• 亿邦公众号
    • APP下载

复旦女学霸 融了100亿

作者:张楠 文章来源:投中网

2026-07-23 11:10

作者:张楠 文章来源:投中网

2026-07-23 11:10

邦小白快读

EN
全文速览

本文核心介绍了复旦女学霸Lin Qiao创办的AI推理公司Fireworks AI的创业故事与Token工厂赛道的核心情况,干货如下:

1. 项目基本信息:Fireworks AI成立仅四年,最新完成102亿元人民币D轮融资,投后估值达到175亿美元,不到两年估值翻了30多倍,目前ARR突破10亿美元,平台每日处理token超40万亿,95%的token来自客户定制模型。

2. 核心业务与优势:公司不做大模型训练也不做C端AI应用,只做AI推理,帮企业微调、托管开源模型,按调用量收费;自研推理优化技术,速度是同价位对手的5倍,价格比闭源模型便宜5-10倍,已经拿到多个企业级合规认证,接入微软Azure服务,客户包括Uber、Notion、Quora等知名企业。

3. 行业背景:当前AI行业资本已经从比拼大模型能力转向布局能稳定低价运行模型的推理层,Token工厂是全球追捧的新风口。

Token工厂赛道的兴起,给品牌商布局AI、落地业务带来了很多新机会,核心干货如下:

1. 消费与行业趋势:随着agent爆火,AI单次调用的token量大幅上涨,闭源模型的推理成本让很多企业难以承受,同时开源模型的能力已经逼近闭源模型,越来越多品牌倾向于打造贴合自身业务的定制AI,降本定制的需求非常旺盛。

2. 产品研发新路径:品牌商不需要投入高额成本自研大模型,可以依托Token工厂的推理平台,用自身的业务数据、客户数据对开源模型做微调,就能得到专属定制AI,成本仅为使用闭源模型的十分之一到五分之一,还能保障自有数据的安全可控。

3. 用户运营优化:很多AI相关消费产品比如AI眼镜、AI玩具此前因为token成本过高,不敢放开让用户使用AI功能,token工厂可以帮助品牌压低单位token成本,让品牌可以放开AI功能提升用户体验,同时控制成本支出。

Token工厂是当前AI领域的高增长新赛道,相关机会、风险与商业模式干货整理如下:

1. 市场机会:资本已经从大模型训练端逐步转向推理部署端,需求端多重利好叠加:agent爆火推高整体token需求量,开源模型成熟让企业有了替代闭源模型的可行选项,企业持续推理成本压力大,Token工厂的降本定制服务缺口非常大。

2. 最新商业模式:核心是把底层算力、开源大模型、推理引擎打包封装成标准化API,开发者不需要自己采购显卡搭建环境,写几行代码就可以按量调用,本质是做AI模型落地生产的基础设施,对标AI领域的台积电,可以帮企业把部署AI的时间从五年压缩到五天甚至五周。

3. 风险提示:该赛道未来可能会被云厂商和大模型公司两头夹击,如果陷入价格战,很容易变成高收入低利润的转售生意,最终大概率会像云计算一样走向头部收敛,中小玩家只有打造出客户离不开的效率壁垒才能存活。

Token工厂赛道的发展,给传统工厂推进AI数字化转型带来了新启示和新机会,核心干货如下:

1. AI落地生产设计的新路径:传统工厂有大量独有的生产、设计、供应链、客户运营数据,需要贴合自身业务的专用AI,不需要投入高额成本自研大模型,可以依托Token工厂,用自有数据对开源模型做微调,就能快速得到定制AI,成本远低于自研和使用闭源模型,门槛很低。

2. 新商业机会:当前头部Token工厂正处于快速扩张阶段,需要大量采购GPU,还要扩大工程团队规模,深化和上下游供应商的合作,芯片制造、硬件生产相关工厂可以对接Token工厂的采购需求,拓展新的ToB业务增长曲线。

3. 数字化转型启示:工厂做AI数字化不需要盲目跟风跟风自研大模型,应该聚焦自身业务数据优势,依托成熟的第三方AI推理基础设施,低成本快速打造适配自身生产设计、运营流程的专用AI,控制转型成本,提升落地效率。

当前AI推理服务行业的发展趋势、客户痛点与可参考解决方案整理如下:

1. 行业发展趋势:AI技术栈中,推理层正在被资本重新估值,Token工厂也就是第三方AI推理服务已经成为全球热门赛道,中美两地都有大量玩家入场,头部玩家估值短时间内数十倍增长,资本关注度极高,市场需求随产业AI化快速扩张。

2. 核心客户痛点:企业落地AI的核心痛点是缺乏把模型部署进生产环境的基础设施、技术和人才,持续推理成本居高不下,通用大模型不贴合企业自有业务需求,闭源模型的高价让企业难以承受,这些痛点都给第三方推理服务商留出了市场空间。

3. 可参考的解决方案:对标头部玩家Fireworks AI的路径,打造自研的推理优化核心技术,提升算力吞吐效率压低单位token成本,支持多种微调方式适配企业自有数据,拿齐企业级合规资质,对接主流云平台,逐步把客户结构从单一领域拓展到多行业,提升抗风险能力。

Token工厂赛道的发展,给AI平台带来了新需求、新机会和需要规避的风险,核心干货如下:

1. 市场需求与新机会:当前大量企业和AI创业公司都需要低成本高效率的推理部署服务,第三方Token工厂已经成为重要的市场玩家,平台可以引入优质Token工厂玩家入驻,丰富自身AI服务生态,也可以和头部玩家深化合作,给客户提供更完善的一体化AI服务,提升客户粘性。

2. 国内平台的差异化机会:国内芯片生态碎片化,第三方Token工厂在多芯片适配、开源模型部署上有天然优势,平台可以通过投资、合作的方式对接这类玩家,弥补自身在细分领域的不足,更好满足企业客户的定制化需求,拓展企业服务市场。

3. 风险规避要点:推理赛道未来竞争会持续加剧,很可能出现价格战,独立Token工厂还会面临大模型厂商和云平台的双向挤压,平台布局该领域时,需要重点考察玩家的核心技术壁垒和毛利率管控能力,避开低壁垒的玩家,聚焦有核心工程效率优势的标的。

本文披露了AI推理领域Token工厂赛道的最新产业动向,相关研究素材整理如下:

1. 产业新动向:当前AI产业资本正在从大模型训练端向推理部署端转移,Token工厂成为AI领域新的风口,全球中美两地同步快速发展,美国玩家依托资本优势估值增长迅猛,客户以全球化企业软件公司为主;中国玩家针对芯片生态碎片化特点,走出了多芯片适配的差异化路线,多家玩家获得资本大额加持,赛道需求随agent爆火快速增长。

2. 创新商业模式:Token工厂的核心定位是AI推理基础设施,不做大模型训练也不做C端应用,核心是帮企业微调定制开源模型、托管推理服务,按调用量收费,定位是AI领域的台积电,是AI落地的卖水人,目前头部玩家ARR已经突破10亿美元,估值增长速度远超大模型玩家。

3. 值得研究的新问题:该赛道未来是否会像早期云计算一样走向头部收敛,独立推理层能否建立起客户离不开的垄断性效率壁垒,还是最终会被云厂商和大模型厂商整合,如何维持稳定的毛利率,这些都是产业层面值得深入研究的新问题。

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 shares the founding story of Fireworks AI, an AI inference startup founded by Fudan-educated entrepreneur Lin Qiao, and breaks down the "Token Factory" track. Key takeaways:

1. Company fundamentals: Founded just four years ago, Fireworks AI recently closed a Series D round of RMB 10.2 billion, reaching a post-money valuation of $17.5 billion — a more than 30-fold valuation jump in under two years. It currently boasts an annual recurring revenue (ARR) of over $1 billion, processes more than 40 trillion tokens daily on its platform, with 95% of tokens coming from customer custom models.

2. Core business and competitive advantages: Fireworks AI focuses exclusively on AI inference, rather than large model training or consumer-facing AI applications. It helps enterprises fine-tune and host open-source models, charging based on inference volume. Its self-developed inference optimization technology delivers 5x the speed of competitors at the same price point, and costs 5 to 10 times less than closed-source models. The company holds multiple enterprise-level compliance certifications, integrates with Microsoft Azure, and counts well-known clients including Uber, Notion, and Quora.

3. Industry context: AI industry capital has shifted its focus from competing for large model capabilities to investing in the inference layer that enables stable, low-cost model operation, making the Token Factory a red-hot new global growth area.

The rise of the Token Factory track has opened up major new opportunities for brands to adopt AI and integrate it into core business. Key insights for brands:

1. Consumer and industry trends: With the booming popularity of AI agents, the number of tokens consumed per AI call has risen sharply. The inference cost of closed-source models is prohibitive for many enterprises, while open-source models now approach closed-source models in capability. More brands are prioritizing custom-built AI tailored to their unique business needs, creating strong demand for cost-effective customization solutions.

2. A new path for product development: Brands do not need to invest heavily in building large models in-house. Leveraging the inference platform of a Token Factory, brands can fine-tune open-source models with their own business and customer data to get a proprietary custom AI. This approach costs just one-tenth to one-fifth of what using a closed-source model would cost, while keeping brands' proprietary data secure and fully in their control.

3. Optimized user engagement: Many AI-powered consumer products, such as AI glasses and AI toys, previously restricted user access to AI functionality due to high token costs. Token factories help brands lower per-token costs, allowing them to offer full access to AI features to improve user experience while keeping overall spending under control.

The Token Factory is a high-growth new track in the current AI space. Below is a breakdown of opportunities, risks, and business models for sellers:

1. Market opportunities: Capital has gradually shifted from large model training to inference deployment, with multiple tailwinds on the demand side: the AI agent boom has lifted overall token demand, mature open-source models provide a viable alternative to closed-source models, and enterprises face persistent pressure to cut inference costs. This has created a huge unmet demand for the low-cost customization services Token factories offer.

2. The emerging business model: The core model is to package underlying computing power, open-source large models, and inference engines into a standardized API. Developers do not need to purchase GPUs or build their own infrastructure; they can integrate the service with just a few lines of code and pay for what they use. Essentially, Token factories are infrastructure for deploying AI into production, often described as "the TSMC of AI." They can cut the time to deploy AI from five years to as little as five days, or five weeks at most.

3. Risk warnings: The track may eventually face double-sided pressure from cloud providers and large model companies. If a price war breaks out, it could easily become a high-revenue, low-margin resale business. Like the cloud computing industry before it, the Token Factory track will likely eventually consolidate around a small number of leading players. Small and mid-sized players can only survive by building irreplaceable efficiency barriers that keep customers locked in.

The growth of the Token Factory track brings new insights and opportunities for traditional manufacturers advancing AI-driven digital transformation. Key takeaways:

1. A new path to deploy AI in production and design: Traditional manufacturers hold large volumes of unique data across production, design, supply chains, and customer operations, and require custom AI tailored to their specific use cases. Instead of investing heavily to develop large models in-house, manufacturers can leverage Token Factory infrastructure to fine-tune open-source models with their own data to quickly build a custom AI, at a far lower cost than both in-house development and closed-source model services, with very low entry barriers.

2. New business opportunities: Leading Token factories are currently in a phase of rapid expansion, requiring large volumes of GPU purchases, expanding engineering teams, and deepening partnerships with upstream and downstream suppliers. This creates new B2B growth opportunities for chip manufacturing and hardware production factories that can meet these procurement needs.

3. Lessons for digital transformation: When pursuing AI-powered digital transformation, factories should not blindly follow the trend of developing in-house large models. Instead, they should leverage their advantage in proprietary business data, build on mature third-party AI inference infrastructure, and quickly develop custom AI for their production, design, and operational processes at low cost, to control transformation costs and improve deployment efficiency.

Below is a summary of current development trends, core customer pain points, and reference solutions for the AI inference service industry:

1. Industry development trends: Within the AI technology stack, the inference layer is now being revalued by capital. Third-party AI inference services, or "Token Factories," have become a globally hot track with heavy participation from players in both China and the U.S. Leading players have seen their valuations multiply dozens of times in a short period, attracting intense capital attention, and market demand is expanding rapidly alongside the broader adoption of enterprise AI.

2. Core customer pain points: The top pain points for enterprises looking to deploy AI are a lack of infrastructure, technical capability, and talent to deploy models into production environments, persistently high ongoing inference costs, the mismatch between general-purpose large models and enterprise-specific business needs, and the prohibitive cost of closed-source models. All these pain points create significant market space for third-party inference service providers.

3. A reference solution for players: Following the path of leading player Fireworks AI, providers can build proprietary core inference optimization technology to improve computing throughput and lower per-token cost, support multiple fine-tuning methods to adapt to enterprises' proprietary data, obtain all required enterprise-level compliance certifications, integrate with mainstream cloud platforms, and gradually expand their customer base from a single sector to multiple industries to improve resilience.

The growth of the Token Factory track has brought new demand, new opportunities, and key risks to avoid for AI marketplaces. Key insights:

1. New market demand and opportunities: A large number of enterprises and AI startups now need low-cost, high-efficiency inference deployment services, and third-party Token factories have become important market players. Platforms can onboard high-quality Token Factory providers to enrich their AI service ecosystem, or deepen partnerships with leading players to offer customers more comprehensive end-to-end AI services and improve customer retention.

2. Differentiated opportunities for Chinese platforms: China's chip ecosystem is fragmented, and third-party Token factories have natural advantages in multi-chip adaptation and open-source model deployment. Platforms can partner with or invest in these players to fill gaps in their own capabilities for niche segments, better meet the customization needs of enterprise customers, and expand their enterprise service market share.

3. Key risk mitigation: Competition in the inference track will continue to intensify, and price wars are likely. Independent Token factories also face dual pressure from large model developers and cloud platforms. When positioning players in this space, platforms should prioritize evaluating candidates' core technical barriers and gross margin management capabilities, avoid low-barrier players, and focus on targets with core engineering efficiency advantages.

This article outlines the latest industry developments in the Token Factory track in the AI inference space, with key research takeaways below:

1. New industry developments: AI industry capital is shifting from large model training to inference deployment, making Token Factories a new growth area in AI that is expanding rapidly on both sides of the U.S. and China. U.S. players have seen extremely rapid valuation growth driven by capital access, and serve primarily global enterprise software companies; Chinese players have pursued a differentiated strategy focused on multi-chip adaptation to address the country's fragmented chip ecosystem, with multiple players securing large capital injections, and demand growing rapidly alongside the AI agent boom.

2. The innovative business model: Token Factories are positioned as core AI inference infrastructure, choosing not to develop large models in-house or build consumer-facing applications. Their core business is helping enterprises fine-tune custom open-source models and host inference services, charging based on usage volume. Positioned as the "TSMC of AI" and the "water suppliers" of the AI boom, leading players already have ARR exceeding $1 billion, with valuation growth far outpacing that of large model developers.

3. New open research questions: Key unresolved industry-level questions that merit in-depth research include whether the track will eventually consolidate around leading players like early cloud computing, whether independent inference layers can build monopolistic efficiency barriers that lock in customers, or whether they will ultimately be integrated by cloud providers and large model companies, and how players can maintain stable gross margins.

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.

全球都在建的Token工厂。

作者 张楠

本文字数4129

有位计算机系的“复旦学姐”,在美国办了一家公司,被黄仁勋叫作“AI工厂的台积电”,红杉、英伟达、Index、TCV、光速这些顶级VC轮番押注,成立才四年,估值已经达到了175亿美元(约合1185亿元人民币)。

这家公司是Fireworks AI,按现在流行的话说,Fireworks AI是一家Token工厂。既不训练前沿大模型,也不做面向消费者的AI应用,只做推理,帮企业把开源模型微调好、托管好,然后按调用量收钱。今年GTC大会上,黄仁勋和Lin Qiao有场对谈,老黄直言,“In a lot of ways,you're the TSMC of AI factories.”

这家公司刚完成了15.05亿美元(约合102亿元人民币)的D轮,投后估值175亿美元,由Atreides Management、Index Ventures和TCV三家领投,英伟达跟投,老股东Lightspeed、Bessemer、Menlo以及红杉、基准资本等一众机构继续加码。

2022年,复旦学姐Lin Qiao从Meta离开,在ChatGPT发布前一个月,于加州红木城拉起一支七人队伍创办了Fireworks AI。

2023年,红杉、英伟达、AMD投了A轮;2024年B轮估值约5.52亿美元;2025年10月C轮2.5亿美元、估值冲到40亿;再到眼下这轮D轮175亿美元估值。

不到两年时间,估值从5.52亿翻到175亿美元,涨了约30多倍。ARR突破10亿美元,是上一年同期的5倍;平台每天处理的token从15万亿涨到40万亿以上;更关键的一点是,平台上超过95%的token来自客户在自己数据上定制过的模型,而不是现成的通用模型。

随着Kimi K3开源,以及DeepSeek V4满血版即将推出,基础模型的估值和壁垒受到冲击,推理层是AI技术栈里正在被资本悄悄重估的那一层,钱正从“谁模型最强”挪向“谁能把模型稳定便宜地跑起来”。

创始人Lin Qiao也是个容易被忽略的角色,从底层框架PyTorch一路干到175亿美元,存在感却远低于许多前沿模型创始人。更妙的是这门生意全球都在做,美国有Together、Baseten,中国也有硅基流动、无问芯穹、九章云极、趋境科技等,把中美两拨人摆在一起,正好能看清“Token工厂”到底是不是一门好生意。

AI台积电

Fireworks的产品说穿了是一套推理云平台。开发者通过一套兼容OpenAI的API,就能调用平台上200多个开源模型,覆盖文本、图像、多模态。

企业客户可以在自己的专有数据上做微调,支持SFT(监督微调)、DPO(直接偏好优化)和RFT(强化微调),还能同时挂上最多100个LoRA适配器而不额外加价。合规层面它拿到了SOC2、HIPAA、GDPR这些企业入场券,2026年3月又接进了微软的Foundry,意味着Azure上的客户能直接走它的推理。

速度是Fireworks对外讲的核心故事。它自研了一套叫Fire Attention的CUDA内核栈,在DeepSeekV4 Pro上能做到每秒167个token,对外宣称是同价位对手的5倍;内部用了解耦服务、语义缓存和推测解码,把长尾延迟压住。

这套工程能力是它和“随便租几张卡跑开源模型”的本质区别,它卖的不止是算力,是把算力榨出更高吞吐的软件效率。

目前,Fireworks无服务器推理起步价大约每百万token 0.2美元(8B级别模型)到0.9美元(70B级别),另外也按小时出租GPU。一个信号性的变化是:今天平台上95%的token来自定制模型。

换句话说,想要把AI用起来,基础大模型远远满足不了企业需求,必须拥有属于自己的、贴着自己业务数据的模型。Fireworks在融资声明里讲得很直白,企业手里都有别人没有的东西,客户关系、工作流、数据、对质量的定义,Fireworks干的事就是把这些变成能拥有、能持续改进的专用智能。

Fireworks的客户包括Cursor、Notion、Quora、Sourcegraph、Uber、Shopify、Perplexity、等知名公司,还有一家叫Sentient的公司,靠Fireworks24小时扛住了180万人的候补排队。一个转折性细节是,Fireworks早先大约一半收入来自AI编程工具Cursor,现在客户结构已经散开到更广泛的企业,这与它从“尝鲜”走向“生产级渗透”的判断能对上。

Fireworks的投资人之一是Atreides,掌门人Gavin Baker早年重仓过Meta和英伟达,Gavin Baker认为,前沿模型和开源模型会越来越合用,而Fireworks是少数被企业信任、用来在生产里训练并托管开源模型的平台。

Index是连投数轮的老股东,TCV投过Netflix和Spotify这类基础设施型生意。再加上英伟达这个“既是金主又是供货商”的角色,整张股东表几乎全是基础设施和企业软件的长线钱,一边投基础模型,一边拿Fireworks在企业级调用做对冲。

这轮融资,Fireworks主要还是为了扩大算力,目前Fireworks从20多家供应商采购GPU,微软是其中之一。另外,Fireworks还打算把工程团队从约200人扩到年底600人、深化和微软与英伟达的合作。换句话说,它要把跑模型这件事的规模和可靠性,以及服务客户的规模上再往上推一个量级。

掌舵人Lin Qiao

Lin Qiao在中国长大,父亲是一家造船厂的资深机械工程师,从零开始造货船,她从小就会对着船舶蓝图读那些精确的角度和尺寸,迷上了把复杂东西拆成可计算零件的感觉;初中起她就偏在数理化上,高中时第一次用BASIC写了个贪吃蛇小游戏,那时就认定计算机科学是自己的未来。

1995年,Lin Qiao考上复旦大学计算机科学专业,本硕连读,又从这里第一次走进AI的世界。再之后她才赴美,在加州大学圣塔芭芭拉分校拿到计算机科学博士,方向是分布式系统和数据库管理。

博士毕业后她先去了IBM,在硅谷实验室和阿尔马登研究中心做研究员,主导过一套极速数据分析优化器的研发。接着是LinkedIn四年,做分布式数据服务,参与过Gobblin数据摄取框架。

Lin Qiao的职业生涯转折点出现在2015年7月,她加入Meta,一待七年,从数据基础设施的高级经理一路做到工程高级总监,手下团队从5个人长到300多人。

在Meta她干成了一件影响整个行业的事:共同创立并主导了PyTorch。

最早这被她当成“六个月就能搞定的项目”,结果变成一场历时五年的底层重建,把Meta整个AI工作负载从数据加载、分布式推理到跨全球数据中心和数十亿设备的训练全部重写一遍。等到她离开时,这套系统每天要撑起超过5万亿次推理。她后来有个比方:在Meta她建了引擎,现在她想建路。

Lin Qiao离开Meta的时点很妙。2022年10月,就在ChatGPT发布前一个月,她从Meta出走创立Fireworks。

当时她的判断是,企业都想上AI,但缺的是把模型真正部署进生产环境的基础设施、资源和人才,而她在Meta花五年解决的正是这个。Lin Qiao的想法是,把Meta要五年的事,压缩到五周甚至五天。七人创始团队里六人来自Meta的PyTorch组,一人来自谷歌的Vertex AI;其中三位是华人,除了Lin Qiao,还有Benny Chen,UCLA本科,Meta广告基础设施负责人和ChenyuZhao伯克利本科,谷歌Vertex AI技术经理。

一开始,Lin Qiao对外讲的理念叫“复合AI”(compound AI),主张用成百个小专家模型分别解决窄问题,而不是死磕单个大模型。落到Fireworks上,就是帮企业把开源模型调成自己的专用智能。

“AI有两条路。一条是智能属于少数几家大实验室,其他人全去租。另一条是全世界每家公司都建自己的专用智能,由只有它懂的那个领域来塑形。我们在建第二条路。”

据CNBC,同等质量下Fireworks比闭源模型便宜5到10倍。这恰恰是它增长的发动机:当最新模型的账单让老板和CFO们坐立难安,当开源模型能力与闭源模型差距缩小到某种程度,企业就开始认真考虑用开源替代了。

全球都在建的Token工厂

眼下,不少公司都在说自己是“Token工厂”,指的就是把底层算力、各种开源大模型、推理引擎打包封装成标准化API,开发者不用自己买显卡、搭环境,写几行代码、按量付费就能用上AI。谁掌握了服务这一层,谁就收这道过路费。

这门生意眼下同时在中美两边升温。

美国这边,Fireworks不是孤例,Together AI在7月1日刚关了8亿美元C轮,估值83亿美元,由沙特阿美旗下的Aramco Ventures领投,英伟达和Salesforce跟投,年预订额据称已破11.5亿美元,靠一套叫ATLAS的自适应推测解码引擎加速。

Baseten更猛,6月23日宣布15亿美元F轮、估值130亿美元,18个月融了四轮,收入同比涨约20倍,客户名单里甚至有Open Evidence、Harvey、Cursor、Notion,英伟达同样是它的股东。

再往外数还有Anyscale,贾扬清创立的Lepton、Modal、Replicate、Groq、Cerebras一众玩家。而且slogan大差不差:own your intelligence。

中国这边,故事换了个讲法,但内核不一样,流行的是异构架构。说白了,就是怎么把英伟达和其他国产芯片一起用起来,让模型推理更有效率。

比如硅基流动,创始人袁进辉是清华博士,早年创办过深度学习框架OneFlow,2023年8月在经历被并购又分拆的波折后重新出发。

这家公司今年已向港交所递表,想冲“AI Token工厂第一股”,成立不到三年完成七轮融资,投后估值77.4亿元人民币(约11亿美元),其中一笔超20亿元的B轮创下国内第三方MaaS赛道最大单笔纪录;阿里巴巴是最大机构股东,持股7.42%,华为旗下的哈勃科技持股4.07%。

硅基流动的跨芯片、多模型适配能力据称全球第一,平台累计支持170多个主流模型,2025年DeepSeek爆火、官网被打瘫时,它是业内第一家走通国产昇腾芯片部署满血版DeepSeek的公司。

再比如无问芯穹,由清华大学电子工程系教授汪玉发起、夏立雪任CEO,不久前拿到超7亿元融资,国资和产业资本入场,其Agentic MaaS平台上线160多种模型,日均token调用量较2025年底增长超20倍,公司内部提出一个“AI生产力=智能规模×Token生产效率×Token价值转化”的公式,把自己类比成石化产业链里的炼化厂,把能源转成数字石油。

老牌的厂商比如九章云极,由清华大学电子工程系博士方磊创办、此前完成多轮战略融资,北京两大国有AI产业基金联合领投,国资与头部产业资本持续加码。前阵子刚开完“Token工厂”的发布会,其新一代Alaya NeW智算云AI工厂体系汇聚1000+大模型与行业专用模型,平台日均Token流转承载力目标直指10万亿,相较2025年底平台处理量增长超数十倍,首创DCU“一度算力” 计量单位,让算力像电力一样流通、计价、取用。

另外,中美两拨人最大的差别还在站位。美国那边的资本更猛、估值更高,客户是全球化的企业软件公司,很多公司英伟达都是直接下场当股东。国内这边,因为芯片生态碎片化,反而给这些中间层的Token工厂留了空间,但另一个问题是火山、阿里云、百度云这些大云厂就在前面,DeepSeek这类模型方也可能自己把推理做了。

未来如何就看故事要怎么讲了,但至少目前来看,需求很大。

自从年初agent的爆火以来,单次请求的token量暴增;开源模型的质量逼近闭源,企业有了真能用的替代;以及,企业终于开始认真算账,训练是一次性的,推理是持续流血的消耗,产品越受欢迎、调用越频繁、亏得可能越狠。

比如有投资人就说过,一些AI眼镜、AI玩具厂商巴不得用户少用点AI能力,拿回去当个摆设就好,因为“Token真烧不起”。

于是“能不能把每次调用的成本压下来”变成了生死线,而夹在模型、GPU、云厂和应用之间的Token工厂们,恰好是解题的人。

但要做Token,毛利率太重要了,算力采购价、利用率、客户压价,每一项都啃利润。如果未来竞争加剧、大家用低价抢客户,它就可能变成“高收入、低利润”的转售生意。如果未来训练的强度降低,也容易被云厂商和大模型公司两头夹击。

总之,不管Fireworks也好,硅基流动也罢,归根到底Token经济学的时代,Token相当于卖水人的角色,这和云计算早期那批卖服务器的公司有点像。可云计算最后收敛成了AWS、微软、谷歌三家,推理层会不会也走这条收敛路?独立推理层到底是台积电还是租车行,得看它能不能把“跑模型”变成客户离不开的、带垄断性的效率。

投中网原创

注:文/张楠,文章来源:投中网(公众号ID:China-Venture),本文为作者独立观点,不代表亿邦动力立场。

【版权提示】亿邦动力倡导尊重与保护知识产权。未经许可,任何人不得复制、转载、或以其他方式使用本网站的内容。如发现本站文章存在版权问题,烦请提供版权疑问、身份证明、版权证明、联系方式等发邮件至run@ebrun.com,我们将及时沟通与处理。

已点赞

+1

{{ praise_text }}

好文章,值得被更多人看到

FAQ回顾

Token工厂是什么?主要提供什么服务?

Token工厂是AI推理赛道的服务商,将底层算力、开源大模型、推理引擎打包封装成标准化API,可帮助企业完成开源模型微调、托管等工作,开发者无需自行搭建环境,仅需按调用量付费即可使用AI能力,能大幅降低企业AI部署成本。

AI Token工厂模式有什么核心优势?

Token工厂模式可最大化算力利用率,同等质量下比闭源模型便宜5到10倍,支持企业基于自有数据定制专用模型,满足生产级AI部署需求,可有效降低企业持续产生的高频推理调用成本。

全球AI Token工厂赛道有哪些代表性企业?

美国市场代表企业有Fireworks AI、Together AI、Baseten等;中国市场代表企业有硅基流动、无问芯穹、九章云极等,其中硅基流动已向港交所递交上市申请,拟冲击AI Token工厂第一股。

Fireworks AI为什么被称为AI工厂的台积电?

Fireworks AI主打AI推理服务,自研Fire Attention CUDA内核栈推理速度达同价位对手的5倍,可帮企业完成开源模型的微调、托管,拥有SOC2、HIPAA、GDPR等合规资质,已服务Cursor、Notion、Uber等大量知名企业,工程能力远超普通算力租赁商。

广告

下一篇

跨境电商资源对接

《全球跨境电商服务商资源手册》第二期手册现已发布,触达更多资源,了解跨境服务商信息,欢迎点击链接下载。

了解详情

整治“恐怖谷效应” 抖音抬高AI仿真人剧准入门槛

日前,抖音集团短剧版权中心发布公告,宣布自8月3日起,针对都市情感、破局成长、家庭温情、都市逆袭四类题材的AI仿真人剧,正式实施新的内容制作规范。