初创公司尽调
尽调报告 AI foundation models and model APIs Late-stage private 2026-07-21

DeepSeek

DeepSeek 尽调报告

DeepSeek 的技术领先和需求动能值得继续跟踪,但财务不透明,地缘政治、IP 与治理风险偏高,当前定价很难撑住,测算必须保守。

封面要素

成立时间 01
2023 year [CO001]
总部 02
Hangzhou, China [CO001]
创始人 03
Liang Wenfeng [CO006]
据报道估值 04
50000 USDm [CV002]
据报道 2026 年 6 月融资 05
7000 USDm [CO022]
Vercel 路由 token 份额 06
17 % [CU017]

公司概况

DeepSeek 源自 Liang Wenfeng 的 High-Flyer 生态,定位为研究优先的 AI 实验室,主攻开放权重前沿模型、低成本 API 推理,以及编码、聊天、推理工作负载的快速迭代。公开证据显示,公司技术动能异常强、云分发很广,开发者采用也很快;但治理、经审计运营数据和长期商业耐久性仍只露出一部分。

官网
www.deepseek.com
成立时间
2023-07-17
创始人
Liang Wenfeng
创立地点
Hangzhou, China
总部
Hangzhou, China
产品
DeepSeek 发布 V2、V3、V4、R1 等开放权重前沿语言与推理模型,并通过 API 平台和云市场分发变现。
客户
开发者、AI 产品团队,以及通过 API 和云目录采用低成本前沿模型的企业。
商业模式
基于用量的模型 API 收入、间接云渠道分发,以及围绕开放权重发布形成的生态采用。
阶段
Late-stage private
融资情况
据报道,首轮外部融资于 2026 年 6 月完成,估值超过 $50 billion;随后又在 2026 年 7 月据称洽谈新融资,估值约 $71 billion。
[CO001, CO006, CO012, CO016, CO022, CV003]

执行摘要

主要优势

  • 开放权重模型性能接近前沿梯队,API 定价又异常低。
  • GitHub、AWS、Azure 和网关渠道都释放出强开发者与云分发信号。
  • High-Flyer 生态提供创始人背书的算力和人才底座。

主要风险

  • 公开财务披露太薄,无法验证收入质量、利润率或留存。
  • 地缘政治和出口管制压力可能卡住算力供给与全球部署。
  • IP、信任和合规指控可能限制企业采用,或压缩估值。

未决问题

  • 已审阅公开来源均未披露经审计收入、ARR 或毛利率。
  • 客户集中度和留存仍只能推断,没有直接披露。
  • 梁文锋之外的治理纵深以及正式董事会结构记录仍很稀疏。

目录

Chapter 01

01公司概况

1.1 身份、使命与法律形式

DeepSeek 的法律实体为 Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.(中文:杭州深度求索人工智能基础技术研究有限公司),注册地在浙江省杭州市。品牌名「深度求索」(Shen Du Qiu Suo)意为「向深处求索」,体现的是研究优先,而非产品优先。公司于 2023 年 7 月 17 日成立,是创始人 Liang Wenfeng 控制的量化对冲基金 High-Flyer Capital Management 的全资子公司。DeepSeek 的使命表述是「以好奇心解开 AGI 的奥秘」,不像美国前沿实验室常见表述那样强调安全、竞争或社会影响,而是聚焦纯粹科学探索。官网(deepseek.com)和平台(platform.deepseek.com)把公司呈现为一家发布开放权重模型的研究组织,并将 API 平台定价刻意贴近成本。公司在 chat.deepseek.com 运营 AI 聊天机器人,也公开发布技术论文和模型权重。多家独立来源确认其杭州注册;北京运营存在具有合理可能,但已抓取来源没有直接验证。截至本次报告运行日,DeepSeek 仍是一家私人公司,正在为 2027 年可能 IPO 做准备。[CO001, CO002, CO003, CO004, CO005, CO009]

公司快照 KPI 表
指标数值或状态日期或期间置信度缺口或注意事项
法定名称Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.(杭州深度求索人工智能基础技术研究有限公司)当前英文音译;通过 Wikipedia 和新闻审阅中文登记信息
品牌DeepSeek(中文:深度求索)当前未审阅正式商标登记
成立July 17, 20232023-07-17Wikipedia 加多家独立来源
总部中国浙江杭州当前多家独立来源;北京运营存在尚未验证
阶段后期私营公司;截至 2026 年 7 月 IPO 准备活跃2026-07-14Bloomberg 和 TechCrunch 2026 年 7 月 14 日
员工约 160(2025 年估计)2025未公布官方员工数;Wikipedia 和新闻引用约 160
最新轮估值约 $50B(2026 年 6 月);2026 年 7 月讨论中为 $71B2026-07CB Insights $50B;Bloomberg/TechCrunch $71B(谈判尚未确认完成)
累计融资$7B(2026 年 6 月;首次外部融资)2026-06Bloomberg 和 TechCrunch 已确认
IPO 目标2027;可能 2026 年 Q42026-07-14Bloomberg 7 月 14 日;未审阅招股书或交易所文件
deepseek-chat API 定价$0.07/M 输入 tokens;$1.10/M 输出 tokens2026-07-21官方 api-docs.deepseek.com 定价页
收入未公开披露;API 被描述为成本之上的小额利润率2025CEO 通过 ChinaTalk 表述;无经审计财务
反向信号Anthropic/OpenAI IP 盗窃指控;美国和澳大利亚政府禁令2026多家独立来源交叉印证;见风险章节

快照混合了官方表层数据、融资轮新闻报道、分析师数据库数字和 Wikipedia;未经审计或媒体估计的指标标为低置信度,而不是当作已验证事实。

[CO001, CO002, CO008, CO016, CO017, CO018]
FO002: 公司投资假设概览

DeepSeek 的竞争逻辑从 Liang Wenfeng 的量化基金算力基础设施出发,经由开放权重模型创新,落到全球开发者采用、国家资本支持和 IPO 可选性;治理不透明和监管禁令则构成主要风险反馈回路。

[CO004, CO006, CO007, CO008, CO011, CO012]

1.2 创始人、领导层与治理

Liang Wenfeng(1985 年生于广东省)是 DeepSeek 的创始人、CEO 和控股股东。他在浙江大学学习 AI 与电气工程,取得学士和硕士学位。多方来源认为,这一背景直接支撑 DeepSeek 的技术研究议程;量化对冲基金创始人很少有这种履历。Liang 于 2015 年共同创立 High-Flyer Capital Management,将其做成中国四大量化对冲基金之一,并积累了计算基础设施——最典型的是 Fire-Flyer 2 集群,包含 625 个节点上的 5,000 块 A100 GPU,支撑了 DeepSeek 的训练任务。多方来源称 Liang 本人每天都在「读论文、写代码、参加小组讨论」,因此更像一位实践型 CEO。截至 2026 年 5 月,他控制公司约 90% 股权;Bloomberg 在 2026 年 7 月估计其个人净资产为 $36 billion,使他成为全球最富有的 AI 模型公司创始人。治理透明度有限:董事会构成、投资者投票权、优先股堆叠条款或股权结构表细节都没有公开披露。公司截至 2025 年约有 160 名研究人员——按前沿 AI 实验室标准异常精简——也带来关键人物依赖和接班问题,公开记录并未给出答案。据报道,公司决定在 2026 年引入外部投资,是为了向员工提供股权;背后动因是资金充足的竞争对手正在挖角研究员。[CO006, CO007, CO008, CO021, CO024, CO027]

领导层与创始人表
人物公开职务背景创始人-市场匹配尽调备注
Liang Wenfeng创始人兼 CEOZhejiang University AI 与 EE 工学学士、硕士;2015 年联合创立 High-Flyer;控制公司约 90%深厚 AI 与 ML 研究资历,加上量化基金运营者可信度;在中国 AI 领域画像独特极强关键人物依赖;治理权利和接班计划未公开
High-Flyer Capital Management 团队母公司资源基础中国前四大量化基金;Fire-Flyer 2 集群(5,000 块 A100 GPU);最近估值 $8B提供算力、资本和人才漏斗,且没有传统风投稀释关联交易、借调员工数和成本分摊未披露
DeepSeek research team(约 160)工程师和研究员据 The Economist,主要为 Zhejiang University 校友以相对美国同业异常精简的员工数交付前沿模型研究个体研究员姓名未广泛披露;关键人物风险高

除 Liang Wenfeng 外,已审阅英文来源没有公开记录其他高管;治理结构、董事会构成和投资人董事席位仍未披露。

[CO006, CO007, CO008, CO021, CO027, CO037]

1.3 资本结构、融资与 IPO 路径

成立后的前两年半,DeepSeek 作为 High-Flyer 的全资部门运营,没有外部投资者。ChinaTalk 在 2024 年 11 月的专访明确称,DeepSeek 完全由 High-Flyer 出资,且没有融资计划。战略转向发生在 2026 年上半年:Financial Times 报道称,Liang Wenfeng 选择融资,是为了向员工发放股份,以应对激烈的人才挖角。Reuters 和 The Information 报道,2026 年 4 月的初步讨论估值约为 $10 billion;到 2026 年 5 月,FT 和 Bloomberg 都称估值在数周谈判中从 $20 billion 飙升至 $45 billion。该轮融资于 2026 年 6 月完成,约融资 $7 billion,估值 $50 billion——这是公司首次外部融资。领投方为 China Integrated Circuit Industry Investment Fund(Big Fund),这是一只支持国内半导体和 AI 发展的国家基金。Tencent 和 Alibaba 已确认参投;CB Insights 还将 CATL 和 Guozhitou Private Equity Fund Management 列入投资者名单。一个值得注意的治理条款是禁止挖角约定,保护 DeepSeek 员工不被投资者的被投公司挖走。到 2026 年 7 月 14 日——距离本次报告运行日一周——Bloomberg 和 TechCrunch 报道称,DeepSeek 正在洽谈进一步融资 $1.5 billion,估值约 $71 billion,为 2027 年 IPO 目标做准备。估值在三个月内从 $10 billion 推进到 $71 billion,反映了国家资本对主权 AI 的热情,也反映了 DeepSeek 模型的国际采用。[CO017, CO018, CO019, CO020, CO022, CO023]

利益相关方或投资人图谱
利益相关方角色战略重要性证据基础尽调问题
China Integrated Circuit Industry Investment Fund(Big Fund,大基金)2026 年 6 月 $7B 轮领投方支持主权 AI 的国家资本载体;释放与国家 AI 议程政策同向的信号TechCrunch 2026 年 5 月称其为轮次领投方;Bloomberg 报道确认厘清董事会权利、政策约束和任何国家安全义务
Tencent2026 年 6 月轮次战略投资人中国顶级云和消费平台;潜在分发伙伴Bloomberg 和 TechCrunch 2026 年 7 月确认 Tencent 为投资人确认商业集成条款、云依赖和持股比例
Alibaba2026 年 6 月轮次战略投资人中国最大云服务商;竞争与伙伴动态并存Bloomberg 和 TechCrunch 2026 年 5 月称 Alibaba 参与洽谈;后确认参投厘清 Alibaba Cloud 托管条款、模型集成协议和排他性
CATLCB Insights 披露的战略投资人中国领先电池公司;释放工业 AI 应用信号CB Insights 独角兽档案列出 CATL 为投资人核实参投情况;厘清工业 AI 协同逻辑
Guozhitou Private Equity Fund Management(国资投私募基金管理)CB Insights 披露的财务投资人国家支持的金融资本;强化公共部门资本协同CB Insights 独角兽档案确认实体身份和基金授权范围
High-Flyer Capital Management创始母公司和 2026 年前唯一出资方约 90% 所有者;提供算力集群和种子资本ChinaTalk 2024 年 11 月;TechCrunch 2026 年 5 月;多家独立来源需要审阅关联协议和算力成本分摊
未来 IPO 投资人潜在公开市场股东如果 IPO 以 $71B 以上推进,将重设披露义务和治理Bloomberg 和 TechCrunch 2026 年 7 月 14 日跟踪交易所文件;基石买方身份尚未公开

投资人角色从公开融资报道推断;未审阅投资人权利协议或附函。

[CO017, CO020, CO022, CO023, CO024, CO025]
FO003: 融资轮估值推进

据报道,DeepSeek 估值从 2026 年 4 月最初估计的 $10 billion 升至 2026 年 7 月谈判中的 $71 billion, 大约三个月涨了 7 倍,反映出国家资本热情和全球模型采用势头。

所有数值都来自媒体报道或数据库档案,不是经审计财报或已签署投资条款清单;实际交割估值可能不同于媒体报道的谈判价格。

[CO017, CO018, CO019, CO022, CO023, CO041]

1.4 模型架构与产品形态

DeepSeek 的产品逻辑是发布一流开放权重模型,配套详细技术论文,并提供价格可负担的 API。2023 年末的首批发布——DeepSeek-Coder(2023 年 11 月)和 DeepSeek-LLM 67B——让这家实验室成为美国开源模型之外可信的中国替代方案。关键架构突破是 DeepSeek-V2(2024 年 5 月):它采用 Multi-head Latent Attention(MLA)和稀疏 Mixture-of-Experts 设计,将 KV cache 内存降到标准 Multi-Head Attention 的 5-13%,大幅压低推理成本。V2 将 API 访问定价为每百万 token 1 RMB——约为当时 Llama 3 70B 成本的七分之一——点燃了中国 AI 价格战,迫使 ByteDance、Baidu、Tencent 和 Alibaba 降价。DeepSeek-V3(2024 年 12 月)把架构扩展到 671 billion 总参数、每个 token 激活 37 billion 参数,并用 14.8 trillion tokens 训练;CSIS 估算,使用 H800 芯片的最终预训练计算成本约为 $5.6 million。DeepSeek-R1(2025 年 1 月 20 日)在多个基准上达到或超过 OpenAI o1,并引发了 DeepSeek Monday 市场事件。R1 方法论文发表于 Nature(volume 645,pages 633-638,2025)。DeepSeek-V4(2026 年 4 月 24 日)推出 V4-Pro,拥有 1.6 trillion 总参数和 1 million-token 上下文窗口。截至本次报告运行日,API 提供 deepseek-chat,价格为每百万输入 token $0.07。模型权重可从 GitHub 和 Hugging Face 免费下载;模型也可通过 AWS、Azure、Google Vertex AI 和 NVIDIA NIM 使用。[CO011, CO012, CO013, CO014, CO015, CO016]

模型家族与发布时间线
模型发布日期架构亮点参数总量重要性
DeepSeek-Coder2023 年 11 月面向代码的 transformer 解码器33B 旗舰变体首次公开发布;建立编程模型可信度
DeepSeek-LLM 67B2023 年 11 月标准仅解码器 transformer67B首个通用 LLM;当时可与 Llama 2 对标
DeepSeek-V22024 年 5 月MLA 加 DeepSeekMoE 稀疏架构236B 总量 / 21B 激活引发中国 AI 价格战;每百万 tokens 1 RMB;全球获赞
DeepSeek-V32024 年 12 月MoE;671B 总量 / 37B 激活;14.8T token 训练;2.788M H800 GPU-hours671B 总量 / 37B 激活CSIS 估计最终训练运行成本约 $5.6M;挑战西方算力成本假设
DeepSeek-R12025 年 1 月 20 日GRPO 强化学习;推理专用;开放权重蒸馏变体;基础尺寸未披露Nature 第 645 卷;发布当周 DAU 增长 700%;DeepSeek Monday 市场事件
DeepSeek-V4 Flash 和 Pro2026 年 4 月 24 日下一代 MoE;1M 上下文窗口;针对 Huawei 芯片优化284B Flash / 1.6T ProV4-Pro 是最大开放权重模型;可在 NVIDIA、AWS、Azure、Google 上使用

参数和日期来自已审阅 arXiv 论文、Wikipedia、GitHub README 和新闻;截至运行日期,V4 内部训练细节尚未在完整技术论文中发表。

[CO011, CO012, CO014, CO015, CO016, CO029]

1.5 里程碑、市场影响与后续尽调缺口

DeepSeek 从创立到可能成为估值超过 $50 billion 的私人公司,只用了 36 个月;这条路径靠三点支撑:High-Flyer 集群带来的算力优势、来自浙江大学的顶尖研究人才,以及刻意选择的开源策略,后者在全球获得了自下而上的采用。Sensor Tower 数据量化了 R1 的病毒式反应:2025 年 1 月 22-28 日期间,日活用户环比增长超过 700%;该应用 19 天内全球下载量超过 23 million,是 ChatGPT 在相同成熟阶段下载速度的两倍以上。Bloomberg 2025 年 10 月数据显示,DeepSeek 在非洲超过 OpenAI 和 Google。截至 2026 年 6 月,DeepSeek 约占 Vercel 平台企业 AI token 流量的 23%。后续章节必须展开的关键反向信号包括:(1)Anthropic 在 2026 年 2 月指控 DeepSeek 通过 24,000 个欺诈创建账户和 16 million 次交互发起工业规模蒸馏攻击;(2)OpenAI 国会备忘录指称持续 IP 盗取;(3)美国 NDAA FY2026 限制政府使用 DeepSeek;(4)澳大利亚政府范围内禁用 DeepSeek;(5)Stanford HAI 观察到,DeepSeek 在隐私保护、数据来源和版权方面明显不透明。公司不披露收入、毛利率、客户数、流失数据、算力烧钱速度或经审计财务。后续章节应独立检验产品市场契合、客户质量、算力经济性、监管暴露和 IPO 准备度。[CO026, CO031, CO032, CO033, CO034, CO036]

里程碑表
日期事件类型金额或估值来源与置信度
2015Liang Wenfeng 联合创立 High-Flyer Capital Management 量化基金公司N/AWikipedia;Fortune;ChinaTalk — 高
2021High-Flyer 部署 Fire-Flyer 2 集群(5,000 块 A100 GPU;1B yuan 预算)基础设施约 1B yuanCSIS 深度证词 — 高
2023-07-17DeepSeek 在杭州注册为 High-Flyer AI 子公司创立N/AWikipedia;新闻 — 高
2023-11DeepSeek-Coder 和 DeepSeek-LLM 67B 开源发布产品N/AarXiv 2401.14196;GitHub — 高
2024-05DeepSeek-V2 引发中国 AI 价格战;MLA 加 MoE 架构获全球好评产品每百万 tokens 1 RMB APIChinaTalk;FT;MIT Tech Review — 高
2024-12DeepSeek-V3 发布;671B MoE 用 14.8T tokens 训练,算力成本约 $5.6M产品约 $5.6M 训练成本(CSIS)arXiv 2412.19437;CSIS;GitHub V3 仓库 — 高
2025-01-20DeepSeek-R1 开源;推理模型在基准测试上追平 OpenAI o1产品N/AarXiv 2501.12948;Nature 第 645 卷;GitHub R1 — 高
2025-01-27DeepSeek Monday:Nvidia 股价下跌约 17%;市值蒸发 $589B市场约 $589B Nvidia 市值损失Guardian;NPR;New Yorker;SensorTower;Wikipedia — 高
2025-10Bloomberg 报道 DeepSeek 在非洲击败 OpenAI 和 Google市场N/ABloomberg 2025 年 10 月专题 — 中
2026-02-13OpenAI 向国会提交备忘录,指控 DeepSeek 盗窃 IP反向N/AFDD 分析 2026 年 2 月 — 高
2026-02-24Anthropic 指控 DeepSeek 发起工业规模蒸馏攻击反向超过 16M 次交互;24,000 个账户CNBC 2026 年 2 月 24 日 — 高
2026-04-24DeepSeek-V4 预览版发布(V4-Flash 284B;V4-Pro 1.6T 参数)产品N/AWikipedia;NVIDIA 构建页;新闻 — 高
2026-05-06TechCrunch 报道轮次谈判中估值数周内从 $20B 飙至 $45B融资$20B 到 $45B 区间TechCrunch 2026 年 5 月 6 日 — 高
2026-06首次外部融资轮完成 $7B 融资,估值约 $50B融资$7B 融资;估值约 $50BTechCrunch 2026 年 7 月 14 日;CB Insights;Bloomberg — 高
2026-07-14Bloomberg:正洽谈以 $71B 估值融资 $1.5B;IPO 目标 2027 年融资讨论 $1.5B;$71B 估值TechCrunch 和 Bloomberg 2026 年 7 月 14 日 — 高

金额和估值来自已审阅新闻报道;未审阅经审计股权结构表或内部里程碑文件。

[CO001, CO007, CO011, CO012, CO013, CO015]
FO001: 公司里程碑时间线

DeepSeek 从 2023 年 High-Flyer 实验室起步,走到 2026 年 $7B 独角兽级别以上融资轮和 IPO 准备;这一进程由连续开放权重模型发布推动,逐步挑战西方前沿模型的主流假设。

[CO001, CO004, CO011, CO012, CO013, CO014]

1.6 图表与论据

Chapter 02

02市场分析

2.1 市场边界:DeepSeek 竞争的是模型支出、平台支出和 AI 应用预算

不应把 DeepSeek 简单归入一个没有差异的「生成式 AI 市场」。已抓取证据至少支持四个相互关联的支出池。第一是前沿模型与 API 推理支出,买家会在各实验室之间比较价格、延迟、上下文和模型质量。第二是 AI 平台支出,企业为评估、治理、路由和应用开发层付费,以便在生产中管理多个模型。第三是智能体与编码工作负载支出,长上下文推理质量和工具使用能力决定一个便宜模型是能部署,还是只值得测试。第四是消费者与专业个人 AI 应用支出,应用下载和订阅带来高量但低治理的漏斗。Gartner 2026 年 7 月预测和 Goldman Sachs 的 $150 billion 软件 TAM 视角都确认总市场很大,但这两个数字都不是 DeepSeek 近期自己的合适 SAM。更好的工作定义是:AI 工作负载中,那些开放权重、低价格、具备推理能力的模型能达到质量门槛,且买家愿意多供应商并用的子集。这个边界让 DeepSeek 的机会明显窄于标题级 TAM,但也可执行得多。[CM001, CM002, CM003, CM004, CM005, CM006]

市场定义表
细分 / 类别纳入支出排除支出买方 / 付款方与 DeepSeek 的关系
前沿模型与推理支出按 token 计量的 API 用量、推理调用、工具使用、长上下文推理原始 GPU 基础设施采购和无关云工作负载开发者、AI 产品团队、平台工程DeepSeek 近期变现的核心轨道
AI 平台与路由层评测、治理、可观测性、用量跟踪、应用开发工具没有部署软件的纯模型研究支出企业 AI 平台负责人、CIO/CTO 预算重要,因为 DeepSeek 可以在第三方平台内部赢得用量
智能体与编码工作负载长周期任务执行、编码智能体、工具驱动推理永远不能变成生产工作的简单聊天机器人流量开发者、企业自动化团队DeepSeek 的低成本推理位置在这里最有价值
消费者和专业消费者 AI 应用聊天订阅、应用使用、创作者试验、自助网页流量不依赖 AI 的一般娱乐支出终端用户和小团队带来认知和漏斗顶端采用,但不必然带来持久收入
现状替代品人工劳动、既有 SaaS、既有美国模型 API、内部模型n/a现有产品和运营预算决定 DeepSeek 是扩大支出,还是只是替代另一家供应商

本章采用四个市场池加现状替代品的市场定义,避免夸大 DeepSeek 的可服务市场。

[CM001, CM002, CM003, CM004, CM005, CM006]
TAM / SAM / SOM 或规模测算视角表
发布方年份地区数值CAGR / 增长方法置信度局限
Goldman Sachs2023全球$150B 生成式 AI 软件 TAM宏观软件 TAM 视角有用的上限,不是 DeepSeek 专属 SAM
Gartner2026全球$64.252B AI 模型与平台终端用户支出较 2025 年 YoY 增长 63.4%分析师支出预测跟踪品类支出,不是一家实验室的可服务份额
Gartner2026全球$23.356B 基础 GenAI 模型支出较 2025 年 YoY 增长 104.2%分析师子细分预测仍宽于 DeepSeek 近期可落地的覆盖范围
Gartner2026全球$4.910B 专用 / DSLM 生成式 AI 模型支出较 2025 年同比 +210%分析师细分市场预测指向买家偏好调优或垂直领域层
State of AI 调查2025美国调查受访者44% 的美国企业为 AI 工具付费高于 2023 年的 5%开放调查 / 商业采用综合采用率调查,不是市场收入
State of AI 调查2025美国调查受访者$530,000 平均 AI 合同规模n/a调查 / 行业综合平均合同数据只能指示方向,不针对 DeepSeek
Vercel AI Gateway2026生产路由样本2026 年 5 月,AI Gateway tokens 环比 +20%,支出环比 +43%月度观测到的路由数据网关样本反映已路由的生产工作负载,不代表完整市场需求
OpenRouter2026开发者网关样本2026 年 1 月至 6 月初,DeepSeek token 份额从 9% 升至 18%六个月变化观测到的请求日志份额token 份额不等于厂商收入份额

这些视角有意混合宏观 TAM、细分市场预测、采用率调查和观测到的路由数据,因为 没有任何单一公开估算能覆盖 DeepSeek 实际可变现市场。

[CM007, CM008, CM009, CM010, CM023, CM024]
FM001: 市场估算区间

公开市场口径从 2026 年模型与平台支出基准 $64.3B,到生成式 AI 软件上限 $150B;DeepSeek 实际可服务市场(SAM) 落在低成本推理和路由层,而不是整个总可用市场(TAM)。

第四行混合的是采用率百分比,而不是美元金额;纳入它是为了给需求设一个有边界的观察口径,因为公开资料没有给出 DeepSeek 精确可服务市场规模。

[CM007, CM008, CM009, CM010, CM011, CM012]

2.2 买家分层:开发者、企业 AI 团队和自助用户行为完全不同

市场证据显示,DeepSeek 的买家不是同一类人。开发者和 AI 原生产品团队最在意兼容性、token 价格、上下文长度,以及模型是否「足够好到可以上线」。DeepSeek 自己的 API 文档通过模拟 OpenAI 和 Anthropic 风格端点降低迁移摩擦;Alibaba Model Studio、BigModel 和 Baidu Qianfan 也说明,中国平台正在把多模型访问和工具编排常态化。企业平台负责人是另一类买家:Gartner 称,预算正在转向能证明成本透明、用量追踪、性能控制和可衡量结果的供应商。这既利好平台和路由层,也利好底层模型实验室。第三类买家是消费者或专业个人用户,习惯形成和自助试验比治理功能更重要。State of AI 的 2025 年调查数据表明,采用已经足够广,消费者和个人专业用户不再只是无关紧要的旁路;它会把认知和试验反馈到企业试点。DeepSeek 的市场位置最强的地方,是这些漏斗重叠之处:低成本开发者采用、智能体试验,以及足够高的质量,使其能升级进入受管理的企业路由,而不是只停留在爱好者流量。实践中,最有价值的买家往往不是终端用户,而是决定流量如何路由的团队。路由负责人通常决定 DeepSeek 会成为持久支出,还是一阵短暂好奇。[CM011, CM012, CM013, CM014, CM015, CM016]

细分市场 / 买家地图
细分市场买家用户付款方工作流 / 目标预算负责人采用触发因素
个人开发者本人开发者自费或报销用低价推理模型构建并测试应用个人或团队工具预算切换成本低,性价比收益立刻可见
AI 原生初创公司 / SMB 团队工程或产品负责人开发者和运营人员CTO 或产品预算上线 AI 功能、AI 助手和任务智能体CTO / 工程副总裁模型通过评测,成本仍远低于前沿实验室
企业 AI 平台团队CIO、CTO 或 AI 平台负责人开发者、分析师、业务用户中央 AI 平台或转型预算在治理、可观测性和策略控制下路由多个模型高级技术预算负责人厂商能接入托管路由,而不只是提供直接 API 调用
消费者 / 专业消费者用户本人本人本人聊天、搜索、创作或实验个人软件支出低价或免费下仍有快速性能和有吸引力的输出质量
云 / 平台中介云或开发者平台运营方自身下游客户平台运营方上架第三方模型,并通过托管目录把用量变现平台 P&L 负责人需求和质量足够支撑平台把 DeepSeek 与同类模型一起上架

同一个人可能会从自助实验逐步转向掌握团队或企业预算;关键区别在于,哪一笔预算 需要批准重复支出。

[CM011, CM012, CM013, CM014, CM015, CM016]
FM002: 买方决策标准矩阵

不同 DeepSeek 买方群体会在价格、治理和渠道触达之间做不同取舍。

[CM032, CM034, CM035, CM036, CM039]
FM003: 采用漏斗 / 价值链图

DeepSeek 的典型路径从低成本试用开始,之后才升级为有治理的平台使用。

[CM018, CM019, CM020, CM023, CM024]

2.3 增长驱动与约束:效率打开市场,但也推动商品化

DeepSeek 受益于当前最重要的市场驱动:买家想要更多 AI 输出,却不愿接受前沿实验室定价。Vercel 2026 年 6 月生产指数和 OpenRouter 的采用说明都显示同一模式——只要低成本模型达到质量门槛,客户愿意把大量生产 token 流量路由过去。DeepSeek 定价页、MiniMax 按用量付费定价、Kimi 的 256k 长上下文定位,以及 Anthropic 的高端定价栈,都说明市场正在按路由策略分层,而不是收敛到一个赢家。与此同时,帮助 DeepSeek 的同一股力量也会限制它的护城河。State of AI 称竞争已经加剧,中国的 DeepSeek、Qwen 和 Kimi 在推理和编码上缩小了差距;Artificial Analysis 强调,买家可以把质量、价格、速度和开放性并排比较;Alibaba 的模型工作室更是把第三方模型直接放在一个目录里售卖。监管和地缘政治又叠加了一层约束。State of AI 讨论的 MOFCOM 出口管制框架和中国特定的国产硅野心都意味着,DeepSeek 所处市场里技术可以快速扩散,但信任、政策和供应获取仍会决定谁能规模化购买。结果是一个需求异常强、商品化压力同样异常强的市场。[CM023, CM024, CM025, CM026, CM027, CM028]

增长驱动因素与约束表
驱动因素 / 约束方向时间影响尽调问题
低成本、前沿质量的推理正向当前前沿实验室定价难以自洽时,DeepSeek 可赢下生产用量买家完成定制评测后,目标工作负载上的质量有多稳定?
OpenAI / Anthropic API 兼容性正向当前降低迁移摩擦和集成成本生产流量中有多少来自即插即用式兼容迁移?
多云与平台分发正向当前借助 Azure、AWS、Google、Alibaba、NVIDIA 和网关层扩大触达哪些渠道带来最多净新增付费用量?
企业支出审查混合当前利好可衡量价值和路由工具,而不只是原始模型质量DeepSeek 或其合作伙伴能否提供可观测性、策略和成本控制?
中国同业开放权重竞争负向当前Qwen、GLM、Kimi、MiniMax 等削弱定价权哪些工作负载仍有足够差异化,能支撑持久利润率?
平台货架化销售第三方模型负向当前让买家多栖成为常态,并削弱锁定效应DeepSeek 拥有终端客户关系,还是只是占据目录里的一个位置?
电力与算力基础设施约束混合当前至 24 个月用量上升拉动模型需求,但推理和供给经济性仍是战略变量哪些算力来源和国产替代能支撑扩张?
隐私与对华政策担忧负向当前至 36 个月即便性价比强,也可能拖慢国际企业采用哪些地区或受监管行业今天基本关闭?
出口管制与芯片地缘政治负向当前至 36 个月可能限制先进硬件获取,或放大政策波动未来模型质量多大程度依赖可能受限的硬件?
专用模型增长混合未来 24 个月垂直领域模型既可能通过合作伙伴扩大 DeepSeek 的 TAM,也可能压缩通用模型价值DeepSeek 应在哪些场景保持通用,哪些场景转向伙伴专用化?

DeepSeek 最大的驱动因素——足够便宜、也足够好的推理——也是这个市场价格压缩和 护城河侵蚀的主要来源。

[CM019, CM020, CM021, CM022, CM027, CM028]
FM004: 市场规模口径

DeepSeek 的机会从宏观 AI 需求延展到更窄的低成本生产推理赛道,受路由、治理和政策过滤塑形。

[CM001, CM007, CM019, CM027, CM031, CM032]

2.4 图表与论据

Chapter 03

03竞争格局

3.1 竞争集合:前沿实验室、中国同行和平台中介都重要

DeepSeek 的竞争版图必须按商业模式拆分,而不能只看基准分数。第一层是前沿实验室:OpenAI、Anthropic 和 Google 仍定义市场高端,尤其在企业更看重信任、分发广度和全栈平台功能,而不是最低价格时。第二层是中国开放权重和偏开放群体——Alibaba/Qwen、通过 Qianfan 分发的 Baidu/ERNIE、Z.ai/GLM、Moonshot/Kimi 和 MiniMax——共同模式是模型迭代快、API 兼容性高,并且很愿意打价格战。CNBC 2026 年 1 月对中国实验室的调查支持一个判断:DeepSeek 的突破加速了整个国内同行的发布周期,而不是为单家公司创造持久垄断。第三层是渠道层:Alibaba Cloud、AWS、Azure 和 Google Cloud 都通过决定哪些模型进入托管目录和企业工具来塑造买家访问。按这个意义,DeepSeek 最重要的一些「竞争对手」其实是路由环境,它们降低了买家的搜索和切换成本。最终市场比纯模型对比更宽,也比普通创业公司对抗在位者的框架更严酷。这也意味着,任何干净的排行榜都会漏掉一个事实:买家经常同时比较渠道、治理层、部署便利性和模型质量。[CP001, CP002, CP003, CP004, CP005, CP006]

竞争格局表
竞争对手类别有证据支撑的定位定价姿态分发备注对 DeepSeek 的意义
OpenAI美国前沿实验室高端通用与企业 AI 栈高端品牌和生态触达极广设定 DeepSeek 被拿来比较的高端参照系
Anthropic美国前沿实验室面向安全与企业的前沿模型高端企业和开发者采用强锚定高信任的高端层
Google Gemini前沿平台在位者广泛模型家族,加上开发者和云工具中端至高端与 Google 开发者和云触点集成靠分发广度和工具竞争
Alibaba / Qwen中国平台在位者Model Studio 上架 Qwen 和大量第三方模型灵活 / 目录驱动区域云覆盖大靠平台触达和模型广度竞争
Baidu / Qianfan中国平台在位者企业一站式模型与智能体开发平台平台主导企业云与搜索集成靠企业工作流集成竞争
Z.ai / GLM 模型中国模型挑战者发布节奏快,并主打长上下文编码能力未知至有竞争力开发者文档驱动分发显示长上下文和智能体上的功能追平压力
Moonshot / Kimi中国模型挑战者强调长上下文和智能体工作流有竞争力消费者品牌加 API 平台争夺编码、搜索和知识工作流量
MiniMax中国模型挑战者token 定价极低,并以 1M 上下文旗舰定位切入低成本开发者平台和消费者品牌从 DeepSeek 下方压缩价格伞

这张表把全球前沿实验室和中国同业放在一起,因为 DeepSeek 的可切换集合取决于工作负载和 地区,而不是一个干净的市场边界。

[CP001, CP002, CP003, CP004, CP005, CP006]
FP001: 竞争定位图

主要竞争对手在两个相关轴上的序位图:价格效率与企业触达。

轴得分是作者基于公开定价姿态、云 / 平台上架和开发者平台证据给出的序位估计,不是经审计的市场份额数据。

[CP009, CP010, CP023, CP024, CP027, CP028]

3.2 功能与价格位置:DeepSeek 价值领先,但同质化正在快速上升

DeepSeek 的公开材料和代码库说明了它为什么会成为如此有冲击力的对照组。DeepSeek-V2 强调经济训练和高效推理,DeepSeek-V3 扩展为 671B 参数 MoE、每个 token 激活 37B 参数,DeepSeek-R1 则建立了强势的公开推理叙事。这个组合重要,因为它满足了严肃竞争的买家下限:能力足够高、开放权重可信、成本极低。但相对同行的功能差距正在收窄。Anthropic 保持高端模型层级,Google 继续扩大 Gemini API 和工具表面,中国同行也已经把长上下文、多模态、智能体工具和兼容性当成基本配置来宣传。Z.ai 称 GLM-5.2 支持 1M 无损上下文;Kimi 主打 256k 上下文和智能体用例;MiniMax 推低单 token 价格;Alibaba 的模型工作室公开在一个目录里售卖多个竞争模型。含义是:DeepSeek 最适合被看作一个正在变拥挤的细分市场中的价值领导者。重视性价比的买家会继续测试 DeepSeek,但更在意广泛企业功能、政策舒适度或高端支持的买家仍有可信替代。换言之,功能趋同扩散得比独特定位更快。OpenAI 自己的模型文档也强调广泛多模态能力,而 Alibaba 定价页显示,平台在位者竞争靠分层折扣和打包,而不只是原始基准宣称。[CP013, CP014, CP015, CP016, CP017, CP018]

功能对比矩阵
厂商开放性 / 权重策略上下文信号兼容性 / 工具信号定价信号有证据支撑的差异化优势
DeepSeek开放权重发布 + APIV4 定价页标注 1M 上下文兼容 OpenAI / Anthropic 的 API公开标价极低推理口碑强,推理成本异常低
Anthropic闭源高端前沿上下文档位完整 Claude API + 工作台高端 API 档位企业信任与高端模型质量
OpenAI闭源产品触点广大型生态和商业工作区工具付费商业和 API 栈许多开发者和企业的默认参照
Google Gemini闭源广泛 API 模型家族Google 开发者生态和云集成公开 API 定价和分层覆盖 Google 多个触点的分发
Qwen / Alibaba开放与托管混合目录覆盖第一方和第三方模型兼容 OpenAI 和 Anthropic 的区域端点取决于目录 / 平台平台广度和区域云分发
GLM / Z.ai 模型偏开放 / 开发者主导GLM-5.2 声称支持 1M 上下文发布节奏快,并定位智能体抓取到的发布说明未重点强调长上下文和编码功能追平压力
Kimi偏开放 / API 主导K2.6 声称支持 256k 上下文支持工具调用和智能体任务公开定价有竞争力长上下文知识工作叙事强
MiniMax偏开放 / API 主导官网突出 1M 上下文开发者文档和模型目录公开按 token 定价低在中国市场形成激进价格伞

这张矩阵有意做比较而非穷尽;只纳入抓取到的公开材料中明确露出的能力。

[CP013, CP014, CP015, CP016, CP017, CP018]
FP002: 市场位置条形图

序位竞争力得分显示,在可比买方口径下,DeepSeek 的价值位置相对主要同行如何。

得分是序位和综合判断;它们汇总了价格、触达和功能姿态等已获取信号,并不声称实际市场份额。

[CP013, CP017, CP021, CP022, CP025, CP026]

3.3 护城河与切换动态:分发有帮助,但多供应商并用削弱锁定

DeepSeek 竞争耐久性最强的论点是,它已经跨过 AI 基础设施市场最难的门槛:它不只是基准出名,而是出现在开发者和企业真正购买模型的地方。Google Cloud 将 DeepSeek 记录为托管或自部署模型选项,Azure 在模型目录中列出 R1,AWS 也把 DeepSeek-R1 加入 Bedrock Marketplace 和 SageMaker JumpStart。这些渠道胜利意味着 DeepSeek 在正常企业采购路径中竞争,而不是只活在研究社区里。问题在于,同样的分发轨道也会压平护城河。Alibaba、Baidu、BigModel 和其他中国平台训练买家期待多模型访问、评估、路由和轻松替换。Artificial Analysis 也通过质量、价格、速度和开放性来定义竞争,而不是只看品牌。因此,DeepSeek 的防御性较少来自排他性,更多来自反复跑通一个困难但脆弱的公式:在推理和编码上贴近前沿,同时明显便宜于高端实验室,并且仍与低成本中国同行保持足够差异。这在今天是真实优势,但它是移动靶,而非永久护城河。因此,公司身处的市场里,分发是必要条件,却永远不充分。买家切换可以快得惊人。每天都在发生。[CP027, CP028, CP029, CP030, CP031, CP032]

分发与切换表
渠道或机制DeepSeek 位置竞争含义切换影响结论
Google Cloud 模型目录列为托管 API / 自部署选项与 Gemini 和第三方同业在同一企业采购界面内竞争降低采用摩擦,也降低独占性分发扩大触达,也把访问商品化
Azure AI Foundry 目录R1 上架模型目录把 DeepSeek 放在许多可替代模型旁边鼓励模型评测和替换提升认知多于形成锁定
AWS Bedrock / JumpStartR1 可通过市场和 JumpStart 使用让 AWS 客户在既有工作流里与在位模型对测让比较成为常规动作利于企业漏斗顶部试用
Alibaba Model Studio 目录DeepSeek 在区域多模型目录中竞争Qwen 和其他对手同列展示让买家多栖常态化区域平台力量与原始模型质量同样重要
开发者网关和排名网站流量份额会随性价比快速移动OpenRouter 和 Vercel 显示厂商之间的份额迁移路由可能逐月变化DeepSeek 必须持续守住用量
API 兼容性DeepSeek 对齐 OpenAI / Anthropic 格式许多对手也这样做,或提供迁移文档代码可移植性削弱锁定效应切换成本低于传统企业软件

只有当渠道存在能比比价购物侵蚀差异化更快地产生重复需求时,它才是竞争优势。

[CP027, CP028, CP029, CP030, CP031, CP032]
替代品与相邻竞争对手表
替代品或相邻力量示例集合买家考虑原因对 DeepSeek 的压力评估
高端前沿 APIOpenAI、Anthropic、Gemini信任、支持和生态深度更强可能把受监管或高风险工作负载从 DeepSeek 拉走DeepSeek 必须让质量足够接近,价格优势才会起作用
中国低成本模型同业Qwen、Kimi、MiniMax、GLM买家同样价格敏感,发布节奏也快压缩价格伞和差异化空间这是 DeepSeek 最难打的日常仗
云端模型目录AWS、Azure、Google Cloud、Alibaba 目录让现有采购路径里的比较更容易把 DeepSeek 变成众多选项之一有利于触达,不利于独占
开发者网关与路由层Vercel AI Gateway、OpenRouter把流量优化到当下表现最好的模型相对价值变化时,能迅速改派用量短周期性能提升会直接受益
内部或微调后的企业栈基于第三方底座自有调优模型降低对单一供应商的依赖压缩通用 API 可触达的经常性支出DeepSeek 必须始终是最经济的底座选择

这些力量重要,因为许多 DeepSeek 评估本质上是在决定「用 DeepSeek 还是路由到别处」, 而不是干净的一对一供应商替换。

[CP001, CP011, CP026, CP031, CP032, CP035]
FP003: 竞争护城河逻辑

DeepSeek 的竞争优势取决于在多模型渠道里持续更强、更便宜、也足够可见;这些渠道既帮它分发,也在商品化它。

[CP029, CP030, CP031, CP032, CP033, CP038]

3.4 图表与论据

Chapter 04

04财务情况

4.1 收入模式与定价:即便 P&L 不透明,变现轨道也清晰

DeepSeek 披露的信息足以理解资金应如何流入,尽管它没有披露最终收入。API 文档称计费基于 token,区分缓存与未缓存输入,并拆分输入和输出定价。这很重要,因为它让变现与用量绑定,而非与订阅绑定:收入会随推理量、模型组合、输出强度和缓存行为扩张。同一定价页显示两条主要 V4 路径——Flash 和 Pro——单价与并发上限差异明显,暗示公司有意把低成本、高吞吐场景和更高要求工作负载分层。DeepSeek 的使用条款还表明,费用会从预付或赠送余额中扣除,公司保留调价权。这些机制更像现代云 API 业务,而不是消费者订阅创业公司。除直接 API 收入外,DeepSeek 通过 Google Cloud、Azure、AWS 以及 Vercel 等网关获得托管可用性,暗示还可通过渠道流量和企业路由变现。未知的是收入结构:公开来源没有说明直接 API 支出、间接云渠道,或免费消费者使用后来转化为付费需求,各自贡献多少。官方变更日志还显示,模型名称和受支持接口正在被主动下线和替换,这让版本管理成为商业系统的一部分,而不只是技术细节。[CI001, CI002, CI003, CI004, CI005, CI006]

收入模式表
收入来源定价基础买方证据财务含义
直接 API 推理按百万 token 计费,输入和输出分别定价开发者和产品团队DeepSeek 定价和 token 用量文档收入随用量规模和模型组合直接变化
缓存推理 / 重复使用缓存输入价格更低重复或已优化工作负载DeepSeek 定价文档缓存命中率会显著改变实际单位经济性
高端模型层级用量Pro 模型价格更高,并发上限更低价值更高的推理或智能体工作负载DeepSeek 定价文档组合向 Pro 倾斜可抬高每 token 收入
云 / 目录分发用量经托管云入口路由企业和平台买方Google、Azure、AWS 上架记录可扩大企业触达,但可能稀释直接利润率
消费者 / 自助漏斗免费或低成本使用,为后续付费需求播种终端用户和小团队采用度代理指标和官方入口品牌带来的流量后续可能转成 API 变现,但公开估值很难

DeepSeek 未披露分部收入,因此本表刻画变现通道, 而非各收入流的已报告收入贡献。

[CI001, CI002, CI003, CI004, CI005, CI006]
定价与单位经济杠杆表
杠杆公开证据为何重要对经济性的方向性影响仍未知的事项
输入 token 量计费基于输入和输出 token 数大提示词会同时推高收入和计算成本用量更高会增加总账单额,也会推高推理成本每 token 实际毛利率
缓存输入与未缓存输入缓存输入价格远低于未缓存价格优化会降低重复工作流的实际收入高缓存命中率可能压缩收入,但提升工作负载效率实际缓存命中结构
输出 token 强度V4 层级的输出价格高于输入价格如果计算成本守得住,重推理输出会很赚钱长输出会同时抬高账单和成本暴露按客户类型划分的平均输出长度
模型组合:Flash vs ProPro 价格更高,并发上限更低高端工作负载占比提升可改善收入质量Pro 用量增加可能提升单客户变现各层级流量占比
渠道组合直接 API 对比云或网关路由间接渠道可能用利润率换触达渠道更丰富的组合可能加速增长,但降低抽成率扣除伙伴经济后的净收入

本表强调从公开定价机制能看到的单位经济杠杆; 不声称实际已实现利润率或贡献经济性。

[CI002, CI003, CI004, CI005, CI007, CI025]
需求与渠道代理指标表
代理指标已报告信号为何对财务重要局限
Vercel AI GatewayDeepSeek token 份额一个月内从不到 1% 跃升至 17%,支出仍接近 1%显示用量增长强劲,但相对高端供应商变现密度低Gateway 样本不是公司完整收入
OpenRouterDeepSeek token 份额从 1 月到 2026 年 6 月初由 9% 翻倍至 18%表明开发者和智能体路由需求上升Token 份额不等于利润率或企业合同价值
Google Cloud 上架托管 API 和自部署上架扩大企业变现入口不披露价格、抽成率或用量
Azure 模型目录R1 上架 Azure AI Foundry增加企业发现和采购入口出现在目录里不等于付费用量
AWS Bedrock / JumpStartR1 可通过 AWS 渠道使用新增一条企业漏斗顶部变现路径未披露公开收入分成

这些是收入邻近代理指标,不是已报告财务报表; 它们帮助判断需求形态,但不能判断盈利能力。

[CI009, CI010, CI011, CI029, CI030]
FI001: 收入模型逻辑

DeepSeek 主要靠按 token 计费的推理变现,经济性受层级组合、缓存和渠道路由影响。

[CI001, CI002, CI003, CI004, CI006, CI007]

4.2 资本结构、融资历史与现金跑道:看得到的钱多于经营表现

资本故事比经营报表更容易观察。Forbes 和 Fortune 都把 DeepSeek 早期融资能力与创始人 Liang Wenfeng 及 High-Flyer 创造的财富相连,这意味着公司起步时似乎拥有对这一规模 AI 实验室来说异常充足的创始人资金。但到 2026 年中,DeepSeek 已明显从创始人单独融资转向外部资本。TechCrunch 5 月报道称,DeepSeek 首轮投资估值可能达到 $45 billion;CNBC 6 月报道称首轮外部融资已以超过 $50 billion 的估值完成;TechCrunch 7 月又报道称,公司在据称仅一个月前融资 $7 billion 后,正在探索以约 $71 billion 估值再融约 $1.5 billion。CB Insights 补充了更多但部分冲突的元数据,将 DeepSeek 列为 Series A 阶段公司,累计融资 $7.546 billion。共同结论是:如果这些报道方向正确,DeepSeek 短期不再受资本约束。更难的问题是现金跑道。公司以低价发布前沿规模模型;这一经营披露缺口,是判断新报道资本基础到底充裕还是仅仅必要的最大障碍。[CI013, CI014, CI015, CI016, CI017, CI018]

融资历史表
日期 / 时段事件已报告金额已报告估值状态来源说明
2023 年启动期创始人出资搭建初创公司,资金能力与 Liang Wenfeng / High-Flyer 收益相关未披露创始人控制据报道Forbes 和 Fortune 将 DeepSeek 早期资金能力与 High-Flyer 财富相联系
May 2026媒体报道首次外部融资轮在讨论未披露~$45B据报道在洽谈TechCrunch 引述 FT 和 Bloomberg 报道
June 2026据报道首次外部融资轮已完成>$7B>$50B据报道已完成CNBC 称首次外部融资轮完成时估值超过 $50B
July 2026据报道探索后续融资和 IPO 准备~$1.5B~$71B据报道在洽谈TechCrunch 引述 Bloomberg 关于新资金和 IPO 时间的报道
CB Insights 档案私营公司元数据快照累计融资 $7.546B估值隐藏冲突数据集CB Insights 将 DeepSeek 列为 Series A,累计融资 $7.546B

公开报道混合了已完成轮次、据报道的洽谈和数据集快照; 投资者应把资本记录视为方向强劲,但尚未完全对齐。

[CI013, CI014, CI015, CI016, CI017, CI018]
披露可见度表
财务主题公开可见度所审最佳证据承销含义
定价机制DeepSeek API 文档和条款变现通道清晰
收入 / ARRNone未看到公开披露无法用基本面锚定估值
毛利率None未看到公开披露无法判断低价能否持久
现金 / runwayNone未看到公开披露无法干净评估融资是否足够
月度 burnNone未看到公开披露无法区分健康投入和压力
融资额部分可见但信号强TechCrunch、CNBC、CB Insights 均报道了大额资本事件资本获取能力看起来强,但仍需对齐核验

可见度仅基于截至运行日期已抓取并审阅的公开来源集评估。

[CI017, CI021, CI027, CI028, CI037, CI038]
FI002: 融资时间线

DeepSeek 在 2026 年从创始人自有资金支持,走向大额外部融资报道。

多个时间线条目是融资事件报道,不是公司确认的文件,因此该序列信号强,但还不是完全对齐的交易历史。

[CI013, CI015, CI016, CI017, CI018, CI021]

4.3 单位经济与财务风险:DeepSeek 可能很高效,但效率不等于耐久性

DeepSeek 的财务上行来自同一组事实,而这些事实也制造主要风险。Vercel 和 OpenRouter 暗示,低标价和快速采用可以带来巨大的 token 量;但低价也留下更少空间来吸收算力冲击、渠道费用或激进竞争性折扣。DeepSeek 的公开材料和合作伙伴列表显示,它可以通过直接 API 和云中介变现;但这些渠道也给客户更多路由灵活性,让收入黏性下降。成本侧更难判断。CSIS、CNBC 和 FDD 都指向一些会在财务上产生影响的风险因素,即便它们常被描述为政策或 IP 故事:出口管制会影响硬件获取,所谓蒸馏或 IP 纠纷会抬高法律和声誉成本,禁令或安全担忧会关闭高信任客户细分。CNBC 6 月关于融资中「禁止挖角」条件的报道也提示,人才稀缺是经济变量,不只是 HR 问题。按模型性能看,DeepSeek 确实可能比许多同行更高效;但投资者不应把技术效率误当成已验证的现金生成耐久性。没有收入披露、流失数据或烧钱数据,基于公开信息的最佳判断是:DeepSeek 拥有可见的变现引擎、充足新资本,以及仍未证明的将规模转化为可防御长期经济性的能力。[CI025, CI026, CI027, CI028, CI029, CI030]

财务风险表
风险证据财务路径严重性缓释或抵消因素投资者尽调问题
激进定价DeepSeek 公开压低相对高端实验室的价格如果计算成本仍高,高用量也可能只带来薄利润率规模和高效推理各模型层级的毛利率是多少?
硬件 / 出口管制暴露政策和芯片限制仍是活跃议题可能推高 capex / 推理成本,或拖慢模型改进国产替代和资本获取下一代模型背后的算力来源是什么?
渠道依赖云和网关伙伴扩大分发间接渠道可能拿走利润率并掌握客户关系更广触达和企业信任收入中直接渠道与伙伴路由各占多少?
IP / 蒸馏指控OpenAI 和 Anthropic 已公开提示中国蒸馏活动法律或声誉成本可能损害客户信任或变现所审来源未证明责任成立DeepSeek 是否为法律或有事项计提准备?
人才留存据报道的互不挖角条款凸显核心研究员稀缺薪酬压力会推高 burn 和执行风险新资本可支撑薪酬年化 R&D 薪酬和流失率是多少?
披露不透明未看到公开且经审计的收入、burn 或 runway 指标估值承销和现金耐久性更难判断据报道的大额融资缓冲近期不确定性提供经审计收入、现金和月度 burn

这些是面向投资者的财务风险,不是完整法律或运营风险清单。 第 7 章展开非财务风险图谱。

[CI027, CI028, CI029, CI030, CI031, CI032]
FI003: 财务压力驱动因素

最大的财务问题不在需求,而在 DeepSeek 能否在保持低价的同时守住利润率、拿到算力并留住人才。

得分是基于公开证据的序位风险权重,不是模型化概率或量化下行情景。

[CI027, CI028, CI029, CI030, CI031, CI032]

4.4 图表与论据

Chapter 05

05产品与技术

5.1 产品表面:DeepSeek 已成为平台,而不只是一次模型发布

已审阅的产品证据显示,DeepSeek 是一个拥有多个用户入口的栈。DeepSeek 公开网站和透明度页面描述的是一连串重大模型发布,而不是单一静态旗舰。API 文档展示了与 OpenAI 和 Anthropic 兼容的访问、模型特定定价、token 用量机制,以及 JSON 输出、工具调用、上下文管理等产品化功能。V2、V3 和 R1 的 GitHub 与 Hugging Face 页面进一步表明,DeepSeek 把开放权重分发视为产品策略的一部分,而不只是研究副产品。这很重要,因为它扩大了可服务用户群:有些客户要直接托管推理,有些要托管云访问,有些要权重或论文来评估或自部署。AWS、Azure、Google Cloud 和 NVIDIA 的合作伙伴来源显示,DeepSeek 的产品表面也延伸到第三方分发层,企业用户在那里发现并测试模型。结果是一套产品架构:多扇门通向同一核心能力基础。技术上,这是优势,因为它增加采用路径;运营上,它制造更多需要维护的合约,涉及兼容性、性能、可靠性和文档。这也意味着,产品尽调不能只看原始模型质量,还必须看包装和生态行为。隐私政策也说明,DeepSeek 运营的是持续软件服务,而不只是发布权重;英文透明度页面则像公开发布台账,包含模型卡和报告链接。[CE001, CE002, CE003, CE004, CE005, CE006]

产品模块 / 资产矩阵
模块 / 资产主要用户有证据支持的能力分发入口战略角色
DeepSeek 网页 / 聊天入口终端用户交互式模型访问和探索DeepSeek 网站认知和自助使用漏斗
API 平台开发者和产品团队通过兼容 OpenAI 和 Anthropic 的格式提供托管推理DeepSeek API 文档核心变现和集成入口
开放权重仓库研究人员、开发者、自托管用户V2、V3、R1 的模型仓库和文档GitHub 和 Hugging Face可信度、实验和生态触达
透明度中心研究人员、评估者、伙伴发布时间线,以及模型卡 / 技术报告指引DeepSeek 透明度页面公开产品台账和信任辅助
云 / MaaS 上架企业平台买方托管 API 或目录访问AWS、Azure、Google Cloud、NVIDIA 渠道企业触达和分发杠杆

矩阵捕捉所审公开材料中直接可见的用户侧产品资产; 不推断内部工具或未披露企业模块。

[CE001, CE002, CE003, CE004, CE005, CE041]
工作流 / 用例表
工作流有证据支持的功能可能用户DeepSeek 为何适配需关注的约束
通用聊天与推理R1 推理家族加 V4 托管层级普通用户和分析师推理口碑强,成本低安全 / 政策约束
编码和智能体任务工具调用、JSON 输出,以及伙伴 / 网关材料中的智能体定位开发者和 AI 产品团队兼容性和结构化输出支持长周期任务可靠性
长上下文文档或知识工作V4 支持 1M 上下文,同行也在比拼 256k–1M 上下文知识工作者和企业团队大上下文经济性可能有吸引力延迟和检索质量
自托管或评估工作流开放权重仓库和模型卡研究人员和基础设施团队低摩擦实验和基准测试自托管用户的运营复杂性
托管企业测试云目录和 MaaS 上架企业 AI 平台负责人易嵌入现有采购和治理轨道渠道依赖和差异化有限

用例来自已发布产品资产和伙伴分发入口, 而非私人客户披露。

[CE006, CE007, CE008, CE009, CE010]
FE001: 产品架构图

DeepSeek 的产品栈从模型家族延伸到托管 API、开放权重仓库和合作伙伴托管的企业渠道。

[CE001, CE002, CE003, CE004, CE011, CE012]
FE002: 客户工作流 / 运营流程

DeepSeek 典型用户旅程从认知和试用,走向集成、路由和持续适配发布。

[CE006, CE007, CE008, CE009, CE019, CE028]

5.2 架构与研发:效率、MoE 设计和推理专精是定义性主题

DeepSeek 的技术身份在公开来源中异常清晰。V2 被明确定位为经济高效;V3 被呈现为大型 MoE 架构,拥有 671B 总参数、每个 token 激活 37B;R1 则被定位为第一代推理家族。R1 仓库强调由强化学习驱动的推理,V2 和 V3 材料则强调低成本训练和高效推理。透明度页面补上了发布节奏,让外界能把这套系统视为演进中的平台,V4 在 2026 年 4 月加入这条弧线。定价文档增加了面向生产的约束和能力,例如上下文长度、输出限制和并发上限。Z.ai、MiniMax 和 Kimi 来源是有用对照,因为它们展示了 DeepSeek 身处的技术竞赛:长上下文、编码、智能体任务和多模态工作流都在常态化。因此,DeepSeek 的技术挑战不只是再发布一个有能力的模型。它必须守住成本性能优势,同时保持 API 稳定,支持托管和伙伴分发部署,并让推理质量在生产中可用,而不仅仅是在演示中惊艳。现有证据支持一个判断:DeepSeek 的 R&D 引擎速度快、架构意识强;但内部测试或基础设施细节披露不足,无法完全验证规模化鲁棒性。持续技术优势取决于执行,而不只是架构口号。[CE013, CE014, CE015, CE016, CE017, CE018]

技术 / 运营架构表
层级公开证据角色技术含义未知事项
底座模型架构V2/V3 仓库和论文核心模型能力和效率MoE 设计是规模和成本叙事的核心实际训练和推理集群经济性
推理层R1 仓库和文档改善推理行为和智能体适配度RL 驱动的推理成为产品差异化点生产环境失败模式和护栏
托管推理 APIAPI 文档和定价页面运营交付入口稳定端点和功能兼容性很关键观察到的 SLA 和客户间质量差异
伙伴托管打包AWS、Azure、Google、NVIDIA企业访问层部署抽象层扩大触达面共同工程深度和支持义务
文档 / 透明度层透明度中心、状态页、条款信任、发布管理和开发者上手快速发版需要文档纪律内部 QA 和发布治理

这是一张外部可观察的架构图,不是内部系统图;它强调公开证据中可见的产品运营层。

[CE013, CE014, CE015, CE016, CE017, CE018]
信任 / 质量 / 合规表
主题证据技术上为什么重要当前判断缺口
API 兼容性兼容 OpenAI/Anthropic 的文档降低集成成本,加快迁移端点演进时需要回归纪律
发布透明度透明度页面链接模型卡和报告帮助评估者追踪变更内容和时间中到强不能替代完整安全披露
服务状态可见性有状态页体现生产运营思维未查阅到公开运行时间历史
条款和命名弃用通知定价页注明旧名称将弃用说明发布管理在持续推进沟通不到位可能打断集成
安全 / 滥用担忧政策和安全评论仍偏负面关系到企业信任和抗滥用能力正负混合未查阅到完整公开红队披露

这张表把积极的运营信号和未解的技术信任缺口放在一起,因为企业采用取决于两者。

[CE019, CE020, CE027, CE028, CE029, CE030]
FE003: 关键依赖图

DeepSeek 的产品质量取决于模型研发、API 稳定性、合作伙伴包装以及算力 / 政策条件之间的互动。

[CE013, CE014, CE015, CE016, CE033, CE034]

5.3 技术风险与路线图:加速真实存在,安全、集成和依赖风险也同样存在

让 DeepSeek 激动人心的同样速度,也制造技术风险。公开文档描述了强大的推理和智能体导向行为,但对抗性与政策评论显示,市场持续担心安全控制、滥用风险,以及发布开放或容易访问的高能力模型带来的更广泛影响。DeepSeek 的条款、状态页和合作伙伴列表显示,公司运营的是真实服务业务,因此可靠性和文档纪律与研究新颖性同样重要。从透明度页面和合作伙伴公告可见的路线图显示,这家公司在发布、渠道和兼容层之间快速推进。这利好采用,但也提高了运营回归、文档漂移或破坏生态的废弃变更概率;定价页关于旧模型名称退役的通知已经给出暗示。关键依赖仍然重要。云列表、网关集成以及 NIM 或 MaaS 分发都意味着,公司部分触达和企业包装依赖外部渠道;出口管制和硬件政策叙事也提醒投资者,模型进步仍依赖算力获取。综合看,DeepSeek 的产品和技术姿态先进且具有商业相关性,但仍带有快速移动的前沿模型平台固有的脆弱性。速度与控制之间的平衡,就是核心产品技术张力。[CE027, CE028, CE029, CE030, CE031, CE032]

路线图 / 发布 / 开发阶段表
发布 / 产物日期有证据支持的状态技术意义阶段判断
DeepSeek-V22024-05代码库和论文已发布经济高效的 MoE 基线已发布 / 历史
DeepSeek-V32024-12代码库和论文已发布大规模 MoE 扩展与效率证明已发布 / 历史
DeepSeek-R12025-01代码库已发布以推理为核心的产品线已发布 / 历史
主要云上的 DeepSeek-R12025-01AWS 和 Azure 公告面向企业试用的伙伴封装已分发 / 扩展中
DeepSeek-V42026-04-24透明度页面已发布,附模型卡 / 报告链接当前旗舰发布主线当前 / 活跃
Vercel AI Gateway 上的 DeepSeek V42026-07网关更新日志扩大智能体开发者分发当前 / 活跃

路线图表使用公开可见的发布和分发里程碑,不是保密的前瞻产品路线图。

[CE021, CE022, CE023, CE024, CE025, CE031]
FE004: 产品成熟度 / 能力图

DeepSeek 在核心模型交付和 API 兼容性上显得成熟,但公开信任可见度和长期运营透明度只属中等。

单元格是作者基于审阅过的公开产品材料和反向评论作出的序位判断,不是基准测试得分或 SLA 承诺。

[CE020, CE027, CE029, CE036, CE037, CE038]

5.4 图表与论据

Chapter 06

06客户情况

6.1 客户细分:自助用户、开发者、企业平台团队和中介都重要

DeepSeek 的公开客户基础必须从产品表面和采用证据中推断,而不是来自一份干净的企业客户名单。第一类是通过聊天、移动应用或病毒式模型发布发现 DeepSeek 的自助用户。Sensor Tower 和 Appfigures 显示,2025 年初这类受众形成得非常快。第二类是通过直接 API 或兼容路由层使用 DeepSeek 的开发者和 AI 原生产品团队。第三类是企业平台团队,他们在 AWS、Azure、Google Cloud 等云目录或托管 AI 平台中遇到 DeepSeek。第四类是中介——网关、云和平台运营商——它们实际上成为客户或渠道客户,因为它们把第三方需求路由给 DeepSeek。这种分层重要,因为各细分的采用质量不同。消费者下载证明知名度;网关 token 份额证明生产试验;云目录存在证明采购可达性。单看任何一个都不能证明黏性的企业账户,但合在一起说明 DeepSeek 的客户足迹比一次病毒式消费者时刻更宽。因此,公司客户故事的强项是广度,弱项是直接披露。[CU001, CU002, CU003, CU004, CU005, CU006]

客户分层表
细分客群主要用户有证据支持的接入路径证据证明什么主要未知
自助消费者 / 专业消费者个人聊天界面、App、网站知名度和漏斗顶端快速采用向付费、持久使用的转化
开发者 / AI 原生团队开发者和产品团队直接 API、开放权重仓库、网关技术试用和集成兴趣账号级留存和支出深度
企业 AI 平台团队集中式 IT / AI 负责人AWS、Azure、Google Cloud、NVIDIA MaaS 入口采购可达性和评估路径赢单率和生产规模
网关 / 平台中介云、网关、目录Vercel、OpenRouter、云目录下游路由需求和分发杠杆利润分成和渠道依赖
研究 / 自托管评估者研究人员和基础设施团队GitHub、Hugging Face、模型卡开放权重可信度和实验评估后的商业转化

这些细分客群来自本报告审阅的公开产品和分发入口;DeepSeek 未披露正式客户分层。

[CU001, CU002, CU003, CU004, CU005]
具名客户证明表
证明点来源类型证明什么客户 / 渠道判断强度
AWS Bedrock Marketplace 和 SageMaker JumpStart 上架伙伴证明DeepSeek 已面向企业云买家完成封装企业分发证明
Azure AI Foundry 目录上架伙伴证明DeepSeek 出现在 Microsoft 模型目录工作流中企业分发证明
Google Cloud 托管 API / 自部署上架伙伴证明DeepSeek 可在 Google 企业智能体平台内采用企业分发证明
NVIDIA NIM 封装伙伴证明DeepSeek 正在面向部署生态落地基础设施生态证明
Vercel AI Gateway 支持客户证明DeepSeek 可用于下游客户使用的生产路由环境开发者 / 生产证明
OpenRouter token 份额报告客户证明开发者网关上的真实路由使用量在上升重复使用证明

这些是采用环境和路由使用的证明点,不等同于已披露的签约终端客户名单。

[CU006, CU007, CU008, CU009, CU019, CU020]
FU001: 客户旅程图

DeepSeek 的典型路径从发现开始,经过试用、路由,最后进入持续的工作负载选择。

[CU001, CU010, CU011, CU023, CU026]
FU003: 客户证明矩阵

DeepSeek 的公开客户证明在渠道和路由环境最强,在具名终端客户披露上较弱。

单元格是基于各类来源能证明客户质量到什么程度作出的序位判断,不是量化得分。

[CU006, CU007, CU008, CU019, CU020, CU027]

6.2 采用与扩张:公开用量代理指标指向真实需求和扩大的部署路径

已审阅材料中最强的客户证据来自衡量真实行为的采用代理指标。Sensor Tower 称,DeepSeek 在最初 19 天获得约 23 million 全球下载,是 ChatGPT 可比发布窗口的两倍多;爆发期内平均移动 DAU 环比增长超过 700%。Appfigures 也称,该应用很快突破 1 million 下载,并即将挑战 ChatGPT。这些移动信号是漏斗顶部,不是持久收入证明,但它们确立了异常快的获客速度。更新近的证据来自基础设施渠道。Vercel 2026 年 6 月生产指数称,DeepSeek 的路由 token 份额一个月内从低于 1% 跃升至 17%,而支出仍接近 1%;OpenRouter 称,2026 年上半年 DeepSeek 的 token 份额从 9% 翻倍至 18%,智能体工作负载贡献了大部分增长。这些信号重要,因为它们把故事从炒作推进到反复生产路由。同时,AWS、Azure、Google Cloud、NVIDIA 和 Vercel 都作为具名证据点,显示 DeepSeek 正在被打包给主流开发者和企业环境中的下游用户。因此,公开可见的扩张最可信地体现为渠道扩张和工作流扩张,而不是已披露的具名账户扩张。这个区别重要,因为路由用量比单纯应用榜单动能更接近可变现客户行为。[CU013, CU014, CU015, CU016, CU017, CU018]

客户增长 / 采用轨迹表
时期 / 信号指标报道值指向什么局限
App 上线后前 19 天全球 App 下载量>23M漏斗顶端消费者获客异常强劲下载量不等于留存活跃用户
App 上线后前 19 天美国下载量~2M虽然后续出现政策担忧,美国市场仍迅速起量仅是短窗口指标
2025 年 1/22–1/28,较前一周移动 App 平均 DAU 增长>700%爆发式用户增长爆发增长未必持续
同一爆发期网站访问量周环比 +650%桌面端发现量随 App 采用一起激增仅是短窗口指标
2026 年 6 月 Vercel AI Gateway路由 token 份额从不足 1% 升至 17%生产路由采用已有实质规模网关份额不等于公司直接收入
2026 年 1–6 月 OpenRoutertoken 份额9% 到 18%2026 上半年重复使用增加仅是特定平台样本

这张轨迹表有意混合消费者和生产代理指标,因为 DeepSeek 未披露统一的客户增长指标。

[CU013, CU014, CU015, CU016, CU017, CU018]
留存 / 重复使用 / 满意度表
维度最佳公开证据判断为什么重要当前缺口
重复路由使用OpenRouter token 份额在 2026 上半年翻倍正向代理指标表明工作负载在持续选择它,而不是一次性猎奇没有队列或 logo 级留存
生产路由持续性Vercel token 份额在 2026 年 6 月大幅上升正向但还早意味着生产实验正在转化为流量没有支出留存或账号扩张数据
移动端重复使用Sensor Tower 报道上线后 DAU 增长强劲正向但呈爆发式证实爆发期用户很快回访未披露长期 App 留存
客户满意度 / NPS未查阅到公开 NPS 或 CSATUnknown对粘性账号扩张很重要未查阅到公开调查或客户访谈
企业重复购买云 / 网关重复上架且持续可用中度正向表明渠道看到足够需求,愿意保持 DeepSeek 在线没有具名续约数据或企业客户背书

这张表使用代理信号,因为 DeepSeek 不披露传统 SaaS 式留存或满意度指标。

[CU023, CU024, CU026, CU027, CU028]
FU002: 采用 / 部署漏斗

DeepSeek 的公开客户漏斗从广泛认知收窄到较小的生产路由核心。

这些阶段有意混合数量、增长率和平台份额,因为 DeepSeek 不发布统一的获客到留存漏斗;该图是从广泛认知走向生产路由的代理路径。

[CU013, CU014, CU015, CU016, CU017, CU018]

6.3 留存、集中度与满意度:最大的客户未知仍藏在水面下

客户质量是公开证据最薄的地方。已审阅来源没有披露净收入留存、logo 留存、队列留存、NPS,甚至没有给出可对账的付费企业客户数。这迫使我们区分可见需求和可核保的客户耐久性。网关和云信号暗示重复使用,因为没有一些经常性工作负载,token 份额不会上升;但它们没有说明需求是否集中在少数超级用户或中介手中。集中度风险真实存在,原因有两点。第一,DeepSeek 大量可见企业触达经由第三方平台实现,这意味着云和网关可能成为关键渠道。第二,TechCrunch、CNA 和 Al Jazeera 通过禁令和限制记录显示,监管和隐私担忧已经把一些公共部门和高信任细分从可触达客户池中移除。实际含义是,DeepSeek 很可能拥有宽广且快速增长的客户表面,但还没有公开可读的客户质量画像。正确的尽调问题不再是「人们是否在用 DeepSeek?」公开证据说是。更难的问题是,用量是否分散、留存、在账户内扩张,并且能以一种抵抗路由灵活性和政策摩擦的方式变现。公开证据仍未达到真正账户分析的层级。[CU026, CU027, CU028, CU029, CU030, CU031]

扩张和集中风险表
风险证据客户层面影响严重性尽调问题
渠道集中许多可见企业证明来自云和网关关键关系可能掌握在中介手里,而不是 DeepSeek 直接掌握付费使用中,直接使用与伙伴路由各占多少?
监管排除一些政府和机构已在部分场景禁止或限制 DeepSeek压缩可触达的高信任客群哪些地区或垂直行业已实质关闭?
付费客户数未知未查阅到公开企业付费客户数难以判断收入或客户 logo 多元化各细分客群有多少活跃付费账号?
留存 / NRR 未知未查阅到公开留存指标扩张质量未验证各细分客群的 GRR/NRR 和扩张率是多少?
消费者向企业转化不确定消费者下载不保证企业变现漏斗顶端可能夸大持久经济性自助使用中有多少比例转为付费 API 使用?
隐私 / 信任担忧多起禁令和安全报道仍在发酵可能拖慢受监管或公共部门客户采购输单中有多大比例出现信任担忧?

这张表聚焦客户质量风险,不覆盖第 7 章里的更广泛公司风险清单。

[CU029, CU030, CU031, CU032, CU033, CU034]
FU004: 留存 / 复用队列代理

由于 DeepSeek 没有披露正式留存指标,该队列是估算代理,显示不同渠道的复用耐久性可能如何分化。

这些百分比是作者估计,锚定 Sensor Tower、Vercel 和 OpenRouter 信号所暗示的相对持续性。它们不是公司披露的队列,只能作为持久性判断线索解读。

[CU023, CU024, CU025, CU026, CU037]

6.4 图表与论据

Chapter 07

07风险

7.1 法律与监管风险:隐私、司法辖区、禁令和 IP 纠纷站在最前线

DeepSeek 的法律与监管风险在公开材料中异常清晰。其隐私政策称,该政策适用于 DeepSeek 应用、网站、软件及相关服务,并将 Hangzhou DeepSeek Artificial Intelligence Co., Ltd. 认定为控制者,还称个人数据会在中国境内被直接收集、处理和存储,以提供服务。其使用条款称,随着法律、法规或技术演进,服务可能变化、因司法辖区而异,或被暂停、终止。任何全球 AI 创业公司都必须重视这些事实,但在这里分量更重,因为多个公开来源记录了与隐私和国家安全担忧相关的政府限制。TechCrunch、CNA、Al Jazeera 和 Conference Board 都描述了国家、机构或州及联邦主体的禁令或限制。在此之上,CNBC 和 FDD 总结了围绕模型蒸馏或 IP 滥用的指控,CNIPA 的商标通知则显示 DeepSeek 品牌本身也成了机会主义注册目标。整体法律/监管判断不只是「中国暴露」。它是一组相互叠加的问题:跨境信任、公共部门排除风险、演进中的 AI 治理,以及法律叙事可能比技术反驳传播更快。这组问题可以同时影响采购和估值。[CR001, CR002, CR003, CR004, CR005, CR006]

监管 / 法律风险登记表
风险证据为什么重要严重性观察点
PRC 数据存储和控制者管辖权隐私政策称数据直接在 PRC 收集、处理和存储可能劝退跨境和受监管买家是否出现特定监管机构行动或采购排除
服务可用性随司法辖区而变条款称服务可能在某些司法辖区不可用带来区域客户和合规不确定性新的特定地区限制
政府和机构禁令TechCrunch、CNA、Al Jazeera、Conference Board 记录了禁令或限制压缩公共部门和高信任需求池禁令扩展到盟友或企业场景
IP / 蒸馏指控CNBC 和 FDD 概述了美国 AI 公司提出的指控可能带来法律、声誉或采购摩擦正式投诉、诉讼或更强公开证据
商标 / 品牌滥用CNIPA 驳回 63 件「DEEPSEEK」商标申请显示该品牌面临品牌保护和仿冒压力升级为所有权争议或高成本维权
AI 治理变化风险条款明确预设法律演进时服务会变化规则可能比产品路线图变得更快AI、数据或出口管制规则的重大变化

这份登记表聚焦外部可见的法律和监管风险;这些风险可能损害客户触达、声誉或平台连续性。

[CR001, CR002, CR003, CR004, CR005, CR006]
FR001: 风险热力图

DeepSeek 优先级最高的风险集中在「高可能性 / 高严重性」象限,政策、信任与平台依赖在这里叠加。

位置是作者基于已审阅来源作出的序位判断,目的是给尽调排优先级,不是预测精确概率。

[CR005, CR010, CR012, CR023, CR029, CR033]

7.2 运营与平台风险:安全、可靠性和分发依赖紧密耦合

DeepSeek 的运营风险和产品分发、市场推广方式绑在一起。CSIS 的负面评论强调滥用和越狱担忧;DeepSeek 自己的条款也提醒用户,不要把输出当成专业意见,并说明输出可能包含错误或遗漏;隐私政策则写明,使用下游应用的开发者并不完全落在 DeepSeek 直接隐私范围内。这些问题重要,是因为 DeepSeek 已经不只是发论文;它在运营在线服务、合作伙伴打包产品和经网关路由的工作负载。因此,可靠性、文档准确性、兼容性稳定性都成了风险面的一部分。公开状态页是正面信号,但不能替代事故历史或 SLA 记录。分发也有两面性。AWS、Azure、Google Cloud、NVIDIA 和 Vercel 都扩大了企业触达;但每多一个入口,就多一层对外部打包、治理或定价逻辑的依赖。Azure 的 V4-Pro 目录页凸显了托管支持和统一计费的好处,也说明企业运营语境有多大一部分可能掌握在平台而非模型实验室手里。结果是,模型质量、抗滥用能力、服务运营和渠道依赖彼此叠加。[CR015, CR016, CR017, CR018, CR019, CR020]

运营 / 质量 / 安全风险登记表
风险证据技术路径严重性缓释措施
越狱 / 滥用暴露CSIS 强调护栏有限和滥用担忧可能引发声誉、客户信任和政策后果更可见的安全治理和红队测试
输出不准确条款称输出可能包含错误或遗漏,且不构成专业建议可能降低企业信任,并引发下游责任担忧人工审核和产品标注
下游 App 的隐私范围复杂性隐私政策称下游开发者应用不在部分政策范围内让终端用户和企业买家难以分清边界更清晰的伙伴 / 开发者义务
服务可靠性不透明有状态页,但未看到公开 SLA 或事故历史难以评估生产环境耐用性运营披露和事故报告
发布变更风险模型和 API 更新太快,可能打断集成迭代速度可能超过客户适配节奏版本管理纪律和下线政策

运营风险不只是宕机,还包括信任、抗滥用能力、文档,以及面向客户接口是否稳定。

[CR015, CR016, CR017, CR018, CR019, CR020]
合作伙伴 / 依赖风险登记表
依赖项证据上行空间风险严重性
云市场目录(AWS、Azure、Google)公开上架和模型目录可用性加快采购,扩大触达将打包、计费或政策控制权推向合作伙伴
网关路由层(Vercel、OpenRouter)生产 token 份额和上架证据让 DeepSeek 更快接触真实使用也让客户同样快速切走
部署生态(NVIDIA、Azure 托管方案)合作伙伴打包降低企业运维负担提升企业就绪度增加外部界面依赖
算力 / 出口管制环境政策和分析师来源持续凸显硬件约束可倒逼效率创新若获取受限,可能拖累性能或训练计划
区域竞争平台千帆、Volcengine、Kimi、MiniMax 等持续建设替代方案印证中国生态需求深度加大预算份额争夺和替代压力

依赖是双刃剑:既扩大触达,也把客户体验控制权和切换成本分散出去。

[CR023, CR024, CR025, CR026, CR033, CR034]
FR002: 风险传导图

多个独立风险节点都可能传导到同样商业结果:客户采用放慢、利润率走弱、可投资性下降。

[CR013, CR014, CR017, CR023, CR024, CR037]

7.3 人员、执行与缓释:快速迭代有价值,但会带来脆弱的协同要求

DeepSeek 的执行风险不只是能不能快速出货;关键在于在高强度审视下协调研究、产品、政策和渠道。CNBC 2026 年 6 月报道称,投资人被要求不要挖人,说明人才留存重要到足以写进融资动态。V3.1 发布说明显示,DeepSeek 可以很快更新模型、模式、API 映射、定价节奏和智能体功能;从产品角度看很亮眼,但从变更管理看也有风险。快迭代会抬高文档漂移、用户困惑和向后兼容错误的成本。外部生态压力又加了一层。State of AI 认为中国开放权重生态在变强,Volcengine、Baidu Qianfan、Kimi 和 MiniMax 的材料也显示,同业正在迅速为客户搭建相邻路径。也就是说,DeepSeek 一边要管内部执行,一边要追上外部市场。缓释风险应当是有条件的判断,而不是绝对判断。更好的披露、更强的安全防护和更有纪律的发布管理,可以降低部分风险。另一些风险——比如地缘政治信任缺口或出口管制制度——只能作为打破投资逻辑的条件持续监控。投资人应假设 DeepSeek 能缓释很多运营风险,但不能假设它能完全控制周围的政策环境。执行风险会在团队之间快速复利。DeepSeek 至少已经发布过一份面向透明度和安全使用的一手模型机制披露;这不是完整的保证体系,但公开释放了一定缓释意图。[CR029, CR030, CR031, CR032, CR033, CR034]

人才 / 执行风险登记表
风险证据重要性严重性监测信号
人才挖角 / 留任CNBC 报道投资人附带不得挖角条件核心研究连续性可能依赖少数稀缺人才更多公开离职或投资人侧限制
快速发布节奏V3.1 发布说明显示功能、定价和映射均有重大变化可能带来兼容性漂移或运营方困惑更多废弃项或突发 API 变更
跨职能协同负荷政策、伙伴和产品都在快跑执行失误可能出在接口处,不只发生在模型上文档、定价和合作伙伴页面相互矛盾
竞争执行压力State of AI 和同行资料显示竞争对手迭代很快DeepSeek 既要改进,也要守住份额同行缩小功能 / 价格差距
品牌 / 山寨压力CNIPA 商标通知显示机会主义仿冒可能造成用户混淆,并增加维权负担更多仿冒或欺诈事件

执行风险本质是协同问题,横跨研究、产品、信任和渠道。

[CR029, CR030, CR031, CR032, CR035, CR036]
缓释措施和终止标准表
风险领域可行缓释措施哪些公开证据会提升信心终止标准
隐私 / 司法辖区风险说明区域控制、企业数据处理和合规立场更细的隐私、数据传输和企业控制披露核心商业市场出现重大新增政府禁令
安全 / 滥用风险发布更强的护栏、红队或滥用响应证据公开安全测试和事故响应透明度大量有记录的有害使用事件直接指向 DeepSeek
合作伙伴 / 渠道依赖渠道多元化,并保留直接客户关系直接客户案例和渠道结构披露主要分发渠道大面积丢失或暂停
执行 / 发布风险更严格的版本管理、变更日志纪律和废弃沟通多个版本持续保持政策、文档和变更日志整洁反复破坏性变更,侵蚀开发者信任
人才风险留任计划和更厚的领导层梯队少数明星之外,高级人才梯队稳定的证据核心研究 / 平台团队出现可见人才外流
政策 / 出口管制风险情景规划和算力多元化更多供应和部署韧性细节实质拖慢下一代模型进展的限制

终止标准是投资人投资假设破裂条件,不是对这些事件会发生的预测。

[CR037, CR038, CR039, CR040, CR041, CR042]
FR003: 依赖关系图

DeepSeek 的执行风险集中在一张依赖网络上,横跨人才、合作伙伴、政策和算力。

[CR020, CR026, CR030, CR031, CR034, CR040]

7.4 展示项

Chapter 08

08估值

8.1 估值方法:看报道轮次、公开可比公司和风险折价,不做伪精确 DCF

传统内在价值估值无法由已审阅的公开记录支撑,因为关键输入缺失。没有经审计收入、未披露毛利率、没有公开烧钱速度,也没有可靠客户队列数据。公开记录能提供的是一组估值锚点:DeepSeek 在 2026 年 5 月被报道估值约 $45 billion,2026 年 6 月超过 $50 billion,2026 年 7 月融资讨论约 $71 billion。它还提供了其他中国模型实验室的可比公司信号,以及说明市场为何愿意关注的分发代理指标。因此,正确方法是三角校验。先看报道中的市场出清价格区间。再用 Moonshot、MiniMax、Z.ai 和 StepFun 的同业估值和公开上市信号交叉检查。然后针对披露不透明、政策悬顶,以及路由用量未必能顺畅转化为持久高毛利收入的风险打折。Gartner、Goldman Sachs、Artificial Analysis 等分析师市场规模来源能解释战略溢价为何存在,但不能单独证明某一轮的精确价格合理。结论必然是区间估值判断,而不是单点数字。保守方法重要,因为这里的表面精确会误导。[CV001, CV002, CV003, CV004, CV005, CV006]

建议摘要表
维度当前判断证据基础含义
当前报道估值~$50B 已交割 / ~$71B 后续融资洽谈TechCrunch、CNBC、CB Insights表面报价高
建议跟踪 / 继续研究公开证据质量与价格不匹配不足以盲投
置信度战略信号强,财务披露弱建议应随新数据调整
风险评级政策、披露和渠道依赖风险叠加仅有需求不够
估值立场偏高溢价叙事跑在公开基本面前面需要数据室验证

本摘要有意把估值立场和业务质量分开;DeepSeek 可以具有战略重要性,但仅凭公开证据看仍可能太贵。

[CV001, CV002, CV027, CV028, CV029]
可比估值表
公司公开估值信号阶段 / 路径相关性注意事项
DeepSeek~$45B 洽谈(2026 年 5 月),>$50B 交割(2026 年 6 月),~$71B 洽谈(2026 年 7 月)私营 / 报道轮次被估值的核心资产公开运营指标仍稀少
Moonshot / Kimi$20B 估值,融资 $2B私营 / 后期私募轮中国模型实验室可比标的,拥有受欢迎的消费端和 API 界面产品组合和已披露牵引力不同
MiniMax2024 年私募估值 $2.5B;2026 年港股首日市值 >$11.5B私营转公开市场中国 AI 实验室,显示跨时间重估潜力时点和市场条件不同
Z.ai / Zhipu(智谱)2026 年在香港上市;首家上市 LLM 公司公开市场为中国 LLM 资产提供公开市场情绪映射仅有上市状态不等于可干净比较倍数
StepFun2026 年接近 $2.5B 的 pre-IPO 轮后期私募 / IPO 路径具备 IPO 轨迹的中国前沿模型可比标的全球知名度低于 DeepSeek

可比信号很异质——有的是轮次,有的是公开市场市值,有的是 IPO 路径融资轮——因此最好用作相对情绪锚,而不是干净的倍数可比。

[CV003, CV004, CV005, CV006, CV007, CV008]

8.2 情景与敏感性:有上行空间,但取决于能否把战略稀缺性转成基本面

DeepSeek 的乐观情景很容易讲清。Vercel 和 OpenRouter 的需求代理指标显示,真实生产兴趣存在。公开定价显示,公司可以把自己放在远低于高端西方实验室的位置。MiniMax、Z.ai 和 StepFun 的信号显示,中国 AI 生态仍在吸引资本和公开市场通道。如果 DeepSeek 能持续提升技术,同时用新资本扩大渠道覆盖和企业信任,高溢价私人估值可以说得通。基准情景更克制:DeepSeek 仍然重要、使用广泛且具战略稀缺性,但估值大部分仍靠叙事支撑,因为收入质量和利润率持久性仍被遮住。悲观情景不是 DeepSeek 消失,而是公开使用量变现低于假设,政策和信任逆风封顶可触达需求,更快的同业压缩溢价。因此,敏感性不主要由市场规模论证驱动,而更受四个变量支配:变现效率、渠道依赖、政策摩擦和证据质量。任何一个变量的小变化,都可能显著移动公允价值区间,因为当前公开数据相对于醒目的估值数字太薄。简言之,假设比电子表格更重要。同一家公司看起来便宜还是贵,取决于这些缺失变量最终是顶级,还是只是普通。[CV014, CV015, CV016, CV017, CV018, CV019]

投资假设 / 反假设表
思路支撑证据反论点净判断
战略稀缺性具备全球相关性的中国 AI 龙头稀缺稀缺性不能消除执行或政策风险正面,但不足够
需求动能Vercel 和 OpenRouter 显示路由采用增长使用代理指标不是经审计收入正面,但不完整
性价比护城河DeepSeek 定价远低于高端实验室同行也在持续挤压同一价格带喜忧参半
资本获取报道中的轮次显示融资能力强能融到钱不证明单位经济模型好偏正面
退出可选性中国同行在融资、上市,并推进 IPO 路径若政策或变现不及预期,公开市场也会快速重估喜忧参半
公开证据质量有不少有用的公开代理指标核心运营指标仍缺失拖累投资判断信心

反假设不是“DeepSeek 不好”,而是“公开证据对当前传闻价格来说太弱”。

[CV014, CV015, CV016, CV017, CV018, CV019]
乐观 / 基准 / 悲观情景表
情景假设估值区间(USD bn)概率判断含义
乐观需求代理指标转化为高质量企业收入;政策风险受控;渠道扩张延续60–80低到中溢价定价可以守住,甚至被超过
基准DeepSeek 仍重要且高增长,但收入质量只是良好、谈不上惊艳,信任拖累仍在35–55只有拿到非公开尽调材料,2026 年 6 月价格才可能说得通
悲观使用变现差,政策摩擦扩大,同行挤压溢价20–35报道估值上沿显著偏高

情景区间是判断范围,不是交易价格。它们锚定报道估值、同行信号和风险折扣,而不是模型化 DCF。

[CV021, CV022, CV023, CV024, CV025, CV026]
FV002: 估值敏感性

估值最敏感的是财务证据缺口和政策折价,而不是宏观 TAM 本身。

分数是估值判断中的序位重要性权重,不是回归输出或概率估计。

[CV012, CV018, CV019, CV024, CV031, CV038]
FV003: 估值 / 回报区间

除非私下尽调发现强得多的基本面,否则基于公开证据的合理价值区间低于最高传闻价格。

这些区间是判断范围,锚定据报道的轮次价格、同业估值信号和风险折价。它们不是市场报价,也不是投资建议。

[CV021, CV022, CV023, CV024, CV025, CV029]

8.3 建议与尽调问题:观察或继续研究,而不是盲目投资

基于公开证据的建议,是观察 DeepSeek 或继续研究,而不是按报道价格上沿激进投资。估值约 $50 billion 时,如果投资人能进入数据室、并相信渠道和质量叙事会转化为持久收入,公司作为稀缺资产仍有可辩护空间。估值约 $71 billion 时,举证责任就高得多。核心原因不是市场机会小,而是证据质量仍远低于这一定价通常要求的水平。Conference Board 和禁令追踪来源显示,政策与信任担忧仍然真实存在。CNIPA 的备案通知和隐私政策披露也强化了一个判断:非技术问题可以影响价值创造。出资前,投资人需要看到客户集中度、收入质量、队列留存、毛利率和算力来源数据。如果这些数字非常出色,DeepSeek 可能配得上溢价。如果只是不错,今天报道的估值大概率偏高。因此,当前建议是谨慎好奇,而不是高信念资本。公开势头应被视为输入,而不是判决。保持谨慎有必要。[CV027, CV028, CV029, CV030, CV031, CV032]

投资假设破裂和终止触发表
触发因素为何重要意味着什么严重性
主要商业市场出现更广泛禁令或信任限制会显著扩大可触达需求折扣客户池和退出可选性受到结构性损伤
证据显示路由使用量无法转化为持久收入会削弱核心增长叙事需求质量弱于标题所暗示
算力或政策约束实质拖慢模型进展会削弱技术稀缺性和战略溢价叙事压缩,倍数支撑下降
重大人才或执行扰动会损害发布节奏和服务质量运营风险上升,信心走弱中高
数据室披露显示经济模型普通,并非顶级会击穿稀缺溢价假设当前定价很可能过高

这些是投资前后都应监测的投资假设破裂条件,不是对近期失败的预测。

[CV030, CV031, CV032, CV033, CV034]
最终尽调问题表
尽调问题为何重要最影响
经审计收入和分部结构将估值锚定到基本面建议和情景基准
按模型层级和渠道拆分的毛利率检验低价策略是否可持续敏感性和终止标准
NRR / 留存 / 头部客户集中度区分需求和可持续经济性情景概率
算力来源和预留容量计划澄清在政策或供应压力下的执行韧性风险折扣
按地域划分的政策 / 隐私企业控制决定可触达市场质量监管折扣
股权结构、条款和清算优先权决定新投资人的真实入场经济性回报区间

如果这些问题答得好,建议可能实质上调;如果被拒绝或答案薄弱,估值很可能应被视为偏高。

[CV035, CV036, CV037, CV038, CV039, CV040]
FV001: 建议逻辑

投资建议来自战略稀缺性和需求证明,但缺失运营指标与政策风险需要打很大折扣。

[CV001, CV010, CV011, CV027, CV028, CV029]
FV004: 投资 KPI 评分卡

这张 IC 风格评分卡概括了为什么 DeepSeek 值得跟踪;但在最高传闻价格下,信息太不透明,无法激进押注。

[CV014, CV015, CV020, CV027, CV028, CV029]

8.4 展示项

免责声明

本报告仅基于截至 2026-07-21 已审阅的公开来源。

证据索引

结论
编号陈述可信度来源
CO001 DeepSeek was incorporated on July 17, 2023 as a wholly owned subsidiary of High-Flyer Capital Management, headquartered in Hangzhou, Zhejiang Province, China. SO002, SO009, SO032
CO002 DeepSeek's full legal name is Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., written in Chinese as 杭州深度求索人工智能基础技术研究有限公司. SO001, SO002
CO003 The brand name 深度求索 (Shen Du Qiu Suo) translates to "seek depth," reflecting a research-first rather than product-first identity distinct from commercially oriented Chinese AI peers. SO002, SO009
CO004 DeepSeek's stated mission is "unraveling the mystery of AGI with curiosity," which distinguishes it from US frontier labs that frame missions around safety or societal benefit. SO015, SO001, SO016
CO005 At founding DeepSeek was 100% owned by High-Flyer Capital Management, Liang Wenfeng's quantitative hedge fund, with no external investors or venture capital. SO002, SO015, SO032
CO006 Liang Wenfeng, born 1985 in Guangdong Province, is DeepSeek's founder and CEO; he earned a bachelor's and master's degree in AI and electrical engineering at Zhejiang University. SO003, SO008, SO015
CO007 Liang Wenfeng co-founded High-Flyer Capital Management in 2015, which grew into one of China's top four quantitative hedge funds, last valued at approximately $8 billion. SO003, SO015, SO032
CO008 DeepSeek employed approximately 160 researchers and engineers as of 2025, an exceptionally lean team by frontier-AI laboratory standards globally. SO002, SO009
CO009 DeepSeek operates with a research-first posture, stating it has no immediate plans for commercialisation beyond API access priced intentionally close to cost. SO015, SO017, SO001
CO010 DeepSeek releases all major models as open-weight on GitHub and Hugging Face, making weights freely downloadable as core strategy to build developer adoption. SO007, SO021, SO022
CO011 DeepSeek-V3, released December 2024, has 671 billion total parameters with 37 billion active per token using a mixture-of-experts design, trained on 14.8 trillion tokens using 2.788 million H800 GPU-hours. SO005, SO022, SO025
CO012 DeepSeek-R1 was open-sourced on January 20, 2025, as an open-weight reasoning model that matched or exceeded OpenAI o1 on multiple reasoning and coding benchmarks. SO004, SO021, SO009, SO010
CO013 On January 27, 2025, Nvidia's stock fell approximately 17%, wiping roughly $589 billion in market capitalisation in a single session following the viral spread of DeepSeek R1; the event was widely labelled DeepSeek Monday. SO002, SO009, SO010, SO011, SO016
CO014 The DeepSeek-R1 methodology paper was published in Nature, volume 645, pages 633-638, 2025, representing peer-reviewed academic recognition for an open-weight reasoning model. SO004, SO002
CO015 DeepSeek-V4 was released approximately April 24, 2026, with a V4-Flash variant at 284 billion total parameters and V4-Pro at 1.6 trillion total parameters; both support a 1 million-token context window. SO002, SO027
CO016 As of the run date, DeepSeek's official API prices deepseek-chat at $0.07 per million input tokens and $1.10 per million output tokens, among the lowest globally for a frontier-class model. SO006, SO015
CO017 By May through June 2026, multiple sources cited DeepSeek's valuation in the range of $45 to $50 billion, with CB Insights recording $50 billion in June 2026. SO020, SO026, SO027
CO018 Bloomberg and TechCrunch reported on July 14, 2026 that DeepSeek was in talks to raise approximately $1.5 billion at a $71 billion valuation, with a 2027 IPO targeted. SO027, SO002
CO019 The July 14, 2026 reporting indicated a 2027 IPO timeline with a possibility of Q4 2026 debut, following the June 2026 closure of a $7 billion first external funding round. SO027, SO002
CO020 ChinaTalk's November 2024 profile stated DeepSeek was fully funded by High-Flyer and had no plans to fundraise, a posture that reversed in 2026 when competitive talent poaching drove the fundraising decision. SO015, SO026
CO021 High-Flyer's Fire-Flyer 2 compute cluster, deployed in 2021 with a budget of approximately 1 billion yuan, comprised 5,000 A100 GPUs in 625 nodes and provided the training infrastructure for DeepSeek's early models. SO032, SO002
CO022 DeepSeek's June 2026 funding round raised approximately $7 billion and was the company's first-ever external funding, closing at approximately $50 billion valuation. SO027, SO020, SO026
CO023 The June 2026 round was led by China's Integrated Circuit Industry Investment Fund (Big Fund), with Tencent, Alibaba, CATL, and Guozhitou Private Equity Fund Management among confirmed or reported participants. SO026, SO027, SO020, SO029
CO024 The Financial Times and TechCrunch reported that Liang Wenfeng's decision to raise outside capital was driven by competitive poaching of DeepSeek researchers by well-funded rivals, with equity-sharing as the solution. SO026, SO029
CO025 The June 2026 funding agreement included a no-poach covenant protecting DeepSeek employees from being hired by portfolio companies of the investors. SO029
CO026 TechCrunch reported in July 2026 that DeepSeek's cloud service runs on chips made by Huawei Technologies rather than Nvidia hardware, consistent with China's domestic semiconductor independence strategy. SO027, SO014, SO032
CO027 As of May 2026, Liang Wenfeng controlled approximately 90% of DeepSeek per the Financial Times as reported by TechCrunch. SO026, SO003
CO028 Bloomberg reported in July 2026 that Liang Wenfeng's personal net worth had reached approximately $36 billion, making him the wealthiest founder of an AI model company globally. SO002, SO003
CO029 DeepSeek-V2, released May 2024, priced API access at 1 RMB per million tokens—approximately one-seventh of Llama 3 70B cost at the time—triggering a China AI price war that forced ByteDance, Baidu, Tencent, and Alibaba to cut rates. SO015, SO016, SO012
CO030 V2's Multi-head Latent Attention architecture reduced KV cache memory requirements to 5-13% of standard Multi-Head Attention, enabling inference cost reductions that made near-cost API pricing viable. SO015, SO005
CO031 Sensor Tower data showed DeepSeek's DAU grew more than 700% week-over-week in the period January 22-28, 2025; the app accumulated over 23 million global downloads in 19 days, more than twice ChatGPT's pace at comparable maturity. SO030, SO002
CO032 As of June 2026, DeepSeek accounted for approximately 23% of all enterprise AI token traffic on the Vercel platform, compared to Anthropic's 32%, according to TechCrunch reporting. SO027
CO033 CNBC reported on February 24, 2026 that Anthropic accused DeepSeek, Moonshot AI, and MiniMax of conducting industrial-scale distillation attacks, generating 16 million exchanges via 24,000 fraudulently created accounts to extract training data from Claude. SO028, SO031, SO002
CO034 OpenAI submitted a memo to the US Congress in February 2026 alleging that DeepSeek used fraudulent account networks and third-party routers to distill ChatGPT and other US frontier models for training purposes. SO031, SO028
CO035 DeepSeek-Coder was the company's first public model release in November 2023, establishing its initial position in the open-source coding-model segment. SO024, SO007
CO036 Stanford HAI published a February 2025 analysis concluding that DeepSeek is "noticeably opaque when it comes to privacy protection, data-sourcing, and copyright," raising concerns for enterprise and regulatory adoption. SO033, SO012
CO037 The Economist reported in February 2025 that DeepSeek's research team is predominantly composed of Zhejiang University alumni, reflecting a domestically oriented talent strategy. SO018, SO015
CO038 Bloomberg reported in October 2025 that DeepSeek was outperforming OpenAI and Google in Africa, illustrating its expansion beyond Chinese and Western developer communities. SO002
CO039 DeepSeek does not publicly disclose revenue, gross margin, customer count, compute burn, or audited financial statements; the FT characterised its posture as "research over revenue." SO017, SO015, SO033
CO040 Liang Wenfeng described in a public ChinaTalk-translated interview that he spends his days "reading papers, writing code, and participating in group discussions," characterising himself as a practitioner-CEO rather than a figurehead. SO015
CO041 Reuters citing The Information reported in April 2026 that DeepSeek was raising funds at a $10 billion valuation, an early indication that preceded the round closing at a substantially higher $50 billion mark. SO002, SO026
CM001 DeepSeek’s market should be bounded as a subset of AI model, platform, agent, and consumer-application spending rather than treated as the whole AI economy. SM001, SM002, SM005
CM002 Goldman Sachs estimated a $150 billion total addressable market for generative AI software. SM002
CM003 Gartner forecast worldwide end-user spending on AI models and platforms at $64.252 billion in 2026. SM001
CM004 Gartner said that 2026 AI models and platforms spending would be up 63.4% from 2025. SM001
CM005 Gartner forecast foundation generative-AI model spending at $23.356 billion in 2026. SM001
CM006 Gartner forecast specialized or DSLM generative-AI model spending at $4.910 billion in 2026. SM001
CM007 State of AI 2025 said OpenAI retained a narrow frontier lead while DeepSeek, Qwen, and Kimi closed the gap on reasoning and coding tasks. SM003
CM008 State of AI 2025 reported that 44% of U.S. businesses now pay for AI tools, up from 5% in 2023. SM003
CM009 State of AI 2025 reported average AI contracts of $530,000. SM003
CM010 State of AI 2025 reported that 95% of surveyed professionals use AI at work or home and 76% pay for AI tools out of pocket. SM003
CM011 DeepSeek publishes OpenAI- and Anthropic-compatible API formats. SM005
CM012 DeepSeek’s current pricing page lists a 1M context window for the V4 generation. SM006
CM013 DeepSeek’s pricing page lists very low headline inference prices relative to premium frontier vendors. SM006, SM007
CM014 Anthropic’s public pricing page includes premium API rates such as $5/$25 per MTok for Opus 4.8. SM007
CM015 OpenAI’s business pricing page lists a $20 per user per month Business plan and custom-priced Enterprise tier. SM008
CM016 Kimi’s K2.6 pricing page says the model supports a 256k context window and long-horizon reasoning. SM018
CM017 MiniMax’s pay-as-you-go pricing page lists M2.7 input at $0.3 per million tokens and output at $1.2 per million tokens. SM016
CM018 Alibaba Cloud Model Studio offers both Qwen models and third-party models in one platform. SM010, SM011
CM019 Alibaba’s model catalog includes DeepSeek, Kimi, GLM, and MiniMax entries alongside Qwen. SM011
CM020 Alibaba publishes OpenAI-compatible and Anthropic-compatible base URLs across multiple regions for several models. SM011
CM021 Baidu Qianfan positions itself as an enterprise one-stop large-model and application-development platform. SM012, SM013
CM022 BigModel’s documentation describes a one-stop large-model platform with fine-tuning, evaluation, web search, knowledge retrieval, and OpenAI SDK compatibility. SM015
CM023 Vercel’s June 2026 AI Gateway index said total tokens grew 20% month over month while spend grew 43% month over month. SM024
CM024 Vercel said DeepSeek’s share of routed tokens jumped from under 1% to 17% in a single month while spend stayed near 1%. SM024
CM025 OpenRouter said DeepSeek doubled its token share from 9% to 18% between January and early June 2026. SM025
CM026 OpenRouter said DeepSeek had been the top model author on its platform since mid-May 2026. SM025
CM027 Multi-cloud distribution puts DeepSeek inside managed buying environments instead of forcing direct API procurement. SM021, SM022, SM023
CM028 Google Cloud documents DeepSeek models as managed APIs and self-deployed models on its Gemini Enterprise Agent Platform. SM021
CM029 State of AI said power supply had emerged as a new constraint in the industrial era of AI. SM003
CM030 State of AI said China expanded its open-weights ecosystem and domestic-silicon ambitions. SM003
CM031 MOFCOM’s 2025 update to China’s prohibited-or-restricted export-technology framework shows that national technology policy remains an active market variable. SM020
CM032 The same platformization that helps DeepSeek distribute also makes buyer multi-homing normal. SM011, SM012, SM015
CM033 Artificial Analysis positions model competition around quality, price, speed, and openness rather than brand alone. SM004
CM034 DeepSeek’s most practical buyer is the developer or AI team willing to route specific workloads to a low-cost model that still clears evals. SM005, SM024, SM025
CM035 Enterprise AI platform buyers care about cost transparency, usage tracking, performance, and reliability in addition to raw model quality. SM001
CM036 Cloud and platform intermediaries are material buyers because they can list DeepSeek as catalog inventory and monetize downstream usage. SM021, SM022, SM023
CM037 DeepSeek benefits when low-cost models become production-worthy, not merely benchmark-worthy. SM024, SM025
CM038 DeepSeek’s moat weakens when rivals make compatibility, long context, and low prices table stakes. SM014, SM016, SM018, SM019
CM039 The most realistic near-term SAM for DeepSeek is the subset of routed inference and agent workloads where price-sensitive buyers are willing to multi-home. SM001, SM024, SM025
CP001 DeepSeek’s competitive set includes premium U.S. labs, Chinese open-weight peers, and cloud/catalog intermediaries rather than one single vendor cohort. SP001, SP010, SP012
CP002 OpenAI remains a premium reference point for developers and enterprise buyers through its business and API stack. SP008
CP003 Anthropic remains a premium reference point through publicly posted Claude API pricing tiers. SP007
CP004 Google competes through the Gemini developer API and related tooling surface. SP009
CP005 Alibaba competes as both a model owner and a catalog operator through Model Studio. SP010, SP011
CP006 Baidu Qianfan positions itself as a one-stop enterprise large-model and application-development platform. SP012
CP007 BigModel documents a one-stop MaaS platform with fine-tuning, evaluation, search, and knowledge retrieval. SP013
CP008 CNBC reported in January 2026 that Chinese AI firms from Alibaba to Moonshot were racing to release new models one year after DeepSeek’s breakout. SP019
CP009 Moonshot AI raised $2 billion at a $20 billion valuation in May 2026 according to TechCrunch. SP020
CP010 DeepSeek’s own API is formatted to be compatible with OpenAI and Anthropic conventions. SP002
CP011 API compatibility lowers code-porting friction between DeepSeek and competing vendors. SP002, SP011, SP013
CP012 DeepSeek’s current public pricing page lists a 1M context window for its V4 generation. SP003
CP013 DeepSeek-V2 was introduced as a strong, economical, and efficient MoE language model. SP006
CP014 DeepSeek-V3 is described on GitHub as a 671B-parameter MoE model with 37B activated parameters per token. SP005
CP015 DeepSeek-R1 is described by the company as its first-generation reasoning model family. SP004
CP016 Anthropic publishes premium API pricing including Opus 4.8 at $5 input and $25 output per million tokens. SP007
CP017 OpenAI publishes paid API and business pricing that reinforces its premium market positioning relative to low-cost challengers. SP008
CP018 MiniMax publicly lists M2.7 pricing at $0.3 input and $1.2 output per million tokens. SP015
CP019 Kimi K2.6 markets a 256k context window and long-horizon reasoning support. SP017
CP020 Z.ai says GLM-5.2 supports 1M lossless context and long-horizon task improvements. SP014
CP021 MiniMax’s site markets MiniMax M3 as a frontier coding and agentic model with 1M context. SP024
CP022 Kimi’s consumer site promotes K3 for agent programming and knowledge work. SP025
CP023 Alibaba’s model catalog offers Qwen and third-party models with OpenAI-compatible and Anthropic-compatible endpoint conventions. SP011
CP024 Baidu’s Qianfan marketing emphasizes search, agent, and enterprise workflow capabilities rather than only raw model access. SP012
CP025 BigModel documents OpenAI SDK compatibility as part of its platform positioning. SP013
CP026 Artificial Analysis frames model competition across quality, price, output speed, and latency. SP001
CP027 Google Cloud documents DeepSeek as available via managed APIs and self-deployed options on Gemini Enterprise Agent Platform. SP021
CP028 Azure announced DeepSeek R1 in Azure AI Foundry’s model catalog. SP022
CP029 AWS announced DeepSeek-R1 in Bedrock Marketplace and SageMaker JumpStart. SP023
CP030 Cloud-catalog availability puts DeepSeek inside normal enterprise procurement and evaluation paths. SP021, SP022, SP023
CP031 The same catalogs that broaden DeepSeek’s reach also place it beside substitutes that are easy to compare. SP011, SP021, SP022, SP023
CP032 Multi-model platforms in China train buyers to expect routing and substitution across vendors. SP010, SP012, SP013
CP033 DeepSeek’s most credible moat today is the combination of strong reasoning reputation, open-weight credibility, and low cost. SP003, SP004, SP005, SP006
CP034 The clearest limit on DeepSeek’s moat is that Chinese rivals increasingly advertise similar context, agent, and pricing features. SP014, SP015, SP017, SP024, SP025
CP035 DeepSeek competes against both premium labs above it and cheaper challengers below it, compressing room for error. SP007, SP015, SP017
CP036 Moonshot/Kimi, MiniMax, GLM, and Qwen together show that DeepSeek did not freeze the Chinese market after its breakout. SP019, SP020, SP014, SP015, SP017
CP037 No reviewed public source provides audited cross-vendor market-share data that would settle competitive ranking cleanly. SP001, SP019
CP038 DeepSeek is best described as the current value leader in a crowded and rapidly converging low-cost model segment. SP001, SP003, SP015, SP017, SP019
CP039 Vercel added DeepSeek V4 Pro and DeepSeek V4 Flash to AI Gateway in 2026, reinforcing DeepSeek’s presence in model-routing workflows. SP026
CP040 OpenAI’s official models documentation says its latest models support text and image input, text output, multilingual capabilities, and vision, underscoring broad multimodal breadth versus low-cost challengers. SP027
CP041 Alibaba Cloud’s Model Studio pricing page documents pay-as-you-go billing, tiered token pricing, and discounts for supported batch or context-cache usage, showing how platform operators compete on commercial packaging as well as model quality. SP028
CI001 DeepSeek bills API usage based on token consumption. SI001, SI002
CI002 DeepSeek distinguishes cached and uncached input pricing on its pricing page. SI001
CI003 DeepSeek separately prices input and output tokens on its V4 pricing page. SI001
CI004 DeepSeek’s pricing page lists separate Flash and Pro tiers for the V4 generation. SI001
CI005 DeepSeek’s pricing page lists different concurrency limits for Flash and Pro. SI001
CI006 DeepSeek’s terms say fees are deducted from recharge balances or gifted balances and gifted balances are deducted first when both exist. SI001, SI003
CI007 DeepSeek’s terms state that product prices may change. SI003
CI008 DeepSeek monetizes direct API inference usage through token billing rather than a flat subscription disclosed in reviewed sources. SI001, SI002
CI009 Google Cloud lists DeepSeek as a managed API and self-deployed model option. SI018
CI010 Azure lists DeepSeek R1 in Azure AI Foundry’s model catalog. SI019
CI011 AWS lists DeepSeek-R1 in Bedrock Marketplace and SageMaker JumpStart. SI020
CI012 DeepSeek’s visible revenue model likely includes a mix of direct API traffic and partner-routed enterprise usage. SI001, SI018, SI019, SI020
CI013 Token volume is a first-order unit-economics driver because revenue and compute both scale with usage. SI001, SI002
CI014 Cache hit rates are economically important because cached-input prices are materially lower than uncached-input prices. SI001
CI015 Output intensity matters economically because output tokens are priced separately and above input prices on V4 tiers. SI001
CI016 Model mix matters economically because DeepSeek publishes higher prices for Pro than Flash. SI001
CI017 No reviewed public source disclosed DeepSeek revenue or ARR. SI006, SI007, SI008, SI015
CI018 Forbes says Liang Wenfeng funded DeepSeek in part with proceeds from High-Flyer. SI013
CI019 Fortune identifies Liang Wenfeng as coming from quantitative finance through High-Flyer. SI014
CI020 TechCrunch reported in May 2026 that DeepSeek’s first outside round could value it at $45 billion. SI006
CI021 CNBC reported in June 2026 that DeepSeek closed its first external funding round at over a $50 billion valuation. SI008
CI022 CNBC reported the June 2026 financing as DeepSeek’s first external funding round. SI008
CI023 TechCrunch reported in July 2026 that DeepSeek was exploring about $1.5 billion in new funds at roughly a $71 billion valuation. SI007
CI024 TechCrunch reported that the July 2026 financing talks followed a reported $7 billion raise only a month earlier. SI007
CI025 CB Insights lists DeepSeek as Series A, with $7.546 billion total raised and founded year 2016. SI015
CI026 If the May-through-July 2026 financing reports are directionally right, DeepSeek likely has substantial near-term capital access. SI006, SI007, SI008
CI027 No reviewed public source provided audited cash-on-hand or a disclosed runway figure for DeepSeek. SI006, SI007, SI008, SI015
CI028 No reviewed public source provided audited monthly burn for DeepSeek. SI006, SI007, SI008, SI015
CI029 Vercel reported that DeepSeek’s share of routed tokens jumped from under 1% to 17% in a single month while spend stayed near 1%. SI016
CI030 OpenRouter reported that DeepSeek doubled token share from 9% to 18% between January and early June 2026. SI017
CI031 Gartner forecast worldwide AI models and platforms spending at $64 billion in 2026, supporting a fast-growing demand backdrop. SI021
CI032 State of AI 2025 reported that 44% of U.S. businesses now pay for AI tools, supporting the idea that demand is monetizing. SI022
CI033 Export controls and hardware access remain financially material because DeepSeek competes in compute-intensive frontier-model markets. SI007, SI010, SI011
CI034 CNBC reported that Anthropic joined OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms. SI009
CI035 FDD summarized OpenAI allegations that DeepSeek had stolen intellectual property to train its models. SI012
CI036 CNBC’s June 2026 funding report said a no-poaching promise was presented as a condition of investing in DeepSeek. SI008
CI037 The single biggest public financial diligence gap is the absence of audited revenue, margin, burn, and cash disclosure. SI015, SI006, SI007, SI008
CI038 DeepSeek can be monetizing rapidly and still be difficult to underwrite fundamentally because demand proxies do not reveal margin durability. SI016, SI017, SI021
CI039 China’s intellectual-property authority said it rejected 63 trademark applications tied to “DEEPSEEK,” indicating brand protection and legal enforcement costs around the franchise. SI026
CI040 DeepSeek’s official API changelog shows that V4 introduced new model names while retiring legacy aliases on a set timetable, indicating that monetization and customer migration depend on active release-management discipline. SI027
CE001 DeepSeek publicly exposes an API platform and a web/chat-facing surface. SE001, SE004
CE002 DeepSeek publishes a transparency page that tracks major model releases. SE004
CE003 DeepSeek distributes model artifacts through GitHub repositories for V2, V3, and R1. SE006, SE007, SE008
CE004 DeepSeek also distributes model artifacts through Hugging Face pages for V2, V3, and R1. SE009, SE010, SE011
CE005 DeepSeek’s product surface therefore includes hosted access, open-weight access, and partner-managed access. SE001, SE004, SE015, SE016, SE017
CE006 DeepSeek’s API docs state that the API uses a format compatible with OpenAI and Anthropic. SE001
CE007 DeepSeek’s pricing docs show support for JSON output and tool calls. SE002
CE008 DeepSeek’s pricing docs show both thinking and non-thinking modes for V4 Flash. SE002
CE009 DeepSeek’s pricing docs list a 1M context window and a maximum 384K output length for current V4 tiers. SE002
CE010 The public materials imply core workflows spanning chat, reasoning, coding, and agent tasks. SE002, SE008, SE019
CE011 AWS lists DeepSeek-R1 in Bedrock Marketplace and SageMaker JumpStart. SE015
CE012 Azure lists DeepSeek R1 in Azure AI Foundry’s model catalog. SE016
CE013 DeepSeek-V2 is described as a strong, economical, and efficient MoE language model. SE006, SE012
CE014 DeepSeek-V3 is described as a 671B-parameter MoE model with 37B activated parameters per token. SE007, SE013
CE015 DeepSeek-R1 is described as a first-generation reasoning model family. SE008, SE014
CE016 The DeepSeek-R1 paper frames the family around reinforcement learning to incentivize reasoning capability. SE014
CE017 The transparency page shows DeepSeek-V4 as a major release dated 2026-04-24. SE004
CE018 DeepSeek’s public technical identity is centered on efficiency, MoE design, and reasoning specialization. SE006, SE007, SE008, SE012, SE013, SE014
CE019 The externally visible operating stack includes base models, hosted APIs, documentation, and partner-packaged distribution. SE001, SE004, SE015, SE016, SE017, SE020
CE020 A public status page indicates that DeepSeek operates a production service surface, not only static research artifacts. SE020
CE021 The transparency page and linked model-card / report references improve public release traceability. SE004
CE022 The terms of use identify Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the service operator. SE005
CE023 Google Cloud documents DeepSeek as available for managed APIs and self-deployed models on Gemini Enterprise Agent Platform. SE017
CE024 NVIDIA described DeepSeek-R1 as an open model with state-of-the-art reasoning capabilities when packaging it in NIM. SE018
CE025 Vercel added DeepSeek V4 Pro and V4 Flash to AI Gateway in July 2026. SE019
CE026 Peer technical materials from Z.ai, Kimi, and MiniMax show that long context, coding, and agent positioning are becoming table stakes. SE025, SE026, SE027
CE027 CSIS argued that DeepSeek’s open-source structure increases misuse and jailbreak risk relative to more controlled Western API approaches. SE021
CE028 The pricing page notes that the legacy model names deepseek-chat and deepseek-reasoner will be deprecated after 2026-07-24 for compatibility reasons. SE002
CE029 Deprecation notices imply an ongoing integration-management burden for developers building against DeepSeek. SE002
CE030 Adverse commentary continues to frame DeepSeek as carrying unresolved safety and abuse concerns. SE021, SE022, SE023
CE031 The public roadmap from V2 to V4/R1 shows a visible release cadence across 2024, 2025, and 2026. SE004, SE006, SE007, SE008
CE032 Cloud and gateway packaging show that DeepSeek is optimizing for product distribution, not only research publication. SE015, SE016, SE017, SE018, SE019, SE029
CE033 Compute access and policy conditions are meaningful technical dependencies because frontier-model progress still depends on deployment and infrastructure availability. SE017, SE018, SE021
CE034 Documentation quality and compatibility stability are critical because DeepSeek relies on standard-format APIs to reduce switching friction. SE001, SE002
CE035 No reviewed public source provides a full internal systems diagram or robust public QA / red-team disclosure for DeepSeek. SE004, SE021, SE022
CE036 The most important technical diligence question is whether DeepSeek can keep reasoning quality and low cost while preserving API stability across fast releases. SE002, SE014, SE019
CE037 Nature/Stanford commentary treated DeepSeek as genuinely disruptive rather than a trivial copycat release. SE024
CE038 DeepSeek’s product moat depends increasingly on operational execution across docs, channels, and service reliability rather than on one benchmark snapshot. SE020, SE025, SE026, SE027
CE039 Managed cloud availability means some enterprise users can adopt DeepSeek without directly operating its infrastructure. SE015, SE016, SE017, SE029
CE040 DeepSeek’s product posture is advanced enough for production experimentation but still carries unresolved trust and robustness questions. SE020, SE021, SE022, SE024
CE041 DeepSeek’s privacy policy identifies Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the controller of its apps, websites, software, and related services, underscoring that the company operates a live service platform in addition to publishing models. SE030
CE042 DeepSeek’s English transparency page lists released models with release dates, model cards, and technical reports, including V4 dated April 24, 2026. SE031
CU001 DeepSeek’s public customer footprint spans self-serve users, developers, enterprise platform evaluators, intermediaries, and researchers. SU001, SU002, SU004, SU012, SU013, SU014
CU002 The DeepSeek app and website prove a consumer and prosumer access surface exists. SU001, SU005
CU003 The API docs prove a developer-facing access surface exists. SU002, SU003
CU004 Cloud catalog listings prove that enterprise AI teams can evaluate DeepSeek inside familiar procurement environments. SU012, SU013, SU014, SU015
CU005 Intermediaries matter because gateways and clouds can route downstream customer demand into DeepSeek. SU009, SU010, SU011, SU012, SU013, SU014
CU006 AWS made DeepSeek-R1 available in Bedrock Marketplace and SageMaker JumpStart. SU012
CU007 Azure made DeepSeek R1 available in Azure AI Foundry’s model catalog. SU013
CU008 Google Cloud documents DeepSeek as available via managed APIs and self-deployed models. SU014
CU009 NVIDIA packaged DeepSeek-R1 in NIM. SU016
CU010 Vercel added DeepSeek V4 to AI Gateway. SU010
CU011 Open-weight repositories and documentation create an additional evaluation path for researchers and infrastructure teams. SU002, SU004
CU012 Status-page visibility supports the view that DeepSeek operates an ongoing service for users, not only one-off model drops. SU006
CU013 Sensor Tower reported that DeepSeek received more than 23 million downloads in its first 19 days. SU007
CU014 Sensor Tower reported about 2 million U.S. downloads in that same launch window. SU007
CU015 Sensor Tower reported average mobile app DAUs increased by more than 700% week over week during the breakout period. SU007
CU016 Appfigures reported that DeepSeek crossed one million downloads quickly after launch. SU008
CU017 Vercel reported that DeepSeek’s share of routed tokens jumped from under 1% to 17% in a single month. SU009
CU018 OpenRouter reported that DeepSeek doubled token share from 9% to 18% in the first half of 2026. SU011
CU019 Cloud and gateway listings show that DeepSeek’s deployment path expanded beyond direct usage into managed enterprise and developer environments. SU010, SU012, SU013, SU014, SU015, SU016
CU020 Vercel AI Gateway is customer-proof because it reflects a production routing environment used by downstream applications. SU009, SU010
CU021 OpenRouter’s token-share report is customer-proof because it reflects routed model usage by gateway users rather than a mere announcement. SU011
CU022 DeepSeek’s most visible public expansion today is channel expansion and workflow expansion, not named-logo expansion. SU009, SU010, SU012, SU013, SU014
CU023 Public evidence suggests repeat usage exists because routed token share rose across multiple periods and environments. SU009, SU011
CU024 Public evidence does not disclose DeepSeek’s net revenue retention, logo retention, or cohort retention. SU009, SU011, SU020
CU025 Public evidence does not disclose NPS or CSAT for DeepSeek. SU001, SU002, SU006
CU026 Public evidence does not disclose a reconciled count of paying enterprise customers. SU001, SU002, SU004
CU027 Strong usage signals do not by themselves prove high-quality, diversified, retained revenue. SU007, SU009, SU011
CU028 A channel-heavy customer footprint can create dependence on intermediaries that own discovery, routing, or procurement. SU009, SU010, SU012, SU013, SU014
CU029 Regulatory and privacy concerns have already led some governments and agencies to ban or restrict DeepSeek. SU020, SU021, SU022
CU030 Such bans can remove public-sector or high-trust segments from the reachable customer pool. SU020, SU021, SU022
CU031 Competing Chinese model labs increase customer wallet competition and can weaken DeepSeek’s share of future routed demand. SU025, SU017, SU018, SU019
CU032 Consumer downloads can overstate durable monetization if conversion to API or enterprise spend is weak. SU007, SU008
CU033 The customer story is strongest on reach and weakest on satisfaction, concentration, and account expansion quality. SU007, SU009, SU011, SU020
CU034 AICPB’s user-ranking methodologies show that customer-attention markets are being tracked monthly across website visits and app MAU, even when DeepSeek-specific rank detail is not fully recoverable in fetched text. SU023, SU024
CU035 DeepSeek’s terms identify a single service operator, which matters because trust and service accountability affect enterprise customer willingness to buy. SU026
CU036 The most important next customer diligence request is a segment-by-segment breakdown of paying accounts, NRR, churn, and channel concentration. SU026, SU009, SU011
CU037 The cohort figure in this chapter is an author estimate because DeepSeek discloses no formal retention metrics; it is only a durability heuristic anchored to public usage proxies. SU007, SU009, SU011
CU038 DeepSeek likely has a broad and fast-growing customer surface, but its customer-quality profile remains largely opaque in public sources. SU007, SU009, SU011, SU020
CU039 The Conference Board said multiple state and federal government bodies moved to ban DeepSeek on government devices because of national-security and privacy concerns. SU027
CR001 DeepSeek’s privacy policy applies to DeepSeek apps, websites, software, and related services. SR001
CR002 DeepSeek’s privacy policy identifies Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the data controller / service provider. SR001
CR003 DeepSeek’s privacy policy says personal data is directly collected, processed, and stored in the PRC to provide services. SR001
CR004 DeepSeek’s terms say services may vary by jurisdiction and may be modified, suspended, or terminated. SR002
CR005 TechCrunch documented that DeepSeek’s tech had been banned by a growing number of countries and government bodies. SR003
CR006 CNA framed DeepSeek bans around privacy concerns, geopolitics, and wider AI-tech implications. SR004
CR007 Al Jazeera also documented countries banning DeepSeek and questioned the reasons. SR005
CR008 The Conference Board said multiple state and federal government bodies moved to ban DeepSeek on government devices. SR006
CR009 CNBC reported that Anthropic joined OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms. SR009
CR010 FDD summarized OpenAI’s allegation that DeepSeek stole intellectual property to train its models. SR010
CR011 CNIPA said it rejected 63 trademark applications tied to “DEEPSEEK.” SR013
CR012 Legal and regulatory narratives can affect procurement and valuation simultaneously because they shape both reach and trust. SR001, SR002, SR003, SR009
CR013 CSIS argued that DeepSeek’s open-source structure increases misuse and jailbreak risk. SR007
CR014 DeepSeek’s terms say outputs may contain errors or omissions and should not be treated as professional advice. SR002
CR015 DeepSeek’s privacy policy excludes downstream applications built by developers from parts of its direct policy scope. SR001
CR016 A public status page exists, indicating a live operated service surface. SR014
CR017 A status page alone does not disclose SLA quality or incident history. SR014
CR018 The transparency page records major release milestones such as DeepSeek-V4 on 2026-04-24. SR015
CR019 The V3.1 release note shows major feature, pricing, and API changes arriving in a single update cycle. SR026
CR020 Rapid release cadence increases compatibility and change-management risk for customers and partners. SR015, SR026
CR021 MOFCOM’s 2025 export-control update shows technology policy remains an active strategic variable. SR012
CR022 Cornell’s “DeepSeek problem” framing shows legal-policy scrutiny extending into U.S. policy debate. SR011
CR023 AWS, Azure, and Google listings expand reach but make DeepSeek partly dependent on third-party distribution rails. SR016, SR017, SR018
CR024 Azure’s V4-Pro catalog page emphasizes Microsoft-managed support, unified billing, and reduced integration effort. SR025
CR025 Microsoft-managed packaging can reduce buyer friction while also relocating part of the customer operating context to Azure. SR025
CR026 NVIDIA packaging and cloud catalog availability show that deployment ecosystems are critical dependencies for DeepSeek’s enterprise reach. SR019, SR018
CR027 Vercel and OpenRouter prove that gateway routing can swing meaningful token share quickly. SR020, SR021
CR028 Gateway routing dependence is strategically risky because the same surfaces that create adoption can accelerate switching away. SR020, SR021, SR028
CR029 CNBC’s June 2026 reporting of a no-poaching investor condition points to unusually intense talent-retention pressure. SR022
CR030 People risk matters more in frontier AI because a small number of researchers or platform operators can disproportionately affect output quality and velocity. SR022
CR031 State of AI reported that competition intensified as Chinese labs closed the gap on reasoning and coding tasks. SR023
CR032 Volcengine, Baidu Qianfan, Kimi, and MiniMax materials all show that adjacent Chinese platforms are moving fast on agent and platform features. SR027, SR028, SR029, SR030
CR033 Compute access remains a structural dependency because model progress and enterprise packaging still rely on hardware and platform availability. SR012, SR018, SR025
CR034 Kimi’s migration guide demonstrates how low code-switching friction can be in this market. SR028
CR035 Brand-copycat pressure is not theoretical: CNIPA documented a wave of attempted “DEEPSEEK” trademark registrations. SR013
CR036 The combined privacy, ban, and IP narratives create a structural trust gap for public-sector or highly regulated buyers. SR001, SR003, SR006, SR009, SR010
CR037 A plausible legal kill criterion would be expansion of bans or restrictions into additional major commercial markets. SR003, SR004, SR006
CR038 A plausible operational kill criterion would be repeated breaking changes or service incidents that materially erode developer trust. SR014, SR015, SR026
CR039 Visible mitigation paths include stronger privacy disclosures, clearer versioning, more safety transparency, and more explicit channel-governance discipline. SR001, SR002, SR014, SR015
CR040 Public evidence still lacks quantified direct-versus-channel mix, detailed incident history, and compute contingency data. SR014, SR018, SR025
CR041 Some risks are structural—geopolitical trust and hardware policy—while others are partially controllable through operations and disclosures. SR001, SR002, SR012, SR026
CR042 The highest-consequence risk today is a compound scenario where policy/trust narratives reduce adoption while dependency and switching dynamics weaken monetization resilience. SR003, SR006, SR020, SR021, SR025
CR043 MiniMax publishes architecture-specific technical narratives as it competes in the same agent and model market, reinforcing the pace of external execution pressure on DeepSeek. SR029, SR031
CR044 DeepSeek has published a first-party model-mechanism and training-methods disclosure that frames transparency and user right-to-know as mitigation against improper model use, providing some visible governance effort even if it does not eliminate broader trust concerns. SR032
CV001 TechCrunch reported in May 2026 that DeepSeek’s first outside round could value the company at $45 billion. SV006
CV002 CNBC reported in June 2026 that DeepSeek closed its first external funding round at over a $50 billion valuation. SV007
CV003 TechCrunch reported in July 2026 that DeepSeek was exploring roughly $1.5 billion in new funds at about a $71 billion valuation. SV008
CV004 Moonshot AI raised $2 billion at a $20 billion valuation in May 2026 according to TechCrunch. SV024
CV005 SiliconANGLE reported that MiniMax raised $600 million at a $2.5 billion valuation in 2024. SV013
CV006 TechNode reported MiniMax’s market capitalization briefly topped $11.5 billion on its Hong Kong debut in January 2026. SV014
CV007 KrASIA reported that StepFun was nearing a USD 2.5 billion pre-IPO round in 2026. SV015
CV008 The Standard also reported StepFun completed a new US$2.5 billion funding round for a Hong Kong IPO push. SV016
CV009 Qiming said Z.ai listed in Hong Kong in January 2026 as the world’s first listed large language model company. SV017
CV010 Yicai likewise reported Zhipu AI as the first LLM company to go public. SV018
CV011 No reviewed public source disclosed audited revenue or ARR for DeepSeek. SV006, SV007, SV008, SV009
CV012 A valuation approach for DeepSeek must therefore rely on reported rounds, comparable valuation signals, and risk-adjusted scenario ranges rather than a fully modeled DCF. SV006, SV007, SV008, SV009
CV013 Forbes says Liang Wenfeng funded DeepSeek in part with proceeds from High-Flyer, supporting a founder-backed scarcity narrative before outside rounds. SV012
CV014 Gartner’s $64 billion 2026 models-and-platforms spending forecast supports the existence of a very large category prize. SV001
CV015 Goldman Sachs estimated a $150 billion generative-AI software TAM, providing an upper-bound strategic context rather than a DeepSeek-specific SAM. SV002
CV016 Artificial Analysis frames model competition around quality, price, speed, and latency, supporting a value-driven premium for winners that score well on multiple axes. SV003
CV017 Vercel reported that DeepSeek’s share of routed tokens rose from under 1% to 17% in a single month, indicating unusually fast production interest. SV010
CV018 OpenRouter reported that DeepSeek doubled token share from 9% to 18% in the first half of 2026. SV011
CV019 Demand proxies support upside, but they do not reveal revenue quality, margin, or retention. SV010, SV011
CV020 Google Cloud’s published Gemini pricing shows why DeepSeek’s low-price positioning can support strategic value creation versus premium U.S. rivals. SV019, SV004
CV021 MiniMax’s published token-plan quick start and pay-as-you-go pricing show that low-switching and low-price Chinese alternatives also compress DeepSeek’s valuation umbrella. SV020, SV021
CV022 Kimi’s pricing page reinforces that long-context Chinese rivals are also competing for the same price-sensitive workloads. SV022
CV023 State of AI said competition intensified as DeepSeek, Qwen, and Kimi closed the gap on reasoning and coding tasks. SV023
CV024 A $50 billion valuation is aggressive but arguable only if private diligence confirms exceptional revenue quality, retention, and margin structure. SV002, SV010, SV011, SV023
CV025 A $71 billion valuation appears substantially more demanding given the absence of public audited fundamentals. SV008, SV009, SV011
CV026 Public Chinese AI comps imply that DeepSeek is being priced at a very substantial premium to most named peers. SV013, SV014, SV015, SV016, SV017, SV018, SV024
CV027 Conference Board and DeepSeek’s privacy disclosures support applying a policy and trust discount in valuation. SV025, SV027
CV028 DeepSeek’s terms support an operating-risk discount because the company expressly reserves the right to modify, suspend, or terminate services. SV028
CV029 The most defensible public-evidence recommendation is track or research more rather than invest blindly at current reported pricing. SV001, SV006, SV007, SV008, SV025
CV030 Confidence in that recommendation should be medium because strategic upside is real but core economics remain undisclosed. SV001, SV010, SV011, SV009
CV031 Channel-heavy adoption deserves a valuation haircut because intermediaries can control customer experience and accelerate switching. SV010, SV011, SV029
CV032 Microsoft’s managed V4-Pro catalog page shows enterprise-readiness upside, but also implies that part of the value proposition can sit with the platform owner. SV029
CV033 A major thesis-break would be further expansion of bans or trust restrictions into important commercial markets. SV025, SV027
CV034 Another thesis-break would be evidence that routed usage converts poorly into durable, direct, high-margin revenue. SV010, SV011
CV035 The most important diligence asks are audited revenue, gross margin by tier and channel, retention, compute sourcing, and financing terms. SV009, SV027, SV028
CV036 The absence of gross margin and retention data is the single biggest public-data weakness in the valuation case. SV009, SV010, SV011
CV037 CNIPA’s filing notice shows that non-technical legal issues can impose real brand and enforcement costs that warrant a discount. SV026
CV038 CB Insights lists DeepSeek as Series A with $7.546 billion total raised, which reinforces the scale of the reported capital story even if metadata conflicts remain. SV009
CV039 The public-evidence fair-value band is most defensibly framed in a roughly $35B–$55B range, with upside to $60B–$80B only under a strong private-diligence bull case. SV006, SV007, SV008, SV010, SV011
CV040 The business may merit strategic fascination today, but the valuation still needs private proof of elite economics. SV006, SV010, SV025, SV027
来源
编号出版方标题引文
SO001 DeepSeek DeepSeek Official Website Official brand surface for DeepSeek models, chatbot, and research lab.
SO002 Wikipedia contributors DeepSeek Wikipedia article Comprehensive encyclopedia article with inline citations through July 2026.
SO003 Wikipedia contributors Liang Wenfeng Wikipedia article Born 1985 Guangdong; Zhejiang University AI and EE; co-founded High-Flyer 2015.
SO004 DeepSeek DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning Published in Nature vol 645 pp 633-638 2025; introduces GRPO reinforcement learning for open-weight reasoning.
SO005 DeepSeek DeepSeek-V3 Technical Report 671B total parameters; 37B active per token; trained on 14.8T tokens; 2788000 H800 GPU-hours.
SO006 DeepSeek DeepSeek API Pricing deepseek-chat input $0.07 per million tokens; output $1.10 per million tokens.
SO007 DeepSeek deepseek-ai GitHub Organisation Official GitHub organisation hosting open-weight model releases.
SO008 Fortune DeepSeek Founder Liang Wenfeng Is the Hedge-Fund Manager Who Could Shake Up Silicon Valley Former quant fund manager whose lab is shaking up the AI world.
SO009 The Guardian Who is behind DeepSeek and how did it achieve its AI Sputnik moment Sputnik moment framing; covers Liang Wenfeng and High-Flyer background in depth.
SO010 NPR DeepSeek: Did a little-known Chinese startup cause a Sputnik moment for AI Covers the January 2025 R1 launch and its geopolitical AI significance.
SO011 The New Yorker Is DeepSeek China's Sputnik Moment Frames R1 launch as potential reset of AI competition assumptions.
SO012 MIT Technology Review How a top Chinese AI model overcame US sanctions Covers DeepSeek training using H800 chips despite US export restrictions.
SO013 MIT Technology Review What is next for Chinese open-source AI 2026 analysis of Chinese open-source AI ecosystem following DeepSeek impact.
SO014 CSIS DeepSeek, Huawei, Export Controls, and the Future of the US-China AI Race H800 training cost estimated at approximately $5.6M for final pre-training run.
SO015 ChinaTalk Deepseek: The Quiet Giant Leading China's AI Race Deepseek is fully funded by High-Flyer and has no plans to fundraise.
SO016 Financial Times How small Chinese AI start-up DeepSeek shocked Silicon Valley FT account of V2 disruption and DeepSeek origins.
SO017 Financial Times DeepSeek focuses on research over revenue in contrast to Silicon Valley DeepSeek focuses on research over revenue, contrasting with Silicon Valley commercialisation drive.
SO018 The Economist Behind DeepSeek lies a dazzling Chinese university Zhejiang University alumnus culture drives DeepSeek research team composition.
SO019 Liberation News DeepSeek sends shock waves across Silicon Valley Covers R1 shock wave to Silicon Valley AI investment orthodoxy.
SO020 CB Insights DeepSeek Products Competitors Financials Employees CB Insights unicorn profile listing $50B June 2026 valuation with Alibaba, CATL, Guozhitou as investors.
SO021 DeepSeek DeepSeek-R1 GitHub Repository Open-source release of DeepSeek-R1 model weights and documentation.
SO022 DeepSeek DeepSeek-V3 GitHub Repository Open-source release of DeepSeek-V3 model weights and training documentation.
SO023 DeepSeek DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models GRPO algorithm first introduced here; later scaled for R1 reasoning training.
SO024 DeepSeek DeepSeek-Coder: When the Large Language Model Meets Programming DeepSeek-Coder technical paper establishing the first flagship coding product line.
SO025 Third-party researchers Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Independent analysis of DeepSeek-V3 training challenges and hardware efficiency strategies.
SO026 TechCrunch DeepSeek could hit $45B valuation from its first investment round Potential valuation soared from $20B to $45B; round led by China Integrated Circuit Industry Investment Fund.
SO027 TechCrunch DeepSeek reportedly in talks to raise $1.5B, then IPO DeepSeek raised $7B at $50B valuation; in talks for $1.5B more at $71B; IPO targeting 2027.
SO028 CNBC Anthropic joins OpenAI in flagging 'industrial-scale' distillation campaigns by Chinese AI firms Anthropic accused DeepSeek of 16M-plus exchanges via 24,000 fraudulent accounts in coordinated distillation attacks.
SO029 CNBC No poaching our people: China's AI behemoth DeepSeek tells investors DeepSeek included a no-poaching clause protecting employees in its investment agreement.
SO030 Sensor Tower MMM: DeepSeek Outpaces AI Competitors In DAU Growth DeepSeek DAU growth over 700% WoW in Jan 22-28 2025; 23M-plus global downloads in 19 days.
SO031 Foundation for Defense of Democracies OpenAI Alleges China's DeepSeek Stole its Intellectual Property to Train its Own Models OpenAI Congressional memo alleges DeepSeek used fraudulent accounts and routers to extract training data from ChatGPT.
SO032 CSIS DeepSeek: A Deep Dive (Congressional Testimony) High-Flyer's roots in AI-enabled trading provided technical foundation in computing infrastructure and talent.
SO033 Stanford HAI How disruptive is DeepSeek? Stanford HAI faculty discuss China's new model DeepSeek is noticeably opaque when it comes to privacy protection, data-sourcing, and copyright.
SM001 Gartner Gartner Forecasts Worldwide AI Platforms and Models Market to Grow 63% in 2026 Worldwide end-user spending on AI models and platforms is projected to total $64 billion in 2026, up 63.4% from $39 billion in 2025.
SM002 Goldman Sachs Generative AI could raise global GDP by 7% GS Research estimates the total addressable market for generative AI software to be $150 billion.
SM003 State of AI Report State of AI Report 2025 OpenAI retains a narrow lead at the frontier, but competition has intensified as Meta reliquinshes the mantle to China’s DeepSeek, Qwen, and Kimi close the gap on reasoning and coding tasks.
SM004 Artificial Analysis AI Model & API Providers Analysis | Artificial Analysis Independent benchmarks across key performance metrics including quality, price, output speed & latency.
SM005 DeepSeek Your First API Call | DeepSeek API Docs The DeepSeek API uses an API format compatible with OpenAI/Anthropic.
SM006 DeepSeek 模型 & 价格 | DeepSeek API Docs 上下文长度 1M
SM007 Anthropic Plans & Pricing | Claude by Anthropic Opus 4.8 ... $5 / MTok ... $25 / MTok
SM008 OpenAI Business Pricing Business ... $20 / user / month
SM009 Google AI for Developers Gemini Developer API pricing
SM010 Alibaba Cloud What is Alibaba Cloud Model Studio - Alibaba Cloud Alibaba Cloud Model Studio offers Qwen and third-party models for text, image, audio, and video.
SM011 Alibaba Cloud Recommended models - Alibaba Cloud Alibaba Cloud Model Studio offers Qwen and third-party models for text, image, audio, and video.
SM012 Baidu Cloud 千帆大模型平台-企业级一站式大模型开发及应用开发平台-百度智能云 百度智能云千帆大模型平台是百度智能云推出的一站式企业级大模型平台
SM013 Baidu Cloud 百度千帆·大模型服务及Agent开发平台 -百度智能云
SM014 Z.AI New Released - Overview - Z.AI DEVELOPER DOCUMENT GLM-5.2 supports 1M lossless context.
SM015 BigModel 平台介绍 - 智谱AI开放文档 OpenAI SDK 兼容
SM016 MiniMax Pay as You Go MiniMax-M2.7 ... $0.3 / M tokens input and $1.2 / M tokens output.
SM017 MiniMax Models - MiniMax API Docs
SM018 Kimi API Platform Kimi K2.6 模型定价 - Kimi API 开放平台 模型上下文长度 256k,支持长思考擅长深度推理
SM019 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models
SM020 MOFCOM 商务部新闻发言人就调整《中国禁止出口限制出口技术目录》应询答记者问
SM021 Google Cloud Documentation DeepSeek models | Gemini Enterprise Agent Platform | Google Cloud Documentation DeepSeek models are available for use as managed APIs and self-deployed models on Gemini Enterprise Agent Platform.
SM022 Microsoft Azure Blog DeepSeek R1 is now available on Azure AI Foundry and GitHub | Microsoft Azure Blog DeepSeek R1 is now available in the model catalog on Azure AI Foundry and GitHub.
SM023 Amazon Web Services DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart
SM024 Vercel DeepSeek enters the fight for token volume, Anthropic continues to dominate spend DeepSeek’s share of tokens jumped from under 1% to 17% in a single month, while its share of spend stayed near 1%.
SM025 OpenRouter Blog DeepSeek V4 Is Earning Agentic Token Share — OpenRouter Blog A direct comparison between January and June 2026 shows just how quickly preferences can shift between model authors. DeepSeek effectively doubled its token share over the period (from 9% to 18%).
SP001 Artificial Analysis AI Model & API Providers Analysis | Artificial Analysis Independent benchmarks across key performance metrics including quality, price, output speed & latency.
SP002 DeepSeek Your First API Call | DeepSeek API Docs The DeepSeek API uses an API format compatible with OpenAI/Anthropic.
SP003 DeepSeek 模型 & 价格 | DeepSeek API Docs 上下文长度 1M
SP004 GitHub GitHub - deepseek-ai/DeepSeek-R1 We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1.
SP005 GitHub GitHub - deepseek-ai/DeepSeek-V3 We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token.
SP006 GitHub GitHub - deepseek-ai/DeepSeek-V2 DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
SP007 Anthropic Plans & Pricing | Claude by Anthropic Opus 4.8 ... $5 / MTok ... $25 / MTok
SP008 OpenAI API Pricing Business ... $20 / user / month
SP009 Google AI for Developers Gemini Developer API pricing
SP010 Alibaba Cloud What is Alibaba Cloud Model Studio - Alibaba Cloud Alibaba Cloud Model Studio offers Qwen and third-party models for text, image, audio, and video.
SP011 Alibaba Cloud Recommended models - Alibaba Cloud OpenAI-compatible ... Anthropic-compatible
SP012 Baidu Cloud 千帆大模型平台-企业级一站式大模型开发及应用开发平台-百度智能云 一站式企业级大模型平台
SP013 BigModel 平台介绍 - 智谱AI开放文档 OpenAI SDK 兼容
SP014 Z.AI New Released - Overview - Z.AI DEVELOPER DOCUMENT GLM-5.2 supports 1M lossless context.
SP015 MiniMax Pay as You Go MiniMax-M2.7 ... $0.3 / M tokens input and $1.2 / M tokens output.
SP016 MiniMax Models - MiniMax API Docs
SP017 Kimi API Platform Kimi K2.6 模型定价 - Kimi API 开放平台 模型上下文长度 256k,支持长思考擅长深度推理
SP018 Kimi API Platform Models - Kimi API 开放平台
SP019 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models.
SP020 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets Moonshot AI, the Beijing-based AI lab developing the popular Kimi series of open-weight models, has raised $2 billion at a $20 billion valuation.
SP021 Google Cloud Documentation DeepSeek models | Gemini Enterprise Agent Platform | Google Cloud Documentation DeepSeek models are available for use as managed APIs and self-deployed models on Gemini Enterprise Agent Platform.
SP022 Microsoft Azure Blog DeepSeek R1 is now available on Azure AI Foundry and GitHub | Microsoft Azure Blog DeepSeek R1 is now available in the model catalog on Azure AI Foundry and GitHub.
SP023 Amazon Web Services DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart
SP024 MiniMax MiniMax MiniMax M3 A frontier coding & agentic model built on a novel attention architecture (MSA) with 1M context
SP025 Kimi Kimi AI 官网 - K3 上线,专为智能体编程与知识工作打造 K3 上线,专为智能体编程与知识工作打造
SP026 Vercel Deepseek V4 on AI Gateway - Vercel DeepSeek V4 is now available on Vercel AI Gateway. There are 2 model variants: DeepSeek V4 Pro and DeepSeek V4 Flash.
SP027 OpenAI Models | OpenAI API All latest OpenAI models support text and image input, text output, multilingual capabilities, and vision.
SP028 Alibaba Cloud Model inference pricing - Alibaba Cloud Model API calls are billed on a pay-as-you-go basis by default.
SI001 DeepSeek 模型 & 价格 | DeepSeek API Docs 我们将根据模型输入和输出的总 token 数进行计量计费。
SI002 DeepSeek Token & Token Usage | DeepSeek API Docs Tokens are the basic units used by models to represent natural language text, and also the units we use for billing.
SI003 DeepSeek DeepSeek Terms of Use Product prices may change, and DeepSeek reserves the right to modify prices.
SI004 DeepSeek DeepSeek | 深度求索 DeepSeek-V4 New 发布时间 2026年4月24日
SI005 DeepSeek DeepSeek All systems are operating as expected.
SI006 TechCrunch DeepSeek could hit $45B valuation from its first investment round DeepSeek is in talks to raise its first round of venture capital, and in just a few weeks, its potential valuation has soared from $20 billion to $45 billion.
SI007 TechCrunch DeepSeek reportedly in talks to raise $1.5B, then IPO DeepSeek ... looks to raise around $1.5 billion in new funds at about a $71 billion valuation.
SI008 CNBC ‘No poaching’ our people, China's AI behemoth DeepSeek reportedly tells investors DeepSeek reportedly closed its first external funding round this week, which valued the AI lab at over $50 billion.
SI009 CNBC Anthropic joins OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms Anthropic joins OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms.
SI010 CSIS DeepSeek: A Deep Dive
SI011 CSIS Delving into the Dangers of DeepSeek DeepSeek exploded onto the AI scene in late January of this year.
SI012 Foundation for Defense of Democracies OpenAI Alleges China’s DeepSeek Stole its Intellectual Property to Train its Own Models OpenAI publicly released a memo ... alleging that DeepSeek had stolen its intellectual property to fuel its own models.
SI013 Forbes Liang Wenfeng Liang launched DeepSeek in 2023 and funded it in part with proceeds from High-Flyer.
SI014 Fortune Meet the hedge fund manager who founded DeepSeek DeepSeek founder Liang Wenfeng ... hails from the world of finance.
SI015 CB Insights DeepSeek - Products, Competitors, Financials, Employees, Headquarters Locations DeepSeek raised a total of $7.546B.
SI016 Vercel DeepSeek enters the fight for token volume, Anthropic continues to dominate spend DeepSeek’s share of tokens jumped from under 1% to 17% in a single month, while its share of spend stayed near 1%.
SI017 OpenRouter Blog DeepSeek V4 Is Earning Agentic Token Share — OpenRouter Blog DeepSeek effectively doubled its token share over the period (from 9% to 18%).
SI018 Google Cloud Documentation DeepSeek models | Gemini Enterprise Agent Platform | Google Cloud Documentation DeepSeek models are available for use as managed APIs and self-deployed models on Gemini Enterprise Agent Platform.
SI019 Microsoft Azure Blog DeepSeek R1 is now available on Azure AI Foundry and GitHub DeepSeek R1 is now available in the model catalog on Azure AI Foundry and GitHub.
SI020 Amazon Web Services DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart
SI021 Gartner Gartner Forecasts Worldwide AI Platforms and Models Market to Grow 63% in 2026 Worldwide end-user spending on AI models and platforms is projected to total $64 billion in 2026.
SI022 State of AI Report State of AI Report 2025 44% of U.S. businesses now pay for AI tools, up from 5% in 2023.
SI023 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models.
SI024 GitHub GitHub - deepseek-ai/DeepSeek-R1 We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1.
SI025 GitHub GitHub - deepseek-ai/DeepSeek-V3 DeepSeek-V3 ... 671B total parameters with 37B activated for each token.
SI026 China National Intellectual Property Administration 关于依法驳回抢注“DEEPSEEK”等相关商标注册申请的通告 依法对第82848449号“DEEPSEEK”等63件商标注册申请予以驳回。
SI027 DeepSeek Change Log | DeepSeek API Docs The two legacy API model names, deepseek-chat and deepseek-reasoner, will be discontinued in three months (2026-07-24).
SE001 DeepSeek Your First API Call | DeepSeek API Docs The DeepSeek API uses an API format compatible with OpenAI/Anthropic.
SE002 DeepSeek 模型 & 价格 | DeepSeek API Docs 支持非思考与思考模式(默认)
SE003 DeepSeek Token & Token Usage | DeepSeek API Docs Tokens are the basic units used by models to represent natural language text.
SE004 DeepSeek DeepSeek | 深度求索 DeepSeek-V4 New 发布时间 2026年4月24日
SE005 DeepSeek DeepSeek Terms of Use Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
SE006 GitHub GitHub - deepseek-ai/DeepSeek-V2 DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
SE007 GitHub GitHub - deepseek-ai/DeepSeek-V3 671B total parameters with 37B activated for each token
SE008 GitHub GitHub - deepseek-ai/DeepSeek-R1 first-generation reasoning models
SE009 Hugging Face deepseek-ai/DeepSeek-V2
SE010 Hugging Face deepseek-ai/DeepSeek-V3
SE011 Hugging Face deepseek-ai/DeepSeek-R1
SE012 arXiv DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
SE013 arXiv DeepSeek-V3 Technical Report
SE014 arXiv DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
SE015 Amazon Web Services DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart
SE016 Microsoft Azure Blog DeepSeek R1 is now available on Azure AI Foundry and GitHub DeepSeek R1 is now available in the model catalog on Azure AI Foundry and GitHub.
SE017 Google Cloud Documentation DeepSeek models | Gemini Enterprise Agent Platform | Google Cloud Documentation DeepSeek models are available for use as managed APIs and self-deployed models on Gemini Enterprise Agent Platform.
SE018 NVIDIA DeepSeek-R1 Now Live With NVIDIA NIM DeepSeek-R1 is an open model with state-of-the-art reasoning capabilities.
SE019 Vercel Deepseek V4 on AI Gateway - Vercel DeepSeek V4 is now available on Vercel AI Gateway.
SE020 DeepSeek DeepSeek All systems are operating as expected.
SE021 CSIS Delving into the Dangers of DeepSeek DeepSeek’s open-source structure means that anyone can download and modify the application.
SE022 CNBC Anthropic joins OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms Anthropic joins OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms.
SE023 Foundation for Defense of Democracies OpenAI Alleges China’s DeepSeek Stole its Intellectual Property to Train its Own Models OpenAI alleges China’s DeepSeek stole its intellectual property to train its own models.
SE024 Nature How disruptive is DeepSeek?
SE025 Z.AI New Released - Overview - Z.AI DEVELOPER DOCUMENT GLM-5.2 supports 1M lossless context.
SE026 Kimi API Platform Kimi K2.6 模型定价 - Kimi API 开放平台 模型上下文长度 256k,支持长思考擅长深度推理
SE027 MiniMax MiniMax MiniMax M3 A frontier coding & agentic model built on a novel attention architecture (MSA) with 1M context
SE028 MiniMax Pay as You Go
SE029 Google Cloud Blog DeepSeek R1 is available for everyone in Vertex AI Model Garden DeepSeek R1 is available for everyone in Vertex AI Model Garden.
SE030 DeepSeek DeepSeek Privacy Policy The Services are provided and controlled by Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
SE031 DeepSeek Transparency Below are DeepSeek's released models, including names, release dates, technical reports and model cards.
SU001 DeepSeek DeepSeek | 深度求索
SU002 DeepSeek Your First API Call | DeepSeek API Docs The DeepSeek API uses an API format compatible with OpenAI/Anthropic.
SU003 DeepSeek 模型 & 价格 | DeepSeek API Docs Json Output 支持 Tool Calls
SU004 DeepSeek DeepSeek | 深度求索 DeepSeek-V4 New 发布时间 2026年4月24日
SU005 DeepSeek DeepSeek App DeepSeek App
SU006 DeepSeek DeepSeek All systems are operating as expected.
SU007 Sensor Tower MMM: DeepSeek Outpaces AI Competitors In DAU Growth DeepSeek has now received over 23mn downloads, more than 2x of ChatGPT.
SU008 Appfigures DeepSeek Crossed a Million Downloads and is About to Challenge ChatGPT DeepSeek crossed a million downloads and is about to challenge ChatGPT.
SU009 Vercel DeepSeek enters the fight for token volume, Anthropic continues to dominate spend DeepSeek’s share of tokens jumped from under 1% to 17% in a single month, while its share of spend stayed near 1%.
SU010 Vercel Deepseek V4 on AI Gateway - Vercel DeepSeek V4 is now available on Vercel AI Gateway.
SU011 OpenRouter Blog DeepSeek V4 Is Earning Agentic Token Share — OpenRouter Blog DeepSeek effectively doubled its token share over the period (from 9% to 18%).
SU012 Amazon Web Services DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart
SU013 Microsoft Azure Blog DeepSeek R1 is now available on Azure AI Foundry and GitHub DeepSeek R1 is now available in the model catalog on Azure AI Foundry and GitHub.
SU014 Google Cloud Documentation DeepSeek models | Gemini Enterprise Agent Platform | Google Cloud Documentation DeepSeek models are available for use as managed APIs and self-deployed models on Gemini Enterprise Agent Platform.
SU015 Google Cloud Blog DeepSeek R1 is available for everyone in Vertex AI Model Garden DeepSeek R1 is available for everyone in Vertex AI Model Garden.
SU016 NVIDIA DeepSeek-R1 Now Live With NVIDIA NIM DeepSeek-R1 is an open model with state-of-the-art reasoning capabilities.
SU017 Alibaba Cloud What is Alibaba Cloud Model Studio - Alibaba Cloud Alibaba Cloud Model Studio offers Qwen and third-party models for text, image, audio, and video.
SU018 Baidu Cloud 千帆大模型平台-企业级一站式大模型开发及应用开发平台-百度智能云 一站式企业级大模型平台
SU019 BigModel 平台介绍 - 智谱AI开放文档 一站式模型即服务
SU020 TechCrunch DeepSeek: The countries and agencies that have banned the AI companys tech DeepSeek’s viral AI models and chatbot apps have been banned by a growing number of countries and government bodies.
SU021 Channel NewsAsia CNA Explains: Are countries banning DeepSeek for legitimate reasons? What are the main concerns, how big of a factor is geopolitics, and what are the implications for global AI and tech?
SU022 Al Jazeera Which countries have banned DeepSeek, and why? Which countries have banned DeepSeek, and why?
SU023 AICPB AI ChatBot Rankings by Users — Jun 2026 Edition The AI ChatBot Rankings for Website are based on Website Visits in Jun 2026.
SU024 AICPB China AI Rankings by Users — Jun 2026 Edition The China AI Rankings for App are based on App MAU in Jun 2026.
SU025 CNBC One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models One year after DeepSeek, Chinese AI firms from Alibaba to Moonshot race to release new models.
SU026 DeepSeek DeepSeek Terms of Use Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
SU027 The Conference Board State and Federal Governments Move to Ban DeepSeek on Government Devices Multiple states and Federal agencies have issued orders banning the use of the Chinese AI platform DeepSeek on government devices because of national security and privacy concerns.
SR001 DeepSeek DeepSeek Privacy Policy To provide you with our services, we directly collect, process and store your Personal Data in People's Republic of China.
SR002 DeepSeek DeepSeek Terms of Use We may add, upgrade, modify, suspend, or terminate services.
SR003 TechCrunch DeepSeek: The countries and agencies that have banned the AI company's tech DeepSeek’s viral AI models and chatbot apps have been banned by a growing number of countries and government bodies.
SR004 Channel NewsAsia CNA Explains: Are countries banning DeepSeek for legitimate reasons? What are the main concerns, how big of a factor is geopolitics, and what are the implications for global AI and tech?
SR005 Al Jazeera Which countries have banned DeepSeek, and why? Which countries have banned DeepSeek, and why?
SR006 The Conference Board State and Federal Governments Move to Ban DeepSeek on Government Devices Multiple states and Federal agencies have issued orders banning the use of the Chinese AI platform DeepSeek on government devices because of national security and privacy concerns.
SR007 CSIS Delving into the Dangers of DeepSeek DeepSeek’s open-source structure means that anyone can download and modify the application.
SR008 CSIS DeepSeek: A Deep Dive
SR009 CNBC Anthropic joins OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms Anthropic joins OpenAI in flagging industrial-scale distillation campaigns by Chinese AI firms.
SR010 Foundation for Defense of Democracies OpenAI Alleges China’s DeepSeek Stole its Intellectual Property to Train its Own Models OpenAI alleges China’s DeepSeek stole its intellectual property to train its own models.
SR011 Cornell Journal of Law and Public Policy U.S. AI Policy and the DeepSeek Problem
SR012 MOFCOM 商务部新闻发言人就调整《中国禁止出口限制出口技术目录》应询答记者问
SR013 China National Intellectual Property Administration 关于依法驳回抢注“DEEPSEEK”等相关商标注册申请的通告 依法对第82848449号“DEEPSEEK”等63件商标注册申请予以驳回。
SR014 DeepSeek DeepSeek All systems are operating as expected.
SR015 DeepSeek DeepSeek | 深度求索 DeepSeek-V4 New 发布时间 2026年4月24日
SR016 Amazon Web Services DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart DeepSeek-R1 model now available in Amazon Bedrock Marketplace and Amazon SageMaker JumpStart
SR017 Microsoft Azure Blog DeepSeek R1 is now available on Azure AI Foundry and GitHub DeepSeek R1 is now available in the model catalog on Azure AI Foundry and GitHub.
SR018 Google Cloud Documentation DeepSeek models | Gemini Enterprise Agent Platform | Google Cloud Documentation DeepSeek models are available for use as managed APIs and self-deployed models on Gemini Enterprise Agent Platform.
SR019 NVIDIA DeepSeek-R1 Now Live With NVIDIA NIM DeepSeek-R1 is an open model with state-of-the-art reasoning capabilities.
SR020 Vercel DeepSeek enters the fight for token volume, Anthropic continues to dominate spend DeepSeek’s share of tokens jumped from under 1% to 17% in a single month, while its share of spend stayed near 1%.
SR021 OpenRouter Blog DeepSeek V4 Is Earning Agentic Token Share — OpenRouter Blog DeepSeek effectively doubled its token share over the period (from 9% to 18%).
SR022 CNBC ‘No poaching’ our people, China's AI behemoth DeepSeek reportedly tells investors Founder Liang Wenfeng has a non-negotiable term for investors: no poaching DeepSeek’s staff.
SR023 State of AI Report State of AI Report 2025 Competition has intensified as ... DeepSeek, Qwen, and Kimi close the gap on reasoning and coding tasks.
SR024 Gartner Gartner Forecasts Worldwide AI Platforms and Models Market to Grow 63% in 2026 Worldwide end-user spending on AI models and platforms is projected to total $64 billion in 2026.
SR025 Microsoft Foundry AI Model Catalog | Microsoft Foundry Models Secure and managed by Microsoft: Purchase and manage models directly through Azure with a single license, consistent support, and no third-party dependencies.
SR026 DeepSeek DeepSeek-V3.1 Release | DeepSeek API Docs Introducing DeepSeek-V3.1: our first step toward the agent era!
SR027 Baidu Cloud 百度千帆·大模型服务及Agent开发平台 -百度智能云 全新的“百度千帆”以Agent为核心,为企业提供模型、Agent开发及数据智能服务等一站式服务。
SR028 Kimi API Platform OpenAI API 协议兼容性提示 - Kimi API 开放平台 只需要将 base_url 和 api_key 替换成 Kimi 大模型的配置,即可无缝将你的应用和服务迁移至使用 Kimi 大模型。
SR029 MiniMax Aligning to What? Rethinking Agent Generalization in MiniMax M2 Rethinking Agent Generalization in MiniMax M2
SR030 Volcengine 火山引擎-你的AI云 Agent适配 豆包大模型 1.8
SR031 MiniMax Why Did MiniMax M2 End Up as a Full Attention Model? Why Did MiniMax M2 End Up as a Full Attention Model?
SR032 DeepSeek Model Mechanism and Training Methods of DeepSeek This will help you use DeepSeek more effectively while ensuring your right to know and control during usage, thereby mitigating risks associated with improper use of the model.
SV001 Gartner Gartner Forecasts Worldwide AI Platforms and Models Market to Grow 63% in 2026 Worldwide end-user spending on AI models and platforms is projected to total $64 billion in 2026.
SV002 Goldman Sachs Generative AI could raise global GDP by 7% GS Research estimates the total addressable market for generative AI software to be $150 billion.
SV003 Artificial Analysis AI Model & API Providers Analysis | Artificial Analysis Independent benchmarks across key performance metrics including quality, price, output speed & latency.
SV004 DeepSeek 模型 & 价格 | DeepSeek API Docs 百万tokens输入(缓存未命中) 3元 / 百万tokens输出 6元
SV005 DeepSeek Token & Token Usage | DeepSeek API Docs Tokens are the basic units ... and also the units we use for billing.
SV006 TechCrunch DeepSeek could hit $45B valuation from its first investment round DeepSeek is in talks to raise its first round of venture capital ... at $45 billion.
SV007 CNBC ‘No poaching’ our people, China's AI behemoth DeepSeek reportedly tells investors DeepSeek reportedly closed its first external funding round this week, which valued the AI lab at over $50 billion.
SV008 TechCrunch DeepSeek reportedly in talks to raise $1.5B, then IPO DeepSeek ... looks to raise around $1.5 billion in new funds at about a $71 billion valuation.
SV009 CB Insights DeepSeek - Products, Competitors, Financials, Employees, Headquarters Locations DeepSeek raised a total of $7.546B.
SV010 Vercel DeepSeek enters the fight for token volume, Anthropic continues to dominate spend DeepSeek’s share of tokens jumped from under 1% to 17% in a single month, while its share of spend stayed near 1%.
SV011 OpenRouter Blog DeepSeek V4 Is Earning Agentic Token Share — OpenRouter Blog DeepSeek effectively doubled its token share over the period (from 9% to 18%).
SV012 Forbes Liang Wenfeng Liang launched DeepSeek in 2023 and funded it in part with proceeds from High-Flyer.
SV013 SiliconANGLE Report: Chinese AI startup MiniMax raises $600M at $2.5B valuation led by Alibaba MiniMax raises $600M at $2.5B valuation.
SV014 TechNode MiHoYo-backed AI firm MiniMax jumps on Hong Kong debut market capitalisation above HK$90 billion ($11.5 billion)
SV015 KrASIA StepFun nears USD 2.5 billion pre-IPO round as industrial investors join StepFun is set to complete a funding round of nearly USD 2.5 billion as it accelerates its listing process.
SV016 The Standard Stepfun, China's AI Six Tigers, finishes new US$2.5b funding round for HK IPO completed a new US$2.5 billion funding round
SV017 Qiming Venture Partners China’s AGI Pioneer and Leader Z.ai Listed onHong Kong Stock Exchange becoming the world’s first listed large language model company.
SV018 Yicai Global Zhipu AI Soars in Hong Kong Stock Market Debut as Chinese Startup Becomes World's First LLM Firm to Go Public became the world’s first large language model company to go public.
SV019 Google Cloud Agent Platform Pricing | Google Cloud Gemini 3.1 Pro Preview ... $2 input ... $12 text output per 1M tokens.
SV020 MiniMax API Docs Quick Start - MiniMax API Docs Quickly test MiniMax M3 with the Claude SDK
SV021 MiniMax Pay as You Go MiniMax-M2.7 ... $0.3 / M tokens input and $1.2 / M tokens output.
SV022 Kimi API Platform Kimi K2.6 模型定价 - Kimi API 开放平台 模型上下文长度 256k,支持长思考擅长深度推理
SV023 State of AI Report State of AI Report 2025 competition has intensified as ... DeepSeek, Qwen, and Kimi close the gap on reasoning and coding tasks.
SV024 TechCrunch China's Moonshot AI raises $2B at $20B valuation as demand for open source AI skyrockets Moonshot AI ... has raised $2 billion at a $20 billion valuation.
SV025 The Conference Board State and Federal Governments Move to Ban DeepSeek on Government Devices Multiple states and Federal agencies have issued orders banning the use of the Chinese AI platform DeepSeek on government devices.
SV026 China National Intellectual Property Administration 关于依法驳回抢注“DEEPSEEK”等相关商标注册申请的通告 依法对第82848449号“DEEPSEEK”等63件商标注册申请予以驳回。
SV027 DeepSeek DeepSeek Privacy Policy we directly collect, process and store your Personal Data in People's Republic of China.
SV028 DeepSeek DeepSeek Terms of Use We may add, upgrade, modify, suspend, or terminate services.
SV029 Microsoft Foundry AI Model Catalog | Microsoft Foundry Models Secure and managed by Microsoft: Purchase and manage models directly through Azure with a single license.
SV030 CB Insights DeepSeek - Products, Competitors, Financials, Employees, Headquarters Locations Competitors of DeepSeek include OpenAI, Anthropic, Cognition, OpenRouter, Moonshot AI and 7 more.