工智能系统设计方法学研究，并曾在阿里云从事用户增长产品的开发与管理，经历横跨底层系统研究与产业产品实践。
▲无问芯穹联合创始人兼CEO夏立雪
另一位联合创始人、首席科学家戴国浩同样毕业于清华大学电子工程系，目前任上海交通大学副教授，获多项国际学术会议最佳论文奖；CTO李伯勋则拥有清华大学电子工程系学士、硕士学位，曾任旷视科技算法总监，长期深耕AI算法及软硬件协同优化。
这支团队的共同积累，源自清华大学汪玉老师的“纳米集成电路与系统实验室高能效计算组（NICS-EFC）”。自2008年起，实验组便持续扎根于软硬件协同、AI算力系统研究的前沿，在相关领域积累了十余年的技术经验。这段长期的科研积累，也为无问芯穹的出现奠定了基础。
AI Infra行业发展之初，行业的发展逻辑十分简单，就是堆更多算力支撑更大规模的模型。随着国产芯片不断发展，算力供给逐渐呈现多元化，新的问题随之出现：不同品牌、不同架构的芯片如何协同工作？如何将分散的算力资源转化为稳定、高效、可持续使用的模型训练与推理能力？
这正是无问芯穹团队长期深耕的方向。公开资料显示，无问芯穹在2024年便完成了全球首次“4+2”式700亿参数大模型异构芯片混合训练，运用华为昇腾、天数智芯、沐曦、摩尔线程四种国产芯片与AMD、NVIDIA两种进口芯片，算力利用率达97.6%。
目前其已可以实现2个工作日完成新芯片适配，国产芯片适配成本降低70%以上；在万亿参数级MoE模型训练中，其吞吐较主流开源框架提升41%至75%，在推理上更是实现成本一年内下降10倍。
这些指标的背后，是团队从实验室到产业一线，始终深耕多元异构与软硬件协同创新的技术积累。
从长期看，国产化与算力供给多元化、碎片化，将使异构协同成为持续性的产业需求。无问芯穹提前围绕“M种模型×N种芯片”构建技术体系，恰好覆盖了这一长期需求。
这种先发优势，让它在国产异构布局中占据了有利位置。
二、前店后厂一中心：不碰模型不碰芯片，中立平台卡位Token需求爆发
随着Agent应用加速发展，AI基础设施需要处理的任务更加复杂。无问芯穹在2026年世界人工智能大会中正式提出了“前店后厂一中心”战略布局，分三个部分应对这些挑战。
1、算力集散中心：37000P异构算力，跑通生产级混训
一中心是指“算力集散中心”，对应Agentic Infra自主式基础设施平台，负责汇聚、管理和调度多元异构算力。
针对分散的、碎片化的异构算力现状，无问芯穹打造了面向异构集群的大模型跨域训练系统。该系统已经受过三大类跨域算力聚合训练实践的生产级考验。
第一是超远距离聚合。公司联合中国电信完成国内首例超4000公里距离的大模型跨域混训，在新疆哈密与广东深圳两地集群间实现联合训练性能提升165%。
第二是多数据中心聚合。打通4个不同代际、不同型号的GPU集群，联合训练近百亿参数大模型，四集群协同性能提升41%。
第三是混合云算力聚合。实现本地私有集群与公有云算力的安全互通混训，整体性能提升68%。目前这套跨域训练系统已在全国部署触达超37000P算力、覆盖16种主流芯片。
2、Token工厂：日均调用量40倍增长，推理成本10倍下降
“Token工厂”对应Agentic MaaS大模型服务平台，负责将模型能力转化为可调用的Token服务。
2026年7月，夏立雪在WAIC上披露了无问芯穹首创的新一代推理系统优化成果，跨集群异构PD分离架构PDD。基于Agentic MaaS线上业务长期的大规模实践积淀，该架构在实现首Token延迟降低51.5%的同时，单Token成本可降低37.5%。
过去一年里，无问芯穹的Token工厂通过一系列原始技术创新，推动推理成本10倍下降。公司相信随着跨集群异构PD分离架构的规模化落地，推理成本依然有10倍下降空间。
增长飞轮也已经全面运转。官方数据显示，截至今年7月，该平台单日Token调用量较去年12月实现了40倍爆发增长。
3、AI生产力商店：从3D游戏到律所合伙人，Token在这里变成真金白银
“AI生产力商店”则对应Agentic Infra行业解决方案。对于同一场景，不同模型规模和不同硬件配置可能对应完全不同的Token成本。无问芯穹以智能体场景自适应技术作为提升AI生产力转化效率的核心技术路径，面向行业输出 Agentic Infra 解决方案，把 Token 落地为可交付的产业业务成果。
在文娱游戏场景，无问芯穹携手VAST的Tripo 3D大模型联合打造AGENTIC 3D GAME GEN解决方案；在法律终端场景，公司打造律师事务所专属AI合伙人“律盾芯人”，与多家律所及生态伙伴展开合作；在能源电力场景，公司正与南网能源探索基于智能体技术的智算中心能耗智能预测与协同调度解决方案。
“前店后厂一中心”三者分别覆盖底层算力、中间的模型服务与上层产业应用，形成从资源调度到Token生产，再到生产力转化的业务链条。
无问芯穹还有一个值得关注的特点，它采用轻资产运营模式。无问芯穹自身并不拥有实体数据中心，而是通过技术服务触达管理分布在全国各地的第三方设施中的算力，同时适配华为昇腾、摩尔线程等不同AI芯片，用更轻的资产负债表换取更快的扩张速度。
同时，独立第三方平台的价值正在显现。无问芯穹并不做模型，也不做芯片，保持中立，作为连接模型、芯片和客户需求的独立平台，将分散的模型能力和异构算力组织起来，持续优化从底层资源到AI生产力的转化效率。
纵观当下行业格局，大厂Infra以服务内部业务为主，芯片厂商Infra绑定硬件生态，模型公司自建Infra受竞争关系限制。而选择中立的独立云厂商，正在成长为Token经济的核心枢纽。
目前，无问芯穹的客户与合作伙伴覆盖头部模型企业、科研机构和产业客户。在模型服务方面，公司与Kimi、智谱等企业开展合作，Agentic MaaS平台已上线百余种大模型，支持客户按需调用。
三、融资双料第一：全明星资方，估值仍有空间
43亿元累计融资，一举刷新中国AI Infra融资额新高，这是资本对无问芯穹发展路径的一次集中投票。
从公开披露的投资方来看，公司背后已经形成多元化的股东阵容。自成立以来其投资方阵容堪称豪华，涵盖国资机构、产业投资方、市场化财务投资机构等多个类别。
其中，国资机构包括杭州高新金投、社保中关村基金（君联资本管理）、珠海科技产业集团、厦门国兴投资、徐汇资本、孚腾资本、中保投资、国开科创等；产业投资方包括Dabanc Holding、联想、九坤创投、同歌创投、腾讯、百度等；财务投资机构则包括红杉中国、洪泰基金、惠远资本、启明创投、达晨财智、顺为资本、砺思资本等知名机构。
资方多元，已成为产业共识标的。国家队看长期复利，地方国资看数字经济基础设施，产业资本看生态协同，头部VC看赛道爆发。不同属性的资金同时下注，指向同一个判断：AI基础设施正在从科技项目变成产业底座。
对比海外，国产AI Infra厂商的估值空间仍然明显。
由公开信息可知，Fireworks AI在2025年10月完成2.5亿美元C轮融资，估值达到40亿美元；2026年7月完成15.05亿美元D轮融资后，估值升至175亿美元。Baseten今年连拿两轮融资，估值也在短时间内大幅上升至130亿美元。Together AI则在2026年7月再完成8亿美元融资，估值达到83亿美元。
值得注意的是，英伟达均投资了上述三家公司。对芯片产业链而言，AI基础设施能够帮助更多模型在更大规模上运行，也能进一步释放底层硬件的应用价值。
无问芯穹与它们相似的地方在于，都超前卡位智能体爆发带来的推理市场，都服务模型厂商和AI原生企业。
不同之处在于两点：其一，Fireworks、Baseten、Together AI主要运行在英伟达生态之上，优化对象相对集中；无问芯穹面对的是英伟达、AMD、昇腾、壁仞、摩尔线程等多种芯片并存的现实，必须从底层解决异构协同问题。这在国内是刚需，在海外并非核心命题。
其二，海外Neo Cloud普遍自建或租赁GPU，靠债务扩张，利润依赖算力租赁价差；无问芯穹不拥有数据中心，以软件编排层为核心，赚的是效率提升的钱，资产更轻，扩张更快，对单一硬件体系依赖更低。
这两点共同指向一个海外估值体系尚未充分定价的市场。当前43亿元累计融资，与Fireworks 175亿美元、Baseten 130亿美元的估值相比，仍有明显升值空间。
无问芯穹的融资成功或许释放出一个产业信号：国产AI Infra 玩家，开始叩响全球Neo Cloud的主桌大门。通过输出优质、稳定、可持续的Token服务，中国有能力为全球AI发展贡献独特的中国方案。
结语：需求爆发前夜，国产玩家卡位Token经济枢纽
需求端的爆发已经不再是一个远期判断。Agent应用正在推动Token消耗指数级增长，日均调用量两年增长超过1000倍。谁能以更低成本、更高效率生产Token，谁就掌握了AI时代的基础生产要素。
无问芯穹用“前店后厂一中心”给出了国产玩家的答案。算力集散中心汇聚异构算力扩大可用资源规模，Token工厂把推理成本一年压低10倍，提升Token的生产效率，AI生产力商店让Token在游戏、法律、能源等场景中高效转化为生产力。它以中立姿态连接模型、芯片和客户，以轻资产Neo Cloud模式加速扩张，这种All in Infra的定力，在全球范围内都是有代表性的。
43亿元累计融资，单轮与累计双料第一，众多头部模型厂商与算力供应方已经用合作投票。随着多轮融资完成交割、IPO启动筹备，公司也将进入新的发展阶段。中国AI Infra的定义者，正在从世界工厂走向世界Token工厂，站上全球Neo Cloud的主桌。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
The author, Li Shui-chung
Edit |
He learned on 12 October that the country ' s leading AI primary infrastructure firm had recently completed a round of C finance, amounting to 2.3 billion yuan, with cumulative financing amounting to 4.3 billion yuan, and that both single-wheel financing and cumulative financing had reached the highest levels in the country ' s AI Infra (artificial intelligence infrastructure).
It is understood that the current round of financing has been handed over and that the company is preparing an IPO.
From its inception in 2023 to the present day, the start-up company, which was established in the Department of Electronic Engineering of the University of Tsinghua, has come to the capital market spotlight. Behind them are the stakes on the AI infrastructure tracks of both national and industrial investors and MFIs.
The valuation of AI Infrastructure is also rising rapidly, looking overseas. The latest valuation for Fairworks AI reached $17.5 billion, Baseten $13 billion and Together AI $8.3 billion. The investment of Ingweida in these three companies highlights the importance that the computing chain attaches to AI infrastructure.
At the moment, competition in the global AI industry is extending along logic, modelling, to Token production efficiency and application of productivity. How to transform diversified computing into a stable, efficient and low-cost Token service is becoming an important business.
Behind $2.3 billion in financing, what are the characteristics of the domestic Token plant? Is the competitiveness of the no-questioned core dome ablaze the market, or has a sufficiently deep technological barrier been constructed? Let's find out exactly.
I. Starting with the Qinghua Electronic Engineering Department, eating isomeric hard bones
Many people know that there are many companies in the AZLLM sector in China that have emerged from Qinghua, and that the top companies are represented by Zhipu, the dark side of the moon, and biotechnologies.
The creation of a no-question dome is also deeply rooted in the academic context of the Department of Electronic Engineering at the University of Chinghua.
Wang Yu, the founder of the company, is a professor-in-charge of the Department of Electronic Engineering at the University of Qinghua, then the head of the Department, IEEE Fellow; the co-founder and CEO Bachelor of Science, Ph.D., graduated from the Faculty of Electronic Engineering at the University of Qinghua and has long been engaged in methodological research in the design of artificial intelligence systems and has worked in the development and management of user growth products at Alibaba Cloud and has experienced cross-bottom system research and industrial product practices.
The founder of the Unquestioned Cove and CEO Charles Snow.
Another co-founder and chief scientist, Dai Kuo-Ho, also graduated from the Faculty of Electronic Engineering at the University of Tsinghua, is currently an associate professor at Shanghai University of Transport and has received awards for the best dissertations from international academic conferences; and CTO Lee Bong-hoon, with a bachelor ' s degree and master ' s degree from the Faculty of Electronic Engineering at Tsinghua University, has served as Director General of Invisible Science and Technology and has long-term deep-drive AI algorithms and soft hardware synergetic optimization.
This team has accumulated from the “NiCS-EFC Energy Efficiency Calculatorium for Nano Integrated Circuits and Systems Laboratories” of Teacher Wang Yu of Qinghua University. Since 2008, the team has been firmly rooted in hardware and software synergy, in the front line of the AIS study and has accumulated more than a decade of technical experience in related areas. This long accumulation of scientific research has also laid the foundation for the emergence of a no-question core.
At the beginning of the development of the AI Infra industry, the development logic of the industry was simple: more computing to support larger models. As domestic chips continue to evolve, the availability of computing power becomes more diversified, and new issues arise: how do different brands and different structures work together? How can decentralized computing resources be translated into stable, efficient and sustainable modelling training and reasoning?
This is exactly the direction of long-term deep-droping by the no-questioned dome team. According to publicly available information, the no-question core dome completed its first worldwide “4+2” parameter LLM isomer chip mix training in 2024, using China's four national chips and AMD and NVIDIA imported chips, with 97.6 per cent capacity utilization.
It has now been possible to achieve two working days to complete the new chip adaptation and to reduce the national chip adaptation cost by more than 70 per cent; in the hundreds of billions of parametric MoE model training, it has increased by 41 to 75 per cent over the mainstream open source framework and, logically, by 10 times the cost per year.
Behind these indicators is the technological accumulation of teams working from the laboratory to the industry, always working on multi-isomers and software and hardware synergies.
In the long term, domesticization and diversification of the supply of energy, and fragmentation, will make isomer synergy a continuous industrial demand. This long-term need is well covered by the early construction of the technological system around the “M model x N chip” at the no-question core dome.
This pre-emptive advantage gives it a favourable position in the national isomeric configuration.
II. A centre in the former back shop: without touching the model, without touching the chip, without breaking the demand for the neutral platform token
The task of AI infrastructure needs to be further complicated with the accelerated development of Agent applications. At the World Congress of Artificial Intelligence, 2026, the no-questioned dome formally presented the strategic layout of the “one centre of the front shop”, addressing these challenges in three parts.
1 Calculus disassembly centre: 37,000 P isomeric calculus, running through production-level mixed training
A centre is the “calculation hub”, which corresponds to the Auto-Infra infrastructure platform and is responsible for the assembly, management and movement of multi-isomer algorithms.
In response to the current state of dispersed, debrisized isomeric algorithms, a cross-area LLM training system for isomers was developed without question. The system has been tested at the production level by three types of cross-domain aggregation training practice.
The first is ultra-range aggregation. The company United China Telecommunications completed its first domestic LLM cross-border training at a distance of more than 4,000 kilometres, resulting in a 165 per cent increase in joint training performance between the Homi and Shenzhen clusters in Xinjiang and Guangdong.
The second is multi-data integration. Through four GPU clusters of different generations and models, joint training of nearly 10 billion parameter LLM was conducted, with 41 per cent improvement in the synergy performance of four clusters.
And the third is mixed cloud computing. The overall performance of local private clusters was increased by 68 per cent through the implementation of security training between local private clusters and public cloud computing. The cross-area training system has now deployed over 37,000 P-calculations nationwide, covering 16 mainstream chips.
Token Factory: 40 times the average daily call volume, 10 times the cost of reasoning
The “Token Plant” corresponds to the Agentic Maas LLM service platform and is responsible for converting model capabilities into a service for Token that can be accessed.
In July 2026, Shelley disclosed on WAIC the results of the optimization of the new generation of reasoning systems initiated without question cores, the cross-cluster isomer PD separation architecture, and the PDF. Based on the long-term large-scale accumulation of operations on the Agenic MaaS line, the structure could reduce the cost of single Token by 37.5 per cent while achieving a 51.5 per cent delay in the first Token.
Over the past year, the Token factory, which has no core, has contributed to a tenfold reduction in the cost of reasoning through a series of original technological innovations. The company believes that with the size of the cross-cluster isomer PD separation architecture, there is still 10 times less space for reasoning costs.
The growth wheel is also fully operational. Official data indicate that, as of July this year, Token ' s single-day call for the platform had increased 40 times more than last December.
3-AAI Productivity Shop: From 3D game to law firm partner, Token here becomes real gold White Silver
The "AI Productivity Shop" corresponds to the Agenic Infra industry solution. For the same scenario, different model sizes and different hardware configurations may have completely different Token costs. The no-question core dome uses AI agent scenario adaptation technology as a core technology path to increase the productivity transformation efficiency of AI, with industry-oriented Agentic Infra solutions and token landing as a deliverable industrial business outcome.
In the entertainment game scene, Tripo 3DLLM of VAST has worked together to create the AGENTIC 3D GAME GEN solution; in the legal terminal scene, the company has created the law firm's exclusive AI Partner, “Counsels”, in collaboration with multiple law firms and ecological partners; and in the energy and electricity scene, the company is exploring with the South Net Energy Center for energy-intensive intelligence prediction and coordinated dispatch solutions based on AI agent technology.
The “one centre at the front shop” covers bottom computing, intermediate modelling services and upper-level industrial applications, forming business chains from resource transfer to Token production to productivity transformation.
A further feature of interest is the no-question core, which uses a light asset operating model. The no-question dome itself does not have a physical data centre, but rather through technical services to manage the calculus in third-party facilities spread throughout the country, with a mix of different AI chips, such as a whistling, a moor line, and a lighter balance sheet in exchange for faster expansion.
At the same time, the value of independent third-party platforms is emerging. The no-question core do not model nor chip, remain neutral, act as an independent platform for connecting models, chips and customer needs, and organize decentralized model capabilities and isomer algorithms to continuously optimize the efficiency of conversion from bottom to AI productivity.
Throughout the current industry pattern, the major plant, Infra, is dominated by service in-house operations, the chip manufacturer, Infra, binds the hardware ecology, and the model company, Incra, is constrained by competitive relationships. The choice of neutral independent cloud manufacturers is growing into the core hub of the Token economy.
At present, clients and partners who do not ask the core cover head model businesses, scientific institutions and industry clients. With regard to modelling services, the company works with companies such as Kimi, Zhipu, and the Agenic Maas platform has reached over 100 LLM to support client calls on demand.
III. FINANCING DIVERSITY FIRST: A full-star capital, room for valuation
The cumulative financing of $4.3 billion and the freshening of China ' s AI Infra financing is a centralized vote of capital on the development path without questions.
In terms of publicly disclosed investors, the company has developed a diversified shareholder base behind it. Since its inception, its investor base has been a luxury, encompassing a wide range of categories, such as State institutions, industrial investors and market-based financial investment institutions.
Among these are the Hangzhou New Cash Investment, the China-China Social Security Fund, the Pearl Sea Science and Technology Industry Group, Xiamen State Investment, Xu Seki Capital, Futeng Capital, China Insurance Investment, the State Establishment, etc.; the industrial investors include Dabanc Holding, the Union, Kowloon Genesis, the Cosmopolitik, ZQongxi, and ZBaidu; and the financial investment agencies include China, Hongtai Foundation, Huixiang Capital, the Enlightened Start-up, Daqim, Sun-kyu Capital, and others.
The diversity of resources has become the subject of industry consensus. The country team looked at the long-term recovery, the local state looked at the digital economy infrastructure, industrial capital looked at ecological synergy, and the head VC looked at the runway. Funds of different attributes are being betted to the same conclusion: AI infrastructure is moving from a technology project to an industry base.
Compared to overseas, the valuation space of the national production AI Infra is still evident.
According to publicly available information, Fireworks AI completed the $250 million C round of financing in October 2025, valued at $4 billion, and the value rose to $17.5 billion after the $1.505 billion D round of financing completed in July 2026. Baseten took two rounds of financing this year, and the valuation increased significantly to $13 billion in a short period of time. Together AI completed a further $800 million in financing in July 2026, valued at $8.3 billion.
It is worth noting that both of these companies are invested in Weida. For the chip industry chain, AI infrastructure can help more models operate on a larger scale and can further release the application value of bottom hardware.
The no-question core domes are similar to them in the reasoning market brought about by the AI agent outbreak, both for model manufacturers and for the AI primary business.
The difference is two: first, the Fireworks, Baseten, Together AI, which is mainly based on the British ecology, optimizes the relative concentration of its objects; the no-question core dome is faced with the reality of the confluence of different chips, such as the British Weeda, the AMD, the Stamp, the Walls, the Mole Line, and so on, which must be addressed from the bottom. This is a domestic demand and a non-core proposition overseas.
Second, overseas Neo Cloud has generally built or leased GPU on its own, debt expansion and profits on the basis of arithmetical lease differentials; unquestioned core domes do not have data centres, with software layers at their core, making more efficient money, lighter assets, faster expansion and lower reliance on a single hardware system.
Together, they point to a market where the system of overseas valuation is not fully priced. There is still room for a significant appreciation of the current cumulative financing of $4.3 billion compared to the valuation of the Fireworks $17.5 billion and the Baseten $13 billion.
The successful financing of the no-question dome may have sent an industrial signal: the national AI Infra player, who began to bow to the global Neo Cloud table. By exporting high-quality, stable and sustainable Token services, China has the capacity to contribute unique Chinese programmes to global AI development.
Conclusion: On the night before the outbreak of demand, the national player card token economic hub
The demand-side outbreak is no longer a long-term judgement. The Agent application is driving up the Token consumption index, with a daily average increase of more than 1,000 times over two years. Those who can produce Token more cheaply and efficiently have mastered the basic factors of production of the AI era.
Unquestioned core dome provides the answer for national players using the "one center of the back shop". The Calculator Cluster has brought together isomeric algorithms to scale up available resources, the Token factory has reduced the cost of reasoning by 10 times a year, improving Token ' s productivity, and the AI Productivity Shop has enabled Token to efficiently convert to productivity in games, laws, energy, etc. It connects models, chips and customers in a neutral manner and accelerates expansion in the light asset Neo Cloud model, the force of All in Infra that is representative globally.
Accumulated financing of $4.3 billion, single-wheel and cumulative double-soldiers, and many head modelers and calculators have voted together. As multiple rounds of finance are completed and the IPO is ready to start, companies will also enter a new stage of development. China ' s definitionator, AI Infra, is moving from a world factory to a world factory in Token, on the global Neo Cloud table.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.