510亿元。
这是DeepSeek首轮外部融资的金额——约74亿美元，投后估值近4000亿元人民币（约500亿美元）。创下中国AI行业单轮融资的历史纪录。
更耐人寻味的是几个细节：创始人梁文锋个人承诺出资200亿元；腾讯和宁德时代是最大的外部投资方；国家人工智能产业投资基金是唯一直接入股并享有投票权的例外；所有外部投资者必须将资金投入梁文锋管理的有限合伙企业，且设置了五年锁定期——这在追求快进快出的风险投资行业几乎是不可想象的条件。
一家坚持了近三年不融资、用内生利润支撑研发的中国AI公司，在美国对H20芯片"无限期"禁售、Anthropic新模型发布72小时即遭出口管制、全球算力供给缺口约3倍的大背景下，完成了这样一轮融资。这不是一次普通的资本事件——这是中国AI独立路线的一次资本公投。
为什么是现在：三年不融资，为什么突然开了口
要理解DeepSeek这轮融资的分量，需要先回到梁文锋的创业哲学。
DeepSeek从成立之初就选择了一条与绝大多数中国AI公司不同的路：不烧钱抢用户、不做ToC超级App、不开新闻发布会融资。它的核心团队来自量化私募幻方，本身就有极强的造血能力——用自营交易的利润养活AI研发，不看VC脸色，不追估值泡沫。
这条路走了将近三年。DeepSeek V2以极致的推理成本震惊行业，V3在多项基准测试中追平GPT-4o，V4完成了六大国产芯片适配——从华为昇腾到寒武纪到平头哥，DeepSeek是最早认真做国产算力适配的大模型公司之一。它的推理成本做到了竞品的几十分之一，这不是营销话术，而是真实的工程能力。
那为什么现在要融资？梁文锋在内部沟通中给出了答案：不是缺钱，是缺战略资源。
大模型竞争到了2026年中，已经从"谁的模型跑分高"变成了"谁能在算力受限的环境下持续迭代"。H20禁售之后，中国大模型公司面临的不再是"买不买得到卡"的问题，而是"能不能在国产芯片集群上完成万亿参数模型的全参数训练"。6月，深圳河套学院/深智城算网完成了DeepSeek-V4-Pro基于国产算力集群的全参数后训练——这是业界首个第三方机构做到这一点，但距离大规模、高效率、低成本的训练还有很长的路。
DeepSeek需要的不只是钱——它需要云计算基础设施的深度协同（腾讯）、硬件供应链的战略合作（宁德时代的能源与制造能力延伸）、以及政策层面的长期支持（国家AI产业基金）。这轮融资本质上不是财务投资，而是战略结盟。
"梁文锋用五年锁定期告诉市场：这不是来赚快钱的，这是来打持久战的。能接受五年锁定期的钱，才是真正的耐心资本。"
—— 一位参与过DeepSeek融资沟通的投资人私下评价
估值500亿美元：贵了还是便宜了
500亿美元估值是什么概念？
做几个对比：OpenAI最新估值约5000亿美元（一级市场私募交易），Anthropic约3500亿美元（据报道5月ARR达470亿美元，B端已反超OpenAI），xAI约1800亿美元。DeepSeek的500亿美元，大约是OpenAI的十分之一，Anthropic的七分之一。
如果单看收入，这个估值不便宜——DeepSeek目前的营收规模远小于上述三家公司。但如果看战略价值，这个估值可能被低估了。原因有三：
第一，中国市场的独立性溢价。当Anthropic的Mythos模型发布72小时就被美国政府出口管制，当谷歌对Meta的Gemini调用设置每周封顶额度，当H20芯片禁售导致英伟达计提45亿美元库存损失——全球AI市场正在分裂成两个体系。在拥有14亿人口、全球第二大经济体、数字经济规模超过50万亿人民币的中国市场，一个不依赖美国技术栈、能在国产芯片上跑万亿参数模型的AI公司，其战略价值怎么高估都不为过。
第二，成本效率的复利效应。Kimi B端负责人黄震昕在近期的一场沟通会上说了一句大实话："今年以来所有模型厂商都在涨价，核心原因是算力成本在全球范围内上涨。"但DeepSeek是个例外——它的推理成本只有竞品的几十分之一，而且最早完成了国产芯片适配。当算力成本持续上涨，成本效率高的公司将获得越来越大的毛利空间，这是一种结构性优势。Anthropic已经实现了季度盈利——这说明AI大模型不是不能赚钱，而是大部分公司在成本结构上就没想清楚怎么赚钱。
第三，从模型到生态的想象力。DeepSeek此前主要聚焦模型层，但融资后人员规模将扩充一倍，涵盖7大类33个岗位，多模态和Agent方向是重点投入方向。7月下旬将停用旧版V4模型倒逼用户升级——这是平台化的信号。从一个"模型提供商"进化为一个"AI基础设施平台"，估值逻辑会完全不同。
不只是DeepSeek：中国AI的资本变局
DeepSeek的融资不是孤立事件。把它放在2026年上半年中国AI的资本版图中看，几条清晰的脉络浮现出来。
具身智能成为最热赛道。上半年国内具身智能及机器人领域共发生288起融资事件，涉及226家企业，披露融资额超460亿元。千寻智能三个月四轮融资45亿元、估值突破200亿元；无界动力成立仅一年即完成超2亿美元天使轮，Pre-A轮近2亿美元也接近完成；港大教授李弘扬创立的源策未来成立仅两个月即获数亿元种子轮。Neura Robotics在德国获得英伟达和亚马逊参投的14亿美元C轮。宇树科技从受理到过会仅73天，创下科创板纪录。
算力基础设施进入国家工程级别。中国正在起草一份约2950亿美元（约2.1万亿人民币）的国家级AI数据中心网格建设计划，目标使用80%国产芯片。6月29日，首届全球太空算力大会在北京中关村开幕，揭牌北京太空算力创新中心，配套三年50亿产业资金。国家数据局发布方案，提出探索词元（Token）交易等新型交易模式——这是全球首次有国家层面提出以Token为基础构建数据价值体系。
大模型价格战降温。字节跳动推出豆包专业版付费服务，Kimi承认行业有泡沫但基本面扎实，所有模型厂商都在涨价——这标志着竞争从"免费烧钱抢用户"转向"价值付费"。黄震昕说得直白："Anthropic已经实现了季度盈利，收益是实实在在的。泡沫存在，但基本面并不空洞。"
"每一代模型，我们都在押注一个非共识。"
—— Sand.ai创始人曹越（前光年之外联合创始人）
真正的硬仗在后面
但需要冷静的是，510亿融资不是胜利的号角，而是更大战役的开始。
第一个硬仗是算力。尽管DeepSeek完成了国产芯片适配，但国产芯片的产能、互联效率、软件生态与英伟达最新一代Vera Rubin（性能比Blackwell Ultra提升3.3倍，下半年出货）仍有代际差距。全球高端AI GPU供给增速15%-20%，需求增速45%-60%，算力缺口约3倍——这意味着未来两三年，算力约束将是所有大模型公司的头号瓶颈。
第二个硬仗是生态。OpenAI正在从模型公司进化为全栈公司：自研芯片Jalapeño（9个月完成流片，计划年底规模化部署）、ChatGPT重构为Agent超级应用、Codex定位企业级核心、与HP达成Frontier战略合作部署AI到全企业栈。Anthropic虽然不做芯片，但在B端已经反超OpenAI，ARR达470亿美元，MCP协议月下载量达9700万次，成为Agent工具互联的事实标准。DeepSeek在模型层很强，但在芯片层、应用层、协议层的生态布局才刚刚开始。
第三个硬仗是人才。John Jumper（AlphaFold核心负责人、2024年诺贝尔化学奖得主）离开Google DeepMind加入Anthropic，Transformer架构核心作者Noam Shazeer流向OpenAI，谷歌AI人才持续外流。顶级AI人才的争夺已经全球化，中国公司能否吸引和留住顶尖人才，是长期竞争力的关键。
一个时代的分水岭
2026年6月，AI行业正在经历一个微妙的分水岭时刻。
在美国，前沿模型首次被纳入出口管制——Anthropic的Mythos 5发布仅72小时就被白宫以国家安全为由限制出口，加拿大总理公开警告过度依赖少数美国AI供应商的风险。OpenAI、Google、Anthropic三家首次在知识产权保护和出口管制上形成统一战线，硅谷三巨头从竞争对手变成了"竞争但同盟"的关系。
在中国，DeepSeek的510亿融资、七部门联合印发AI行动方案（2026-2028）、科创板第五套上市标准扩容接纳AI硬科技、国家级AI数据中心网格计划、Token交易探索——这些信号汇聚在一起，指向一个清晰的方向：中国AI正在从"跟跑"转向"并跑"，在某些领域甚至开始"领跑"，但这条"独立路线"需要的资金、算力、人才和政策支持，都是万亿级别的投入。
梁文锋用五年锁定期筛选了真正的长期资本，DeepSeek用三年时间证明了成本效率路线的可行性。但资本给的是弹药，不是胜利。真正的考验是：在算力受限的环境下，能不能持续迭代出世界一流的模型？能不能从一个模型公司成长为一个生态平台？能不能在全球AI市场分裂为两个体系的大背景下，建立起自己的技术护城河和商业闭环？
这些问题，510亿回答不了。只有时间能回答。
但可以确定的是：中国AI的独立战争，已经从游击战转入了正面战场。
51 billion RMB.
That's the size of DeepSeek's first external funding round -- approximately $7.4 billion, with a post-money valuation nearing 400 billion RMB (about $50 billion). A historic record for the largest single funding round in Chinese AI.
More telling are the details: founder Liang Wenfeng personally committed 20 billion RMB; Tencent and CATL are the largest external investors; the National AI Industry Investment Fund is the only exception allowed direct shareholding with voting rights; all external investors must put capital into a limited partnership managed by Liang Wenfeng, with a five-year lock-up period -- terms almost unthinkable in the fast-in, fast-out world of venture capital.
A Chinese AI company that held out for nearly three years without funding, sustaining R&D through internally generated profits, pulled off this round against a backdrop of "indefinite" U.S. bans on H20 chips, export controls on Anthropic's new model within 72 hours of release, and a global compute supply gap of roughly 3x. This is no ordinary capital event -- it is a capital referendum on China's independent AI path.
Why Now: Three Years Without Funding, Why Open the Door?
To understand the weight of this round, you need to go back to Liang Wenfeng's founding philosophy.
From day one, DeepSeek chose a path unlike nearly every other Chinese AI company: no cash-burning user acquisition, no consumer super-app, no press-conference fundraising. Its core team came from the quantitative hedge fund High-Flyer, which had strong self-sustaining revenue -- using proprietary trading profits to fund AI research, answering to no VCs, chasing no valuation bubbles.
That path held for nearly three years. DeepSeek V2 stunned the industry with extreme inference cost efficiency; V3 matched GPT-4o across multiple benchmarks; V4 completed adaptation to six domestic chips -- from Huawei Ascend to Cambricon to T-Head. DeepSeek was among the first LLM companies to seriously adapt to domestic compute. Its inference costs reached a fraction of competitors -- not marketing fluff, but real engineering capability.
So why fundraise now? Liang Wenfeng gave the answer in internal communications: it's not about money, it's about strategic resources.
By mid-2026, the LLM race has shifted from "whose model scores higher on benchmarks" to "who can iterate continuously under compute constraints." After the H20 ban, Chinese LLM companies no longer face a "can we buy chips" problem, but a "can we complete full-parameter training of trillion-parameter models on domestic chip clusters" problem. In June, the Shenzhen Hetao Academy/Shenzhen Smart City Compute Network completed full-parameter post-training of DeepSeek-V4-Pro on a domestic compute cluster -- the first third-party institution to do so, but large-scale, high-efficiency, low-cost training is still a long way off.
DeepSeek needs more than money -- it needs deep coordination with cloud infrastructure (Tencent), strategic partnerships in hardware supply chains (CATL's energy and manufacturing capabilities), and long-term policy support (the National AI Industry Fund). This round is fundamentally not financial investment, but strategic alliance.
"Liang Wenfeng used the five-year lock-up to tell the market: this isn't about quick money, this is about a protracted war. Capital that accepts a five-year lock-up is true patient capital."
-- An investor who participated in DeepSeek funding discussions, speaking privately
$50 Billion Valuation: Expensive or Cheap?
What does a $50 billion valuation mean?
Some comparisons: OpenAI's latest valuation is roughly $500 billion (private market secondary trades), Anthropic is around $350 billion (reportedly reaching $47 billion ARR in May, surpassing OpenAI on the enterprise side), xAI is around $180 billion. DeepSeek's $50 billion is roughly one-tenth of OpenAI, one-seventh of Anthropic.
Looking purely at revenue, this valuation isn't cheap -- DeepSeek's current revenue scale is far smaller than those three. But looking at strategic value, the valuation may actually be understated. Three reasons:
First, the independence premium of the China market. When Anthropic's Mythos model was hit with export controls 72 hours after release, when Google capped Meta's weekly Gemini API calls, when the H20 ban forced Nvidia to take a $4.5 billion inventory write-down -- the global AI market is splitting into two systems. In a market with 1.4 billion people, the world's second-largest economy, and a digital economy exceeding 50 trillion RMB, an AI company that doesn't depend on the U.S. tech stack and can run trillion-parameter models on domestic chips has strategic value that can hardly be overstated.
Second, the compounding effect of cost efficiency. Kimi's B2B lead Huang Zhenxin put it bluntly at a recent event: "Every model vendor has been raising prices this year, the core reason being that compute costs are rising globally." But DeepSeek is the exception -- its inference costs are a fraction of competitors, and it was the first to complete domestic chip adaptation. As compute costs keep rising, companies with superior cost efficiency gain expanding gross margins -- a structural advantage. Anthropic has already achieved quarterly profitability -- proving that LLMs can make money; most companies simply haven't figured out their cost structure.
Third, the upside from model to ecosystem. DeepSeek previously focused mainly on the model layer, but post-funding headcount will double across 7 categories and 33 positions, with multimodal and Agents as key investment areas. In late July, it will discontinue the old V4 model to force user upgrades -- a platformization signal. Evolving from a "model provider" to an "AI infrastructure platform" would completely change the valuation logic.
Not Just DeepSeek: China's AI Capital Landscape Shift
DeepSeek's funding is not an isolated event. Placed within the capital landscape of Chinese AI in H1 2026, several clear threads emerge.
Embodied AI becomes the hottest track. In H1, domestic embodied AI and robotics saw 288 funding events involving 226 companies, with disclosed funding exceeding 46 billion RMB. Qianxun Intelligence raised 4.5 billion RMB across four rounds in three months, valuation breaking 20 billion; Wujie Dynamics completed a $200M+ angel round just one year after founding, with a nearly $200M Pre-A close; HKU professor Li Hongyang's Yuance Future raised hundreds of millions in seed funding just two months after launch. Neura Robotics secured a $1.4B Series C in Germany with Nvidia and Amazon participating. Unitree Robotics went from IPO filing to approval in a record 73 days on the STAR Market.
Compute infrastructure enters national-engineering scale. China is drafting a roughly $295 billion (~2.1 trillion RMB) national AI data center grid construction plan, targeting 80% domestic chip usage. On June 29, the first Global Space Compute Conference opened in Beijing's Zhongguancun, inaugurating the Beijing Space Compute Innovation Center with 5 billion in three-year industrial funding. The National Data Bureau released a framework proposing exploration of Token trading and other novel transaction models -- the first time any nation has proposed building a data value system on a Token basis.
LLM price wars cool down. ByteDance launched Doubao Pro paid service; Kimi acknowledged industry bubbles but solid fundamentals; all model vendors are raising prices -- marking a shift from "free cash-burning for users" to "value-based payment." Huang Zhenxin put it plainly: "Anthropic has achieved quarterly profitability, the returns are real. Bubbles exist, but the fundamentals aren't hollow."
"With every generation of models, we're betting on a non-consensus view."
-- Cao Yue, founder of Sand.ai (former co-founder of Light Years Away)
The Real Battles Lie Ahead
But a cool-headed view is essential: 51 billion is not a victory trumpet, but the start of a much larger campaign.
The first hard battle is compute. Although DeepSeek has completed domestic chip adaptation, domestic chips still have a generational gap in production capacity, interconnect efficiency, and software ecosystem compared to Nvidia's latest Vera Rubin (3.3x performance over Blackwell Ultra, shipping in H2). Global high-end AI GPU supply grows 15-20% annually, demand grows 45-60%, and the compute gap is roughly 3x -- meaning compute constraints will be the #1 bottleneck for all LLM companies over the next two to three years.
The second hard battle is ecosystem. OpenAI is evolving from a model company into a full-stack company: custom chip Jalapeño (taped out in 9 months, planning scale deployment by year-end), ChatGPT restructured as an Agent super-app, Codex positioned as enterprise core, Frontier strategic partnership with HP to deploy AI across the full enterprise stack. Anthropic doesn't make chips but has surpassed OpenAI in B2B, with $47B ARR and MCP protocol hitting 97 million monthly downloads, becoming the de facto standard for Agent tool interconnection. DeepSeek is strong at the model layer, but ecosystem positioning at the chip layer, application layer, and protocol layer is just beginning.
The third hard battle is talent. John Jumper (AlphaFold core lead, 2024 Nobel Chemistry laureate) left Google DeepMind for Anthropic; Transformer architecture co-author Noam Shazeer moved to OpenAI; Google's AI talent continues to bleed. The competition for top AI talent has gone global; whether Chinese companies can attract and retain top talent is key to long-term competitiveness.
A Watershed Moment for an Era
In June 2026, the AI industry is experiencing a subtle watershed moment.
In the U.S., frontier models were placed under export controls for the first time -- Anthropic's Mythos 5 was restricted by the White House on national security grounds just 72 hours after release, and Canada's Prime Minister publicly warned about the risks of over-reliance on a handful of American AI providers. OpenAI, Google, and Anthropic formed a united front for the first time on IP protection and export controls; Silicon Valley's big three shifted from competitors to "competitive but allied."
In China, DeepSeek's 51 billion funding, seven ministries jointly issuing the AI Action Plan (2026-2028), the STAR Market's fifth listing standard expanding to accommodate AI hard tech, the national AI data center grid plan, Token trading exploration -- these signals converge on a clear direction: Chinese AI is shifting from "catching up" to "running neck-and-neck," and in some areas starting to "lead," but this "independent path" requires trillion-level investments in capital, compute, talent, and policy support.
Liang Wenfeng used a five-year lock-up to screen for truly long-term capital; DeepSeek spent three years proving the viability of the cost-efficiency path. But capital provides ammunition, not victory. The real test is: under compute constraints, can you keep iterating world-class models? Can you grow from a model company into an ecosystem platform? Against a backdrop of the global AI market splitting into two systems, can you build your own technical moat and commercial flywheel?
51 billion cannot answer these questions. Only time can.
But one thing is certain: China's AI war for independence has moved from guerrilla warfare to the front lines.
DeepSeek融资纪录 · 中国AI独立路线 · 国产算力适配 · 五年锁定期耐心资本 · AI地缘政治分裂 · 大模型价格战降温 · 具身智能融资爆发 · 算力缺口3倍
DeepSeek funding record, China AI independence, domestic chip adaptation, 5-year lock-up patient capital, AI geopolitical split, LLM price war cooling, embodied AI funding boom, 3x compute gap
