广告
加载中

清华博士卖Token 公司成立三年 冲刺IPO

漫地 2026-07-08 10:57
漫地 2026/07/08 10:57

邦小白快读

EN
全文速览

本文核心是清华博士袁进辉二次创业打造硅基流动,成立不到三年冲击港股“AI token工厂第一股”,核心干货信息如下:

1. 核心业务定位:打造AI时代的Token工厂,把底层算力和各类开源大模型打包成可直接调用的API,用户不需要自行采购显卡搭建环境,仅写几行代码就能使用AI,定位独立生态词元供应商,按2025年词元年吞吐量计,硅基流动排中国所有词元供应商第四,独立生态供应商第一,跨芯片多模型适配能力全球第一。

2. 创业与融资:袁进辉师从AI奠基者张钹院士,首次创业做深度学习框架OneFlow被并购后,带领原团队二次创业,不到三年完成7轮融资,估值从2.8亿元涨到77.4亿元,涨幅超33倍,阿里巴巴是最大机构股东。

3. 商业化现状:目前仍在早期,2025年营收5533万元,同比增长653%,公有云服务已经成为第一大收入来源,但仍处于亏损状态。

本文为AI领域品牌商梳理了第三方MaaS赛道头部玩家的发展经验,可参考的干货如下:

1. 用户与市场趋势:当前AI市场客户分层需求明确,开发者和中小企业看重低成本、开箱即用的AI服务,大型企业更看重数据安全和私有化部署能力,不同层级客户需求差异清晰,可针对性布局。

2. 产品研发与竞争方向:可走差异化竞争路线,通过支持国产芯片构建差异化供应链,避开在英伟达稀缺算力上和巨头直接竞争;提前适配潜力新模型可抓住流量红利,硅基流动就因提前适配DeepSeek,成为首家基于华为昇腾提供满血版服务的平台,顺利承接流量,用户破千万。

3. 商业化结构:可采用公有云服务加本地部署双轮驱动模式,公有云按Token消耗量计费面向中小客户,本地部署针对大型企业收取软件许可、实施及维保费用,匹配不同客户需求。

本文给AI赛道相关卖家梳理了独立词元供应赛道的机会与风险,干货内容如下:

1. 赛道机会:随着AI应用规模化落地,推理需求快速增长,独立词元供应是新兴增长赛道,这类玩家不绑定任何云提供商、模型或应用场景,差异化生存空间大,硅基流动成立不到三年就做到独立生态供应商第一,赛道增长潜力明确。

2. 可借鉴的经营经验:竞争上走国产算力差异化路线避开巨头打压,提前布局潜力模型获取流量;商业化采用分层模式,公有云服务跑量做规模,本地部署做高客单,2025年硅基流动本地部署单客户平均收入已经达到130.3万元,盈利空间可观。

3. 风险提示:当前赛道仍处于发展早期,需要大量采购算力、投入营销与研发,很容易出现负毛利亏损,硅基流动2025年毛利率为-24%,全年亏损3.45亿元,创业者需要提前做好资金储备与规划,平衡扩张与成本。

本文给工厂数字化转型、拓展新业务提供了相关干货,具体如下:

1. 转型商业机会:当前AI大模型已经逐步应用到生产、设计等环节,工厂做数字化转型不需要自行采购高额算力、搭建大模型环境,可以通过硅基流动这类第三方Token服务平台,低成本快速获得可用的AI能力,大幅降低转型门槛。

2. 需求匹配方案:如果工厂对数据安全有较高要求,可以选择私有化本地部署方案,将AI推理能力部署在工厂自有基础设施内,既满足生产、设计环节的AI使用需求,又能保障核心生产数据不对外泄露,符合工厂的实际需求。

3. 新增长曲线机会:有相关资源的实体工厂可以通过产业合作切入AI赛道,比如拥有绿色电力资源的晶科能源就作为产业资本投资硅基流动,实现绿色电力与算力的协同,为自身开拓了新的增长方向,也给同类工厂提供了可参考的路径。

本文给AI相关服务商梳理了赛道趋势和客户解决方案,干货内容如下:

1. 行业发展趋势:当前AI大模型落地进入加速阶段,市场对便捷、低成本、中立开放的AI推理服务需求快速增长,独立第三方词元供应已经成为一个新兴赛道,资本认可度高,多家头部玩家已经进入B轮及以后融资阶段,赛道发展速度快。

2. 核心技术与客户痛点:当前客户的核心痛点是想用AI但不想承担高额的前期硬件和研发投入,同时存在多芯片适配的需求,硅基流动的解决方案是打造专有推理引擎加算力资源编排系统,高效调度异构算力,输出标准化的Token服务,很好解决了行业痛点。

3. 分层客户解决方案:针对不同客户需求推出对应服务,面向开发者和中小企业推出按Token消耗量计费的公有云服务,降低使用门槛;面向有数据安全需求的大型企业和机构推出本地私有化部署方案,满足个性化和安全需求,可覆盖全层级客户。

本文给AI相关平台商梳理了第三方平台的发展经验和风险点,干货内容如下:

1. 市场需求方向:当前市场对中立开放、不绑定特定云厂商或特定模型的AI基础设施平台需求强烈,第三方中立平台不依托自有大模型和云资源,差异化竞争优势明显,存在很大的市场发展空间。

2. 可参考的运营做法:定位中立开放的AI基础设施平台,聚焦推理侧工程化能力,适配多芯片和上百种主流大模型;走差异化竞争路线,依托国产芯片避开和互联网巨头的直接竞争,提前适配潜力大模型抓住流量红利,用免费Token代金券做推广获取用户,快速做大规模。

3. 需要规避的运营风险:当前赛道处于商业化早期,需要大量投入算力采购、研发和营销获客,很容易出现毛利率下滑、持续亏损的问题,平台需要平衡规模扩张和成本控制,提前对接多类资本,引入产业资本对接场景资源,保障资金链稳定。

本文给AI产业研究者提供了独立词元供应赛道的最新产业动向,干货内容如下:

1. 产业新动向:随着AI大模型应用落地加速,在AI基础设施和大模型应用之间,诞生了独立词元供应商这个全新细分赛道,目前赛道已经聚集了多家头部玩家,其中硅基流动、无问芯穹、潞晨科技三家核心玩家均为清华系,都进入B轮及以上融资阶段,硅基流动已经递交招股书冲刺IPO,赛道进入快速商业化阶段。

2. 代表性商业模式:目前赛道已经分化出三种清晰的差异化商业模式,分别是硅基流动的“推理优先+中立Token服务”模式,潞晨科技的“自研大模型+推理平台”闭环模式,以及无问芯穹的“异构混训算力运营+芯片IP变现”模式,三种模式定位各有不同,为产业研究提供了完整样本。

3. 产业待解问题:目前整个赛道仍处于商业化早期,玩家普遍需要高额投入研发、算力和营销,头部玩家也出现了毛利率转负、持续亏损的问题,规模效应能否覆盖成本、商业模式能否盈利仍待验证,是值得后续研究的核心问题。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

This article centers on SiliconFlow, a startup founded by Yuan Jinhui, a Tsinghua PhD, for his second entrepreneurial venture. Less than three years after its founding, the company is now competing to become the first "AI token factory" listed on the Hong Kong Stock Exchange. Key takeaways are as follows:

1. Core business positioning: SiliconFlow positions itself as an AI-era "token factory" that packages underlying computing power and various open-source large models into directly callable APIs. This eliminates the need for users to purchase graphics cards or build development environments on their own—users can access AI capabilities with just a few lines of code. As an independent ecosystem token provider, it ranks 4th among all Chinese token providers and 1st among independent ecosystem providers in terms of 2025 projected annual token throughput, and holds the top global position in cross-chip multi-model adaptation capability.

2. Entrepreneurship and financing: Yuan studied under academician Zhang Bo, a founding pioneer of AI. After his first startup, deep learning framework OneFlow, was acquired, Yuan led the original team to launch this second venture. In less than three years, the company has completed 7 rounds of financing, with its valuation surging from RMB 280 million to RMB 7.74 billion, an increase of more than 33 times. Alibaba is its largest institutional shareholder.

3. Commercialization status: The business is still in an early stage. It projected revenue of RMB 55.33 million for 2025, representing a year-on-year increase of 653%. Public cloud services have become its largest revenue source, but the company still remains unprofitable.

This article summarizes development insights from leading players in the third-party MaaS track for AI industry brands, with key takeaways as follows:

1. User and market trends: Current AI market customer demand is clearly stratified. Developers and small-to-medium enterprises (SMEs) prioritize low-cost, out-of-the-box AI services, while large enterprises place more emphasis on data security and private deployment capabilities. With clear demand differences across customer segments, brands can pursue targeted layout.

2. Product R&D and competitive strategy: Brands can adopt a differentiated competition strategy by building a differentiated supply chain through supporting domestic chips, avoiding direct competition with tech giants over scarce NVIDIA computing power. Proactively adapting to promising new models allows brands to capture traffic dividends. SiliconFlow, for example, became the first platform to offer full-configuration DeepSeek services based on Huawei Ascend chips by adapting the model early, successfully captured incoming traffic, and grew its user base to over 1 million.

3. Commercial structure: Brands can adopt a dual-driven model combining public cloud services and on-premise deployment. Public cloud services, billed by token consumption, target SMEs, while on-premise deployment serves large enterprises with software licensing, implementation and maintenance fees, matching the needs of different customer groups.

This article outlines opportunities and risks in the independent token supply track for AI-related sellers, with key insights as follows:

1. Track opportunities: With the large-scale adoption of AI applications, inference demand is growing rapidly. Independent token supply is an emerging high-growth track. Since players in this track are not bound to any specific cloud provider, model or application scenario, they enjoy large room for differentiated survival. SiliconFlow became the top independent ecosystem provider in less than three years, signaling clear growth potential for the track.

2. Transferable operational experience: Compete through a differentiated strategy centered on domestic computing power to avoid direct confrontation with giants, and capture traffic by laying out promising models in advance. Adopt a tiered commercialization model: public cloud services drive volume and scale, while on-premise deployment delivers high average order value. In 2025, SiliconFlow’s average revenue per on-premise deployment client reached RMB 1.303 million, indicating considerable profit potential.

3. Risk warnings: The track is still in an early development stage, requiring massive spending on computing power procurement, marketing and R&D, which often leads to negative gross margins and losses. SiliconFlow posted a -24% gross margin and a full-year net loss of RMB 345 million in 2025. Entrepreneurs need to make adequate capital reserves and planning in advance to balance expansion and cost control.

This article shares key insights for factories pursuing digital transformation and new business expansion, as follows:

1. Transformation opportunities: Large AI models are now gradually being applied to production, design and other factory links. To pursue digital transformation, factories do not need to purchase high-cost computing power or build large model environments on their own. They can access usable AI capabilities quickly at low cost through third-party token service platforms like SiliconFlow, greatly lowering the transformation threshold.

2. Demand-aligned solutions: For factories with high requirements for data security, private on-premise deployment is available, which places AI inference capabilities within the factory’s own infrastructure. This meets the demand for AI in production and design links while preventing core production data from leaking outside the organization, aligning well with factories’ actual needs.

3. New growth curve opportunities: Physical factories with relevant resources can enter the AI track through industrial cooperation. For example, JinkoSolar, a solar manufacturer with access to low-cost green power, invested in SiliconFlow as an industrial capital, enabling synergy between green power and computing power, opening up a new growth direction for itself, and providing a reference path for similar factories.

This article summarizes track trends and customer solutions for AI-related service providers, with key takeaways as follows:

1. Industry development trends: The adoption of large AI models is now accelerating, driving rapid growth in market demand for convenient, low-cost, neutral and open AI inference services. Independent third-party token supply has become an emerging track with high capital recognition; multiple leading players have already completed B-round or later financing, indicating the track is growing rapidly.

2. Core technology and customer pain points: The core pain point for customers is that they want to leverage AI but do not want to bear high upfront costs for hardware and R&D, while also requiring multi-chip adaptation. SiliconFlow addresses this pain point by building a proprietary inference engine paired with a computing power orchestration system, which efficiently schedules heterogeneous computing power and delivers standardized token services, effectively solving the industry’s core pain points.

3. Tiered customer solutions: Providers can launch tailored services for different customer needs: for developers and SMEs, offer public cloud services billed by token consumption to lower the barrier to entry; for large enterprises and institutions with data security requirements, offer on-premise private deployment to meet personalized and security needs, enabling coverage of the full customer spectrum.

This article summarizes development experience and risk points for third-party AI platform operators, with key takeaways as follows:

1. Market demand direction: There is currently strong market demand for neutral, open AI infrastructure platforms that are not bound to specific cloud vendors or models. As third-party neutral platforms do not rely on proprietary large models or cloud resources, they have clear differentiated competitive advantages and large room for market growth.

2. Reference operational practices: Position yourself as a neutral and open AI infrastructure platform, focus on inference-side engineering capabilities, and support adaptation for multiple chips and hundreds of mainstream large models. Pursue differentiated competition by leveraging domestic chips to avoid direct competition with Internet giants, adapt promising large models in advance to capture traffic dividends, and acquire users through promotions such as free token coupons to scale up quickly.

3. Operational risks to avoid: The track is still in the early stage of commercialization, requiring massive investment in computing power procurement, R&D and customer acquisition, which often leads to declining gross margins and sustained losses. Platforms need to balance scale expansion and cost control, secure funding from multiple types of capital in advance, and bring in industrial capital to access scenario resources, to ensure stable cash flow.

This article shares the latest industry developments in the independent token supply track for AI industry researchers, with key insights as follows:

1. New industry developments: With the accelerating adoption of large AI model applications, a brand-new niche segment—independent token suppliers—has emerged between AI infrastructure and large model applications. The track now hosts multiple leading players, three of the core players—SiliconFlow, Wuwen Xinqiong and Luchen Technology—are all founded by teams with Tsinghua University backgrounds, and all have completed B-round or later financing. SiliconFlow has already filed its prospectus to prepare for an IPO, marking that the track has entered a phase of rapid commercialization.

2. Representative business models: Three clear differentiated business models have emerged in the track: SiliconFlow’s "inference-first + neutral token service" model, Luchen Technology’s closed-loop "proprietary large model + inference platform" model, and Wuwen Xinqiong’s "heterogeneous hybrid training computing operation + chip IP monetization" model. The three models have distinct positioning, providing a complete sample set for industry research.

3. Unresolved industry questions: The entire track is still in an early commercialization stage, and players generally require massive investment in R&D, computing power and marketing. Even leading players face negative gross margins and sustained losses. Whether economies of scale can offset costs and whether the business model can achieve profitability remain unproven, and are core questions worthy of follow-up research.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

近日,硅基流动(SiliconFlow)提交了港股招股书,冲击“AI token工厂第一股”。

围绕“独立生态词元供应商”这一核心定位展开,硅基流动正在打造一个AI时代的“Token工厂”。简单说,就是把底层算力和各种开源大模型打包成可直接调用的API,让用户不用自己买显卡、搭环境,写几行代码就能用上AI。

按2025年词元年吞吐量计,硅基流动目前已跻身中国所有词元供应商第四,并在独立生态词元供应商中排名第一,市场份额1.5%。

据弗若斯特沙利文数据,截至目前,硅基流动的平台已累计支持超过170个主流AI模型,其跨芯片、多模型适配能力位列全球第一。在商业化落地方面,硅基流动采用了公有云服务与本地部署解决方案双轮驱动的业务结构:公有云服务面向开发者与中小企业,按Token消耗量计费;本地部署解决方案则主要面向有数据安全和私有化需求的大型企业与机构客户,提供软件许可、实施及维保服务。

硅基流动创始人袁进辉博士毕业于清华大学计算机系。2016年,他创立一流科技(OneFlow),打造了开源深度学习框架。2023年,经历被光年之外和美团并购的波折后,袁进辉放弃丰厚的留任邀约,带领核心团队于8月创立硅基流动。

自2023年12月成立至2026年6月递表港交所,硅基流动在不到3年时间内累计完成了7轮上市前融资,估值从天使轮后的2.8亿元一路攀升至B+轮后的77.4亿元,涨幅超过33倍。其中,B轮融资超20亿元,创下了当年国内第三方MaaS(模型即服务)赛道最大单笔融资纪录。

阿里巴巴是最大机构股东,持股7.42%。其他主要外部股东包括哈勃科技持股4.07%、创新工场持股4.01%、华创资本持股3.01%、耀途资本持股3.01%、王慧文持股2.06%、普华资本持股1.81%。

清华博士后的两次创业

2008年,袁进辉在清华大学计算机系获得博士学位,师从中国人工智能奠基者张钹院士。学术功底深厚的袁进辉,其博士论文也获得了校级优秀荣誉。

随后,袁进辉继续在清华大学从事计算神经科学的博士后工作。博士后阶段与脑科学等交叉学科的碰撞,让他在深度学习尚未进入大众视野的年代起,便在心中埋下了大模型的种子。

袁进辉前瞻性地判断,人脑拥有约千亿个神经元,连接规模远超当时的大多数模型——智能若要真正“涌现”,模型必须做得足够大。带着这份判断,他曾作为早期成员加入360搜索团队,后又在微软亚洲研究院担任主管研究员。

在微软研究院的第二年,袁进辉便发明了当时世界上最快的主题模型训练算法LightLDA。这意味着只用数十台服务器即可完成以前数千台服务器才能实现的大规模主题模型,该技术成功应用于微软在线广告系统。

2016年,怀着“在AI领域重要问题上做出贡献”的初心,袁进辉创立了一流科技(OneFlow),专注于研发深度学习的底层框架。在当时,这是一场极为“反共识”的“非典型”探索,因为大模型还不是主流,业界更多流行的深度学习模型还是卷积神经网络,模型都很小,只有几千万参数。而OneFlow的目标却直指未来超大规模模型的底层架构。

但是,与其说OneFlow是一个公司,其实它更像一个实验室。因为在其成立的前四年没有任何商业化收入,员工大部分工作是在写代码。甚至于,“CEO写代码是对公司不负责任的做法”这个道理,也是袁进辉在创业后期才领悟的。

不过即便如此,OneFlow凭借先进的设计理念,不仅做出了产品,其静态调度与流式执行架构甚至影响了包括PyTorch在内的国际主流框架。

然而不幸的是,随着ChatGPT引爆全球AI热潮,在大模型从少数人的共识变成全行业焦点之时,OneFlow却因资源限制,在商业转化上未能同步推进。

2023年,这艘技术之船被美团联合创始人王慧文创立的光年之外并购,随后又因光年之外被美团收购而一同并入美团。面对美团的留任邀请与同行抛出的超千万年薪橄榄枝,袁进辉却犹豫了。

在AI浪潮刚掀起之时,难道要放弃创业走下牌桌吗?在思考了一个多月之后,袁进辉的答案是:继续前进。为了让原来的团队不被肢解,他和原班人马于2023年8月再度出发,创立了硅基流动。

相较于上一次的创业,创建硅基流动时期的袁进辉显然更为成熟与务实。他曾坦言,OneFlow的经历“有点像是一次行为艺术”,技术上的好奇心在前一次的创业过程中已满足,但是还未完成的商业层面,他想在硅基流动实现。

因此,硅基流动不再自研框架,而是定位于AI时代的“Token工厂”,即一个中立、开放的AI基础设施平台,旨在将各种算力资源和大模型“消化”并统一输出成标准化的Token服务,让开发者像用水电一样使用AI模型。

硅基流动的核心技术基础由专有推理引擎及算力资源编排系统组成,二者共同构成专为AI推理工作打造的一体化技术栈。硅基流动的技术能够高效编排异构算力资源,并将底层算力转化为高吞吐量、低延迟且具备成本效益的词元生产能力。

这一回,袁进辉描绘的商业路径也更清晰:通过支持国产芯片构建差异化供应链,避免在稀缺的英伟达算力上与巨头直接竞争,从而率先占领开发者和中小客户市场。

2025年初,当DeepSeek在全球爆火,验证了硅基流动的前瞻性。凭借此前对DeepSeek潜力的判断和适配,硅基流动成为业内首家基于华为云昇腾算力提供满血版DeepSeek服务的平台,成功接住了“泼天”流量,平台总用户数随之突破千万。

两年融资七轮

阿里、华为押注

在硅基流动成立四个月后,2024年1月,硅基流动完成了天使轮融资。此轮融资金额为5000万元人民币,由创新工场领投,耀途资本、奇绩创坛以及美团联合创始人王慧文等跟投。

值得注意的是,在天使轮融资中,王慧文的身份不仅是投资人,更作为创始股东深度参与,在公司战略、管理、组织乃至高管招聘等层面给予了诸多指导。本轮融资后,硅基流动的估值从投前的2.3亿元升至约3.06亿元。

2024年7月的天使+轮融资中,智谱AI和360率先以产业方身份进入,华为也通过旗下哈勃科技入局。美团则通过天津三快科技入股。

互联网巨头也开始入场。阿里巴巴通过杭州多项网络科技有限公司在A轮融资(2025年6月前后)中以领投方身份进入。

2026年6月的B轮融资是产业资本集中入场的节点,单轮融资金额超20亿元。本轮新增产业资本包括:携程战投、壁仞战投、蔚来资本、商汤战投(AI技术协同),晶科能源控股(绿色电力与算力协同)、金蝶股份(企业级SaaS场景)、巨人网络(游戏娱乐场景)、联通新沃及联通资本旗下联创创新基金(通信运营商算网融合),以及盛奕资本(润泽集团产业资本,代表算力基础设施方)。

从硅基流动的整体融资机构背景来看,财务投资机构贯穿多轮融资。创新工场在早期天使轮即领投,目前持股4.01%;而耀途资本则是唯一从天使轮起连续三轮跟投的机构,持股3.01%;另外普华资本在Pre-A轮跟投入局,纪源资本、华登国际等美元基金均在B轮进入。

国资平台与政府基金方面,国开金融代表国家级金融资本、中关村科学城为地方产业引导基金、北工投资为北京市属国资平台、国泰君安创新投等均在B轮进入。

硅基流动自2023年8月成立至递表港交所,已累计完成7轮上市前融资,估值从天使轮后的2.8亿元攀升至B+轮后的77.4亿元。

目前,创始人袁进辉及四个员工激励平台共同构成公司的控股股东层,合计控制已发行股份中约44.48%的投票权。其中,袁进辉博士直接持有公司股份的14.35%,为单一最大个人股东;四个员工激励平台:硅基创新、硅基未来、硅基探索各持股7.85%,硅基卓越持股6.59%,四个平台合计持有约30.14%的股份。

仍在商业化早期

硅基流动成立的2023年,其收入仅为6000元;2024年全年营收735万元;2025年营收大幅增长至5533万元,同比增幅高达653%。

短短三年时间,按年度词元吞吐量计,硅基流动已经为2025年中国第四大词元供应平台,市场份额1.5%,在所有独立生态词元供应平台中排名第一。

不过,营收步步趋高的背后,毛利率却在快速下降。2024年,硅基流动毛利率为39.4%,2025年则骤降至-24%,即每提供1元词元服务需倒亏0.24元,2025年的负毛利主要是由于公司大量采购算力资源。

其销售及营销开支从2024年的639万元暴增1210%至2025年的8374万元,其中推广型算力资源成本(免费Token代金券)占比高达64.7%。研发投入方面,2025年研发费用达2.09亿元,是当年营收的3.78倍。

同期,硅基流动在年内亏损分别为8192万元和3.45亿元,经调整净亏损(剔除股份支付与赎回负债利息)分别为5401万元和1.87亿元。

从硅基流动的业务模式来看,公司采用公有云服务与本地部署解决方案双轮驱动。2024年,本地部署解决方案收入占比85.4%,公有云服务仅占14.6%;2025年,公有云服务收入2926万元,占比跃升至52.9%,首次成为第一大收入来源。

公有云服务方面,2025年,硅基流动面向开发者与中小客户的无服务器词元服务收入1430万元,付费客户为71.6万人;中大型企业客户为49家,单客户平均收入为30.5万元。

相较之下,其本地部署业务主要是将推理引擎及算力资源编排系统在客户自有的数据中心或私有算力环境内部署。具体而言,这类部署是在客户的数据中心内有效创建类似私有云的AI基础设施,使员工能透过集中、受控的接口使用多种AI模型,同时确保所有数据及推理活动完全保留在客户的自有基础设施内。

目前这部分业务在2025年占据硅基流动收入的47.1%。2025年收入2607万元,客户虽然从28家降至20家,但单客户平均收入从22.4万元大幅提升至130.3万元。

“清华系”企业领跑MaaS

在AI应用规模化落地及推理需求快速增长的背景下,行业内涌现了像硅基流动、无问芯穹、潞晨科技等这一构建于AI基础设施和大模型之上的独立词元供应商。不同于阿里云等往往依托自有云资源、模型、算力基础提供服务,独立词元供应商通常不绑定任何云提供商、模型或应用场景。

巧合的是,硅基流动、无问芯穹和潞晨科技同属“清华系”,无问芯穹的发起人汪玉是清华电子系本硕博、长聘教授/系主任;CEO夏立雪是汪玉门下博士生;潞晨科技创始人尤洋是清华大学计算机系硕士。另外,这三家公司融资进程都处于B轮、B+轮阶段。但是,从技术路线与业务边界来看,这三家企业呈现出显著分野。

硅基流动走的是典型的“推理优先”路线,通过聚焦推理侧的工程化能力,延续了袁进辉在OneFlow时期从头自研底层框架的基因。硅基流动的推理引擎独立于主流的vLLM和TensorRT-LLM,强调对单一芯片的极致性能压榨,并通过异构算力调度系统将多元算力转化为标准化的Token服务。

相较之下,潞晨科技则选择了一条差异更大的路径。其创始人尤洋认为,如果推理平台上没有自己的优质模型,将难以建立核心竞争力,因此潞晨科技在自研并开源了大规模深度学习分布式训练与推理加速框架Colossal-AI之外,入局了文生视频大模型(Open-Sora、Video Ocean),试图通过“自研模型+推理平台”的闭环来推动商业化。

无问芯穹则走出了第三条路,不局限于推理MaaS,而是定位为“算力运营商”,主打软硬协同的“异构混训”能力,能将英伟达与华为昇腾、壁仞、摩尔线程等多种国产芯片混合起来共同训练同一个大模型,在百卡和千卡级别已完成验证。在此基础上,无问芯穹还计划向硬件层延伸,布局端侧大模型推理处理器LPU,试图通过芯片IP变现。

注:文/漫地,文章来源:创业邦(公众号ID:ichuangyebang ),本文为作者独立观点,不代表亿邦动力立场。

文章来源:创业邦

广告
微信
朋友圈

FAQ回顾

硅基流动是做什么的?

硅基流动定位为AI时代的“Token工厂”,是国内头部独立生态词元供应商,核心是将底层算力与各类开源大模型打包为可直接调用的API,采用公有云服务+本地部署双轮驱动的业务模式,可帮助用户低成本、低门槛使用各类AI模型服务。

国内独立AI词元供应商赛道有哪些代表企业?

国内独立AI词元供应商赛道核心代表企业均为清华系,包括主打推理优先路线的硅基流动、走“自研模型+推理平台”闭环路线的潞晨科技、定位算力运营商主打异构混训能力的无问芯穹,三家企业目前融资均处于B轮及B+轮阶段。

硅基流动的营收情况怎么样?

硅基流动2023年营收仅6000元,2024年营收735万元,2025年营收大幅增长至5533万元,同比增幅达653%;2025年公有云服务收入占比52.9%,首次超过本地部署业务成为第一大收入来源。

硅基流动的服务适合哪些客户使用?

硅基流动公有云服务按Token消耗量计费,对使用门槛要求低,适合开发者与中小客户选择;本地部署解决方案可保障数据安全与私有化需求,适合对数据管控要求高的大型企业、机构客户采购使用。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0