近日，硅基流动官宣完成B+轮二期和C轮融资，2026年累计股权融资额接近29亿元。把日历往回翻82天，这家公司刚把招股书放上港交所的案头——成立34个月，7轮融资，投后估值77.4亿元，恰好越过港交所18C章对未商业化公司80亿港元的市值门槛。
653%的增长，海浪把"卖水人"推上浪头
袁进辉的履历干净得像一行代码。清华博士，师从张钹院士，微软亚研院研究员，2017年做深度学习框架，公司后被光年之外收购再并入美团。50天后他带着原班人马出走，创办硅基流动，让算力在软件之上跑起来。开张后的头4个月，收入6000元。
转折在2025年春节前，梁文锋提醒他提前备20台H800部署DeepSeek V3，月均算力成本约600万元，机器还未必跑得满，他没下注。DeepSeek爆火后，他公开认错，欲哭无泪。懊悔变成行动力，整个春节团队无休，2月1日首发满血版DeepSeek R1和V3，官网流量暴涨40倍，一度压过腾讯云与阿里云。
数字自此陡峭起来，营收从2024年的734.6万元涨到2025年的5533万元，日均词元吞吐从478亿爬到5785亿，单峰破万亿，注册用户超1000万，企业客户1.3万家，170多款模型在平台上跑。只是653%的增速不能全记在自己头上，2025年中国词元供应市场膨胀了1602.6%，企业级MaaS调用量翻了约16倍。
MaaS跑通了量，还没跑通利
招股书里最诚实的一行数字是毛利率轨迹：2023年83.3%，2024年39.4%，2025年-24.0%。2025年销售成本6863.2万元，超过了5533万元的全年营收，其中算力租赁一项就花掉5962.7万元，占销售成本的86.9%。换算下来，每卖出1元钱的词元，算力等成本要花1.24元。
亏损的大头还有获客：2025年销售及营销开支8370万元，其中5421.3万元拿去给用户发免费词元代金券，占64.7%。这批补贴拉来的无服务器词元客户，全年只贡献了1430万元收入，不到代金券的三成。再叠加2.09亿元研发投入，全年净亏损3.45亿元，经营活动现金流净流出1.72亿元，平均每月烧掉约1480万元。
账本也有另一面：本地部署业务的毛利率稳定在82.5%到92.4%之间，软件许可、实施和维护费用构成一条安静的现金流水线。公司的设计是个漏斗——让开发者先在公有云上低成本试用，用量做大、数据安全和合规要求变高之后，再升级为专属实例或本地部署。公有云负责造势引流，本地部署负责造血养家。
不站队是最贵的资产，也是最险的赌注
硅基流动从成立第一天起就不绑定任何一家云、芯片或模型厂商，却拿到了横跨互联网、芯片、算力、能源、运营商的全产业链资本。阿里、华为哈勃、美团、商汤、智谱、携程、晶科、联通系，这些彼此存在竞合关系的名字挤在同一份股东名单里。7轮融资下来，估值从天使轮的2.8亿元涨到77.4亿元，两年半涨了33倍。
招股书显示，2025年阿里和华为哈勃分别向硅基流动供应1970万元和1262万元算力，合计占采购额近20%，而阿里百炼和华为云MaaS恰是它在公有云市场上的直接对手。按IDC统计，2025年MaaS市场按调用量计，字节火山引擎份额超过40%，阿里云约30%，硅基流动位列第四，份额1.5%。
中立让它拿到了全产业链的入场券，却没有给它议价权。它真实的护城河是推理工程能力——同等算力压榨出更多词元。工程能力又是折旧极快的资产，竞对每发布一代新框架，护城河就得重新挖一遍。
一枚Token背后，是比模型更重的"工厂"
招股书给自己的定义是"开放独立的词元供应平台"，它做的事情颇像AI时代的发电厂和电网调度商：上游接GPU、NPU和不同算力集群，中间铺推理引擎、资源调度、模型适配、负载均衡等系统软件，下游通过API把模型能力交付给开发者和企业，最后按Token、实例或本地部署方案收钱。
弗若斯特沙利文进行的行业研究把市场分成"封闭生态"和"独立生态"两类，按该研究的2025年Token调用量口径，中国市场总量约2426.3万亿Token，硅基流动约36.7万亿Token，份额约1.5%，总市场位列第四；在所谓"独立生态词元供应平台"中排名第一。
接下来的财报要证明的，是这门生意配得上"工厂"两个字。工厂二字自带一条铁律：投入必须小于产出。港交所的锣声，只是把这条铁律从会议室，挂到了交易大厅。
本文来源：https://mp.ofweek.com/ee/a456714177637
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Recently, Silky announced the completion of a second Series B+ and a Series C financing round, with cumulative equity funding approaching CNY 2.9 billion in 2026. Turning back the calendar 82 days, this company had just placed its prospectus on the desk of the Hong Kong Stock Exchange — after 34 months of operation and 7 rounds of financing, its post-money valuation reached CNY 7.74 billion, just crossing the HKEX’s Chapter 18C market cap threshold of HK$ 8 billion for uncommercialized companies.
A 653% growth thrust the “water seller” to the forefront.
Yuan Jinhui’s resume is clean as code: PhD from Tsinghua, mentored by Academician Zhang Chuo; researcher at Microsoft Research Asia; founder of a deep learning framework company in 2017, later acquired by Lightyear and merged into Meituan. 50 days after that merger, he left with his original team to found Silky, bringing computing power to life above software. The first 4 months saw revenue of just CNY 6,000.
The turning point came just before the 2025 Lunar New Year, when Liang Wenfeng suggested pre-purchasing 20 H800 for DeepSeek V3 deployment — with monthly compute costs around CNY 6 million, even if machines aren’t fully utilized — but he didn’t take the bet. After DeepSeek exploded, he publicly apologized, tears holding back. Regret became drive: the team worked non-stop through the holiday, launching the first full-spec DeepSeek R1 and V3 on February 1, spiking website traffic 40x, overtaking Tencent Cloud and Alibaba Cloud temporarily.
Numbers soared: revenue jumped from CNY 7.346 million in 2024 to CNY 55.33 million in 2025; daily token throughput climbed from 4.78 billion to 57.85 billion, peaking over 1 trillion; registered users surpassed 10 million, with 13,000 enterprise clients and 170+ models running. But the 653% growth isn’t all self-made — the Chinese token supply market itself grew 1602.6% in 2025, with enterprise-level MaaS calls increasing roughly 16x.
MaaS succeeded in volume, but not yet in profit.
The most honest line in the prospectus is the margin trajectory: 83.3% in 2023, 39.4% in 2024, and -24.0% in 2025. In 2025, sales costs reached CNY 68.632 million, exceeding full-year revenue of CNY 55.33 million, with compute rental alone accounting for CNY 59.627 million — 86.9% of total costs. This translates to every CNY 1 of token sales incurring CNY 1.24 in compute and other costs.
A major loss driver is customer acquisition: sales and marketing expenses hit CNY 83.7 million in 2025, of which CNY 54.213 million was spent on free token vouchers for users — 64.7% of the budget. The serverless token customers acquired via these subsidies contributed only CNY 14.3 million in revenue, less than 30% of the voucher value. Coupled with CNY 209 million R&D investment, the net loss for the year was CNY 345 million, with negative cash flow from operations of CNY 172 million — averaging about CNY 14.8 million burned per month.
There’s an upside: local deployment services maintain margins between 82.5% and 92.4%, providing a quiet pipeline of cash via software licenses, implementation, and maintenance fees. Silky’s business model resembles a funnel: developers try the service cheaply on public cloud, then upgrade to dedicated instances or local deployments as usage grows and compliance/security requirements increase. Public cloud serves as a promotional channel, while local deployment generates the real profit.
Neutrality is its most valuable asset — and riskiest gamble.
Since day one, Silky has avoided binding itself to any single cloud, chip, or model vendor, yet secured investment from the full tech industrial chain: Internet, semiconductors, computing, energy, and telecom. Names like Alibaba, Huawei Hubble, Meituan, Sensing, Zhipu, Ctrip, GCL, and China Unicom jostle in the same shareholder list — despite their competitive relationships. Over 7 rounds, valuation soared from CNY 280 million to CNY 7.74 billion — a 33x increase in 2.5 years.
In 2025, Alibaba and Huawei Hubble supplied CNY 19.7 million and CNY 12.62 million in compute power respectively, comprising nearly 20% of Silky’s procurement — while simultaneously competing against it in the public cloud MaaS space. According to IDC, ByteDance’s Volcano Engine leads the Chinese MaaS market with >40% share, followed by Alibaba (30%), with Silky ranking fourth at 1.5%.
Neutrality opened doors across the industry, but didn’t grant pricing leverage. Its true moat is inference engineering — extracting more tokens per unit of compute — but this is quickly depreciating expertise. Every time a competitor releases a new framework, the moat must be re-dug.
Behind every token lies a “factory” heavier than the model.
Silky defines itself as an “open, independent token delivery platform,” acting like an AI-era power plant and grid dispatcher: integrating GPUs, NPUs, and compute clusters upstream; layering in inference engines, resource scheduling, model adaptation, and load balancing; and delivering model capabilities via API to developers and enterprises downstream, charging by token, instance, or local deployment.
Frost & Sullivan segments the market into “closed” and “independent” ecosystems. In 2025, the Chinese token market totaled ~24.3 trillion tokens, with Silky accounting for ~36.7 billion (1.5% share), ranking fourth overall but first among independent providers.
Moving forward, financial reports will need to prove that this business deserves the label “factory.” Factories must produce more than they consume — a rule that the Hong Kong Stock Exchange has elevated from boardroom to trading floor.