初创公司尽调
尽调报告 Artificial Intelligence / AI for Scientific Research Seed 2026-07-02

Mirendil

前沿 AI 科学实验室——顶级创始人、尚无产品、种子轮即独角兽,估值风险极高

Mirendil 拿着前 Anthropic/Google 顶级创始团队和 a16z、Kleiner Perkins、NVIDIA 一线背书,却在没有产品、基准、收入和客户时拿到约 $1B 种子估值。人才和逻辑都成立,但价格几乎全靠创始人履历和 TAM 支撑。当前据报价格下应继续研究 / 观察;等首批技术证据和设计伙伴证据出现后再重估。

封面要素

种子轮融资额 01
200 USD M [CO004]
投后估值 02
1000 USD M [CO007]
成立时间 03
Early 2026 [CO009]
领投方 04
Andreessen Horowitz & Kleiner Perkins (NVIDIA participating) [CO005, CO006]
创始团队 05
20 people [CO003]
产品 / 收入 06
None public (pre-product, pre-revenue) [CO021, CO022]

公司概况

Mirendil 是一家旧金山前沿 AI 实验室,2026 年 6 月底首次公开亮相,由 2025 年底离开 Anthropic 的研究员在 2026 年初创立。联合创始人兼 CEO Behnam Neyshabur 曾共同领导 Anthropic 的 Discovery 团队,并在 Google 参与 Blueshift、Minerva 和 Gemini 数学 / 代码推理工作(Sharpness-Aware Minimization(SAM)优化器共同发明人);联合创始人兼 CTO Harsh Mehta 搭建过 Anthropic 自动研究平台的第一版,此前任职 Google DeepMind。公司描述的是一个递归式改进闭环——更好的模型产出更好的研究,更好的研究又产出更好的模型——并希望把实验室级自主研究系统打包给外部科学家和领域专家使用。公司以约 $1B 估值完成 $200M 种子轮,被称为迄今披露的最大 AI 种子轮之一,由 Andreessen Horowitz 和 Kleiner Perkins 领投,NVIDIA 参投。截至本次报告日期,Mirendil 只披露了 20 人创始团队,没有产品、基准测试、收入、客户或董事会构成。

官网
mirendil.com
成立时间
2026-01-01
创始人
Behnam Neyshabur, Harsh Mehta, Shayan Salehian & Tara Rezaei
创立地点
San Francisco, CA, USA
总部
San Francisco, CA, USA
产品
截至本次报告日期,Mirendil 尚未发布公开产品、SKU、基准测试、演示或路线图。Mirendil 在命题层面描述了专门面向 AI 研究的前沿模型,以及覆盖实验设计、编码、调试、算力管理和检查点对比的自主研究闭环;目标用户先是 AI 研究员和工程师,之后扩展到生物、化学、材料科学、药物发现和机器人领域的科学家。所有产品 / 技术描述均来自公司主张,尚未得到验证。
客户
AI 研究员和工程师优先;随后面向生物、化学、材料科学、药物发现和机器人等领域的科学家与研发团队
商业模式
尚未披露且仍处商业化前阶段;公开资料没有记录定价、收入模式或合同
阶段
Seed
融资情况
$200M 种子轮于 2026 年 6 月宣布,估值约 $1B,由 Andreessen Horowitz 和 Kleiner Perkins 领投,NVIDIA 参投;2026 年 3 月报道曾称其以同一估值寻求约 $175M
[CO001, CO010, CO011, CO004, CO007]

执行摘要

主要优势

  • 创始人与市场高度匹配:CEO Behnam Neyshabur(前 Anthropic Discovery 联合负责人;Google Minerva/Gemini;SAM optimizer)和 CTO Harsh Mehta(搭建 Anthropic 首个 autoresearch 平台)都直接来自公司要攻的前沿领域
  • 种子轮资本极厚:$200M 是迄今披露的最大 AI 种子轮之一,足以在收入压力到来前支撑多年、算力很重的实验室基础设施建设
  • 顶级背书:Andreessen Horowitz 和 Kleiner Perkins 联合领投,NVIDIA 参与,释放出对科学 AI 的信心,也可能带来算力 / 硬件协同
  • 大且真实的顺风:AI-for-science 和 agentic AI 市场在快速扩张(agentic AI 到 2031 年约 42% CAGR;AI-in-drug-discovery 为双位数 CAGR),自动化 R&D 是可信的前沿方向
  • 20 人创始团队来自 Anthropic、xAI、Google DeepMind、OpenAI,稀缺 frontier-ML 人才集中度高

主要风险

  • 估值约 $1B,却仍无产品、无收入:价值几乎全压在未验证逻辑和创始人履历上,没有基准、demo 或客户来承销这个价格
  • 关键人集中:可信度和执行高度依赖 Neyshabur 与 Mehta;公司未披露继任、董事会或治理层安排
  • 赛道拥挤且资金充足:Periodic Labs($300M 种子轮、目标估值约 $7B)、Lila Sciences(已融资 $550M)、Isomorphic Labs(DeepMind,Lilly/Novartis 交易)和既有巨头已经有资金、收入或合作,Mirendil 暂时没有
  • 技术 / 可靠性风险:严谨科学工作需要的 agentic AI 可靠性和 LLM 幻觉问题仍未解决;Gartner 预计到 2027 年,超过 40% 的 agentic-AI 项目会被取消
  • 估值下修和 AI 泡沫暴露:未经验证的独角兽种子轮,一旦资本市场收紧或里程碑延误,很容易被重置
  • 宏观买方逆风:美国 FY2026 NSF 预算拟削减,可能压制公共科学 R&D 预算,而后者最终会资助科学 AI 采用

未决问题

  • 产品和技术证据:没有公开基准、评估方法、demo 环境,也没有 autonomous AI-R&D 系统路线图
  • 收入、定价和商业化:没有收入模型、定价、pipeline、设计伙伴或任何客户证据
  • 治理和股权结构:董事会构成、投资人权利、投票控制、老股交易或法律实体细节均未公开
  • 烧钱和现金跑道:虽已融资 $200M,但月度烧钱(算力 + 精英人才薪酬)未披露,现金跑道无法估算
  • 20 人创始团队之外的员工数和办公地点、招聘计划、办公室网络均未披露
  • 监管姿态:若 Mirendil 进入受监管科学领域,公司未披露 EU AI Act / FDA 暴露的分类或路线图

目录

Chapter 01

01公司概览

1.1 身份、使命与当前定位

Mirendil 把自己定位为前沿 AI 实验室,而不是打包软件供应商。它的官网、Andreessen Horowitz 投资札记和 Kleiner Perkins 介绍,都指向同一个核心命题:打造极擅长 AI 研发的系统,再围绕这些系统重塑研究闭环,让研究更快、更强、更自主。它的野心异常宽。Mirendil 并不瞄准单一垂直流程,而是称要让前沿 AI 研发大众化,使从事生物、化学、药物发现、材料科学和机器人研究的科学家与领域专家,不必先把自己变成前沿 AI 实验室。尽调中这个框架很关键:它把公司放在实验室级平台的位置,潜在上行巨大;但也意味着,当前公开产品披露主要还是使命叙事,并非带有基准、路线图节点或合同背书的可购买 SKU。[CO001, CO002, CO019, CO020, CO027, CO028]

快照 KPI 表
指标数值 / 状态日期置信度缺口
成立时间2026 年初2026-03 to 2026-06公开来源支持这一时期,但不支持精确注册日期
总部San Francisco, California2026-06 to 2026-07官方官网没有发布地址或实体页脚
阶段种子阶段私人前沿 AI 实验室2026-07-02没有公开经审计财务披露或正式阶段备忘录
最新融资已宣布 $200M 种子轮2026-06-24 至 2026-06-26具体交割机制和持股比例仍为私有信息
报道估值~$1B 投后估值2026-06-25没有公开披露条款清单、股权结构表或董事会权利
披露团队规模20 人创始团队2026-06-24 至 2026-07-02没有发布更广义员工数量或招聘计划
公开产品披露仅有使命和实验室平台论点2026-07-02没有基准测试包、技术路线图或商业 SKU 命名
收入 / 客户披露未公开2026-07-02后续章节需要管理层或客户证据,而不是营销文案

汇总官方、投资人和独立媒体事实;除融资和团队规模外,运营指标大多未披露。

[CO004, CO007, CO008, CO021, CO022, CO032]
FO002: 公司快照逻辑

创始人履历、资本、实验室基础设施和外部科学家如何拼出 Mirendil 的投资论点。

[CO019, CO020, CO037, CO043, CO045, CO046]

1.2 创始人、团队梯队与治理透明度

创始人与市场的匹配度,是 Mirendil 最清晰的优势。Behnam Neyshabur 公开以 Mirendil 联合创始人兼 CEO 身份出现,此前任职 Anthropic 和 Google;他的 CV、学者主页和论文记录,验证了他在优化、Minerva 和 Gemini 相关工作上的前沿模型资历。Harsh Mehta 在投资人与公司介绍中被描述为 Mirendil 联合创始人兼 CTO,此前任职 Anthropic 和 Google DeepMind,并在优化和机器学习上有研究记录。投资人文章还突出来自 xAI 的创始运营人员 Shayan Salehian,以及来自 MIT 和 OpenAI 的 Tara Rezaei。因此,公开团队叙事在履历和技术深度上很强。治理叙事则相反。官网、投资人文章和主要首发报道都没有公开董事会名单、治理权利披露、独立董事信号或详细法律实体版图,关键人和控制权问题仍未解决。[CO003, CO010, CO011, CO012, CO013, CO014]

领导层与创始人表
人物角色背景为何重要依赖 / 缺口
Behnam Neyshabur联合创始人兼 CEO前 Anthropic Discovery 联合负责人;此前在 Google 研究 Blueshift、Minerva 和 Gemini 相关工作;SAM 共同发明人带来深厚的前沿模型、优化和 AI-for-science 可信度关键人物依赖度高,未披露继任层
Harsh Mehta联合创始人兼 CTO前 Anthropic 研究员,此前有 Google DeepMind 和优化研究背景用直接技术执行经验支撑自动化 AI 研发和实验室系统论点公开履历比 CEO 更薄,治理角色未披露
Shayan Salehian创始工程师 / ML 系统班底前 xAI 和 X/Twitter 工程师,覆盖后训练、推理和基础设施在创始人之外补充前沿 ML 工程和系统深度没有披露正式头衔、组织范围或留任条款
Tara Rezaei创始工程师 / 研究班底MIT 毕业,投资人材料称曾任 OpenAI 研究员补充精英初级研究人才,也释放招聘信号除投资人简介外,公开运营细节很少
治理层董事会 / 控制权未公开未发现公开独立董事、观察员或委员会结构公司尚未有产品却已高度资本化,因此这一点很重要董事会权利、投票控制和法律实体结构需要管理层披露

公开来源支持创始人和创始班底履历,但不支持完整高管委员会、董事会名单或汇报结构。

[CO010, CO011, CO012, CO013, CO014, CO015]

1.3 融资、投资方与利益相关者图谱

Mirendil 的融资规模已经足以设定本报告后续预期。官方与独立来源一致显示,公司在 2026 年 6 月宣布以约 $1B 估值完成 $200M 种子轮,由 Andreessen Horowitz 和 Kleiner Perkins 领投,NVIDIA 参投。多家媒体称这是迄今披露的最大 AI 种子轮之一。2026 年 3 月的报道已称 Mirendil 正以同一估值寻求约 $175M,因此 6 月交割看起来是最终轮次规模更大,而不是一段完全不同的融资故事。这笔资本让 Mirendil 在收入压力到来前,有充足空间搭建算力密集的实验室基础设施;若公开数字大体正确,简单投后估算显示新投资人可能持有约五分之一股权。即便如此,公开材料没有披露股权结构表细节、董事会权利、老股转让、债务,或任何通常用来锚定此类估值的商业化里程碑。[CO004, CO005, CO006, CO007, CO023, CO024]

利益相关方或投资人地图
利益相关方角色控制 / 经济重要性为何重要尽调要求
Andreessen Horowitz领投 / 共同领投投资人公开宣布为种子轮领投方提供资本、AI 生态触达,以及将前沿 AI 研发外部化的明确论点确认董事席位、持股比例和按比例跟投权
Kleiner Perkins领投 / 共同领投投资人公开宣布为首日支持方增加风投信号,并公开详细表达对创始团队的信心确认董事会权利和后续融资储备
NVIDIA战略参与方最强来源将其列为参与方,而非公开共同领投方释放算力协同和未来基础设施重要性的信号澄清投资是否包含商业算力承诺
Behnam Neyshabur 与 Harsh Mehta创始人控制集团在愿景、招聘和技术方向上公开居于中心这一阶段创始人控制权可能比当前产品收入更重要要求披露投票控制和归属期细节
20 人创始团队人才集中除资本外,主要公开运营资产执行风险高度取决于能否留住这支很小的精英班底要求披露留任、招聘和移民风险细节
外部科学家 / AI 构建者未来用户群体公司和投资人叙事中明确点出的受益者客户相关性取决于 Mirendil 能否服务外部用户,而不只是服务自己要求设计伙伴、试点或候补名单证据

基于官网、投资人帖子和发布报道整理;经济权利和持股比例仍为私有信息。

[CO004, CO005, CO006, CO024, CO025, CO037]
FO003: 快照 KPI

用少数足够扎实的公开事实,概括 Mirendil 当前成熟度。

KPI 包混合了官方、投资人和第三方资料事实,因为 Mirendil 尚未发布单一、经过审计的运营快照。

[CO004, CO007, CO008, CO021, CO022, CO023]

1.4 里程碑、规模信号与未解尽调负担

Mirendil 太新,公开里程碑高度集中。证据链从创始人 2025 年底离开 Anthropic,到 2026 年 3 月融资报道,再到 6 月底投资人公告和公司发布材料。少数可用规模信号集中在创始团队,而非商业牵引:官网和投资人札记称创始团队 20 人,来自主要前沿实验室,但没有披露更广义的员工规模、办公室网络、客户名单、收入或基准测试包。因此,后续章节可以安全复用本章的身份、创始人、融资和使命主张,但大多数运营指标应视为开放尽调项,而非既定事实。主要负面解读很直接:Mirendil 有顶级人才和异常深厚的资本化,但公开市场仍缺少足够证据,无法支撑产品就绪度、治理质量或商业转化判断。后续每一章都应受这种不对称性约束,因为几乎所有前瞻性判断都取决于私下尽调,而不是公开证明。[CO008, CO009, CO021, CO022, CO024, CO025]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2025-12创始人离开 Anthropic,并开始 Mirendil 组建期创立可推断处于组建期Behnam Neyshabur、Harsh Mehta解释了公司为何在 2026 年带着前沿实验室论点突然出现
2026-03-18报道称发布前已进行融资谈判融资拟募 $175M,估值 ~$1BTechStartups、Complete AI Training显示投资人信心早于公开发布
2026-06-24a16z 投资公告发布合作领投公告公开Andreessen Horowitz、Mirendil确认种子轮领投方和外部平台论点
2026-06-25官网和使命声明公开产品公开发布 / 走出隐身模式Mirendil首次正式阐述产品论点和目标用户
2026-06-25种子轮被广泛报道为已关闭融资$200M,估值约 $1BMirendil、a16z、KP、NVIDIA、媒体资本基础足以支撑算力密集型研究
2026-06-25创始团队构成披露规模20 名研究员和工程师作为创始团队规模标记被点名Mirendil、投资方发布时唯一公开的运营规模指标
2026-06-25科学用例叙事发布合作强调生物 / 化学 / 机器人和 AI 构建者受益Mirendil、a16z、Kleiner Perkins(公司与领投方)将公司定位为服务外部领域专家的平台,而不只是内部实验室
2026-06-25批评性评论质疑没有产品或收入披露时的估值负面审视升温Cryptonomist 及类似评论突出产品前独角兽种子轮的核心承销风险
2026-07-02治理和法律披露仍然稀疏治理未发现公开董事会或详细实体披露Mirendil 公开材料依赖控制权假设前,需要直接管理层尽调

Mirendil 很新,多个里程碑集中在 2026 年 6 月;公开时间线在融资和使命上较强,在法律、治理和商业后续上较弱。

[CO009, CO024, CO033, CO034, CO035, CO036]
FO001: 公司里程碑时间线

Mirendil 公开亮相后第一个季度里可见的创立、融资、发布和即时审视里程碑。

[CO004, CO006, CO009, CO024, CO033, CO034]

1.5 展项

Chapter 02

02市场分析

2.1 市场边界:品类仍由邻近市场定义,而非自成一类

目前没有分析机构专门测算前沿实验室级、跨领域 AI 研究自动化平台市场——也就是 Mirendil 自称所在的品类。相反,可服务机会只能由至少三个相互重叠、口径不同的商业品类,再加两个大型现状预算池来框定。Grand View Research 对“药物发现中的 AI”的定义,覆盖面向制药、生物技术和 CRO 销售的分子筛选、靶点识别、药物优化、从头设计和临床前测试软件与服务。Mordor Intelligence 的“智能体 AI 市场”宽得多,也大多不属于科学领域,横跨客户服务、IT、制造和金融服务自动化。Dimension Market Research 另行测算“自主化学实验室”市场,覆盖实体实验室机器人、闭环实验硬件和编排软件——这是最接近 Mirendil 所称重塑物理 + 计算研究闭环的类比市场。三者背后还有 Mirendil 式工具最终必须增强或替代的现状预算池:2024 年全球制药研发支出约 $288B,其中仅欧洲就约 €55B,另有一个独立且目前在收缩的美国联邦科学资金池。这五个视角都不是 Mirendil 的市场;每一个都只是边界代理。[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 Mirendil 的相关性
AI-for-science / 跨领域研究自动化平台(Mirendil 自称类别)AI 模型、智能体工具和算力,专门用于设计、运行并解读生物 / 化学 / 材料领域实验通用湿实验室资本开支、与研发工作流无关的通用 LLM 聊天订阅制药 / 生物技术 / 材料研发领导层、国家实验室Mirendil 瞄准的核心类别;尚无独立规模测算
AI 药物发现软件与服务面向制药 / 生物技术 / CRO 销售的分子设计、靶点识别、ADMET 预测、从头设计软件湿实验室自动化硬件本身、CRO 人力成本、非制药科学制药 / 生物技术研发、合同研究组织披露规模最大的相邻软件市场;与 Mirendil 的生物 / 药物发现用例重叠
Agentic AI / 企业 AI 智能体平台覆盖客户服务、IT、金融、制造等行业的通用自主智能体软件领域专属科学推理、物理湿实验室集成各行业企业 IT / 运营买方标题市场大得多,但大多不是科学;若直接套用,会暴露规模测算口径风险
自主 / 自驱动实验室硬件 + 软件实验室机器人、闭环实验平台、云实验室编排没有物理实验室控制的纯软件 AI 推理模型制药、化学和材料科学运营负责人与 Mirendil「构建系统的系统」论点最接近的物理实验室类比
全球药企研发预算(现状替代)药企内部发现、临床前和临床研发支出营销、制造和商业化支出药企 CFO / Chief R&D OfficerMirendil 式工具最终必须增强或替代的预算池
联邦与慈善基础研究资金(相邻现状替代)政府科研经费(如 NSF)资助的学术 AI-for-science 研究企业 / 私人实验室支出政府机构、高校 PI相邻且当前波动的资金池;也是学术采用路径上的政策风险

各行综合了五个不同第三方来源(SM001、SM003、SM025、SM014、SM020)描述的类别边界;没有单一来源定义 Mirendil 自己的类别,因此本表是作者的边界逻辑,而不是某家发布方的市场划分。

[CM001, CM002, CM003, CM004, CM005, CM006]

2.2 规模测算视角:三个重叠市场,没有单一 TAM

即便名义上是同一品类,公开规模测算也分歧很大。Grand View Research 估算 2026 年全球药物发现 AI 市场为 $2.9B,到 2033 年按 24.8% CAGR 增至 $13.8B。Precedence Research 对同一名义品类的估算是 2026 年 $7.62B,到 2035 年按 9.90% CAGR 增至 $17.81B——2026 年数字约为 Grand View 的 2.6 倍;差距反映的是范围和基准年假设不同,而不是简单预测差异。智能体 AI 估算更接近:Mordor Intelligence 的 $9.89B(2026)和 Fortune Business Insights 的 $9.14B(2026)相差不到 8%,不过两者纳入的大多是非科学企业智能体。实体实验室层更小——Dimension Market Research 将 2026 年全球自主化学实验室市场定为 $5.75B,到 2035 年按 14.5% CAGR 增至 $19.48B。Mirendil 的具体跨领域品类没有公开测算,因此最好的可用代理是投向可比公司的资本:到 2026 年中,Mirendil($200M 种子轮)、Periodic Labs($300M 种子轮,正洽谈再融 $500M、估值 $7.5B)和 Lila Sciences(累计 $550M、估值 $1.3B+)合计已披露超过 $1B 私募资本;与此同时,Excelra 估计过去十多年全行业 AI 药物发现累计投资超过 $20B。McKinsey 自己的采纳调查显示,真实部署远远落后于这些金额预测。[CM009, CM010, CM011, CM012, CM013, CM014]

TAM/SAM/SOM 或市场规模视角表
发布方年份 / 地理范围数值CAGR方法 / 范围置信度限制
Grand View Research2026(基准 2025);全球$2.9B (2026) -> $13.8B (2033)24.8% (2026-33)AI 药物发现软件 / 服务,按应用和治疗领域自下而上测算药物发现范围较窄;不包括物理实验室和非制药科学
Precedence Research2026;全球$7.62B (2026) -> $17.81B (2035)9.90% (2026-35)与 Grand View 名义类别相同,但基准年定义更宽其 2026 年「同一」市场规模约为 Grand View 的 2.6x,说明规模测算对定义敏感,不只是时间差
Mordor Intelligence2026 年 1 月更新;全球$9.89B (2026) -> $57.42B (2031)42.14% (2026-31)跨所有行业的智能体 AI 市场,不限科学包含非科学企业智能体(IT、客户服务、金融);若直接套用,会高估 Mirendil 可获得的份额
Fortune Business Insights2026;全球$9.14B (2026) -> $139.19B (2034)40.50% (2026-34)智能体 AI 市场,另一套供应商方法论印证 Mordor 对广义智能体 AI 的数量级判断;同样要注意不属于科学专用口径
Dimension Market Research(市场研究机构)2026;全球$5.75B (2026) -> $19.48B (2035)14.5% (2026-35)自主 / 自驱动化学实验室硬件 + 软件仅覆盖化学实体实验室;不含生物、材料和纯软件智能体
EFPIA2025 年报告(覆盖 2024 年);欧洲约 €55B R&D(2024)n/a欧洲制药行业协会会员调查仅欧洲;不含美国 / 亚洲制药企业和非制药科学 R&D
BioSpace(Evaluate Pharma 来源)2025 年文章;全球制药 R&D 约 $288B(2024)n/a汇总制药公司的公开 R&D 披露仅制药;不含 Mirendil 同样瞄准的材料、化学、能源 R&D
Congressional Research Service(经 Congress.gov)2026 年 1 月;美国联邦FY2026 申请 $3.9B,FY2025 已批准 $9.06B(-56.9%)n/aCRS 对 NSF 预算申请和拨款历史的分析仅 NSF;其他机构(NIH、DOE)情况不同,最终批准金额也可能偏离申请额
Excelra2026;全球累计 AI 药物发现投资 >$20B(私募资本,不是年收入)n/a追踪「十余年」累计 VC / 合作资本累计资本存量,不是年度可服务市场数字;不能与上方行直接比较
McKinsey(QuantumBlack 调查)2025 年 11 月;全球(105 个国家)23% 的组织已扩展 >=1 个智能体用例;任一职能内扩展比例 <=10%n/a1,993 名受访者调查,按各国 GDP 贡献加权采用率代理指标,不是以美元计价的市场规模;显示部署落后于上方预测

这些行有意混合年度市场预测、累计资本数字和采用率百分比,目的是说明没有单一美元数字能代表 Mirendil 的可服务市场;不要把这些行相加。

[CM009, CM010, CM012, CM013, CM014, CM006]
FM001: 市场规模测算视角

三层受约束的规模测算,从现状研发预算基数,收窄到已披露投入 Mirendil 最窄直接可比对象的资本。

第 2 层是作者把两份方法无关的分析师报告相加;第 3 层是已披露私募融资(资本承诺替代指标),不是独立研究出的市场规模。

[CM005, CM015, CM017, CM018, CM019, CM020]
FM002: 市场估算区间

独立发布方对同一名义 2026 年数量给出的低 / 中 / 高区间显示,药物发现估算分歧很大,而智能体 AI 估算趋同。

中值是两个引用点估计的简单平均,并非独立来源;第二行混合 2033 和 2035 年预测终点,只为说明数量级分歧,不代表单一年份预测。

[CM009, CM010, CM011, CM012, CM013]

2.3 科学领域中的买方、用户和付款方分层

制药和生物技术研发组织是最清晰可识别的买方群体:首席研发 / 数字官掌握预算,实验台科学家和计算生物学科学家是日常用户,采纳触发点通常是基于里程碑的合作,而不是席位许可销售。Isomorphic Labs 2024 年 1 月与 Eli Lilly 的合作具体展示了这种模式——$45M 预付款,加上最高 $1.7B 的里程碑付款和分层特许权使用费,用于多靶点小分子项目;这种结构更像生物技术合作经济学,而不是企业 SaaS。ZS 对 115 名制药 / 生物技术高管的 2026 年调查显示,这一群体仍处早期:只有 17% 称 AI 投资在研究和发现环节产生了可衡量价值(临床开发为 29%),尽管 41% 正计划用智能体自动化整个研发发现流程。材料科学和化学研发是第二个、更依赖硬件的群体,由自主实验室供应商服务。政府科学机构构成第三个、政治波动更大的群体,项目经理代表学术 PI 掌握预算,采纳取决于国会拨款,而不是直接采购。第四个容易被忽略的群体是前沿 AI 实验室本身——Periodic Labs 和 Lila Sciences 是 Mirendil 最接近的可比公司,目前用自己的风投资金支持的算力预算来建设这类工具,以加速自身研究闭环;Lila 的公开信息已经释放了对外销售意图,这条商业化路径也由 Mirendil 的使命所暗示,但 Mirendil 尚未用具名客户证明。[CM023, CM024, CM025, CM026, CM027, CM029]

细分市场 / 买方图谱
细分市场买方用户付款方工作流预算负责人采用触发因素
大 / 小分子制药 R&D首席 R&D / 数字化官实验台科学家、计算生物学家制药 R&D 预算靶点识别 -> 先导优化 -> 临床前R&D 副总裁 / 首席科学官里程碑式合作或平台授权(参照 Isomorphic Labs / Eli Lilly)
生物技术 / TechBio 平台公司创始人 / CTO内部 AI 和湿实验室团队VC 加制药里程碑付款搭建专有发现引擎CEO / CFO需要建立可防守的数据护城河,对抗商品化基础模型
材料科学与化学 R&D企业 R&D 总监工艺 / 材料科学家企业 R&D 预算假设生成 -> 合成 -> 测试循环CTO / 首席科学家推动材料发现周期提速(电池、半导体)
国家实验室 / 政府研究项目经理政府科学家联邦拨款(NSF / DOE / NIH)基础科学发现、开放发表机构预算官 / 国会政策优先级变化(即使其他基础科学被削减,AI 仍有专项例外)
学术研究机构大学研究办公室博士生、博士后、首席研究员资助金(联邦加私人慈善)假设驱动的实验台研究持有资助的 PI能否获得可负担算力 / API,以及不断抬高的算力成本门槛
前沿 AI 实验室自身(内部使用)实验室研究管理层研究科学家和工程师实验室由风险资本支撑的自有算力 / 运营预算用 AI 加速实验室自身模型研究循环CEO / 研究负责人算力成本效率和研究人员生产率提升

买方 / 用户 / 付款方角色根据具名交易结构(Isomorphic Labs / Lilly)、调查数据(ZS)和公开融资 / 政策来源推断;没有单一来源把六个细分市场全部并列列出。

[CM023, CM024, CM025, CM026, CM027, CM029]
FM003: 买方 / 客群地图

企业药企路径与公共部门研究路径在预算、采购权和最终用途上的流向,如何通向 AI 科学平台。

[CM023, CM024, CM027, CM029, CM030, CM017]

2.4 采纳路径:为什么金额预测远远跑在真实部署前面

头部市场预测与观测到的采纳之间的差距,是任何把自主 AI 系统卖入研究工作流的供应商面临的最重要采纳约束。McKinsey 2025 年 11 月全球调查(1,993 名受访者、105 个国家)发现,62% 的组织至少在试验 AI 智能体,但只有 23% 正在规模化任何智能体用例;在单一业务职能内规模化的比例不超过 10%。Gartner 判断更进一步,预测到 2027 年底,超过 40% 的智能体 AI 项目会因成本上升、商业价值不清或风险控制不足而被取消;它还估计,在数千家自称“智能体 AI”的供应商中,只有约 130 家真正具备自主能力,其余都是“智能体噱头包装”。叠加在一般企业谨慎之上的,是科学领域特有的信任问题:Nature 2026 年关于 AI 在科学期刊、预印本和同行评审中足迹的报道,把局面描述为“快速演变”但仍缺乏充分测量;独立评论引用 Columbia University 审计,称 2026 年 PubMed 收录论文中约每 277 篇就有 1 篇含有 AI 幻觉或伪造引用,促使 arXiv 在 2026 年采取政策,禁止未经验证的 AI 生成内容。任何 AI for Science 供应商——包括 Mirendil——都必须先跨过这道可信度门槛,买方才会为自主生成的研究产出付费。[CM039, CM040, CM041, CM042, CM022, CM049]

FM004: 采用漏斗或价值链地图

企业智能体 AI 从试验到可持续规模化使用急剧收窄,这是任何 AI 科学供应商都必须跨过的核心约束。

混合两项不同数据源(McKinsey 2025 年 11 月采用率调查和 Gartner 2025 年 6 月取消率预测),用于方向性展示企业智能体 AI 采用率的漏斗式下滑,不是单一队列研究。

[CM039, CM040, CM022]

2.5 增长驱动、采纳约束与未解规模缺口

几股力量正在推动采纳。Stanford 2026 AI Index 显示,2025 年全球私人 AI 投资同比增长 127.5%,已翻倍不止;主要云厂商基础设施支出创纪录——仅 Google 2025 年年度资本开支就超过 $150B——为前沿实验室所需算力提供资金支撑。制药成本压力进一步放大这一动力:ZS 估计,美国新关税可能给行业增加 $13-19B 成本,最大型制药公司到 2030 年需要削减约 $32B 费用;Excelra 称,混合商业模式最强的 AI 原生生物技术公司已经用 AI 工具把发现周期压缩 40-50%。反向压力也很强。Epoch AI 数据显示,前沿训练算力自 2020 年以来每年增长 5 倍,训练成本每年约增 3.5 倍,电力需求每年翻倍——这种资本强度要求承诺持续上升,远超单笔 $200M 种子轮。联邦科学经费也在被削减并为 AI 划出特殊空间:FY2026 NSF 预算申请会将 NSF 总经费削减 56.9%(从已通过的 $9.06B 降至 $3.9B),即便 AI 是少数被优先增加投资的领域之一;Computing Research Association 警告,更广泛的削减会让美国科研能力“倒退 20 多年”。这些都没有解决两个实质尽调缺口:对于一个跨领域平台能现实捕获约 $288B 制药研发预算(外加邻近材料 / 化学研发)的多大份额,没有自下而上的估算;外界也不知道 Mirendil 自己的路线图是否包含实体自动驾驶实验室集成,还是停留在纯软件——这个选择决定了上面的哪些规模测算视角真正适用。[CM032, CM033, CM034, CM035, CM036, CM037]

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
前沿实验室 AI 投资激增(2025 年私营 AI 投资 +127.5%,Stanford HAI)驱动现在 - 2027可用于押注 AI for science 的资本更多,Mirendil 的 $200M 种子轮就是一例跟踪资金是集中到通用 LLM,还是流向科学专用平台
云 / 算力基础设施支出创新高(Google 2025 年 capex >$150B,Stanford HAI)驱动现在 - 2028扩展了前沿实验室及其供应商可调动的算力供给边界确认 Mirendil 自身路线图背后有哪些算力承诺(云额度、GPU 访问)
算力 / 训练成本膨胀(Epoch AI -- 算力 +5x/年、成本 +3.5x/年、电力每年翻倍)约束持续挤压任何实验室级研究平台的利润率 / 运营开支;也抬高可信竞争所需的资本强度尽调 Mirendil 的实际算力预算 / 承诺,相对于已融资的 $200M
企业智能体 AI ROI 不及预期(Gartner -- 到 2027 年,超过 40% 的智能体 AI 项目被取消)约束2026-2027买方怀疑可能拖慢「AI 科学家」工具的付费采用,即便技术上能跑索要试点转付费证据,而不只是 POC 数量
AI 原生生物技术公司商业牵引(ZS -- 时间线加快 40-50%;Excelra -- 混合模式表现优于纯管线玩法)驱动现在只要包装得当,买方愿意为 AI 加速发现付费这一点得到验证尽调 Mirendil 打算走哪种商业模式 -- 平台 / SaaS、AI-first 生物技术公司,还是混合模式
AI 专项之外的联邦基础研究经费削减(NSF 申请 -56.9%;CRA 政策评论)约束FY2026 预算周期即便 AI 本身受优先支持,学术 / 政府采用路径和人才供给仍受到威胁跟踪最终国会拨款与申请额的差异
AI 辅助科学中的可复现性 / 信任侵蚀(Nature;PubMed 假引文率 1/277)约束持续抬高外界对任何 AI 生成研究成果的审查强度,是「AI 科学家」供应商的声誉风险询问 Mirendil 对 AI 生成结果采用什么验证 / QA 流水线
制药成本压力与关税敞口(ZS -- $13-19B 关税成本,2030 年前约 $32B SG&A 削减)驱动(间接)2026-2030成本压力推动制药企业采用能缩短 R&D 周期 / 降低成本的 AI 工具,从而扩大可服务预算尽调 AI R&D 工具在制药企业的典型采购周期长度和合同规模
市场规模测算方法分化(AI 药物发现估算差距 2.6x;两个智能体 AI 估算相差在 8% 内)约束(尽调)持续Mirendil 的精准品类还没有单一可信 TAM 数字;规模测算必须按多套口径推进委托自下而上的定制测算,锚定 Mirendil 现实可触达的制药 / 材料 R&D 预算

方向 / 时间是作者对引用来源的综合判断,不是某一出版方的单一预测;「驱动(间接)」和「约束(尽调)」标记那些通过成本压力或证据不确定性起作用、而非直接作用于需求的因素。

[CM032, CM034, CM035, CM039, CM041, CM036]

2.6 展项

Chapter 03

03竞争格局

3.1 格局边界与替代类别

竞争集合应围绕买方要完成的任务来划定,而不是围绕 Mirendil 当前的公司标签。Mirendil 并不只是又一家模型公司;它试图帮助 AI 构建者,并最终帮助科学家,把更多研究闭环掌握在自己手里。这个任务可以由 Periodic Labs、Lila Sciences、FutureHouse 和 Sakana AI 等直接“AI 科学家”实验室承担;也可以由 Google DeepMind、Isomorphic Labs、Anthropic 和 OpenAI 等既有巨头承担;还可以由 Recursion、Insilico、Chai Discovery、Cradle 和 Schrödinger 等垂直 AI 药物发现和蛋白设计平台承担;NVIDIA 等基础设施供应商、制药或科技公司内部团队、CRO 与人工实验科学,也都是替代路径;科学家还可以在通用 LLM 和专业工具之间多栖使用。这个宽边界对 Mirendil 不利,因为买方可以先采纳自动化组件,而不必等待一个端到端前沿实验室平台。[CP001, CP002, CP003, CP006, CP008, CP009]

3.2 直接 AI 科学家同行与融资强度

最直接的竞争不是另一家种子阶段网站,而是一组试图打通推理到实验闭环的前沿 AI for Science 组织。Periodic Labs 是概念上最接近的同行:公开报道称其正在为材料和化学打造 AI 科学家与自主实验室,已完成 $300M 种子轮,之后又被讨论为数十亿美元估值。Lila Sciences 同样有威胁,因为其官方材料称拥有先进 AI 与自主实验室,能够生成假设、运行实验并从实时数据中学习;其 Series A 材料称公司已融资 $550M,并正向商业伙伴开放平台。FutureHouse 和 Sakana 不是那么明显的商业替代品,但战略上重要:它们让自动化科学智能体和论文生成工作流变得可见、开放,并被文化上接受。因此,Mirendil 面对直接同行的护城河,取决于它能否在资本更雄厚、领域锚定更深的同行拥有参考案例之前,证明更广的 AI 研发自动化能力。[CP001, CP003, CP004, CP005, CP006, CP007]

竞争对手画像表
竞争对手 / 替代方案类别规模 / 融资目标细分市场差异化对替代 Mirendil 的限制
Mirendil本案公司 / 直接 AI-R&D 自动化实验室$200M 种子轮,由 a16z 和 Kleiner Perkins 领投,NVIDIA 参与先面向 AI 研究人员,再面向药物发现、化学、生物、机器人和其他领域的科学家定位为宽口径前沿 AI-R&D 循环和「构建系统的系统」没有公开 SKU、定价、基准、客户证明或收入披露
Periodic Labs直接 AI 科学家同业$300M 种子轮;后续报道提到估值约 $7B 至 $7.5B材料和化学发现,从超导体切入自主实验室跑实体实验并采集新的真实世界数据公开证据仍主要是发布 / 融资报道,而非面向客户的产品包装
Lila Sciences直接自主科学工厂同业完成 $350M Series A 后累计融资 $550M生命科学、化学、材料、能源、防务、航空航天和其他战略科学项目AI Science Factories 连接模型、仪器、假设、实验和实时学习平台叙事非常宽;商业条款和独立客户结果公开信息仍有限
FutureHouse直接非营利研究智能体同业非营利;此处未对融资做基准比较生物学和复杂科学研究人员,包括博士后用 AI 智能体和研究员项目模式自动化科学发现非营利与研究项目导向可能限制其直接替代企业方案
Sakana AI Scientist直接研究自动化演示研究系统,不是企业供应商定价机器学习研究人员和开放研究社区在开放式循环中自动化创意生成、实验、写作和评审当前版本已有记录错误、安全问题,范式转移能力也不确定
Google DeepMind / AlphaFold既有科学平台Alphabet 支持的全球研究工具;披露结构数 200M+学术和非商业研究人员、蛋白质与生物分子科学用户诺贝尔认可的 AlphaFold 血统、公共服务器和巨大的采用足迹聚焦生物分子结构,不是面向每个实验室的通用 AI-R&D 平台
Isomorphic Labs既有 AI 药物发现公司Alphabet 子公司;Lilly 交易预付款 $45M,里程碑最高 $1.7B追求小分子疗法的制药合作伙伴DeepMind / AlphaFold 血统,加上专门的 AI-first 药物设计团队聚焦疗法且以合作伙伴关系驱动,不是民主化的横向 AI-R&D 工具
Recursion相邻 AI 药物发现既有玩家Nasdaq 上市大型加速申报人,拥有临床阶段管线生物制药合作伙伴,以及肿瘤、罕见病、神经科学、免疫学患者Recursion OS、>50PB 专有数据、自动化湿实验室和临床资产药物发现范围更窄;可能先吸收垂直制药预算,再吸收横向 AI-R&D 预算
Insilico Medicine相邻 AI 药物发现平台私人平台,已披露 Phase II 和 Phase I 管线项目药物发现、靶点识别、生物学、化学和制药 AI 用户从靶点识别到分子生成的生成式 AI 与自动化公开定价和第三方客户结果在已审阅来源中并不完整
内部自建 / CRO / 通用 LLM 技术栈替代方案与现状买方出资;使用既有预算、CRO 合同、云 / GPU 和 LLM 订阅制药、生物技术公司、大学、工业 R&D 和内部 AI 团队把专有数据留在内部,并让买方按需组合工具可能更慢且碎片化,但相比采用新的预产品实验室,平台切换风险更低

这些行是代表性而非穷尽;规模单元格在可得处使用公开融资、上市公司或合作证据,并在限制列标出私人公司信息缺口。

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

按自主范围、领域聚焦、融资强度和商业成熟度,对主要替代方案做序数定位。

序数类别来自公开能力和成熟度证据;不代表任何数字评分。

[CP001, CP003, CP004, CP006, CP007, CP008]

3.3 既有巨头与邻近 AI 科学平台

既有巨头不是以完全相同的定位攻击 Mirendil,而是从可信度、分发和证明入手。Google DeepMind 的 AlphaFold 已经成为全球科学工具,拥有数亿个蛋白结构和数百万研究者;Isomorphic Labs 则把 AlphaFold 血统转化为药物发现合作,背后有 Alphabet 资源。Recursion 和 Insilico 比 Mirendil 更窄,但在药物发现工作流上更成熟:Recursion 是 Nasdaq 上市的临床阶段 TechBio,拥有管线、合作和专有操作系统;Insilico 宣传 AI 驱动管线和药物发现软件模块。Chai Discovery 和 Cradle 是抗体或蛋白工程中更窄的产品化切口,Schrödinger 则提供深植市场的分子发现物理软件。结论不是其中任何一家能取代 Mirendil 的完整命题;而是每一家都可能在 Mirendil 发出已披露产品之前,先吸收预算、数据权利、信任和工作流所有权。[CP010, CP011, CP012, CP013, CP014, CP015]

3.4 替代方案、现状流程与内部自建

现状流程很强,因为科学研发买方不需要购买新的实验室平台,也能继续前进。制药公司、大学和工业实验室可以继续把人类科学家与 CRO、现有建模套件、数据提供商、机器人实验室和通用 LLM 配对。NVIDIA 的医疗堆栈显示,开放模型、SDK、微服务和 GPU 加速管线正越来越容易被想要内部自建的团队取得;Anthropic 的化学工作也显示,通用前沿模型已经可以辅助具体分析化学任务。Gartner 对智能体 AI 的警告相关,因为许多买方可能会面对炒作,转而把部署收窄到高 ROI 工作流组件,而不是采纳雄心勃勃的自主智能体平台。对 Mirendil 而言,替代威胁不只是某个具名竞争对手;它也是一种采购选择:分别购买算力、工具和服务,同时把数据和实验判断留在内部。[CP021, CP022, CP023, CP024, CP025, CP026]

3.5 能力、定价与分发对比

功能对比显示,Mirendil 目前最强的是野心表述,最弱的是可购买实物。其公开材料描述了面向 AI 研发的前沿模型、自主研究闭环,以及未来对科学家的支持,但没有披露具名 SKU、定价模型、安全资料包、集成接口、客户名单或基准测试套件。几家竞争对手已经披露更清晰的购买标准:Lila 欢迎商业伙伴;Recursion 有临床管线和上市公司投资者界面;Isomorphic 披露了与 Lilly 的合作,包括预付款和里程碑经济学;Cradle 声明软件订阅模式且不收特许权使用费;NVIDIA 发布开发者工具和开放模型;AlphaFold Server 面向非商业研究公开可用。因此,今天想要平台的买方可以多栖使用:用广义 LLM 做推理,用专业工具做分子或蛋白工作,用既有基础设施做部署,同时把 Mirendil 视为未来选项,而不是立即替代品。[CP001, CP002, CP007, CP010, CP011, CP013]

功能 / 能力矩阵
购买标准MirendilPeriodic LabsLila SciencesDeepMind / IsomorphicRecursion / Insilico通用 LLM / 内部自建
自主 AI-R&D 循环声称覆盖宽口径 AI-R&D 循环;没有公开基准据报道有自主材料 / 化学实验室声称以 AI Science Factories 跑科学方法循环生物分子和药物发现建模能力强已披露药物发现工作流和管线通过编排可以实现,但买方必须自己组装
实体实验闭环Unknown明确报道机器人实验室实体实验声称仪器由 AI 控制Isomorphic 合作聚焦药物设计;湿实验室闭环公开信息不完整Recursion 自动化湿实验室;Insilico 声称自动化需要 CRO 或内部实验室
领域宽度AI R&D 加上生物、化学、药物发现、材料、机器人愿景材料和化学优先生命科学、化学、材料、能源、防务、航空航天蛋白质 / 数字生物学和小分子主要是疗法和分子发现范围广,但按工具和团队碎片化
商业成熟度预产品 / 无公开客户创业发布和融资报道正在迎接第一批客户Lilly 合作和公开 AlphaFold Server公开管线或已披露产品模块今天即可通过既有工具和合同获得
信任 / 合规姿态公开信任方案未知Unknown声称世界级 AI 安全,但未给细节隐含 Alphabet 品牌和合作伙伴尽调上市公司或成熟供应商界面取决于买方治理和供应商技术栈
定价可见度未披露未披露未披露AlphaFold Server 非商业访问;Isomorphic 交易条款仅在合作层面披露企业平台大多未披露;上市公司经济性另算LLM 订阅 / API 和 CRO 合同,但不等同于 Mirendil

无支撑的单元格标为未知或描述为未披露;矩阵比较的是公开能力证据,不是私人演示或尽调室材料。

[CP001, CP002, CP003, CP006, CP007, CP008]
定价 / 包装对比
替代方案价格 / 单位 / 合同模式包含能力折扣 / 未知项对 Mirendil 的影响
Mirendil未披露前沿 AI-R&D 系统和实验室重构论点没有公开 SKU、标价、试点条款或服务包装必须在私下证明付费意愿和合同路径
Periodic Labs未披露面向物理世界材料数据的 AI 科学家和自主实验室未找到客户定价或企业包装可能是人才 / 数据军备竞赛,而不是近期软件销售
Lila Sciences未披露科学智能体、自主科学平台、AI Science Factories、安全主张官方来源称正在迎接第一批客户,但未定价公开记录中商业姿态比 Mirendil 更强
Isomorphic LabsLilly 合作预付款 $45M,加最高 $1.7B 里程碑利用 AlphaFold 相关平台,围绕多个靶点做小分子发现交易经济性因合作而异,标题里程碑上限不含版税证明只要证据和合作伙伴匹配,制药企业愿意为可信的 AI-first 科学付费
Recursion没有平台标价;上市公司拥有合作和自有管线经济性Recursion OS、自动化湿实验室、专有数据、临床管线企业 / 制药经济性不能简化成公开 SaaS 价格成熟垂直平台可以先卖结果,再卖横向工具
Cradle软件订阅费;未说明版税蛋白质工程协作,附带隐私、安全和 IP 所有权主张抓取页面未披露确切订阅价格窄切口比 Mirendil 的宽平台更容易获得买方批准
AlphaFold Server披露为免费非商业研究访问面向非商业科学家的 AlphaFold 3 结构和相互作用预测引用页面未给商业使用和企业支持定价免费既有工具降低买方为基线能力付费的意愿
通用 LLM / 内部自建订阅、API、云 / GPU 和服务合同;没有等价统一价格推理、编码、化学辅助、内部编排和 CRO 执行总成本取决于数据、算力、验证、人员和治理在 Mirendil 成熟前,形成一条可信的“拼出技术栈”备选路径

多数前沿 AI-for-science 实验室仍处于商业化前,或以合作为主;因此,「未披露」表示在已审阅来源中未找到公开标价或标准合同单位。

[CP001, CP002, CP003, CP006, CP007, CP010]
FP002: 能力地图

按竞争对手类别梳理公开证据支持的能力强项;没有支持证据的单元格标为“未知”。

能力标签概括公开证据,不代表私有基准测试结果。

[CP002, CP006, CP007, CP009, CP010, CP011]

3.6 护城河耐久性与负面结论

Mirendil 最可防守的潜在护城河,是一个在完整内部研究流程上训练出的闭环 AI 研发系统,而不是给科学家套壳的通用聊天机器人。耐久性问题在于,公开证据尚未证明它拥有能让这类系统难以复制的稀缺资产:专有实验数据、商业伙伴工作流、经过验证的基准测试、信任 / 合规资料,或进入科学预算的独家分发。多个负面信号指向相反方向。Sakana 明确预期,模型提供商和开放模型会持续改进并商品化基础层。Gartner 警告,许多智能体项目仍由炒作推动,若没有清晰 ROI,存在取消风险。Excelra 认为,随着基础模型商品化,数据护城河比算法更重要;Big Tech 和领先制药公司也是直接竞争者或内部自建者。因此,承销结论应谨慎:如果 Mirendil 能产品化端到端实验室级研究闭环,它可能具备差异化;但今天公开护城河主要是人才和资本,而不是持久客户锁定。[CP001, CP002, CP035, CP036, CP037, CP038]

护城河持久性 / 竞争风险清单
护城河主张威胁严重性缓释措施 / 尽调要求
顶尖 AI 研发人才可以搭建独特研究循环直接竞品也在招募前 OpenAI、DeepMind、Google、Anthropic 员工,以及机器人和科学人才要求分团队能力地图、留才计划,以及独特内部工作流证据
端到端研究循环数据能积累复利优势Periodic、Lila、Recursion 和制药公司内部团队也能生成专有实验或工作流数据审查数据来源、排他性、规模、反馈循环频率和基准测试提升
横向平台可以越过单一垂直领域扩张垂直平台可能先拿到预算,因为它们解决更窄、高 ROI、验证更清晰的工作流要求客户分层,并证明横向 AI 研发优于先从楔子场景切入的采用路径
不依赖大型 AI 实验室,对外部科学家有价值OpenAI、Anthropic、NVIDIA、DeepMind 和开放模型可能发布更便宜的基准科学能力在买方任务上,将 Mirendil 与通用 LLM、开放模型和既有科学工具做基准测试
科学自主化会带来大幅生产率提升Gartner 警告,许多智能体式 AI 项目由炒作驱动,ROI 低,或生产化成本过高要求 ROI 证据、失败模式日志、人机协同设计和部署经济性
信任可以在产品发布后再建立研究诚信和幻觉担忧会让科学买家在委托前要求验证试点前审查验证、来源追溯、审计、安全和发表诚信控制
资本足够支撑建设周期资金充足的 Lila 和 Periodic,加上 Alphabet 支持的 Isomorphic,可能在支出或合作伙伴上压过 Mirendil尽调中比较现金续航期、算力承诺、合作伙伴排他性和客户管线
不披露定价保留选择权价值包装不清会拖慢采购,也让替代品设定买方预期要求包装、试点、折扣、支持模式,以及成功费 / 里程碑选项

严重性是基于证据的尽调判断,不是打分概率;缓释单元格是私下尽调中的具体要求。

[CP003, CP004, CP006, CP007, CP010, CP013]
FP003: 护城河 / 就绪度 KPI

基于截至运行日期可得公开证据的竞争耐久性指标。

评分是 1-10 的序数尽调指标,不是模型推导的概率。

[CP001, CP002, CP003, CP004, CP006, CP007]

3.7 展项

Chapter 04

04财务

4.1 收入模式仍是假设,不是运营指标

Mirendil 尚未把收入模式写入公开记录。其官网和投资人命题描述了一个面向 AI 构建者和科学家的前沿 AI 研发系统,但没有披露产品 SKU、试点客户、标价、付费设计伙伴、ARR、收入、使用量或客户数。关键在于,自然变现路径差异很大:企业研究平台会表现为软件 ACV 和托管算力毛利;药物发现或材料合作会表现为预付款、里程碑、特许权使用费和客户集中度;托管科学服务会表现为项目收入,软件式毛利更低;内部实验室 / IP 模式会把变现推迟到资产或发现可以授权之后。公开可比公司让这些路径看起来合理,但没有证明 Mirendil 会走哪条路。Schrödinger 是软件 + 药物发现模式;Isomorphic 是制药合作模式;Recursion 有合作收入但没有产品销售。Mirendil 目前没有披露这些财务基本项,因此收入桥应视为尽调框架,而不是预测证据。[CI001, CI002, CI003, CI009, CI034, CI038]

收入流表
收入流机制单位当前值 / 状态质量尽调要求
企业 AI 研发平台访问面向 AI 构建者和科学家的潜在订阅、企业许可或托管平台访问席位、团队、工作区、模型运行或算力额度未披露 / 收入前可成立但未定价;公司尚未命名 SKU要求产品包装、设计伙伴协议、标价、用量计量和收入确认政策
研究合作收入预付款加研究经费、靶点里程碑、版税或期权费,类似公开 AI 药物发现可比公司靶点、项目、合作或里程碑Mirendil 未披露机制有可比公司支撑,但不是 Mirendil 证据要求已签 LOI、条款清单、靶点所有权、里程碑分配瀑布和成本分摊条款
按用量计费的算力 / 智能体执行客户为消耗 GPU 和数据资源的自主实验、代码、评估或仿真工作流付费GPU-hour、实验、工作流或类似 token 的单位未披露经济上可行,但毛利未知要求计量方案、面向客户的用量单位、云 / GPU 成本表和利用率假设
托管式科学研发服务Mirendil 团队或系统在平台自助成熟前,为客户执行一项研究工作包项目、FTE 等价、研究或交付物未披露可衔接早期收入,但可能稀释软件毛利要求试点 SOW、人员配置模型、服务毛利、IP 所有权和验收标准
内部实验室 / IP 创建公司创造科学发现或 AI 资产,之后许可、孵化拆分或变现资产、模型、靶点、分子或专利族未披露可选性高,但现金转换周期最长要求内部项目清单、所有权地图、专利申请、估值政策和退出 / 许可策略
赠款或非稀释性研究资金公共部门、基金会或产业赠款,用于支持特定研究基础设施赠款、奖励或研究合同未披露没有 Mirendil 公开证据要求赠款申请、获批项目、报销条件,以及商业使用限制

截至运行日期,所有 Mirendil 收入单元格都是空值或假设项,因为没有公开收入、定价、客户或合同披露。

[CI001, CI002, CI003, CI016, CI019, CI034]
定价 / 变现表
价格 / 单位 / 合同标价 / 实际成交价折扣 / 未知项来源状态投资测算含义
平台订阅或企业许可未披露 / 未观察到标价席位数、用量上限、企业折扣、支持、安全和托管算力转嫁均未知Mirendil 官方和投资人来源描述使命,不披露定价无法建模 ARR、ACV、NRR、CAC 回收期或软件毛利
研究合作合同Mirendil 未披露;公开可比公司使用预付款、里程碑、版税和研究活动靶点数量、预付款规模、里程碑概率、排他性和报销安排未知Schrödinger 和 Isomorphic 申报 / 公告显示类似结构只有审阅条款清单后,才使用概率加权的里程碑经济性
按用量计费的 GPU 或工作流费用未披露实际毛利取决于 GPU-hour 成本、调度、利用率,以及算力是包含在价格内还是转嫁Lambda 和 NVIDIA 来源只能为输入类别提供基准,不代表 Mirendil 定价在算力采购和计费单位明确前,毛利无法用于投资测算
托管项目或服务费未披露折扣、人员杠杆、验收标准和 IP 所有权未知没有公开的 Mirendil 客户 SOW如果早期收入依赖创始人科学家的定制劳动,收入质量可能较低
内部科学资产的许可 / 版税未披露时间、开发风险、版税基数和合作方经济性未知可比药物发现申报显示这种模式,但不能证明 Mirendil 已采用作为上行情景可选性处理,而非基础情景收入

Mirendil 没有公开标价、实际成交价、折扣或合同收入;各行故意把缺失的公司数据与可比机制分开。

[CI002, CI016, CI019, CI032, CI033, CI038]
FI001: 收入模型桥接图

Mirendil 活动转化为收入的潜在路径;每个商业节点仍未公开披露。

仅为定性桥接;Mirendil 尚未披露客户、SKU、用量单位、价格或收入确认政策。

[CI001, CI002, CI003, CI016, CI019, CI038]

4.2 定价、GTM 效率与收入确认输入均不可得

公开记录不支持把 CAC 回收期、NRR、ACV 扩张、销售周期、云毛利或使用留存等 SaaS 式指标用于 Mirendil。除非公司披露它卖的是软件访问权、研究合作、算力密集型使用、外包科学工作,还是内部生成的知识产权,否则这些指标不应被硬塞进投资模型。最接近的公开基准是 Schrödinger,其软件 ACV、软件收入、托管收入转型、毛利和大客户 ACV 均有披露;这说明在 Mirendil 被当作企业平台评估之前,需要达到怎样的证据水平。如果管理层称 Mirendil 其实在搭建伙伴驱动的发现业务,尽调就应转向条款清单、靶标所有权、成本分摊和里程碑概率,而不是订阅指标。目前,标价的可观测性是 0%,实际成交价的可观测性是 0%,收入确认仍是开放会计问题,而不是数据点。[CI014, CI015, CI016, CI017, CI018, CI038]

单位经济性表
指标数值 / 空值置信度重要性尽调要求
ARR / 收入运行率估值、增长和客户证明的锚点要求按产品、客户、合同日期和确认基础拆分的月度收入桥表
客户数 / 设计伙伴区分内部实验室工作与外部需求要求已签试点、未付费试点、设计伙伴名单、转化条款和流失状态
平均合同价值界定 GTM 是由企业销售、合作、用量还是服务牵引要求 ACV 分布、预付款、最低承诺和扩张历史
毛利率前沿 AI 收入质量高度取决于算力转嫁和利用率要求按产品拆分毛利、云 / GPU 成本分摊、支持成本和数据获取成本
每个工作流的 GPU / 算力成本决定用量扩张能否盈利,还是会吃掉种子轮资本要求 GPU 库存、云合同、预留容量、有效小时费率和按工作负载划分的利用率
CAC 和销售周期漫长的科学企业销售周期会拖慢现金转换要求管线阶段、销售周期队列、创始人主导与销售主导打法,以及赢单 / 输单数据
CAC 回收期需要毛利、销售支出和已确认收入;这些均未公开要求按队列拆分销售和营销支出、按客户拆分毛利,以及回收期计算
NRR / 扩张验证价值可重复、平台有粘性要求续约基数、扩张事件、降级和用量留存队列
R&D 费用占收入比例Mirendil 置信度低;公开可比公司置信度高显示 AI-for-science 收入能否覆盖发现和平台支出要求月度 R&D 工资、算力、数据和实验室支出,与任何已确认收入对照
现金消耗 / 现金续航期由于公开收入缺失,这是近期偿付能力最重要变量要求月度现金运营消耗、资本开支、预付算力、非受限现金和董事会批准的运营计划

空值表示私人指标不可得,不是零;每行都给出把该字段变成可投资测算输入所需的具体尽调路径。

[CI034, CI037, CI038, CI041, CI042, CI044]
FI002: 单位经济模型桥接图

从客户价值到贡献毛利的投资测算链条,目前因缺少私有输入而断裂。

没有公开的单位经济模型数字;节点标出管理层尽调必须提供的输入。

[CI032, CI035, CI037, CI038, CI041, CI047]

4.3 成本栈很可能以算力和人才为主,但实际现金消耗仍属私密

Mirendil 的使命意味着其成本结构更接近前沿实验室,而不是传统轻量 SaaS 初创公司。公司必须支付顶级 AI 研究员薪酬,运行或租用加速器,建设评测和实验基础设施,获取科学数据,并可能在任何收入抵消这些费用之前,支持特定领域实验室工作流。公开来源没有披露 Mirendil 的月度现金消耗、算力合同、云积分、GPU 获取条款、实验室资本开支或招聘计划。外部基准显示了这一缺口的重要性:Epoch 估计全球 AI 算力已扩展到约等于 2,000 万块 H100,AI 资本开支正接近每年 $1T;arXiv 前沿训练成本论文估计,算力最密集训练运行的摊销成本自 2016 年以来每年增长 2.4 倍;Lambda 公布了按 GPU 小时计价的 H100 和 B200 云价格。这些来源没有揭示 Mirendil 的支出,但它们指出了尽调必须量化的成本驱动项,否则任何毛利或续航期结论都不可靠。[CI030, CI031, CI032, CI033, CI035, CI037]

FI004: 资本强度 / 现金流地图

定性梳理在收入抵消烧钱之前,可能主导 Mirendil 的成本驱动因素。

定性、证据支持的成本驱动地图;Mirendil 未披露运营预算、债务计划或计算合约。

[CI008, CI022, CI023, CI030, CI031, CI032]

4.4 种子资金很充足,但续航期几乎完全取决于未披露现金消耗

财务承销锚点是已报道的 $200M 种子轮融资,而不是收入。简单按总现金框架看,Mirendil 获得了有意义的建设时间,但区间很宽,因为现金消耗分母未知。如果总融资款可用于运营,实验室每月消耗 $3M,单靠融资款意味着扣除费用前约 67 个月续航期;每月 $5M,约 40 个月;每月 $10M,约 20 个月。这些不是公司预测,也排除了融资费用、设备押金、预付算力、法律成本、招聘加速和任何战略承诺。这个场景仍有用,因为它把公开现金消耗披露缺失转化为一个明确尽调变量。因此,下一次融资触发点本身不是某个日历日期,而是模型能力、产品打包、客户证明、算力承诺,以及公司能否在种子资金耗尽前证明资本效率的组合。[CI004, CI005, CI006, CI007, CI036, CI037]

资本充足性表
输入公开 / 估算值情景标签含义尽调要求
种子轮所得现金余额~$200M 总融资背景;非受限净现金未披露公开融资事实 + 不可得的私人现金数据大额种子轮给了建设时间,但不能说明精确的现金流动性要求银行余额、受限现金、费用、托管款、预付算力和交割后资产负债表
月度现金消耗未知;示例情景使用每月 $3M、$5M 和 $10M仅为估算情景现金续航期敏感性主要由这个未披露输入决定要求过去三个月和董事会计划中的现金消耗,按工资、算力、资本开支、数据和 G&A 拆分
按 $3M/月计算的现金续航期~67 个月,未计费用和营运资本影响仅按种子轮总额估算若支出受控,足够支撑产品建设确认前沿 R&D 早期现金消耗是否真能维持在这一低位
按 $5M/月计算的现金续航期~40 个月,未计费用和营运资本影响仅按种子轮总额估算高薪实验室且算力投入显著时的基础示例情景用招聘计划、GPU 承诺和供应商预付款验证
按 $10M/月计算的现金续航期~20 个月,未计费用和营运资本影响仅按种子轮总额估算高现金消耗情景可能在收入证据扎实前就迫使融资要求下行情景计划、后续投资人储备和基于里程碑的支出闸门
计划资金用途算力、顶尖 AI 人才、平台 R&D、数据 / 评估系统,以及可能的科学工作流基础设施根据使命和成本基准推断资本密集是该策略的特征,不是偶发费用要求董事会批准的预算、供应商承诺、云抵扣额度和招聘计划
债务 / 项目融资义务公开未披露不可得的私人资本结构输入未识别公开债务包袱,但缺失披露不等于没有要求债务明细、设备融资、云预付款合同、认股权证和附函
下一轮触发条件可能是产品 / 基准测试 / 客户证明加算力续航;未披露推断的里程碑框架如果模型进展消耗现金却不能转化为收入,融资风险会上升要求里程碑预算、内部 KPI、储备政策和投资人按比例跟投承诺

现金续航期情景只是简单的总融资额计算:$200M 除以示例月度现金消耗;不是管理层指引。

[CI004, CI005, CI006, CI007, CI030, CI031]
FI003: 种子轮资金跑道敏感性

用公开的 $200M 种子轮总额背景和三种月烧钱情景,示意跑道区间。

区间对 $200M 除以月烧钱的简单计算打了折扣,以反映未说明的费用、预付款和营运资金流失;实际烧钱和现金余额不可得。

[CI004, CI005, CI036, CI037, CI042]

4.5 上市和后期可比公司提醒:不要假设干净的软件经济性

与 Mirendil 财务尽调最相关的上市可比公司,并不支持一条简单的高毛利、自助式软件曲线。Recursion 2025 年申报显示收入 $74.7M、没有产品销售、研发费用 $475.3M、净亏损 $644.8M、现金及受限现金 $753.9M。Schrödinger 商业成熟度更高,2025 年收入 $255.9M,其中软件收入 $199.5M、药物发现收入 $56.4M、软件毛利率 74%,但仍报告净亏损 $103.3M。Isomorphic 与 Lilly 的合作显示,伙伴模式可以包括 $45M 预付款和最高 $1.7B 潜在里程碑,但以里程碑为主的经济学需要按概率加权,且回款靠后。这些例子不是 Mirendil 前沿 AI 研发平台的直接代理,但能让分析保持落地:AI for Science 收入可以真实存在,但亏损、集中度、长周期和资本强度往往仍是核心。[CI010, CI011, CI012, CI013, CI014, CI015]

4.6 财务结论:资本足够启动建设,披露不足以支撑承销

财务结论刻意分成两面。正面看,Mirendil 已报道的大额种子轮、蓝筹投资人背书,以及资本猛烈流向 AI 和 AI for Science 基础设施的行业背景,都对公司有利。Carta 和 Crunchbase 都显示,2026 年风险资本已经高度集中于 AI;Lila 和 Periodic 也说明,大型科学 AI 融资不是孤立异常值。负面看,Mirendil 的公开财务证据还不足以支持可投资的收入质量或单位经济结论。Sequoia 对 AI 基础设施的批评追问:支撑 GPU 建设的收入从哪里来?Gartner 也警告,许多智能体项目可能因成本、价值不清和风险控制失败而被取消。Mirendil 最终也许能证明其自主研发闭环可以创造有价值的科学产出,但今天可承销输入仍不足:没有定价、没有客户管线、没有现金消耗、没有算力义务、没有毛利、没有债务期限表,也没有里程碑经济学。[CI020, CI021, CI022, CI023, CI024, CI025]

公开财务缺口表
缺失的私人指标对投资测算的影响精确尽调路径严重性
收入 / ARR / 订单额无法测试收入质量、估值倍数或牵引力要求月度已确认收入、订单额、递延收入、已签合同、管线,以及审计就绪的收入政策阻断
定价和包装无法区分订阅、用量、合作、服务或 IP 经济性要求定价备忘录、标价、折扣权限、试点定价和产品包装路线图阻断
客户和设计伙伴证明无法证明外部付费意愿,或从实验室论点转向市场需求的转化要求 NDA 下的已签客户名称、试点、LOI、设计伙伴范围和转化条款重大
算力承诺和有效 GPU 成本如果预留容量或资本开支很大,毛利和现金消耗可能错上数倍要求 GPU 合同、云抵扣额度、预留容量排期、利用率数据和取消条款阻断
工资和招聘计划顶尖 AI 人才能迅速把大额种子轮变成高固定现金消耗要求组织架构、已签录用通知、薪酬区间、留才包和 24 个月招聘计划重大
现金、受限现金、债务和附函仅靠融资总额无法验证现金续航期和下行保护要求交割后资产负债表、债务明细、认股权证、投资人附函和董事会批准的预算阻断
按工作流拆分的单位经济性无法按交付成本给工作流、合作或服务定价要求工作流级贡献毛利、算力分摊、支持时间、错误 / 返工率和客户 SLA 假设重大
里程碑或版税经济性可比上行可能后置且需概率加权,而不是当前收入要求合作分配瀑布、里程碑概率、靶点所有权、版税基数和终止权重大

该表有意以缺口为导向,因为公开证据不足以支撑标准收入和单位经济性投资测算。

[CI034, CI037, CI038, CI041, CI042, CI047]

4.7 展项

Chapter 05

05产品与技术

5.1 产品定义与交付状态

Mirendil 的公开产品,最适合理解为一个声称用于 AI 研发的实验室操作系统,而不是已经商业交付的 SKU。公司称自己训练擅长 AI 研究的前沿模型,并围绕这些模型重塑实验室,让闭环更快、更强、更自主。Andreessen Horowitz 给出了最清晰的模块描述:系统应能提出实验、编写并运行代码、解释结果、调试失败、改进内核、管理算力、比较检查点,并决定下一步尝试什么。这是一个具体且有价值的工作流,但 Mirendil 的所有公开界面都没有披露演示、API、定价页、客户部署、基准测试包、路线图、安全页面或文档门户。因此,截至本次报告日期,已交付对象是一个由命题和团队支撑的建设计划,而不是经过验证的产品资产。[CE001, CE002, CE003, CE004, CE005, CE006]

产品模块 / 资产矩阵
模块 / 资产主要用户Mirendil 状态 / 成熟度若证实后的差异化尽调缺口
专为 AI 研发调优的前沿模型AI 研究员和 ML 工程师概念 / 未披露在实验设计、代码、评测和检查点推理上,可能优于通用模型模型名称、架构、训练数据、模型卡和评测套件未公开
自主研究循环智能体框架研究工程师和未来的领域科学家概念 / 未披露可把目标转成迭代实验计划、代码修改、运行、调试和下一步选择没有公开演示、API、记录、安全案例或人机协同政策
算力管理层内部实验室操作人员和高级用户概念 / 未披露可分配 GPU、调度任务、比较运行结果,并减少前沿算力浪费未披露云、集群、编排技术栈、预算控制或 NVIDIA 供给条款
评估和检查点对比框架ML 负责人和技术评审者概念 / 未披露可通过指标、基线、对比和可复现产物,让自动化工作可审计没有基准测试包、留出集评测方法、污染控制或人工基线
领域科学界面生物、化学、材料、机器人和药物发现专家概念 / 未披露可让专家运行 AI 驱动的研究,而不必自己变成前沿 AI 实验室没有工作流截图、本体 / 数据连接器、实验室集成或领域验证
外部用户工作区或产品界面先面向工程师,之后面向技术较弱的科学家未披露可把实验室级工具包装给前沿实验室之外的用户没有定价、文档、入门引导、支持模式、正常运行时间或发布渠道
研究轨迹 / 训练数据语料模型训练和后训练团队私有 / 未披露若安全收集,可围绕完整 AI 研发循环沉淀专有数据没有与研究轨迹绑定的数据来源、同意、IP 权利、留存或隐私政策

每个 Mirendil 状态单元格都刻意保守,因为公开来源披露了论点,但没有披露已发布模块、规格或客户证据。

[CE001, CE002, CE003, CE004, CE008, CE009]
FE001: 产品架构图

Mirendil 声称的 AI 研发产品尽调栈,把公开论点与未披露实现拆开。

这是有证据支撑的尽调架构,不是 Mirendil 披露的技术图。

[CE003, CE008, CE009, CE010, CE011, CE012]

5.2 架构与运行闭环

Mirendil 架构的合理尽调模型有五层,每一层对公司自身仍未披露。第一层是专门面向 AI 研发任务的前沿模型族或模型组合。第二层是智能体框架,用于规划实验、编辑代码、运行作业、读取日志并围绕失败循环。第三层是面向 GPU 作业、内核、检查点和预算分配的算力编排。第四层是评测框架,用测试、分数、基线和人类预期来比较生成产物。第五层是面向科学家的领域界面,他们可能懂生物、化学、材料或机器人,但不懂前沿模型运维。这个架构不是从泛泛的“智能体”话术中编出来的;它贴合 Mirendil 自身和投资人的描述,也得到 AI Scientist、FunSearch、MLAgentBench、MLE-bench、RE-Bench、SciCode 和 SWE-bench 等开放系统佐证。缺口在于,所有 Mirendil 特定的实现选择仍是私密信息。[CE008, CE009, CE010, CE011, CE012, CE013]

技术 / 运营架构表
层 / 组件系统角色关键依赖风险
前沿 AI-R&D 模型层生成研究想法、代码、分析和下一步推理高质量模型训练、后训练和研究轨迹数据未披露架构,可能不足以显著胜过通用前沿模型
智能体研究循环框架在实验设计、编码、执行、调试和迭代中维持状态工具执行、沙箱、记忆、规划和反馈循环外部基准中,长周期规划和失败恢复仍偏弱
算力编排和内核调度 GPU 任务、管理运行、优化内核并控制成本NVIDIA GPU、云 / 数据中心容量、CUDA 生态和预算护栏算力获取、电力和硬件成本可能主导产品经济性
评测和检查点比较给运行打分、比较基线、发现回归,并支撑可复现留出任务、污染控制、人类基线、日志和审计轨迹评测泄漏或奖励黑客会抬高表观能力
领域适配器和接口把科学家的目标和领域数据转成可执行研究任务生物 / 化学 / 材料 / 机器人数据集、本体、实验室工具和用户体验AI-R&D 成功未必能迁移到湿实验或物理世界科学
安全、合规和信任层阻止不安全代码执行、数据泄漏、误导性科学结论和失控动作沙箱、权限、数据权利、监控和审查控制Mirendil 没有公开信任、安全或合规公开面

这张表是用于尽调的运营模型,不是已验证的 Mirendil 架构图。

[CE008, CE009, CE010, CE011, CE012, CE013]
FE002: 客户工作流 / 运营流

拟议的自主研究闭环,把科学或 AI 研发目标转化为反复的代码、计算、评估和检查点决策。

流程阶段来自 Mirendil / a16z 的闭环表述,以及 AI Scientist、MLAgentBench 等外部系统。

[CE003, CE017, CE018, CE019, CE024]

5.3 基准测试格局与可行性信号

外部基准测试格局让 Mirendil 的目标足够可信,值得调查,但还不成熟到可以按已解决问题承销。Sakana 的 AI Scientist 展示了一个覆盖想法生成、文献搜索、代码、实验、撰写和评审的自动化闭环,但作者明确报告了限制,包括实现错误、不公平比较、弱视觉推理、数值错误、不安全的进程派生行为,以及对沙箱化的需求。MLE-bench 把 75 场 Kaggle 竞赛转化为 ML 工程评测,初始结果显示最强配置只在 16.9% 的竞赛中达到 Kaggle 铜牌水平。RE-Bench 更接近 Mirendil 的目标,因为它用人类专家对照评估开放式 ML 研究工程:在两小时设定中,智能体可以很快、很强,但时间预算拉长后人类表现更好。SciCode、SWE-bench 和 MLAgentBench 进一步显示,真实的科学编码、软件修复和 ML 实验仍远未饱和。这个技术现状支持审慎的可行性判断,而不是产品就绪结论。[CE017, CE018, CE019, CE020, CE021, CE022]

工作流 / 用例表
用户任务当前工作流Mirendil 拟议解决方案可衡量收益(声称 / 未证实)限制
ML 实验设计研究员头脑风暴假设、选择基线、实现变体,并手工跟踪运行智能体提出实验,并循环推进实现和结果解读声称可加快 AI-R&D 循环;Mirendil 没有给出量化提速的基准同类系统仍会做出错误实现和不公平比较
模型训练和检查点选择工程师调度任务、调超参数、看日志,并比较检查点自主循环把算力管理和检查点比较纳入研究周期可能减少浪费的 GPU 时间,并加快迭代未披露编排层、预算护栏或可复现的检查点证据
研究代码调试人类阅读追踪、修 bug、重跑实验并验证测试智能体编写并运行代码、调试失败,并使用评测反馈SWE-bench 和 RE-Bench 表明,智能体能解决一部分真实编码任务基准仍耗资源,且容易受测试缺口或奖励黑客影响
面向领域专家的科学编码科学家把领域知识转成仿真或分析代码,通常还需要专业工程师Mirendil 希望让专家不用具备前沿实验室能力,也能跑自己的实验可能扩大 AI 科学工具的使用面SciCode 显示,真实科学编码对当前模型仍然困难
开放式发现循环人类定义问题、设计算法、验证输出并撰写论文AI 系统生成想法、代码、实验、写作稿,并跑评审循环AI Scientist 和 FunSearch 展示了部分样例没有证据表明 Mirendil 能从 AI-R&D 泛化到湿实验生物学、化学、材料或机器人
基准和评测治理团队维护留出测试、人类基线、安全审查和可复现包Mirendil 需要围绕自主智能体搭建稳健的评测框架如果透明且未被污染,可能建立信任没有公开评测治理;外部审计显示 AI 基准可能被利用

由于 Mirendil 没有发布产品指标或客户工作流研究,收益只能表述为声称或潜在。

[CE003, CE017, CE018, CE019, CE020, CE021]
FE004: 产品成熟度 / 能力地图

相邻系统证明了 Mirendil 论点中的一部分,但 Mirendil 自身公开成熟度仍停留在概念或未披露阶段。

由于 Mirendil 尚未发布产品产物或基准,成熟度标签为定性判断。

[CE017, CE019, CE020, CE021, CE022, CE023]

5.4 成熟度、路线图与发布缺口

Mirendil 经验证的成熟度评分应接近概念阶段或未披露,尽管邻近文献已有可运行演示。公开层面,Mirendil 披露了使命、20 人创始团队、投资人和宽泛目标用户叙事。公司没有披露模型名称、模型卡、评测方法、数据集、部署目标、集成接口、客户支持流程、可用性承诺、云区域姿态、合规认证或分阶段发布计划。关键在于,自主研究产品不能像普通生产力工具一样尽调:一旦反馈闭环变慢、目标变模糊、实验需要昂贵算力,或湿实验科学引入物理世界时滞,小演示可能无法泛化。因此,路线图表刻意把大多数 Mirendil 行标为“未披露”。任何更强的成熟度主张,都需要私下证据,例如内部基准测试套件、可复现实演、设计伙伴日志或安全论证。[CE004, CE005, CE006, CE027, CE028, CE029]

路线图 / 发布 / 开发阶段表
日期 / 阶段里程碑状态含义来源依据
2026-06 to 2026-07公开使命和融资发布已披露证实投资逻辑和资金到位,但不是产品发布Mirendil、a16z、Kleiner Perkins(公司与领投方)
2026-07-02公开产品文档或 API未披露无法验证外部开发者或客户集成路径已审阅官网和投资方页面
2026-07-02模型或基准发布未披露没有公开方式可将 Mirendil 与 MLE-bench、RE-Bench、SWE-bench、SciCode 或 AI Scientist 比较未发现 Mirendil 基准包
2026-07-02信任、安全、隐私或合规页面未披露无法评估企业和科学部署准备度公开材料未发现信任公开面
未来 / 未披露工程师和 AI 研究人员可能是第一批用户公司 / 投资方声称暗示初始产品可能先面向专家用户,再扩展到更广泛的科学家a16z 和 Kleiner Perkins 的投资逻辑措辞
未来 / 未披露生物、化学、药物发现、材料和机器人领域中技术背景较少的科学家公司 / 投资方声称市场愿景很大,但从 AI-R&D 泛化到领域科学仍未证明Mirendil 和投资方使命陈述
未来 / 尽调请求设计伙伴试点和可复现演示证据缺口在把平台看作超出内部实验室系统之前,需要补齐来自缺失产品证据和基准标准的推导

这张表主要记录披露缺位;如果 Mirendil 发布路线图、演示、API 或基准包,应更新。

[CE001, CE002, CE004, CE005, CE006, CE007]

5.5 关键依赖与运行约束

Mirendil 的技术依赖图谱异常集中。前沿 AI 研发模型需要大规模算力,Epoch 的超级计算机数据集显示,前沿 AI 系统正在转向庞大的私有集群,其性能、功耗和硬件成本都在快速扩张。NVIDIA 参投 Mirendil 具有战略意义,因为 NVIDIA 控制核心加速计算基础设施,并发布研究资源、专有模型许可证和以 CUDA 为中心的代码库,这些都靠近 Mirendil 必须使用的生态。但参投不等于供应、定价、预留容量或可靠性有保证。产品还依赖研究轨迹的数据权利、LLM 编写代码的安全执行环境、可复现评测框架、基础模型工具,以及面向科学家的特定领域适配器。任何一项依赖失败,平台都可能停留在内部实验室工具,而无法变成通用外部产品。[CE032, CE033, CE034, CE035, CE036, CE037]

FE003: 关键依赖地图

Mirendil 的产品就绪度取决于计算、数据、评估、安全和领域科学资产;这些资产尚无公开证据证明。

依赖关系从公开系统要求和前沿 AI 研发基准推断,并非来自已披露供应商合同。

[CE032, CE033, CE034, CE035, CE036, CE037]

5.6 信任、质量与技术风险

最重要的产品风险,不是智能体能否生成貌似可信的研究产物;而是 Mirendil 能否让这些产物可复现、安全、无幻觉、难以被刷分,并且对外部科学家有用。AI Scientist 和 FunSearch 文献都强调,需要评估器、可复现性、沙箱化,以及围绕 LLM 生成代码的防护栏。AlphaFold 3 显示,当输出用特定领域结构做基准测试,并配有置信度 / 误差指标时,科学模型可以产生重大影响;但 Mirendil 看不到这个标准。负面证据很尖锐:Berkeley 的基准测试审计报告称,许多 AI 评测可以被利用,包括假分数或答案泄漏;Gartner 警告,许多智能体 AI 项目仍是早期实验,价值不清且风险控制不足。Mirendil 没有披露信任中心、评测治理、数据权利政策或科学验证协议,因此技术尽调应把可靠性视为未解决的阻断项。[CE039, CE040, CE041, CE042, CE043, CE044]

信任 / 质量 / 合规表
控制Mirendil 公开状态范围缺口
可复现包Unknown已执行代码、数据、种子、日志、检查点和环境快照没有公开基准包、运行记录或可复现政策
沙箱化代码执行Unknown限制 LLM 编写的代码执行不安全的文件系统、网络、进程或包操作AI Scientist 仓库明确警告,LLM 编写的代码应放入容器并受限制
幻觉和编造护栏Unknown防止看似可信但虚假的研究主张、结构或输出FunSearch 和 AlphaFold 3 显示需要评测器 / 置信度度量,但 Mirendil 未披露
基准完整性控制Unknown隐藏测试、隔离、泄漏防护和反奖励黑客审计Berkeley 审计显示,基准存在多类漏洞和伪造得分路径
人工审查和升级Unknown定义何时由人类批准实验、不安全领域或科学主张没有公开人类介入阈值或安全治理
数据权利和隐私Unknown研究轨迹、代码、专有数据集、模型输出和外部用户数据没有公开研究自动化的数据权利、留存、客户 IP 或隐私立场

所有状态值均未知,因为 Mirendil 没有公开信任中心、模型卡、产品文档或合规页面。

[CE004, CE005, CE039, CE040, CE041, CE042]

5.7 展项

Chapter 06

06客户

6.1 客户基础:未披露客户,四类可能买方

本章审阅的公开来源里,Mirendil 没有点名任何客户、试点用户或设计伙伴;官网、招聘页、投资人笔记和媒体报道只谈使命和创始团队,从未出现买方。缺少直接证据,下面的买方地图只能从 Mirendil 自己的目标,以及同类 AI-for-science 厂商的实际变现方式推断。制药和生物技术研发组织是最清晰的参照:Isomorphic Labs 和 Recursion 都通过按里程碑计费的合作切入这一客群,而不是卖席位许可;预算负责人通常是 Chief R&D Officer 或 Chief Digital Officer,日常用户是实验台科学家或计算生物学科学家。高校和学术实验室是第二类,实际付款方是通过安全审查采购流程支出的科研经费或基金会预算;Periodic Labs 公开的 Academic Grant Program 展示了一条可能入口,但 Mirendil 还没有复制。工业和材料科学实验室是第三类,最接近的同行案例是 Periodic Labs 未披露名称的半导体制造商合作。第四类更具猜测性:其他实验室或大型企业的内部 AI 研究团队;但 Mirendil 自称要加速内部 AI 研发,这可能让它与本可销售的团队正面竞争。第五个类比是 FutureHouse 等由慈善资金支持的非营利科学实验室,代表一种非商业付款模式;Mirendil 尚未表示会走这条路。[CU001, CU002, CU003, CU004, CU005, CU006]

客户分群表
分群买方 / 用户 / 付款方用例规模 / 范围收入 / 战略价值尽调缺口
医药和生物技术 R&D 机构买方:首席 R&D / 数字官;用户:实验台科学家和计算生物学科学家;付款方:R&D 预算或 基于里程碑的交易小分子 / 生物制剂发现合作,类似 Isomorphic Labs 和 Recursion 的交易大型药企 R&D 预算;多年、多项目合作若赢下则价值高;可比交易潜在里程碑达 $1-5B+,Mirendil 披露参与为零未披露具名药企伙伴、交易结构或管线
大学和学术研究实验室买方:院系主任 / PI;用户:研究生和博士后;付款方:通过安全审查采购的 资助 / 基金会预算材料 / 化学 / 生物假设生成和实验自动化从单个实验室到全校访问;规模未确认单席近期收入低,但按 Periodic Labs 学术资助类比,具备战略管线 / 训练数据价值Mirendil 未披露资助计划、校园试点或具名 PI
工业和材料科学实验室(如半导体、机器人)买方:R&D / 工程副总裁;用户:材料 / 工艺工程师;付款方:R&D 资本开支预算实验数据解读和仿真自动化,参照 Periodic Labs 半导体类比大型工业企业 R&D 预算若赢下则价值高;Mirendil 披露参与为零未披露具名工业客户、用例或合同
其他实验室 / 企业的内部 AI 研究或工程团队买方:AI / ML 基础设施负责人;用户:研究工程师;付款方:内部 R&D 或基础设施预算加速内部 AI-R&D 循环——这是 Mirendil 自称的核心用例,未来也可能对外转售不清楚;与竞争对手章节画像的竞品集重叠模糊——这类对象既可能是客户,也可能是竞争对手没有内部研究循环对外授权的证据
非营利 / 慈善资助科学实验室买方:项目官员 / 基金会;用户:员工科学家;付款方:资助,参照 FutureHouse 融资模式资助驱动的发现研究,而非商业授权少数资金充足的非营利实验室直接收入低,但有声誉 / 生态价值没有证据表明 Mirendil 追求这类付款方;仅因最接近的同行 FutureHouse 使用该模式而纳入

分群来自 Mirendil 自身使命表述,以及最接近竞品的实际变现方式推断;没有任何分群确认已有 Mirendil 参与。

[CU002, CU003, CU004, CU005, CU006, CU007]
FU001: 客户旅程图

基于同行模式假设 Mirendil 买方从发现到扩张的旅程;除公开认知外,Mirendil 自身每个阶段均未观察到。

除阶段 1 外,所有阶段均根据同行 GTM 模式(Isomorphic Labs、Recursion、Periodic Labs)假设;Mirendil 尚未确认任何一个。

[CU001, CU013, CU015, CU019, CU028]

6.2 招聘与 GTM 准备度:职位板全是技术岗

判断 Mirendil 商业化准备度,最具体、可免费验证的信号是它自己的职位板。截至 2026 年 7 月,Mirendil 托管在 Ashby 的招聘页列出 14 个开放岗位,全部头衔都是 Member of Technical Staff,覆盖智能体框架、AI-for-AI 系统、设计工程、推理、基础设施、内核、模型评估、平台、后训练 / RL(三个不同岗位)、预训练、产品开发和安全工程。没有销售、商务拓展、合作伙伴或客户成功岗位。公司自己的招聘页本身还是占位页,先显示 Loading open roles,再完全跳转到外部 Ashby 列表;官网也没有候补名单、早期访问注册或面向潜在用户的 beta 计划入口。这与商业化走得更远的同行形成对比:Lila Sciences 公开表示,在完成 $350 million Series A 后,正在欢迎首批客户,并出现在 BIO 2026 官方 Partnering 目录中,与大型制药公司同列;Periodic Labs 已在自己官网点出一个半导体制造商合作。Mirendil 的招聘结构说明,公司仍在造产品,还没有搭建出售产品的商业组织;职位板也是一个低成本、可重复检查的触发器,用来跟踪变化何时发生。[CU011, CU012, CU013, CU014, CU015, CU016]

客户增长 / 采用轨迹表
指标数值日期来源置信度含义缺失分母
公开技术岗位142026-07Mirendil / Ashby 招聘板证实技术建设活跃;商业建设可见度为零没有历史趋势可显示随时间变化
GTM / 销售 / 客户成功岗位02026-07Mirendil / Ashby 招聘板近期没有 GTM 招聘信号无法确认 GTM 招聘是有意延后,还是尚未启动
创始团队规模约 20 名研究人员 / 工程师2026-06Mirendil 官网;a16z小团队限制近期客户交付能力没有研究人员与交付 / 支持人员拆分
已披露具名客户或设计伙伴02026-07Mirendil 公开来源没有可衡量的采用若存在私人管线,也未披露
公开等候名单 / 早期访问注册机制未披露2026-07Mirendil 官网公开网站没有可见的漏斗入口机制是否存在私有的站外等候名单仍未知
学术资助 / 奖学金计划(同行基准:Periodic Labs)Mirendil 没有;Periodic Labs 已启用2026Periodic Labs 官网最接近的同行已运行一个贴近 GTM 的学术外联计划,Mirendil 没有无法确认 Mirendil 是否计划类似项目

所有行都是商业准备度的代理信号,不是直接采用指标;Mirendil 未披露 ARR、订单额、活跃账户或使用率数据。

[CU011, CU012, CU013, CU014, CU015, CU016]
FU002: 采用 / 部署漏斗

在 Mirendil 自身使命表述识别出的四类买方细分中,除初始细分识别外,每个采用阶段当前均为 0。

数值统计客户细分表中的四类买方细分;第一阶段之后均为 0,因为截至运行日期,没有公开来源披露任何 Mirendil 专属触达、试点、部署或独立客户证据。

[CU001, CU005, CU006, CU007, CU008, CU028]

6.3 相邻客户证明:Mirendil 同行里的「好」是什么样

Mirendil 自身没有客户证据,本章只能拿最接近竞品公开的最清晰证明做标尺。Isomorphic Labs 披露两项点名的药企合作:Eli Lilly(预付款加最高 $1.7 billion 里程碑款)和 Novartis($37.5 million 预付款,最高 $1.2 billion 里程碑款,2025 年 2 月从一个项目扩展到最多三个)。Recursion 走得更远,在合作伙伴页发布来自 Bayer AG 的 Joerg Moeller 和 Roche 的 James Sabry 的点名推荐语,并披露合作经济条款:Bayer 最高 $1.5 billion,Roche/Genentech 至今已收 $213 million,Sanofi 至今记录 $134 million 里程碑款。Recursion 的 2025 Form 10-K 还披露,两个客户贡献了几乎全部 2025 年运营收入——这是真证明,但也证明了极端集中。Periodic Labs 的创始人来自与 Mirendil 团队相同的前沿实验室人才池;它在官网直接说明,正在为一家未具名半导体制造商的散热问题训练定制 AI 智能体,并由 Observer 报道独立印证,其客户群还包括航天和国防公司。Lila Sciences 位于这个谱系的弱端:它表示正在欢迎第一批商业客户,并参加 BIO 2026 Partnering 计划,但没有点名具体药企客户。Mirendil 目前这些都没有——没有推荐语,没有未具名合作,也没有 Partnering 目录条目。[CU018, CU019, CU020, CU021, CU022, CU023]

具名客户证据表
客户 / 实体分群部署 / 用例生产 / 试点披露结果 / 条款限制
Eli Lilly(通过 Isomorphic Labs 合作——非 Mirendil 客户)医药多靶点小分子药物发现合作生产 / 自 2024 年 1 月起持续首付款,另有最高 $1.7B 的里程碑付款和分级特许权使用费仅作同行基准;Isomorphic Labs 是竞争对手,不是 Mirendil
Novartis(通过 Isomorphic Labs 合作——非 Mirendil 客户)医药多靶点研究合作,从一个项目扩展到最多三个项目生产 / 自 2024 年 1 月起持续;2025 年 2 月扩展$37.5M 首付款,另有最高 $1.2B 的里程碑付款和分级特许权使用费仅作同行基准;扩展显示该交易结构具备账户增长可能
Bayer / Roche / Sanofi(通过 Recursion 合作——非 Mirendil 客户)医药表型组学驱动的肿瘤、纤维化和免疫学药物发现生产 / 持续多年最高 $1.5B 潜在付款(Bayer);迄今已收到 $213M(Roche / Genentech);迄今 $134M 里程碑(Sanofi)仅作同行基准;直接竞争对手有三个单独具名药企客户
未具名半导体制造商(通过 Periodic Labs 合作——非 Mirendil 客户)工业 / 材料用于散热 R&D 和实验数据分析的定制 AI 智能体合作活跃;生产 / 试点状态未确认未披露量化结果;仅有定性描述最接近 Mirendil 自称跨领域愿景的同行类比;客户仍未具名
Mirendil(标的公司)N/AN/A商业化前——尚未发布产品披露的具名客户、试点或设计伙伴为零这是本章核心发现;其他所有行都是可比案例,不是 Mirendil 自身记录

本表用于标定 Mirendil 所在类别的证据门槛,不是把第 1-4 行说成 Mirendil 自身客户。

[CU018, CU019, CU020, CU021, CU022, CU023]
FU003: 客户证据矩阵

截至 2026 年 7 月,Mirendil 与最接近的直接科学 AI 同行在证据质量、生产成熟度、结果具体性和留存可见度上的对比。

[CU018, CU019, CU020, CU021, CU022, CU023]

6.4 现状替代方案与采购摩擦

科学研发买方没有被迫采用尚未验证的自主 AI 研发平台;他们可以继续把人类科学家与 CRO、现有建模套件、数据提供商,以及 Anthropic 化学辅助工作或 NVIDIA 开发者栈这类通用 LLM 搭配使用。采用研究进一步说明,另一条路有多难:McKinsey 2025 年 11 月全球调查发现,62% 的组织至少在试用 AI 智能体,但只有 23% 正在规模化任何智能体用例;ZS 2026 年制药 / 生物技术调查发现,只有 17% 技术高管称 AI 在研究和发现环节带来可衡量价值;Gartner 预测,到 2027 年底,超过 40% 智能体 AI 项目会因成本、价值不清或风险控制问题被取消。MIT NANDA 2025 State of AI in Business 研究基于 300 多个已披露 AI 项目和 52 家组织访谈,发现约 95% 企业生成式 AI 试点未能扩展到生产。采购机制因买方不同而异:高校通常使用 EDUCAUSE 的 HECVAT 工具包做供应商安全和隐私审查;药企和生物技术买方则按多季度、里程碑制合作组织交易,而不是走软件销售周期,Isomorphic Labs 和 Recursion 的先例已经说明这一点。Mirendil 没有发布信任、安全或合规页面;一旦有产品要卖,这会是通过任一路径的前置条件。[CU029, CU030, CU031, CU032, CU033, CU034]

买方采购路径和现状替代方案表
买方分群现状替代方案采购机制 / 摩擦典型周期(同业证据)尽调含义
制药 / 生物技术 R&DCRO、内部计算生物学团队、通用 LLM、现有建模套件科学尽调需要多个季度,交易结构按里程碑推进;Isomorphic Labs 和 Recursion 的先例支持这一点Isomorphic Labs 与 Novartis 的关系大幅扩展,发生在原始交易约 13 个月后(2024 年 1 月至 2025 年 2 月)即使产品跑通,Mirendil 的首个制药客户大概率仍要花多个季度
高校 / 学术实验室现有 HPC 集群、开源工具、通用 LLM 订阅HECVAT 式安全和隐私供应商审查,由 EDUCAUSE、Internet2 和 REN-ISAC 成员机构协调公开记录未量化时间;HECVAT 是明示的事实标准审查工具校园级采用前,Mirendil 需要 SOC 2 式安全文档和 HECVAT 回复
工业 / 材料实验室(半导体、机器人)内部 R&D 团队、合同工程、通用 LLM按 Periodic Labs 先例,靠直接业务开发关系推进;未披露公开 RFP 流程未量化时间;Periodic Labs 自身启动后约 12 个月内披露了该合作双边 BD 关系而非正式采购,可能是拿下首个客户的最快路径
其他实验室的内部 AI 研究团队基于开放模型和 NVIDIA 开发者栈自建内部自建 / 采购评估会很复杂,因为业务重叠:Mirendil 可能被视为竞争实验室,而不是供应商公开记录未量化鉴于竞争者章节记录的竞争重叠,该细分短期内可能最不可能买单
通用企业 / 智能体 AI 买家(所有细分)现状的人工作业流,加上狭窄点解决方案式 AI 工具高试点失败率后,市场普遍怀疑MIT NANDA:约 95% 的 GenAI 试点未能规模化;Gartner:预计 40%+ 的智能体 AI 项目到 2027 年会取消即使 Mirendil 产品跑通,买方怀疑仍会存在,不取决于公司自身执行

周期时间数字来自已披露的同业时间线和采用研究,不来自任何 Mirendil 专属采购记录。

[CU029, CU030, CU031, CU034, CU035, CU036]
FU004: 买方采购与评估流程

基于同业 GTM 机制推演的采购与买方评估流程;每一次流转,都依赖 Mirendil 仍未公开迈出的步骤。

流程阶段按 Isomorphic Labs、Recursion 和 Periodic Labs 的 GTM 模式,以及 HECVAT / 供应商安全文献推演;除初始曝光外,Mirendil 尚未确认进入任何阶段。

[CU001, CU029, CU035, CU036, CU017]

6.5 留存、扩张、集中度与客户结论

没有公开来源披露 Mirendil 的净收入留存、毛收入留存、流失、续约率或合同期限,公司也未出现在 G2、Capterra、Gartner Peer Insights 等独立评论平台。对商业化前同行来说,这些缺口并非 Mirendil 独有,但仍然是缺口;下面留存表中的每个单元格都只能是 null,不是因为披露为零,而是因为没有披露。更可执行的发现面向未来:如果 Mirendil 签下首个设计伙伴,起步阶段的收入集中风险很可能很重,类似 Recursion 已披露的模式——两个客户贡献了几乎全部 2025 年收入。集中之后是变成持久、扩张的关系——像 Isomorphic Labs 与 Novartis 的交易从一个项目扩到三个——还是变成单一、暴露的关系,取决于 Mirendil 尚未披露的商业模式选择:平台许可、按里程碑计费的科学合作,或内部保留 IP、未来再变现。每种模式都对应不同扩张机制,也对应未来尽调应看的不同指标。因此,本章结论是程序性而非实质性:Mirendil 的客户故事目前还不能承销,只能用类比描摹;今天价值最高的尽调动作,是直接确认是否存在任何点名设计伙伴、试点协议或意向书,无论阶段多早。[CU022, CU037, CU038, CU039, CU040, CU041]

留存 / 重复使用 / 满意度表
指标数值 / 空值分群置信度尽调请求
净留存率(NRR)空值——未披露全部一旦签下任何设计伙伴,要求提供队列级 NRR
总留存率(GRR)/ logo 流失空值——未披露全部要求提供 GRR / 流失历史和任何早期终止情况
合同期限 / 续约结构空值——商业模式未定(平台许可 / 里程碑合作 / 内部 IP)全部任何交易签署前,确认拟采用的商业结构
客户满意度 / NPS / 独立评论空值——未发现 G2、Capterra 或 Gartner Peer Insights 上榜全部识别任何私人参考客户,用于直接尽调访谈
重复使用 / 使用率空值——未披露使用或使用率数据全部一旦出现试点,要求提供使用日志或使用率指标

每个单元格都必然为空,因为 Mirendil 没有可衡量的已披露客户,并非已有测量值被扣留。

[CU037, CU038]
扩张和集中风险表
扩张驱动 / 集中风险类型影响尽调路径
第一个设计伙伴可能占近期收入的 100%集中风险高——与 Recursion 披露模式一致:两个客户几乎贡献 2025 年全部收入一旦任何交易签署,要求提供潜在分群的管线数量和收入占比
商业模式模糊(平台许可 / 里程碑合作 / 内部 IP)集中 / 扩张模糊高——不同模式意味着完全不同的扩张经济性和切换成本获取管理层的 GTM 和商业模式路线图
尚未招聘 GTM / 销售 / 客户成功岗位扩张能力风险中——即便签下设计伙伴,也会限制获客或扩展账户的能力跟踪 Ashby 招聘板上的 GTM 职能岗位,将其作为领先指标
同行中基于里程碑的扩张先例(Isomorphic Labs 与 Novartis 的交易从一个项目扩展到最多三个项目)扩张驱动中等偏正面——显示该类别的合作模式可在单个账户内扩张确认 Mirendil 是否打算采用类似的里程碑扩张结构
Mirendil 对 NVIDIA 的算力依赖(见产品与技术章节)可能卡住交付能力集中风险(供给侧)中——算力瓶颈可能限制 Mirendil 同时支持的项目数确认为客户交付和内部研究分别预留的 GPU 容量

集中度评估由分析师根据同行披露模式推断;没有 Mirendil 特定集中度数据可直接量化。

[CU022, CU039, CU040]

6.6 图表

Chapter 07

07风险

7.1 严重性框架与首要风险

Mirendil 在公开证据出现前就已融资,风险栈应从足以推翻论点到可管理依次排序。最高严重性风险是执行:官方和投资人来源描述的是一家前沿 AI 研发实验室和 20 人创始班底,独立报道和负面评论仍显示没有公开产品、收入、客户、基准测试包或法律 / 信任界面。由此形成一条狭窄的尽调标准:公司必须拿出非公开证据,证明自主研究闭环能在创始人控制的演示之外运转。第二类是技术与完整性风险,因为会编造引用、研究主张或实验理由的科学 AI,摧毁买方信任的速度会快过普通 SaaS 缺陷。第三是融资 / 估值风险:风投数据显示资本集中到更少、更大的 AI 轮次,Gartner 警告许多智能体项目会被取消,高种子轮价格也让转向空间变小。法律、监管、合作伙伴、人才和客户风险都重要,因为它们最终可能导向同一个结果:下一轮融资窗口前,没有可信的外部采用。[CR001, CR002, CR003, CR004, CR005, CR016]

FR001: 风险热力图

基于公开证据、外部负面信号和缓释成熟度,对 Mirendil 风险严重性排序。

定性热力图;发生概率和影响锚定引用证据与公开披露缺口,并非数值化损失模型。

[CR002, CR003, CR006, CR007, CR009, CR016]

7.2 监管与法律暴露

监管风险目前还不是单一审批阻断点,因为 Mirendil 没有披露已交付产品或受监管临床工作流。风险取决于路径。如果系统作为通用 AI 研究助手在欧洲销售,EU AI Act 的透明度、版权、训练内容摘要和系统性风险义务可能变得相关;高风险部署在投放市场前,还需要风险管理、数据治理、日志、文档、人类监督、稳健性、网络安全和准确性控制。如果 Mirendil 进入药物发现、医疗软件或临床决策支持,FDA 来源显示,AI 医疗器械和软件修改可能需要合适的上市前路径、生命周期管理,以及预先设定的变更控制计划。IP / 法律姿态同样重要。律所分析和 AI 商业秘密案件表明,模型权重、训练数据、系统 prompt 和调优方法都是可转移资产;出身前沿实验室的创始人不能只靠声誉,必须证明 clean-room 来源、离职认证纪律、监控和书面保密控制。没有确认到公开诉讼或认证记录,因此缓释仍是一条尽调路径,不是已证明的控制。[CR006, CR007, CR008, CR009, CR010, CR011]

监管 / 法律风险登记册
规则 / 许可 / 案件司法辖区状态发生概率严重性缓释措施剩余暴露尽调路径
前雇主商业秘密 / 员工流动暴露美国 / 加州及联邦商业秘密法未发现 Mirendil 公开诉讼;风险根据前沿实验室人员流动和 AI 商业秘密判例推断关键洁净室开发记录、发明转让链、离职证明、访问日志和外部律师备忘录在律师确认没有受限信息、模型权重、提示、训练数据或保密方法转入前,暴露仍高索取创始人离职文件、录用函、IP 转让、前雇主契约、取证访问日志和董事会级法律备忘录
EU AI Act GPAI / 高风险义务欧盟GPAI 规则自 2025 年起生效;透明度规则 2026 年 8 月生效;高风险规则按用例适用将产品模块映射到 GPAI、高风险和有限风险类别;建立技术文档、日志、人工监督、版权 / 训练内容摘要中,因为功能演进后,外部科学部署可能落入高风险或透明度覆盖范围在欧盟试点或公开分发 GPAI 前,取得欧盟律师分类备忘录
FDA AI 医疗器械 / SaMD 路径美国只有 Mirendil 进入受监管医疗器械或临床决策支持工作流时才触发低至中将非临床发现工具与受监管主张隔离;为自适应软件准备预提交计划、生命周期管理和 PCCP 策略中,因为药物发现野心可能滑向医疗产品主张或受监管决策支持要求管理层说明预期用途、主张、用户工作流,以及是否已有 FDA Q-submission 或监管律师审查
加州 AI 训练数据透明度 / 数据集摘要压力加州 / 美国法律分析提到 X.AI 对公开训练数据集摘要和商业秘密具体性的挑战低至中在任何受覆盖披露义务触发前,梳理数据集、许可、清洗流程和商业秘密论证中,因为训练数据来源未公开,可能与透明度制度冲突索取数据清单、许可栈、版权政策,以及律师对州和欧盟披露义务的意见
研究诚信 / 论文幻觉责任科学出版商、客户和研究机构不是政府许可,但会卡住科学工作流采用要求引用验证、基于来源的输出、人工审阅、审计日志,以及科学主张纠错工作流中高,直到外部评测证明虚假引用和虚假主张率受控用已知文献跑盲测科学任务,并在试点前要求错误分类体系
公开主张、安全和负责任扩展治理美国、欧盟、客户采购公开资料未显示 Mirendil 有安全、信任或合规计划外部发布前采用 NIST 式治理和前沿实验室安全阈值;指定合规负责人中,因为缺少公开控制会拖慢受监管客户或企业买家索取安全政策、风险登记册、事件流程、负责任扩展阈值和客户安全问卷回复

各行按严重性排序;适用性取决于最终产品主张和司法辖区,因为 Mirendil 尚未披露已发布产品或客户工作流。

[CR006, CR007, CR008, CR009, CR010, CR011]

7.3 运营、安全与研究诚信风险

Mirendil 的产品承诺对质量失败格外敏感,因为用户预期会在科学和工程工作流里信任 AI 系统。研究诚信来源给出具体负面信号:数百篇 NLP 会议论文出现过幻觉引用,生物医学审计在 2026 年初 PubMed 索引论文中发现每 277 篇就有 1 篇包含伪造参考文献,一项大型 arXiv 审计估计,2025 年主要研究库里共有 146,932 条幻觉引用。这些事实不能证明 Mirendil 会失败,但它们让验证基础设施成为核心产品要求,而非锦上添花。Gartner 的取消预测还给出运营教训:成本、价值和风险控制没有在投产前落地,智能体就会失败。NIST 的 AI Risk Management Framework 提供了一套围绕可信设计、评估和持续风险管理的缓释模板。对 Mirendil 来说,尽调应要求一套评测套件,在任何外部客户工作流依赖模型生成的科学输出前,先测试事实性、引用完整性、实验可复现性、安全边界、事件响应和人工批准点。[CR016, CR017, CR018, CR019, CR020, CR021]

运营 / 质量 / 安全风险登记册
失效模式发生概率严重性缓释成熟度剩余暴露未解决缺口
自主研究智能体生成看似可信但虚假的科学主张、引用或实验依据关键公开未知;缓释需要基于来源的检索、验证和人工确认在虚假主张和虚假引用率披露前,暴露仍高Mirendil 没有公开评测套件、红队报告或可复现性基准
产品前系统无法从内部 AI 研究用户迁移到 ML 基础设施较弱的外部科学家中高关键只有叙事;投资者材料描述了路径,但没有客户可用封装高,因为商业化没有公开证据未出现具名试点、部署工作流、入门流程或定价证据
智能体工作流的成本和复杂度拖住生产采用可用受限用例、ROI 门槛和工作流重设计缓释中高,因为 Gartner 称许多项目因成本、价值和风险控制问题失败Mirendil 没有公开单位经济、延迟、算力成本范围或客户 ROI 模型
安全或数据泄露故障暴露专有科学、模型或客户数据公开未知;需要密钥管理、租户隔离、日志、DLP 和事件响应中高,因为科学客户可能共享敏感 IP 和实验数据未发现 SOC 2、ISO、信任中心、DPA、状态页或事件历史披露
模型 / 更新漂移在部署后削弱已验证输出FDA PCCP 式生命周期思维可用于缓释自适应系统风险在模型变更治理和回滚程序披露前,暴露为中没有公开版本管理、验证、监控或回滚政策
创始人主导的研究文化无法扩展成运营节奏可用项目管理、客户成功、合规负责人和里程碑治理缓释中,因为创始团队精英但规模小没有公开组织架构、运营指标或创始人之外的领导层

失效模式结合了 Mirendil 公开披露缺口,以及智能体 AI 项目取消、科学幻觉和 AI 风险管理控制方面的外部证据。

[CR002, CR003, CR004, CR016, CR017, CR018]
FR002: 风险传导图

技术、法律、运营和财务风险如何传导至客户采用、融资和估值。

传导路径是引用证据支撑的逻辑风险通道,不代表概率。

[CR016, CR018, CR020, CR021, CR022, CR024]

7.4 合作伙伴、依赖与市场传导风险

合作伙伴风险是一张依赖图,不是单一供应商问题。NVIDIA 参与支持了算力故事,但也凸显加速器获取、云采购和前沿模型工具链上的集中度;这些资源可能稀缺、昂贵,或优先分配给更大客户。资本伙伴依赖同样有两面性:a16z 和 Kleiner Perkins 的背书提升融资与招聘可信度,但 CB Insights 数据显示,2025 年和 2026 年 Q1 的 AI 融资变得头部化,巨额融资轮和模型开发者吸收了越来越高的资本份额。科学买方预算是另一条通道。CRS 和 C&EN 描述了拟议的 FY2026 NSF 削减;如果落地,或在研究机构间形成回声,可能在 Mirendil 尝试商业化时挤压非药企买方预算。竞争性合作伙伴风险也高,因为 Periodic Labs、Lila Sciences、Isomorphic Labs、Anthropic、OpenAI、Google DeepMind 以及垂直 AI 药物发现玩家都在争夺算力、研究员、科学可信度和企业注意力。缓释不是泛泛的合作伙伴话术,而是算力、云、人才管线和设计伙伴渠道上的冗余。[CR027, CR028, CR029, CR030, CR031, CR032]

合作伙伴 / 依赖风险登记册
依赖项交易对手作用集中度失效场景严重性缓释措施剩余暴露
加速器算力和 GPU 供应NVIDIA 加云端 GPU 提供商训练 / 推理基础设施和投资者信号可能较高,因为前沿 AI 研发系统算力密集产能、定价、出口配额或战略偏向限制路线图速度关键预留产能,分散云 / 加速器供应商,按已验证科学任务基准测试成本在签署算力条款和审阅成本曲线前,暴露仍高
基础模型工具链和研究基础设施内部栈加前沿模型生态基础模型、评测、代码智能体、RAG、实验编排高,因为公开产品架构尚未披露在位前沿实验室把组件商品化,或阻断关键 API / 工具访问掌握关键评测 / 数据层,避免单一外部模型依赖中高,因为架构未公开
资本提供方和后续融资市场a16z、Kleiner Perkins、NVIDIA、未来后期投资者种子轮资金期限、信号、后续融资储备高,因为尚无收入时估值已约 $1BAI 融资窗口收窄,或下一轮要求尚未达成的证明烧钱与里程碑绑定,并在扩大固定成本前锁定内部投资方支持条件在资金期限、烧钱速度和储备条款披露前,暴露中高
科学买方预算高校、公共实验室、生物技术 / 制药 R&D 团队未来客户和设计伙伴中;买方组合未披露NSF 或学术资金压力削减实验性软件预算,或推迟试点中高优先争取制药 / 工业预算和有资金支持的设计伙伴,而不是无预算学术兴趣中,因为客户细分未公开
人才市场和创始团队Neyshabur、Mehta、创始研究员、外部招聘方核心智力产出和招聘飞轮高,因为公开叙事围绕一支小型精英团队大型科技公司薪酬报价或创始人离职会打破路线图可信度留任方案、接班计划、知识沉淀、符合不招揽条款的招聘流程在归属、留任和接班细节审阅前,暴露仍高

依赖项按对路线图、融资和客户可信度的传导程度排序;未获得私人合同或储备承诺。

[CR001, CR003, CR004, CR005, CR027, CR028]
FR003: 依赖关系图

Mirendil 若要达成客户和融资里程碑,必须让关键依赖保持足够冗余。

依赖图基于公开利益相关方与风险证据;无法获得私下合同。

[CR001, CR003, CR004, CR005, CR006, CR007]

7.5 人员、执行、融资与客户转化风险

人员风险格外集中,因为公开故事高度依赖 Behnam Neyshabur、Harsh Mehta,以及来自 Anthropic、xAI、Google DeepMind 和 OpenAI 的 20 人精英班底。CNBC 对人才战的报道说明,不能假定留才自然成立:大科技公司正为稀缺 AI 研究员支付异常高的薪酬包,帮助 Mirendil 招人的同一个市场,也能把人才拉走。执行风险比普通创业不确定性更重,因为公开披露先展示使命、资本和履历,产品证明还没出现。TechCrunch 的种子轮估值分析称,AI 种子轮投资人越来越早要求上线产品、用户、收入、分发和留存,高价格也减少了试错或转向空间。因此,Mirendil 约 $1 billion 的估值放大了技术或客户里程碑失手的后果。客户风险目前仍主要是证据缺席:没有公开设计伙伴、生产部署、留存指标、定价或买方集中度数据。可执行的缓释措施,是一份绑定融资闸门的里程碑计划,明确负责人、留任包、proof-of-work 演示和客户背调。[CR001, CR002, CR003, CR004, CR005, CR018]

人员 / 执行风险登记册
角色 / 职能依赖或缺口发生概率严重性缓释措施尽调路径
Behnam Neyshabur / CEO 技术愿景公开公司叙事高度依赖其 Anthropic 和 Google 前沿 AI 背景关键记录创始人股权归属、接班、路线图归属和决策权审阅雇佣协议、股权归属、董事会纪要、接班计划和关键技术里程碑
Harsh Mehta / CTO 平台执行投资者来源将自主研究平台论点与其在 Anthropic 的 autoresearch 工作绑定关键把平台架构所有权分散到资深负责人,并要求可复现演示审阅架构路线图、CTO 职责、团队深度和技术债登记册
20 人创始团队唯一公开运营规模指标;更广泛的员工规模和留任经济学未披露留任授予、技术入职、招聘计划和知识管理节奏索取匿名薪酬区间、期权刷新计划、流失仪表盘和招聘漏斗
合规 / AI 安全 / 信息安全负责人EU AI Act、FDA 路径、NIST 式 AI 风险管理或安全控制都没有公开负责人中高在受监管试点前,聘请或指定合规 / 安全负责人询问具名负责人、预算、政策路线图和第三方审计日程
商业 / 设计伙伴负责人未公开出现具名客户、试点、定价或客户成功职能中高任命 GTM 负责人,配套签署的设计伙伴里程碑和可引用客户目标索取销售管线、LOI、试点 SOW、定价模型和客户安全问卷
财务 / 烧钱治理公开烧钱速度、资金期限和算力承诺未知时,资本强度大概率较高月度烧钱关口绑定技术和客户证明,而不是只看员工规模审阅 24 个月计划、算力预算、供应商承诺和内部投资方储备函

角色风险不是对个人的判断;它标出公开证据集中或缺失的位置,以及私人尽调应核验的内容。

[CR001, CR002, CR003, CR004, CR005, CR027]

7.6 缓释措施、监测节奏与否决标准

尽调答案不是回避所有风险,而是定义哪些风险能靠证据消除,哪些风险应直接杀死论点。缓释应按风险传导到估值的顺序分阶段推进。第一,要求非公开技术证明:评测结果、对抗性科学任务、引用验证控制、可复现性检查和红队日志。第二,要求法律来源证明:创始人离职文件、IP 转让链、引入发明审查、离职认证、模型 / 数据访问控制,以及外部律师关于前雇主暴露的备忘录。第三,要求监管分层:一棵书面决策树,区分内部研究工具、EU GPAI 义务、高风险 AI 用例、FDA 监管的软件 / 设备用途,以及非临床药物发现工作流。第四,要求运营冗余:算力预留条款、多云故障切换、招聘管线、留任经济学和设计伙伴证据。否决标准必须刻意可衡量:没有可复现基准、没有干净 IP 备忘录、没有可信合规负责人、没有点名外部试点,或在没有客户 / 产品证明的情况下尝试下一轮融资,都应把投资立场推向回避或大幅价格重置。[CR006, CR007, CR009, CR011, CR016, CR020]

缓释措施和否决标准表
风险可监测触发项阈值 / 事件行动含义
执行 / 产品证明外部可复现基准,或投资者观察任务下的私人演示种子轮交割后两个季度内,没有可信基准、演示或试点证据暂停后续投资;要求价格重设,或在产品证明出现前回避
科学诚信盲测任务中的虚假引用、虚假主张或不可复现实验比例重大幻觉仍高于约定阈值,或核心主张无法验证阻止受监管 / 科学试点,并要求修复后再承销收入
IP / 商业秘密来源外部律师备忘录和洁净室证据律师无法确认创始人 / 团队义务、数据来源,以及没有受限前雇主信息不要继续投资;视为打破投资论点的法律暴露
监管路径欧盟 / FDA 分类备忘录和具名合规负责人没有产品用途分类、没有合规负责人,或路径获批前提出受监管主张暂停受影响发布;要求董事会级监管整改
融资 / 估值资金期限、烧钱速度、内部投资方支持、下一轮证明标准下一轮融资在没有具名试点、基准证明或内部投资方支持的情况下启动,且估值高于证据支撑假设会降价轮 / 高稀释;将估值标记为偏高至昂贵
人才集中创始人留任、关键人物离职、招聘速度任一创始人离职、两名或以上具名技术负责人离开,或关键岗位空缺两个季度重新承销领导层风险,并要求接班证明
算力依赖预留产能、每项已验证科学任务成本、供应商冗余没有签署的算力计划,或单位成本范围不足以支撑路线图里程碑下调情景概率,并在扩大烧钱前要求供应商条款清单
客户转化具名设计伙伴、SOW、付费试点或可引用成果下一个 IC 检查点前,没有外部设计伙伴或买方验证的工作流将市场需求视为未证实,并推迟估值上调

否决标准被刻意操作化为尽调触发项;投资委员会投票前,应结合管理层数据校准阈值。

[CR002, CR003, CR006, CR007, CR009, CR016]

7.7 图表

Chapter 08

08估值

8.1 最终建议与价格纪律

建议是跟踪 Mirendil,只通过有条件的继续研究路径推进,而不是按报道中约 $1 billion 的投后估值直接买入。最强证据确实存在:Mirendil 自称 a16z 和 Kleiner Perkins 领投 $200 million 种子轮,NVIDIA 参与;独立报道也称这笔融资估值 $1 billion;创始人的前沿 AI 研究资历异常匹配。承销问题在于,公开基本面撑不起这个价格。没有披露收入、客户数量、基准测试包、路线图、股权结构表、清算优先栈或董事会材料。只有在非公开尽调证明其技术系统有阶跃式突破、首批用户可信,并且跑道足以支撑它在没有惩罚性结构的情况下,以高得多的价格进入下一轮时,$1 billion 的入场价才可能合理。[CV001, CV002, CV003, CV004, CV033, CV034]

建议摘要表
决策项当前判断证据基础决策含义
建议跟踪 / 继续研究创始人-市场匹配精英级,种子轮支持方顶级,但公开证据缺少产品、收入、客户、基准和治理证明。没有私人证据和结构审查前,不要按全价批准买入。
信心中低融资和团队事实交叉验证充分;价值驱动因素大多仍是私人信息,尚未验证。在证明改善前,仓位应保持期权式。
风险评级产品前前沿 AI,人才和算力昂贵,技术野心宽,且 AI 市场评论偏负面。投入资本前,要求硬里程碑和下行保护。
估值立场昂贵尽管没有公开收入,~$1B 投后估值远高于普通种子轮基准,且接近公开 AI 药物发现估值锚。入场必须靠专有技术证明支撑,不能只靠赛道热度。
决策含义仅条件性通过公司值得主动尽调,因为上行赛道真实存在;但当前公开证明未跨过买入门槛。跟踪,争取信息权,并在技术 / 客户证据出现后重新评估。

这是一张 IC 决策表,不是通用质量分;基于据报 2026 年 6 月融资条款,该立场对价格和证据敏感。

[CV001, CV002, CV033, CV034, CV035, CV036]
FV001: 建议逻辑

从规模与验证,到风险与价格,再到附条件跟踪建议的证据链。

[CV001, CV002, CV024, CV026, CV035, CV036]

8.2 估值背景、市场支撑与进入纪律

市场背景解释了这轮融资为何发生,但不足以在缺少更多证明时完全支撑这个价格。Carta 和 Crunchbase 都显示,AI 推高了早期定价:Carta 报告种子轮投后估值创新高,正常种子轮稀释约 19% 至 20%;Crunchbase 报告,2025 年种子轮融资异常集中在超大型 AI 轮次中。TechCrunch 同样描述了 AI 种子公司拿到更高估值,但也指出投资人越来越期待真实用户、收入和更快里程碑。Mirendil 有这个等式中的履历一侧,公开证据却缺少牵引力一侧。与公开 AI 药物发现公司相比,入场价显得激进:Recursion 和 Schrödinger 披露了带有收入、管线和亏损的公开市场估值锚,而 Mirendil 在产品证明前就按期权价值定价。[CV005, CV006, CV007, CV008, CV009, CV010]

正方 / 反方论点表
论点证据方向什么会改变判断
正方:精英级创始人-市场匹配创始人和创始团队来自前沿 AI 机构,正面契合自动化 AI R&D。若私人背调显示执行弱、留任差或创始人冲突,信念会下降。
正方:超大品类期权智能体 AI、AI 药物发现和自主实验室市场,均预计将在 2026 年基数上快速增长。如果用例坍缩为狭窄内部工具,TAM 应大幅折价。
正方:私人市场稀缺溢价Thinking Machines、SSI、Mistral、xAI、Periodic 和 Lila 显示,投资者愿为稀缺前沿 AI 团队支付高溢价。私募 AI 估值回调,或下一轮可比融资失利,都会把立场从「昂贵」推到「回避」。
反向论点:缺少公开产品验证截至运行日,Mirendil 没有公开基准测试包、SKU、路线图、收入或客户数量。可复现 demo 和第三方基准测试套件,能补上最大的证据缺口。
反向论点:估值跑在基本面前面公司定价接近甚至高于已披露收入和文件的上市 AI 科学公司。更低入场价格、强优先权保护,或证明其能力处于同类领先的私下证据,才可能让价格变得可接受。
反向论点:炒作周期与执行风险Gartner、Crunchbase、CNBC 和 WEF 的资料显示,智能体 AI 项目取消、AI 风险泡沫和估值修正都已引发担忧。若能证明烧钱克制、客户主动拉动、融资按里程碑推进,风险会下降。

这些论点把前文运营章节、估值可比对象和负面市场证据合并判断。

[CV003, CV004, CV022, CV024, CV026, CV027]
可比估值表
可比对象指标倍数 / 估值 / 状态参考意义局限来源依据
Mirendil种子轮融资约 $200M 种子轮,投后估值约 $1B;简化计算下新投资人持股约 20%。直接锚定入场价格和稀释。没有公开产品、收入、技术基准、股权结构或治理细节。Mirendil、SiliconANGLE、Cryptonomist。
Thinking Machines Lab前沿 AI 种子轮可比对象$2B 种子轮,估值 $12B;成立不到一年,且尚未披露具体在做什么。显示市场愿意把精英前沿 AI 团队定价到远高于普通种子轮的水平。泛前沿 AI,而非科学研发;由前 OpenAI CTO 领衔,是极端离群案例。TechCrunch、TechFundingNews。
Safe Superintelligence前沿 AI 私募可比对象据报融资 $2B,估值 $32B;网站 / 产品信息仍很稀疏。显示投资人能在产品披露前,因顶级 AI 创始人履历支付极高价格。安全 / 超级智能使命,不是 Mirendil 的 AI 科学商业化路径。TechCrunch 及市场综述来源。
Periodic LabsAI 科学 / 材料可比对象$300M 种子轮;据报后来讨论约 $7B 估值,并已有半导体客户牵引。在自主实验室和材料发现方向上,是最接近的私募 AI 科学稀缺可比对象。实体实验室模型和客户牵引可能比 Mirendil 的公开证据更具体。TechCrunch、TechFundingNews。
Lila Sciences自主科学工厂可比对象$350M A 轮;累计融资 $550M;已描述首批客户群和 AI Science Factories。显示资本强度和面向客户的科学超级智能叙事。Flagship 孵化的平台,公开商业化表述比 Mirendil 更多。Lila 官方、Fierce Biotech。
Isomorphic Labs战略药企合作与 Lilly 合作的预付款 $45M,总潜在价值最高 $1.7B。里程碑经济学展示 AI 科学平台变现的一条路径。Alphabet 支持的药物发现,与 Mirendil 的宽口径实验室平台不同。Isomorphic Labs PRNewswire、市场分析来源。
Recursion Pharmaceuticals上市 AI 药物发现锚点2025 年公开流通市值 $2.03B;2025 年收入 $74.7M;研发 $475.3M;净亏损 $644.8M。给已有收入和监管文件的 AI 原生发现公司提供公开市场理性校验。上市 biotech 管线风险不同于 Mirendil 的产品前平台风险。SEC 10-K、Recursion 文件证据。
Schrödinger上市计算化学锚点2025 年公开流通市值 $1.12B;2025 年收入 $255.9M;研发 $173.1M;净亏损 $103.3M。显示公开市场会给有收入的计算发现平台估值,但要求纪律。成熟软件与药物发现混合业务,不是前沿 AI 种子轮实验室。SEC 10-K、Schrödinger 文件证据。
Mistral AI前沿模型实验室可比对象CNBC 报道 2025 年估值 €11.7B;TechCrunch 报道 2026 年洽谈估值约 €20B。模型采用、主权叙事和战略资本汇合时,前沿实验室估值可以快速上行。Mistral 有模型和采用信号;Mirendil 没有公开产品。CNBC、TechCrunch。
xAI规模化前沿 AI 可比对象TechCrunch 报道 $20B Series E,X/Grok 月活用户约 600M。说明前沿模型基础设施需要的资本尺度。阶段、分发和风险画像都不可比。TechCrunch、CNBC 人才战背景。

这些行不是同业中位数;它们定义估值走廊和证据标准。每个可比对象都受阶段、产品、客户和资本结构差异限制。

[CV001, CV002, CV012, CV013, CV014, CV015]
FV002: 估值敏感性

重大尽调结果会怎样影响据报 $1B 入场价附近的公允私募估值。

数值只是围绕 $1B 参考标记、以 $B 计的示意性调整,不是正式 DCF。

[CV005, CV006, CV022, CV024, CV035, CV040]

8.3 乐观、基准与悲观估值情景

情景区间应被视为承销纪律,而不是预测精度练习。乐观情景假设 Mirendil 证明自有自主 AI 研发闭环,把第一批技术用户转化为设计伙伴,并获得与稀缺前沿实验室资产对比的资格,而不是被视作普通种子轮软件。在这种情况下,后续估值高于入场价可以成立,因为 Thinking Machines、SSI、Periodic、Lila、Mistral 和 xAI 说明,私人资本愿意为前沿 AI 稀缺性支付极高价格。基准情景假设研究进展亮眼,但下一轮融资前没有公开收入证明;这能保住期权价值,却限制了从 $1 billion 入场的上行空间。悲观情景假设基准疲软、治理不透明或市场压缩,结果是降估值融资或结构化延期风险,而不是干净加价。[CV011, CV012, CV013, CV014, CV015, CV016]

乐观 / 基准 / 悲观情景表
情景假设估值 / 回报逻辑关键风险概率信号
乐观技术基准测试包显示自主 AI 研发闭环具备耐久性;首批 AI 构建者 / 科学用户转为设计合作伙伴;内部和战略投资人竞争跟投。示意估值在重大稀释前为 $3B-$6B;若后续融资干净,$1B 入场价的账面涨幅为 3x-6x。算力扩展、安全、可复现性,以及留住稀缺研究员。需要私下基准测试胜出、签约设计合作伙伴,以及可信的下一轮需求。
基准研究进展和招聘强劲,但商业证据有限,下一轮融资前也没有广泛外部产品。示意估值 $0.8B-$2B;入场上行有限,稀释或优先权之后可能持平。产品化延迟、收入模型不清,还需要更多资本。与当下公开证据最一致。
悲观基准测试不及预期,外部用户未转化,烧钱上升,或 AI 私募市场定价修正。示意估值 $0.2B-$0.7B;主要风险变成下轮降价、结构化过桥或人才收购。资本过剩、优先权堆叠、关键人物流失和市场倍数压缩。若尽调材料薄弱,或 12-18 个月内没有可衡量里程碑,就会触发。

情景估值是基于阶段、稀缺可比对象和公开市场锚点给出的示意 $B 区间,不是管理层指引。

[CV012, CV013, CV014, CV016, CV018, CV021]
FV003: 估值 / 回报区间

下一轮验证周期后,乐观、基准和悲观三种私募市场估值区间($B)。

区间以十亿美元计,刻意给得很宽,因为公司仍处于产品前阶段且未上市。

[CV012, CV013, CV014, CV016, CV018, CV021]

8.4 风险评级与论点破裂触发器

风险评级为高,因为主要承销变量都是非公开的。市场对更好科学研发的需求庞大且可信,但 Gartner 对 agentic AI 的警告、Crunchbase 的风险泡沫框架、CNBC 对估值恐惧的报道,以及 World Economic Forum 的泡沫清算分析,都不支持把当前 AI 资本市场视为稳定出清价。Mirendil 的特定风险被其吸引力来源放大:稀缺人才、昂贵算力、漫长技术闭环和宽平台野心。公司仍可能极其出色,但如果非公开尽调包不能显示基准优势、外部客户拉力、有纪律的烧钱、可防御权利,以及在没有过度清算优先权悬置的情况下,以显著更高估值进入下一轮的可信路径,投资论点就应迅速破裂。[CV022, CV023, CV024, CV025, CV026, CV027]

破坏投资论点和终止触发器表
触发器阈值对投资论点的传导行动含义
技术验证失败没有私下基准测试或 demo 能证明自主 AI 研发表现超过强通用模型。削弱「闭环才是产品」这一核心主张。不要按当前价格投资;等可复现验证后再看。
客户拉动失败12-18 个月内没有签约设计合作伙伴、付费试点或可信外部用户管线。投资论点从平台公司变成内部研究实验室。只保留观察权敞口,或放弃。
现金 runway / 烧钱错配种子轮计划在形成实质验证前就需要再融一大轮,或依赖未承诺算力。制造下轮降价和优先权压顶风险。要求按里程碑分批拨款、更低入场价或保护性结构。
治理不透明董事会权利、投票控制、清算优先权或创始人归属无法通过尽调核实。在异常昂贵的产品前估值上抬高代理风险。在法律和股权结构包审完前阻断。
市场修正可比 AI 种子轮或前沿实验室估值明显压缩,或新一轮要求结构化条款。降低高溢价后续融资的概率。除非入场价实质重置,否则把立场从跟踪改为回避。
人才流失产品验证前,两位创始人任一离开,或多名核心研究员流失。损害正在承销的主要资产。除非有极强接班证据,否则视为投资论点破裂。
安全 / 完整性失败研究输出对科学用户而言不够可复现、可审计或安全。堵住客户信任和战略合作路径。暂停投资,直到控制和验证经过独立核实。

这些触发器被设计成可监控的尽调关口,因为公开证据尚不足以支撑基于基本面的估值。

[CV022, CV023, CV024, CV025, CV032, CV035]
FV004: 投资 KPI

IC 打分在市场规模和团队质量,与验证、经济性、风险、估值、证据质量之间取平衡。

分数是 1-10 的尽调启发式评分,仅基于公开证据和当前私募市场可比标的。

[CV001, CV002, CV022, CV026, CV027, CV028]

8.5 退出准备度、流动性与悬置压力

现在的退出准备度低。Mirendil 没有收入准备度,不具备申报准备度,也还没有充分战略性去风险,无法形成近期 M&A 或 IPO 观点。最干净的正向路径是长期前沿实验室结果:用种子轮资金搭建一个能创造可展示研究加速的系统,再从内部人或战略算力与生命科学伙伴处以溢价融资。负向路径同样清楚:高种子轮估值、稀缺技术披露和宽使命,可能让公司卡在高前轮价格与不足以说服跨阶段或公开市场投资人的证据之间。公开可比公司让纪律可见。Recursion 和 Schrödinger 有真实收入和申报文件,却仍表现出重 R&D 支出和亏损,因此不能假设后期公开市场会在没有证明时自动奖赏 AI-for-science 叙事。[CV017, CV018, CV019, CV020, CV021, CV024]

8.6 最终尽调要求与决策门槛

最终尽调闸门应写清楚。按当前报道价格,投资人在批准任何配置前,应要求完整技术基准测试包、客户或设计伙伴证据、算力承诺、烧钱与跑道模型、股权结构表、董事会权利、清算优先权、创始人归属安排、安全姿态和里程碑计划。团队质量和类别上行的不对称性,可以支撑较小的观察支票或仅限内部人的权利;但全价一级市场入场需要证据证明 Mirendil 已从叙事进入持久能力。改变观点的关键事件,是非公开证据显示外部科学家或 AI 构建者反复使用 Mirendil 系统,比现有工作流更快产出有价值研究结果,并且留存可衡量、商业模式可信。[CV036, CV040, CV041, CV042, CV043, CV044]

最终尽调请求表
主题缺失证据重要性负责人 / 尽调路径
技术基准可复现评测套件、基线对比、失败模式和 demo 日志。估值押注的是技术不连续性,而不只是强团队。由 CTO 牵头技术尽调,并引入外部 AI 科学评审。
客户 / 设计合作伙伴签署 LOI、试点条款、使用日志、续约意向和用户画像。外部拉动把平台价值和内部实验室工具区分开。CEO 和 GTM 负责人;在 NDA 下做客户访谈。
商业模式定价、目标买方、部署模式、服务负载和毛利率路径。$1B 入场价需要证明科学用户能变成有经济价值的客户。财务和产品负责人;审阅模型情景和合同模板。
算力与烧钱已承诺算力、云 / NVIDIA 条款、招聘计划、月烧钱和 runway。在形成验证前,算力和人才市场可能吃掉大笔种子轮资金。CFO / 运营尽调;检查供应商协议和董事会预算。
股权结构与条款所有权、期权池、SAFE、债务、清算优先权、按比例跟投权和 side letter。下行和后续融资经济性同样取决于条款,而不只是账面估值。法律顾问和领投方;审阅章程和融资文件。
治理董事会构成、保护性条款、信息权和创始人归属。高价产品前轮次需要比公开材料显示的更强监督。公司法律顾问和董事会观察员;索取治理包。
安全 / 研究完整性数据控制、实验来源、模型评测政策和滥用缓释。科学自动化必须可审计且安全,才赢得信任。安全 / ML 治理审查,加独立红队范围。
退出和后续融资路径下一轮里程碑、潜在内部支持、战略伙伴兴趣和 IPO / M&A 类比。只有公司能跨过明显更高的融资或战略价值门槛,入场才成立。IC 负责人;梳理投资人储备和战略伙伴尽调。

这些请求优先寻找能改变建议的证据;每一项都关系到「昂贵」能否变成「可忍受」或「回避」。

[CV035, CV036, CV040, CV041, CV042, CV043]

8.7 图表

免责声明

本报告只是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术和合同事实仍未公开;任何投资决策前,都应直接向管理层和原始文件核验。

证据索引

结论
编号陈述可信度来源
CO001 Mirendil publicly describes itself as a frontier lab building systems that excel at AI R&D. SO001, SO002, SO003
CO002 Mirendil says its goal is to democratize frontier AI R&D so scientists in fields such as biology, chemistry, drug discovery, and robotics can use advanced AI without first becoming frontier AI labs. SO001, SO002, SO003
CO003 Mirendil’s public materials say the founding team consists of 20 researchers and engineers from Anthropic, xAI, Google DeepMind, and OpenAI. SO001, SO002, SO003
CO004 Mirendil announced a $200 million seed round in June 2026. SO001, SO002, SO004, SO005
CO005 Andreessen Horowitz and Kleiner Perkins were publicly identified as the lead or co-lead investors in Mirendil’s seed round. SO001, SO002, SO003, SO004
CO006 NVIDIA was publicly identified as a participant in Mirendil’s June 2026 financing. SO001, SO002, SO004, SO006
CO007 Public coverage places Mirendil’s June 2026 valuation at roughly $1 billion. SO004, SO005, SO006, SO007
CO008 Multiple launch reports describe Mirendil as a San Francisco-based startup. SO006, SO018, SO020, SO021
CO009 Mirendil came together in early 2026 after the founders left Anthropic in late 2025. SO011, SO018, SO022
CO010 Behnam Neyshabur publicly identifies himself as Mirendil’s co-founder and CEO. SO010, SO011, SO001
CO011 Harsh Mehta is publicly described as Mirendil’s co-founder and CTO. SO006, SO009, SO017
CO012 Behnam Neyshabur says he co-led Anthropic’s Discovery team with the goal of building an AI Scientist or Engineer. SO010, SO011
CO013 Kleiner Perkins says Behnam co-led Google’s Blueshift effort, contributed to Minerva, and later helped drive Gemini math and code reasoning. SO003, SO010, SO014, SO015
CO014 Behnam Neyshabur is listed on the SAM optimizer paper, corroborating frontier optimization pedigree relevant to Mirendil. SO010, SO012, SO013
CO015 Andreessen Horowitz says Harsh Mehta built the first version of Anthropic’s autoresearch platform and initially scaled it as a team of one. SO002, SO003
CO016 Public profile pages describe Harsh Mehta as having prior Google DeepMind and optimization-research experience before Mirendil. SO016, SO017, SO020
CO017 Investor writeups identify Shayan Salehian as a founding engineer from xAI and earlier X or Twitter ML work. SO002, SO003, SO021
CO018 Investor writeups identify Tara Rezaei as an MIT graduate and former OpenAI researcher on Mirendil’s founding bench. SO002, SO003, SO021
CO019 Mirendil’s product thesis is a lab-grade system that loops over research and engineering problems more autonomously rather than a narrow point solution. SO001, SO002, SO003
CO020 Mirendil’s backers say the first users are likely engineers and AI researchers before the platform expands to less technical scientists and domain experts. SO002, SO003
CO021 As of the run date, Mirendil has not publicly disclosed revenue, customer counts, or other commercial traction metrics. SO001, SO026, SO027
CO022 As of the run date, Mirendil has not publicly released technical benchmarks, a product roadmap, or detailed model specifications. SO001, SO026, SO027
CO023 Several outlets describe Mirendil’s financing as one of the largest AI seed rounds yet disclosed. SO005, SO006, SO008
CO024 March 2026 coverage reported Mirendil seeking roughly $175 million at a $1 billion valuation before later June reports announced a $200 million close. SO018, SO022, SO004
CO025 Public reporting is not fully consistent on investor roles because at least one draft-style article misstated NVIDIA as a lead instead of a participant. SO002, SO005, SO027
CO026 No public board composition, governance rights, or independent-director disclosures surfaced across Mirendil’s website, investor posts, or core launch coverage. SO001, SO002, SO003, SO004
CO027 Mirendil’s public website offers mission copy and a contact path but not a robust legal, trust, or disclosure surface. SO001
CO028 Both Mirendil’s homepage and Kleiner Perkins frame the company as a frontier lab and as the first lab from the future. SO001, SO003
CO029 Behnam Neyshabur’s CV lists Mirendil as his role from 2026-present and Anthropic as his role for 2024-2025. SO011, SO010
CO030 Behnam Neyshabur’s scholarly record includes authorship on Minerva and Gemini-related research. SO012, SO014, SO015
CO031 Harsh Mehta’s scholarly and profile pages show a pre-Mirendil publication record in optimization and machine learning. SO016, SO017
CO032 Third-party company-profile sites classify Mirendil as a seed-stage frontier AI lab rather than a revenue-disclosed software company. SO009, SO025
CO033 On March 18, 2026, press coverage surfaced Mirendil’s fundraising discussions before the round was announced as closed. SO018, SO022
CO034 On June 24, 2026, Andreessen Horowitz publicly announced that it was leading Mirendil’s seed round. SO002
CO035 On June 25 and 26, 2026, Mirendil emerged publicly through its website and a wave of financing coverage. SO001, SO004, SO005, SO006
CO036 Launch coverage consistently ties Mirendil’s use cases to scientific domains such as biology, chemistry, materials science, drug discovery, and robotics. SO001, SO004, SO006, SO020
CO037 Investor theses describe Mirendil as infrastructure for a broader AI ecosystem rather than a single narrow internal lab. SO002, SO003
CO038 Mirendil’s homepage invites candidates to join the company but does not disclose a careers portal, job list, or office network. SO001
CO039 The only clearly disclosed operating-scale metric is a 20-person founding team, while broader headcount remains undisclosed. SO001, SO003, SO009
CO040 San Francisco appears in third-party profiles, but Mirendil’s official homepage does not state a headquarters address. SO001, SO009, SO017
CO041 If the reported $200 million was raised at a roughly $1 billion post-money valuation, incoming investors would own about 20% of the company on a simplified post-money basis. SO007, SO026
CO042 Some public profiles describe Shayan Salehian and Tara Rezaei as co-founding operators, while the official site discloses only the 20-person founding team and not formal titles. SO003, SO021, SO026
CO043 Mirendil’s core thesis is a recursive improvement loop in which better models do better research and better research produces better models. SO001, SO003
CO044 Kleiner Perkins says Behnam Neyshabur and Harsh Mehta first met at Google about seven years before Mirendil’s launch and had been building together since. SO003
CO045 Andreessen Horowitz says the training data for Mirendil’s platform must cover the full AI-research loop, including experiment design, coding, debugging, compute management, and checkpoint comparison. SO002
CO046 Mirendil’s company and investor materials position the first external beneficiaries as AI builders and scientific experts rather than mass-market end users. SO002, SO003, SO019
CO047 No public evidence surfaced of debt facilities, secondary share sales, or other non-equity financing around Mirendil’s June 2026 round. SO004, SO005, SO006
CO048 Mirendil’s talent and financing are well substantiated, but commercialization, governance, and technical disclosure remain thin enough to keep the diligence burden high after chapter 1. SO021, SO026, SO027
CM001 The addressable category closest to Mirendil's stated mission is fragmented across at least three distinct third-party market definitions -- AI-in-drug-discovery software, agentic AI software, and autonomous/self-driving laboratory hardware -- each sized differently by analyst firms. SM001, SM003, SM025
CM002 Grand View Research defines "AI in drug discovery" as software and services covering molecular library screening, target identification, drug optimization/repurposing, de novo drug design, and preclinical testing, sold mainly to pharmaceutical/biotech companies and CROs. SM001
CM003 Mordor Intelligence defines the "agentic AI market" broadly across customer service, IT/technical support, manufacturing, financial services, and other enterprise functions, not as a science- or research-specific category. SM003
CM004 Dimension Market Research sizes a distinct "autonomous chemical laboratory" market covering lab hardware, software, and services for closed-loop, robot-run chemical synthesis and testing, separate from both the drug-discovery-software and agentic-AI-software categories. SM025
CM005 Global pharmaceutical industry R&D spending is a status-quo substitute budget pool -- the industry invested roughly $288 billion in R&D in 2024 across in-house and outsourced discovery, preclinical, and clinical development. SM014
CM006 European pharmaceutical companies alone reported roughly €55 billion in R&D investment for 2024 through their industry association EFPIA, illustrating that even a single region's status-quo R&D budget dwarfs any disclosed AI-for-science software market. SM015
CM007 US federal government science funding is itself under acute pressure -- the FY2026 budget request cuts NSF's total discretionary funding to $3.9 billion from $9.06 billion enacted in FY2025, a 56.9% reduction. SM020, SM009
CM008 Congress had previously authorized NSF funding as high as $17.8 billion for FY2026, 357% above the FY2026 request, showing the executive-branch request and legislative authorization for the same adjacent public R&D pool diverge sharply. SM020
CM009 Grand View Research estimates the global AI-in-drug-discovery market at $2.9 billion in 2026, growing to $13.8 billion by 2033 at a 24.8% CAGR. SM001
CM010 Precedence Research estimates the same nominal AI-in-drug-discovery category at $7.62 billion in 2026, growing to $17.81 billion by 2035 at a 9.90% CAGR -- roughly 2.6x Grand View Research's 2026 figure for what is described as the same market. SM002
CM011 The divergence between Grand View Research and Precedence Research for the same nominal 2026 market size stems from differing scope and base-year assumptions rather than a simple forecasting-horizon difference, since both firms cite 2026 as a base/forecast year. SM001, SM002
CM012 Mordor Intelligence sizes the global agentic AI market (all industries) at $9.89 billion in 2026, growing to $57.42 billion by 2031 at a 42.14% CAGR. SM003
CM013 Fortune Business Insights independently sizes the same global agentic AI market at $9.14 billion in 2026, growing to $139.19 billion by 2034 at a 40.50% CAGR -- closely corroborating Mordor Intelligence's 2026 figure despite different vendors and methodologies. SM024, SM003
CM014 Dimension Market Research sizes the global autonomous/self-driving chemical laboratory market at $5.75 billion in 2026, growing to $19.48 billion by 2035 at a 14.5% CAGR -- the smallest and slowest-growing of the disclosed adjacent-market lenses. SM025
CM015 Combining the 2026 AI-in-drug-discovery estimate ($7.62B, Precedence) and the 2026 autonomous chemical laboratory estimate ($5.75B, Dimension) yields an illustrative ~$13.4 billion combined 2026 "disclosed AI-for-science software and lab hardware" layer, though the two reports use unrelated methodologies and should not be treated as an audited total. SM002, SM025
CM016 No independent research firm publishes a market-size estimate specific to frontier-lab-grade, general-purpose AI research-automation platforms spanning multiple sciences at once -- the category Mirendil says it targets -- so its addressable market must be proxied from adjacent, narrower categories.
CM017 As a capital-committed proxy for the narrowest "AI scientist" tier, at least three venture-backed companies -- Mirendil ($200M seed), Periodic Labs ($300M seed, in talks for $500M more at a $7.5B valuation), and Lila Sciences ($550M total across seed and Series A, $1.3B+ valuation) -- had collectively raised more than $1 billion in disclosed private capital for cross-domain AI-driven scientific-discovery platforms by mid-2026. SM010, SM011, SM019, SM012, SM013
CM018 Periodic Labs, founded by former OpenAI and Google DeepMind researchers, raised a $300 million seed round in September 2025 at a $1.3 billion valuation to build AI systems that run automated physics and chemistry lab experiments. SM010, SM011
CM019 By May 2026, Periodic Labs was in advanced talks to raise at least $500 million more at a $7.5 billion valuation, led by AMP, nearly a sixfold increase in under eight months. SM011, SM019
CM020 Lila Sciences, a Flagship Pioneering venture, raised a $200 million seed round followed by a $350 million Series A (total $550 million) reaching a valuation above $1.3 billion, with Nvidia's venture arm among its backers. SM012, SM013
CM021 Excelra estimates that more than $20 billion in cumulative private capital has been invested in AI-in-drug-discovery companies over more than a decade, alongside partnership economics ranging from $50-100 million upfront payments to billion-dollar milestone structures. SM026
CM022 McKinsey's November 2025 global survey (1,993 respondents, 105 countries) found 23% of organizations were scaling at least one agentic AI use case, but no more than 10% were scaling agents within any single business function -- a much lower penetration rate than the dollar-denominated market forecasts above would suggest. SM027
CM023 Pharmaceutical and biotechnology R&D organizations are the largest identifiable buyer segment for AI-in-drug-discovery tools, with Chief R&D/Digital officers as budget owners and bench/computational-biology scientists as end users. SM001, SM016
CM024 Isomorphic Labs' January 2024 collaboration with Eli Lilly illustrates the milestone-based payer structure common to AI-drug-discovery partnerships -- $45 million upfront plus up to $1.7 billion in milestone payments and tiered royalties for a multi-target small-molecule discovery program. SM023
CM025 ZS's 2026 survey of 115 US-based pharma/biotech technology executives found only 17% report measurable value from AI investment in research and discovery specifically (vs. 29% in clinical development), even though 41% are planning to automate entire R&D discovery workflows with intelligent agents. SM016
CM026 Materials-science and chemicals R&D is a distinct buyer segment served by autonomous/self-driving laboratory vendors, with corporate R&D directors as budget owners and process/materials scientists as end users, adopting closed-loop synthesis-and-testing platforms rather than pure software agents. SM025
CM027 US federal science agencies (led by NSF) are a government buyer/payer segment whose adoption path runs through congressional appropriations rather than direct procurement, with program managers as budget owners and academic PIs as the effective "customers." SM020, SM009
CM028 Even amid overall NSF cuts, the FY2026 budget request singles out artificial intelligence, quantum information science, and technology-innovation partnerships as the agency's "critical activities" prioritized for continued or increased investment, while the Computer and Information Science and Engineering directorate underlying much foundational AI research would still be cut by roughly 65% versus FY2024. SM020
CM029 Academic research institutions represent a smaller-budget, grant-funded buyer segment whose adoption of AI-for-science tools depends on federal and philanthropic grant funding rather than commercial procurement cycles, making them more exposed to public-budget volatility than corporate buyers. SM020, SM021
CM030 Frontier AI labs themselves (including Mirendil's direct comparables Periodic Labs and Lila Sciences) are also an internal buyer segment, funding AI-for-science tooling from their own venture-backed compute and opex budgets to accelerate their own model-research loop rather than to resell externally in the near term. SM010, SM012
CM031 Lila Sciences' public messaging explicitly invites "prospective customers and startups" to build on its platform, signaling an intended shift from internal-only R&D tool to an externally sold platform -- the same commercialization path Mirendil's public mission statement implies but has not yet evidenced with a named customer. SM012
CM032 Global private investment in AI more than doubled in 2025 (127.5% year-over-year growth), with generative AI capturing nearly half of all private AI funding, according to Stanford's 2026 AI Index. SM006
CM033 US private AI investment is roughly 23 times larger than China's disclosed private AI investment, though Chinese state guidance funds reportedly deployed an estimated $184 billion into AI firms between 2000 and 2023, complicating direct country comparisons. SM006
CM034 Major cloud providers are reaching record infrastructure spending levels that support AI-for-science compute demand -- Google alone reported more than $150 billion in annual capital expenditure in 2025, per Stanford's AI Index. SM006
CM035 Epoch AI finds that training compute for frontier language models has grown roughly 5x per year since 2020, training costs are climbing about 3.5x annually, and power requirements for frontier training runs are doubling each year -- a direct capital-intensity constraint on any lab claiming to build frontier-grade AI research systems. SM018
CM036 Pharma R&D leaders report tangible efficiency drivers from AI adoption -- ZS's 2026 outlook cites AI-native biotechs shortening drug discovery and development timelines by 40-50% versus traditional approaches, and generative-AI platforms cutting documentation time by more than 90% in some workflows. SM016
CM037 Cost pressure is also pushing pharma toward AI-driven R&D efficiency -- ZS estimates new US pharmaceutical tariffs could add $13-19 billion in industry costs, and the largest pharma companies need to cut roughly $32 billion in expenses by 2030, creating incentive to substitute AI tooling for slower/costlier traditional R&D processes. SM016
CM038 Excelra reports that hybrid AI-drug-discovery business models (combining software licensing with partnership economics) are outperforming pure AI-pipeline biotechs commercially, citing Insilico Medicine's $85.8 million revenue (+68% year-over-year) against 80-90% valuation declines for pure pipeline-only peers. SM026
CM039 Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls, and estimates only about 130 of thousands of vendors marketed as "agentic AI" actually deliver genuine agentic capability ("agent washing"). SM017, SM028
CM040 McKinsey's November 2025 survey found 62% of organizations are at least experimenting with AI agents, but only 23% report scaling any agentic use case and no more than 10% are scaling within a given business function, showing a large gap between experimentation and durable paid deployment. SM027
CM041 Trust and integrity concerns are rising around AI-generated scientific output -- Nature reports that early 2026 studies of AI's footprint in journals, preprints, and peer review show a "rapidly evolving" but still poorly measured problem, and independent analysis cites a Columbia University audit finding roughly 1-in-277 PubMed-indexed papers in 2026 contained AI-hallucinated or fabricated citations. SM007, SM022
CM042 In response to integrity concerns, arXiv adopted a 2026 policy prohibiting unverified AI-generated content and mandating human oversight of submissions, according to independent commentary on the scientific-publishing integrity crisis. SM022
CM043 US federal science-funding volatility is a structural adoption constraint for any AI-for-science vendor whose buyer base includes academic or government labs -- the FY2026 NSF request would cut the share of funded grant proposals from about 26% to 7% and reduce Graduate Research Fellowship Program scale to about 55% of its current size if enacted as proposed. SM020, SM021
CM044 The Computing Research Association publicly opposed the FY2026 NSF budget request, warning it would "turn back the clock more than 20 years" on US research funding and shrink the pipeline of researchers available to adopt or build AI-for-science tools. SM021
CM045 A US congressional reporter (Chemical & Engineering News/ACS) documented bipartisan political friction over the FY2026 NSF request, quoting Representative Zoe Lofgren's public statement that continued cuts threaten US scientific leadership, indicating the final appropriated NSF budget remains genuinely contested rather than settled. SM008
CM046 Capital intensity is a further constraint specific to Mirendil's stated ambition -- Epoch AI's data showing 5x/year compute growth and 3.5x/year training-cost growth implies that matching frontier labs' pace of model improvement requires continuously rising capital commitments well beyond a single $200 million seed round. SM018
CM047 No public, audited bottoms-up estimate exists for the specific dollar revenue Mirendil-style cross-domain AI research-automation platforms could realistically capture from the ~$288 billion global pharma R&D budget plus adjacent materials/chemicals R&D spend, leaving SAM/SOM effectively unresolved pending private diligence.
CM048 It remains unresolved from public sources whether Mirendil's own roadmap includes physical, self-driving laboratory integration (the ~$5.75 billion 2026 autonomous chemical-lab category) or stays purely software/agent-based (the $9-10 billion 2026 agentic-AI category), a distinction that materially changes which sizing lens applies.
CM049 A January 2025 Gartner survey of more than 3,400 professionals found only 19% of organizations reported significant investment in agentic AI, 42% investing conservatively, and 31% undecided, underscoring how early-stage enterprise commitment still is relative to the multi-billion-dollar market forecasts. SM028
CP001 Mirendil positions itself as a frontier lab building systems and frontier models that excel at AI R&D. SP001, SP002, SP003
CP002 Mirendil’s public materials say the goal is to democratize frontier AI R&D for scientists in drug discovery, chemistry, biology, robotics, and related fields. SP001, SP002, SP003
CP003 Periodic Labs is a direct peer because public coverage says it is building AI scientists and autonomous laboratories for scientific discovery. SP004, SP005
CP004 Periodic Labs publicly raised a $300 million seed round backed by prominent technology investors including a16z, DST, Nvidia, Accel, Jeff Bezos, Eric Schmidt, and Jeff Dean. SP004, SP006
CP005 Later 2026 reports discussed Periodic Labs raising additional capital at roughly a $7 billion to $7.5 billion valuation. SP005, SP006
CP006 Lila Sciences claims to combine advanced AI and autonomous labs that generate hypotheses, design and run experiments, and learn from new data in real time. SP007, SP009
CP007 Lila Sciences announced a $350 million Series A close that brought total capital raised to $550 million and said it was welcoming a first cohort of customers. SP008, SP009
CP008 FutureHouse describes itself as a non-profit building AI agents to automate research in biology and other complex sciences. SP010, SP004
CP009 Sakana AI’s AI Scientist automates idea generation, experiments, manuscript writing, and automated review for machine-learning research. SP011, SP004
CP010 Google DeepMind’s AlphaFold has become an incumbent scientific platform with over 200 million protein structures and millions of researchers using the system or database. SP014, SP013
CP011 AlphaFold 3 and AlphaFold Server provide non-commercial scientists with structure and interaction prediction capabilities. SP014, SP013
CP012 Isomorphic Labs says it is building predictive and generative AI models to transform drug discovery and design novel molecules. SP012, SP013
CP013 Isomorphic Labs disclosed a Lilly collaboration with $45 million upfront and potential total deal value up to $1.7 billion excluding royalties. SP013, SP012
CP014 Recursion is a public clinical-stage TechBio company advancing an AI-native drug discovery and development platform and a pipeline across therapeutic areas. SP015, SP016, SP018
CP015 Recursion says its platform includes more than 50 petabytes of proprietary biological and chemical data and an automated wet lab that captures millions of cell experiments per week. SP015, SP016
CP016 Recursion’s public pipeline includes multiple clinical and candidate-stage programs across oncology, rare disease, and related indications. SP017, SP016
CP017 Insilico Medicine advertises generative AI and automation for target discovery, disease modeling, molecule generation, and a pipeline with Phase II and Phase I programs. SP019, SP029
CP018 Chai Discovery positions Chai-2 around drug-like antibody design against challenging targets with atomic precision. SP020, SP029
CP019 Cradle emphasizes secure protein-engineering collaboration, customer ownership of data and IP, no royalties, and a software subscription fee. SP021, SP029
CP020 Schrödinger offers a physics-based computational platform for therapeutics and materials discovery plus its own collaborative and proprietary pipeline work. SP023, SP029
CP021 Anthropic’s research page shows frontier AI work spanning safety, agents, biology, chemistry, coding, and societal impacts, making Claude a substitute for pieces of scientific workflows. SP024, SP025
CP022 OpenAI describes its o-series as advanced reasoning systems for complex STEM problems, which supports general LLMs as partial substitutes for scientific reasoning workflows. SP026, SP025
CP023 NVIDIA publishes healthcare and life-sciences AI tools, open models, SDKs, and BioNeMo resources that enable internal build paths for scientific AI. SP022, SP015
CP024 Internal build remains a credible substitute because buyers can combine infrastructure, general LLMs, vertical tools, CROs, and proprietary data without adopting a single new platform. SP022, SP023, SP024, SP025, SP029
CP025 CRO-supported and manual research workflows remain status-quo alternatives because scientific organizations already execute experiments through internal teams, vendors, and established software stacks. SP023, SP029
CP026 Anthropic reports Claude is starting to assist chemists with translation, recall, and integration work while still leaving expert judgment necessary. SP025, SP024
CP027 Mirendil has no public list pricing, SKU, customer list, security package, or benchmark suite in the reviewed public materials. SP001, SP002, SP003
CP028 Isomorphic, Recursion, Lila, AlphaFold, and Cradle disclose more concrete commercial or usage surfaces than Mirendil does today. SP008, SP013, SP014, SP015, SP021
CP029 Public evidence supports strong capability claims for competitors in physical experiment closure, protein structure prediction, drug-discovery pipelines, and protein engineering, but not for every matrix cell. SP007, SP014, SP015, SP017, SP019, SP021
CP030 Distribution power favors incumbents and mature vertical platforms because they already have public tools, pharma partnerships, investor surfaces, or specialized workflows. SP013, SP014, SP016, SP017, SP019, SP023
CP031 Most direct frontier AI-for-science labs in the reviewed set do not disclose standard list pricing or a public contract unit. SP001, SP004, SP007, SP008, SP010, SP011
CP032 Isomorphic’s Lilly collaboration provides partnership economics but not a reusable price list for an off-the-shelf platform. SP013, SP012
CP033 Cradle’s page is unusually explicit among reviewed alternatives because it says customers retain IP and pay a software subscription fee without royalties. SP021, SP020
CP034 AlphaFold Server’s disclosed free non-commercial access creates buyer expectations that some baseline scientific AI capabilities should be inexpensive or freely available. SP014, SP013
CP035 Mirendil’s strongest possible moat is the full AI-research loop covering experiment proposal, coding, debugging, compute management, and checkpoint comparison. SP002, SP003
CP036 Gartner warns that more than 40% of agentic AI projects may be canceled by the end of 2027 because of cost, unclear business value, or inadequate risk controls. SP028, SP029
CP037 Sakana AI’s own AI Scientist writeup says current systems can implement ideas incorrectly, make unfair comparisons, and produce misleading results. SP011, SP028
CP038 Sakana AI states that competition among LLMs has led to commoditization and that open models offer lower cost, availability, transparency, and flexibility. SP011, SP026
CP039 Excelra’s 2026 report argues that data moats matter more than algorithms as foundation models commoditize. SP029, SP028
CP040 Excelra identifies Big Tech players and major pharmas scaling internal AI as direct competitive pressure in AI drug discovery. SP029, SP022
CP041 Public sources reviewed for Mirendil do not show proprietary experimental datasets, customer workflow data, or exclusive partner data rights comparable to Recursion’s disclosed data assets. SP001, SP015, SP016
CP042 Trust and validation are competitively important because scientific AI outputs can affect drug discovery, chemical interpretation, and research integrity. SP025, SP028, SP029
CP043 Agentic AI buyers are likely to require clear ROI and workflow redesign rather than adopting broad autonomous systems solely because they are branded agentic. SP028, SP029
CP044 Scientific buyers can multi-home across general LLMs, incumbent scientific tools, vertical platforms, and CRO or internal lab execution while Mirendil matures. SP014, SP019, SP021, SP023, SP024, SP025
CP045 Likely entrants include frontier AI labs, open-model providers, cloud and GPU infrastructure vendors, pharma internal AI groups, and vertical discovery platforms. SP022, SP024, SP026, SP029
CP046 Mirendil’s public differentiation is currently more durable as a talent-and-thesis story than as a proven customer lock-in or data-moat story. SP001, SP002, SP003, SP029
CP047 Capital alone is not a durable moat because Periodic and Lila disclose or report larger financing scale than Mirendil and Isomorphic benefits from Alphabet backing. SP004, SP006, SP008, SP013
CP048 The decisive diligence test is whether Mirendil can outperform a composed stack of general LLMs, vertical scientific tools, internal data, and human-in-the-loop laboratories on real buyer tasks. SP002, SP014, SP021, SP024, SP025, SP028
CI001 Mirendil publicly describes its mission as democratizing frontier AI R&D for scientific and technical users rather than selling a named financial product today. SI001, SI002, SI003
CI002 Mirendil’s public website and launch materials do not disclose a product SKU, paid customer, revenue number, or pricing page. SI001, SI005
CI003 Investor materials position Mirendil’s first users as AI researchers and builders before broader scientists and domain experts. SI002, SI003
CI004 Mirendil announced or was reported to have raised a $200 million seed financing in June 2026. SI001, SI002, SI004, SI005
CI005 Public launch coverage placed Mirendil’s June 2026 financing context at roughly a $1 billion valuation. SI004, SI005
CI006 Andreessen Horowitz and Kleiner Perkins were identified as lead investors, and NVIDIA was identified as a participant in Mirendil’s financing. SI001, SI002, SI003, SI004
CI007 No public launch source reviewed disclosed debt, project-finance obligations, or equipment-financing obligations for Mirendil. SI001, SI004, SI005
CI008 Mirendil’s public materials disclose a 20-person founding team, which is the only visible operating-scale marker relevant to payroll burn. SI001, SI002, SI003
CI009 Adverse launch commentary characterized Mirendil as having no product and no revenue at the time of the financing. SI005
CI010 Recursion’s 2025 Form 10-K says it had no products approved for commercial sale and had not generated revenue from product sales. SI006, SI007
CI011 Recursion reported $753.9 million of cash, cash equivalents, and restricted cash as of December 31, 2025 and said it had at least 12 months of funding under its plan. SI006, SI007
CI012 Recursion reported 2025 total revenue of $74.681 million and 2025 research and development expense of $475.271 million. SI006, SI007
CI013 Recursion disclosed that two customers represented substantially all of its 2025 operating revenue, highlighting customer-concentration risk in comparable collaboration revenue. SI007
CI014 Schrödinger reported 2025 total revenue of $255.9 million and a 2025 net loss of $103.3 million. SI008, SI009
CI015 Schrödinger reported 2025 software revenue of $199.5 million, software ACV of $198.5 million, drug-discovery revenue of $56.4 million, and software gross margin of 74%. SI008, SI009
CI016 Schrödinger says its collaboration agreements typically include upfront consideration, discovery, development, commercial and regulatory milestones, and royalties. SI008
CI017 Schrödinger’s Novartis collaboration included eligibility for up to $2.272 billion in total milestones across initial programs, with no milestone revenue recognized as of December 31, 2025. SI008
CI018 Schrödinger disclosed 27 commercial customers with ACV of at least $1.0 million and average ACV of $3.9 million for that group in 2025. SI008, SI009
CI019 Isomorphic Labs announced an Eli Lilly collaboration with a $45 million upfront payment and potential total deal value up to $1.7 billion, excluding the upfront payment and royalties. SI017
CI020 Lila Sciences announced a $350 million Series A close and $550 million of total capital raised for its AI Science Factories strategy. SI015, SI016
CI021 TechCrunch reported that Periodic Labs raised a $300 million seed round to automate scientific discovery. SI014
CI022 Crunchbase reported that Q1 2026 global venture investment reached $300 billion and that $242 billion, or 80% of the total, went to AI companies. SI013
CI023 Carta reported that more than 60% of venture capital raised by companies on Carta in Q1 2026 went to AI companies and that foundational model startups at Series A had a $300 million median valuation versus $55 million for non-AI startups. SI010, SI025
CI024 Carta reported a $24 million median seed post-money valuation in Q4 2025, a $78.7 million median Series A post-money valuation, and median seed and Series A dilution between 19% and 20%. SI011, SI025
CI025 Carta reported that U.S. pre-seed startups on Carta raised more than $2.3 billion in Q1 2026 and that AI reached about 50% of pre-seed dollars. SI012, SI025
CI026 Excelra’s 2026 AI/ML drug-discovery report states that the category has seen more than $20 billion of cumulative investment. SI019
CI027 Dimension Market Research estimates the autonomous chemical laboratory market at $5.7485 billion in 2026 and $19.4834 billion by 2035. SI018
CI028 Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. SI020
CI029 Sequoia’s AI infrastructure analysis frames the core adverse question as whether enough revenue exists to justify the scale of AI GPU and data-center investment. SI026
CI030 Epoch AI estimates global AI computing power at the equivalent of around 20 million Nvidia H100s and says AI capital expenditure is approaching $1 trillion per year. SI021
CI031 The arXiv frontier-training-cost paper estimates that the amortized cost to train the most compute-intensive models has grown 2.4x per year since 2016 and could exceed $1 billion for the largest runs by 2027. SI022
CI032 Lambda’s public cloud pricing page listed NVIDIA H100 SXM instances at $3.99 per hour and NVIDIA B200 SXM6 instances at $6.69 per hour when fetched. SI023
CI033 NVIDIA’s H100 product page positions the H100 as a data-center GPU for high-performance, scalable AI workloads. SI024
CI034 No public source reviewed disclosed Mirendil ARR, bookings, GMV, active customers, utilization, or recognized revenue. SI001, SI004, SI005
CI035 Mirendil’s likely pre-revenue cost drivers are elite AI talent, frontier compute, data/evaluation infrastructure, and scientific workflow development. SI001, SI002, SI003, SI021, SI022, SI023
CI036 Using the public $200 million gross seed context, simple illustrative runway equals about 67 months at $3 million of monthly burn, 40 months at $5 million, and 20 months at $10 million before fees or working-capital effects. SI002, SI004, SI005
CI037 Mirendil has not publicly disclosed monthly burn, unrestricted cash, prepaid compute, capex, or a board-approved operating budget. SI001, SI004, SI005
CI038 Mirendil has not publicly disclosed list pricing, realized pricing, discounting, contract minimums, or revenue-recognition policy. SI001, SI002, SI003, SI005
CI039 The plausible Mirendil monetization set includes enterprise platform access, usage-based AI R&D workflows, research collaborations, managed scientific services, and internally generated IP, but none is company-confirmed as live revenue. SI001, SI008, SI009, SI017
CI040 Public AI-for-science comparables show that revenue can be collaboration- and milestone-heavy rather than pure SaaS, and therefore revenue quality can vary materially by contract structure. SI007, SI008, SI009, SI017
CI041 Mirendil’s gross margin cannot be underwritten from public evidence because pricing, compute allocation, support labor, and revenue mix are all undisclosed. SI001, SI008, SI023
CI042 Mirendil has high financing-dependency risk because public evidence shows substantial seed capital but no offsetting revenue and no disclosed burn denominator. SI004, SI005, SI021, SI022, SI026
CI043 A reasonable next-round trigger for Mirendil is proof of product capability, design-partner conversion, and compute-efficient scaling before seed cash is depleted, but no formal trigger is public. SI001, SI002, SI003, SI013
CI044 Recursion illustrates that an AI-drug-discovery public comparable can have tens of millions of revenue while still consuming hundreds of millions in R&D expense and reporting large losses. SI007
CI045 Schrödinger illustrates that a more mature computational drug-discovery software business can report meaningful software revenue and gross margin while still being GAAP unprofitable. SI008, SI009
CI046 The 2026 private-market environment supports unusually large AI financings but also raises the evidence bar for capital efficiency because funding is highly concentrated in AI. SI010, SI011, SI013
CI047 If Mirendil commercializes through design partners before self-serve software, the relevant diligence metrics are SOW terms, milestone economics, delivery cost, and IP ownership rather than standard SMB SaaS metrics. SI001, SI008, SI017
CI048 Gross seed proceeds alone are insufficient to prove capital adequacy because fees, restrictions, prepayments, debt, vendor commitments, and the actual burn plan are private. SI004, SI005, SI007
CI049 The minimum financial diligence package should include post-close cash, monthly burn, compute contracts, hiring plan, pricing framework, design-partner proof, revenue-recognition policy, and any debt or side-letter obligations. SI001, SI007, SI008, SI020, SI026
CE001 Mirendil publicly describes itself as a frontier lab building systems that excel at AI R&D. SE001, SE002, SE003
CE002 Mirendil says its goal is to democratize frontier AI R&D for fields including drug discovery, chemistry, biology, robotics, and related science domains. SE001, SE002, SE003
CE003 Andreessen Horowitz says Mirendil’s platform must cover proposing experiments, writing and running code, interpreting results, debugging failures, improving kernels, managing compute, comparing checkpoints, and deciding next steps. SE002, SE003
CE004 As of the run date, Mirendil has not publicly released a product demo, API documentation, pricing page, customer deployment, or benchmark pack. SE001, SE002, SE003
CE005 Mirendil’s public disclosure supports a product thesis and build plan rather than a verified commercial product asset. SE001, SE002, SE003
CE006 No public Mirendil source reviewed disclosed model names, model cards, evaluation methodology, datasets, deployment targets, integration surfaces, or support commitments. SE001, SE002, SE003
CE007 Mirendil’s verified product maturity should be treated as concept or undisclosed until private demos, benchmarks, or design-partner evidence are reviewed. SE001, SE002, SE003, SE020
CE008 A diligence architecture for Mirendil should include a frontier AI-R&D model layer, an agentic research-loop harness, compute orchestration, evaluation/checkpoint comparison, and domain interfaces. SE001, SE002, SE004, SE010
CE009 The AI Scientist architecture includes idea generation, novelty/literature search, experimental iteration, paper write-up, and automated review as a repeating loop. SE004, SE005, SE006
CE010 MLAgentBench frames ML experimentation as agents reading and writing files, executing code, inspecting outputs, and iterating toward a research goal. SE023, SE024
CE011 RE-Bench environments give agents access to a computer, scoring functions, and resources for realistic ML research engineering tasks such as scaling-law fitting and GPU-kernel optimization. SE010, SE011, SE012
CE012 The domain interface for Mirendil remains unproven because AI-R&D automation has not publicly been shown to transfer into Mirendil-specific biology, chemistry, materials, or robotics workflows. SE001, SE004, SE025, SE027
CE013 SWE-bench evaluates language models on real-world GitHub issues where the model must generate a patch that resolves a described problem. SE013, SE014
CE014 SWE-bench Verified is a human-filtered subset of 500 instances, while the full benchmark reports the percentage of instances resolved. SE013, SE014
CE015 SciCode covers scientific coding problems across natural-science subfields including mathematics, physics, chemistry, biology, and materials science. SE025, SE026
CE016 FunSearch pairs an LLM that proposes code with an automated evaluator that scores candidate programs, creating an iterative discovery loop. SE017, SE018
CE017 The AI Scientist paper reports a fully automatic discovery system that generates research ideas, writes code, executes experiments, visualizes results, writes a paper, and runs simulated review. SE004, SE005
CE018 Sakana says the first AI Scientist version can produce a full paper at about $15 per idea, while also warning that generated papers can contain flaws. SE004, SE005
CE019 MLE-bench curates 75 ML engineering-related Kaggle competitions to test data preparation, model training, and experiment-running skills. SE007, SE008, SE009, SE029
CE020 The original MLE-bench report found the best-performing setup, OpenAI o1-preview with AIDE scaffolding, achieved at least Kaggle bronze medal level in 16.9% of competitions. SE007, SE008
CE021 RE-Bench consists of seven challenging open-ended ML research engineering environments and includes data from 71 eight-hour attempts by 61 human experts. SE010, SE011, SE012
CE022 RE-Bench reports that top AI agents score about four times higher than human experts at a two-hour budget, while humans overtake at eight hours and reach about twice the top-agent score at 32 hours. SE010, SE012
CE023 RE-Bench authors report that agents generate and test implementations more than ten times faster than humans but often struggle to react to novel information or build on progress over time. SE010, SE012
CE024 MLAgentBench reports 13 diverse ML experimentation tasks and identifies long-term planning and hallucination reduction as key challenges for language-agent research assistants. SE023, SE024
CE025 SciCode reports 338 subproblems from 80 research-level problems and states that Claude 3.5 Sonnet solved only 4.6% in the most realistic setting in the arXiv version. SE025, SE026
CE026 Benchmark leaderboards and repositories for MLE-bench, SWE-bench, RE-Bench, SciCode, and AI Scientist create developer-signal proxies for AI research automation progress even though Mirendil has no public developer surface. SE006, SE009, SE011, SE013, SE014, SE026, SE028
CE027 Mirendil has not disclosed a release date, phased roadmap, product documentation plan, or deployment timeline on its official and investor launch surfaces. SE001, SE002, SE003
CE028 Mirendil’s likely first users are engineers and AI researchers before less technical scientists, according to investor thesis language. SE002, SE003
CE029 Mirendil and its investors position less technical scientists and domain experts as later beneficiaries of the platform if the AI-R&D loop works. SE001, SE002, SE003
CE030 No public evidence reviewed shows Mirendil has a status page, support SLA, incident history, deployment architecture, or customer support process. SE001, SE002, SE003
CE031 A private benchmark pack, reproducible internal demo, design-partner logs, and safety case would be required to move Mirendil from concept maturity toward verified product maturity. SE004, SE010, SE020, SE022
CE032 Epoch AI reports that frontier AI development relies on powerful AI supercomputers whose leading performance grew about 2.5 times per year in its 2019-2025 dataset. SE015
CE033 Epoch AI reports that power requirements and hardware costs for leading AI supercomputers doubled every year in its dataset. SE015
CE034 NVIDIA Research publishes resources, code, demos, proprietary model licensing paths, and CUDA-oriented code libraries across generative AI, robotics, rendering, and related fields. SE016
CE035 NVIDIA’s public research page supports the relevance of the NVIDIA ecosystem to Mirendil, but it does not disclose Mirendil-specific compute supply, pricing, or reserved-capacity terms. SE001, SE016
CE036 Mirendil’s autonomous research-loop product would depend on safe execution environments because adjacent open-source systems warn that LLM-written code can execute risky packages, web access, or processes. SE006, SE024
CE037 Data rights for research traces, code, proprietary datasets, model outputs, and external user data are material product dependencies that Mirendil has not publicly addressed. SE001, SE002, SE020
CE038 If Mirendil cannot secure compute, research-trace data, evaluation integrity, and domain adapters, the system could remain an internal lab tool rather than a broad external product. SE002, SE015, SE020, SE025
CE039 The AI Scientist authors warn that the system can incorrectly implement ideas, make unfair comparisons, make critical numerical errors, and attempt unsafe self-modifications such as changing execution scripts. SE004, SE006
CE040 FunSearch explicitly uses an automated evaluator to guard against hallucinations and incorrect ideas while evolving LLM-generated programs. SE017, SE018
CE041 AlphaFold 3 includes confidence measures and addresses hallucination behavior in generative structure prediction, illustrating the kind of domain-specific validation expected for AI-for-science systems. SE019, SE027
CE042 Berkeley RDI reports benchmark exploits such as fake scores, answer leakage, weak tests, and evaluation-code manipulation across widely used AI benchmarks. SE020, SE021
CE043 Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. SE022
CE044 No Mirendil public source reviewed disclosed a trust center, model-safety report, benchmark-governance process, privacy posture, or compliance certification for autonomous research workflows. SE001, SE002, SE003, SE020, SE022
CE045 Mirendil’s main technical risk is proving that autonomous research artifacts are reproducible, safe, non-hallucinated, non-gamed, and useful to external scientists rather than merely plausible. SE004, SE018, SE020, SE021, SE022
CU001 Mirendil has not named any customer, pilot user, or design partner in any public source reviewed as of the run date. SU001, SU002, SU003
CU002 Mirendil's homepage states its ambition is to democratize frontier AI R&D so scientists in biology, chemistry, drug discovery, materials science, and robotics do not need to become frontier AI labs themselves. SU003
CU003 Andreessen Horowitz's investment note names the intended workflow (propose experiments, write and run code, interpret results, manage compute) without naming a customer, pilot, or design partner. SU004
CU004 Kleiner Perkins' investment thesis frames Mirendil's ambition as building the system that builds systems for AI R&D without disclosing any commercial engagement. SU005
CU005 Pharma and biotech R&D organizations are a plausible buyer segment for Mirendil because comparable AI-for-science vendors Isomorphic Labs and Recursion monetize primarily through pharma collaborations. SU007, SU008, SU009, SU011
CU006 Universities and academic research labs are a plausible buyer segment because comparable AI-for-science organizations such as Periodic Labs and FutureHouse run named academic grant or fellowship programs targeting university researchers. SU019, SU020, SU022
CU007 Industrial and materials-science labs, such as semiconductor manufacturers, are a plausible buyer segment because Periodic Labs, whose founders come from a similar frontier-lab talent pool to Mirendil's team, discloses an active semiconductor-manufacturer engagement. SU022, SU023
CU008 Internal AI-research teams at other frontier labs or large enterprises are a fourth plausible buyer segment because Mirendil's own stated mission is to give non-frontier-lab teams access to frontier AI-R&D capability. SU003, SU004
CU009 FutureHouse operates as a nonprofit funded by philanthropic and grant sources, including Eric and Wendy Schmidt, Open Philanthropy, the National Science Foundation, and the AI Safety Institute, rather than by selling to paying customers. SU020
CU010 FutureHouse's commercially oriented AI-for-science technology has been transferred to a separate for-profit entity, Edison Scientific, indicating that even a directly adjacent nonprofit lab eventually spins out a commercial entity to sell to paying customers. SU020
CU011 As of July 2026, Mirendil's public job board lists 14 open roles, all titled Member of Technical Staff across agent harness, AI-for-AI-systems, design engineering, inference, infrastructure, kernels, model evaluation, platform, post-training/RL, pretraining, product development, and security engineering. SU001, SU002
CU012 None of Mirendil's 14 public job listings are in sales, business development, partnerships, or customer-success functions. SU002
CU013 Mirendil's public careers page displays a client-side Loading open roles placeholder and routes all role detail to an externally hosted Ashby job board. SU001, SU002
CU014 Mirendil's founding team is described as approximately 20 researchers and engineers drawn from Anthropic, xAI, Google DeepMind, and OpenAI. SU003, SU004
CU015 Mirendil's public site contains no waitlist, early-access signup form, or beta-program mechanism for prospective users. SU003, SU001
CU016 Periodic Labs' public site operates a named Academic Grant Program inviting researchers to apply for funding opportunities, a GTM-adjacent mechanism absent from Mirendil's public site. SU022
CU017 Lila Sciences' public messaging states it is welcoming its first cohort of customers following a $350 million Series A that brought its total raised to $550 million, a more advanced GTM signal than anything Mirendil has disclosed. SU015
CU018 Isomorphic Labs' collaboration with Eli Lilly includes an upfront payment plus up to $1.7 billion in milestone payments and tiered royalties for a multi-target small-molecule discovery program. SU007
CU019 Isomorphic Labs' collaboration with Novartis, expanded in February 2025 to add up to three additional research programs, includes a $37.5 million upfront payment plus up to $1.2 billion in milestone payments and tiered royalties. SU008
CU020 Recursion's own partners page publishes named testimonials from Bayer AG's Joerg Moeller and Roche's James Sabry describing active collaborations in fibrotic-disease and oncology-adjacent biology respectively. SU009
CU021 Recursion has disclosed collaboration economics including up to $1.5 billion in potential payments from Bayer, $213 million received to date from Roche/Genentech, and $134 million in milestone payments logged to date from Sanofi. SU009, SU010
CU022 Recursion's 2025 Form 10-K discloses that two customers represented substantially all of its 2025 operating revenue, evidencing extreme customer concentration in the comparable collaboration-revenue model. SU010, SU011
CU023 Schrödinger disclosed 27 commercial customers with average annual contract value of $3.9 million among the group with ACV of at least $1.0 million, illustrating the depth of named, quantified customer evidence available from a public comparable. SU012, SU013
CU024 Periodic Labs' own homepage states it is training custom AI agents for a semiconductor manufacturer's engineers and researchers to address chip heat-dissipation issues and accelerate iteration on experimental data. SU022
CU025 Independent reporting corroborates that Periodic Labs' customer base also includes companies in the space and defense sectors, though none are named. SU023
CU026 Cradle states that customers retain ownership of their data and intellectual property and pay a software subscription fee without royalties, an explicit commercial model that Mirendil has not disclosed for itself. SU017
CU027 Lila Sciences is listed as a Participating Company in BIO 2026's official Partnering directory alongside major pharmaceutical companies, indicating active commercial business-development engagement even though no named pharma customer has been disclosed. SU016
CU028 No public source reviewed discloses a named customer, pilot, letter of intent, or design partner specifically for Mirendil, in contrast to Isomorphic Labs, Recursion, and Periodic Labs, each of which discloses at least one concrete named or described customer engagement. SU001, SU003, SU007, SU009, SU022
CU029 Universities and colleges commonly use the Higher Education Community Vendor Assessment Toolkit, maintained by EDUCAUSE with Internet2 and REN-ISAC, to run standardized security and privacy reviews of prospective AI vendors before procurement. SU029
CU030 MIT NANDA's State of AI in Business 2025 research, based on a review of over 300 disclosed AI initiatives and interviews with 52 organizations, found that about 95% of enterprise generative-AI pilots fail to scale into production while roughly 5% succeed. SU027, SU028
CU031 Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. SU025, SU026
CU032 ZS's 2026 survey of pharma/biotech technology executives found only 17% report measurable value from AI investment specifically in research and discovery. SU024
CU033 McKinsey's November 2025 global survey found 62% of organizations are at least experimenting with AI agents but only 23% are scaling any agentic use case. SU025
CU034 The status quo for scientific R&D buyers includes continuing to pair human scientists with CROs, existing modeling suites, data providers, and general-purpose LLMs rather than adopting an unproven autonomous AI-R&D lab platform. SU024, SU025
CU035 Pharma and biotech buyer procurement in this category tends to follow a milestone-based collaboration structure, per Isomorphic Labs and Recursion precedent, rather than a standard software sales cycle, meaning a first Mirendil deal is more likely to resemble a multi-quarter scientific partnership negotiation than a seat-license sale. SU007, SU008, SU009
CU036 Mirendil has not published a trust, security, or compliance page, which would be a prerequisite for clearing university HECVAT-style or pharma/biotech vendor-security reviews. SU001, SU003
CU037 No public source discloses net revenue retention, gross revenue retention, churn, renewal rate, or contract length for Mirendil. SU001, SU003
CU038 Mirendil does not appear on independent customer-review platforms such as G2, Capterra, or Gartner Peer Insights as of the run date. SU030
CU039 If Mirendil eventually signs an initial design partner, revenue concentration risk is likely to be severe at first, mirroring Recursion's disclosed pattern in which two customers represented substantially all 2025 revenue. SU011, SU022
CU040 Mirendil's public materials do not specify whether its eventual commercial model will be a software platform license, a milestone-based scientific collaboration, or internally retained IP monetized later, a distinction that changes which retention and concentration metrics are relevant. SU001, SU003
CU041 No public source reviewed for this chapter discloses any failed pilot, churned design partner, blocked deployment, or customer complaint involving Mirendil, but this reflects the absence of any disclosed commercial engagement rather than a clean track record. SU001, SU002, SU003
CU042 As of the run date, the most recent public evidence bearing on Mirendil's commercial or customer activity is its job board, accessed July 2026, which shows no change in hiring pattern toward customer-facing roles since the company's late-June 2026 launch coverage. SU002
CR001 Mirendil announced a $200 million seed round in June 2026 at roughly a $1 billion valuation, led by a16z and Kleiner Perkins with NVIDIA participating. SR001, SR002, SR003, SR004
CR002 Mirendil’s public materials and launch coverage describe a frontier AI-R&D lab, not a publicly shipped product with disclosed revenue, customer count, pricing, or benchmarks. SR001, SR002, SR003, SR005
CR003 The disclosed 20-person founding team is the only clear public operating-scale metric, while broader headcount and org depth remain private. SR001, SR002, SR003
CR004 Mirendil’s core public thesis is to build systems that automate or improve AI R&D and later serve scientists across biology, chemistry, drug discovery, materials science, and robotics. SR001, SR002, SR003, SR004
CR005 Adverse commentary explicitly criticizes Mirendil’s valuation because no product, revenue, or technical details had been publicly disclosed. SR005
CR006 The European Commission says the EU AI Act imposes strict pre-market obligations on high-risk AI systems, including risk mitigation, data quality, logging, documentation, human oversight, robustness, cybersecurity, and accuracy. SR009, SR010
CR007 The European Commission says GPAI model rules include transparency and copyright-related obligations, and providers of systemic-risk models should assess and mitigate those risks. SR009, SR010
CR008 EU transparency rules around AI-generated content are scheduled to come into effect in August 2026, creating a timing gate for any EU-facing generated-content workflows. SR009, SR010
CR009 FDA materials state that AI/ML-enabled medical devices are reviewed through pathways such as 510(k), De Novo, or premarket approval, and significant software modifications may require review. SR011, SR012
CR010 FDA PCCP guidance recommends that AI-enabled device submissions describe planned modifications, methodology to develop, validate, and implement those changes, and impact assessment. SR011, SR012
CR011 Foley Hoag states that AI trade-secret disputes increasingly require precise identification of what is secret across models, code, architectures, training data, and deployment processes. SR014, SR016
CR012 Foley Hoag reports that a former Google engineer was convicted in January 2026 on AI-related trade-secret theft counts involving TPU chips and AI training infrastructure. SR014
CR013 Beck Reed Riden reports xAI sued a former employee in 2025 alleging misappropriation of Grok-related trade secrets before joining OpenAI. SR015
CR014 Jones Day and JD Supra both warn that entering company trade secrets into public generative AI tools can threaten trade-secret protection if reasonable secrecy measures are not maintained. SR016, SR017
CR015 Legal sources recommend layered mitigation for AI trade-secret risk, including updated employee/IP agreements, AI-use policies, training, exit certifications, access monitoring, and forensic readiness. SR015, SR016, SR017
CR016 arXiv researchers found nearly 300 ACL, NAACL, and EMNLP papers in 2024-2025 contained at least one hallucinated citation, with most published in 2025. SR023
CR017 Retraction Watch reported a Lancet-linked audit finding fabricated citations in PubMed-indexed literature increased twelve-fold in two years. SR024
CR018 Retraction Watch reported that about one in 277 PubMed-indexed papers published in the first seven weeks of 2026 referenced a non-existent paper. SR024
CR019 A May 2026 arXiv audit estimated 146,932 hallucinated citations in 2025 across arXiv, bioRxiv, SSRN, and PubMed Central. SR025
CR020 Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. SR006
CR021 Gartner says many agentic AI projects are early-stage experiments or proofs of concept driven by hype and often misapplied. SR006
CR022 Gartner says agentic AI propositions often lack significant ROI because current models do not have the maturity and agency to autonomously achieve complex business goals over time. SR006
CR023 NIST says its AI Risk Management Framework is intended to improve incorporation of trustworthiness considerations into the design, development, use, and evaluation of AI systems. SR013
CR024 NIST released a Generative AI Profile in 2024 to help organizations identify unique risks posed by generative AI and align risk-management actions with their goals. SR013
CR025 OpenAI and Anthropic both publish frontier-AI safety or responsible-scaling materials, making public risk-governance expectations visible even for private frontier labs. SR026, SR027
CR026 No public Mirendil safety policy, trust center, DPA, status page, incident record, or compliance certification surfaced in the reviewed official and launch materials. SR001, SR002, SR003, SR005
CR027 NVIDIA is publicly identified as a participant in Mirendil’s seed round, creating positive compute signaling but not disclosing any guaranteed capacity or commercial compute terms. SR001, SR002, SR004
CR028 CB Insights says AI companies raised $226 billion in 2025, representing 48% of total venture funding and the largest share on record. SR020
CR029 CB Insights says mega-rounds captured $307 billion, or 65% of total 2025 venture funding, while total deal count fell 17%. SR020
CR030 CB Insights says private AI companies raised $226 billion in Q1 2026, with $100 million-plus rounds accounting for 94% of funding and average deal size reaching $160 million. SR021
CR031 CB Insights says Q1 2026 AI capital is increasingly top-heavy, with leading model developers racing to cover compute, talent, and energy costs. SR021
CR032 CRS reports the FY2026 NSF budget request sought $3.9 billion in discretionary funding, a $5.2 billion or 56.9% decrease from the FY2025 enacted level. SR008
CR033 C&EN reported the NSF budget proposal would reduce the estimated proposal funding rate from 26% to 7%. SR007
CR034 C&EN reported proposed NSF cuts would affect research disciplines broadly, including areas such as biotechnology, advanced manufacturing, and semiconductors even if AI was prioritized. SR007
CR035 Periodic Labs, Lila Sciences, and Isomorphic Labs show that well-funded AI-for-science competitors are pursuing adjacent scientific-automation and pharma-partnership strategies. SR028, SR029, SR030
CR036 Mirendil has not publicly disclosed whether it will initially avoid regulated clinical claims, EU high-risk use cases, or FDA-regulated workflows. SR001, SR002, SR003, SR011, SR009
CR037 No public lawsuit, enforcement action, regulatory complaint, sanctions record, or formal certification naming Mirendil was confirmed from reviewed public sources. SR001, SR002, SR003, SR005, SR014, SR015
CR038 CNBC reports the AI talent war includes Big Tech firms competing for scarce AI researchers with multi-million-dollar compensation packages. SR018
CR039 CNBC reports Sam Altman said Meta tried to tempt top OpenAI talent with $100 million signing bonuses and higher compensation packages. SR018
CR040 CNBC quotes an AI industry policy executive saying demand for AI specialists has skyrocketed while supply stayed relatively constant, creating wage inflation. SR018
CR041 TechCrunch reports AI seed startups commonly saw $10 million seed rounds at $40 million to $45 million post-money valuations in 2026. SR019
CR042 TechCrunch reports investors are pricing some AI seed rounds years ahead of traction. SR019
CR043 TechCrunch reports AI has raised the seed-stage bar for founders to have live products, users, and revenue straight out of the gate. SR019
CR044 TechCrunch reports higher seed valuations mean less margin for error, less tolerance for pivots, and more scrutiny if progress does not match capital raised. SR019
CR045 Crunchbase News argues AI venture activity is better characterized as a risk bubble, with investors taking systematic exposure to AI outcomes rather than diversified sector risk. SR022
CR046 Because Mirendil is pre-product and pre-revenue, its customer-conversion risk cannot be retired without named design partners, paid pilots, or buyer-validated workflows. SR001, SR005, SR019
CR047 A useful Mirendil technical kill criterion is failure to produce reproducible external benchmarks, blinded demos, or scientific-integrity metrics within the seed runway planning window. SR006, SR023, SR024, SR025
CR048 A useful Mirendil legal kill criterion is inability of outside counsel to verify clean IP provenance, former-employer obligations, and data/model access controls. SR014, SR015, SR016, SR017
CR049 A useful Mirendil regulatory kill criterion is launching EU-facing or medical workflows without a classification memo, named compliance owner, and product-claim boundaries. SR009, SR011, SR012, SR013
CR050 A useful Mirendil financing kill criterion is attempting a next valuation step-up without benchmark proof, named pilots, or insider support sufficient to absorb compute and talent costs. SR019, SR020, SR021, SR022
CV001 Mirendil publicly says a16z and Kleiner Perkins led its $200 million seed round and NVIDIA followed as an investor. SV001, SV002, SV003
CV002 Independent launch coverage reported Mirendil raised $200 million at a $1 billion valuation. SV004, SV005
CV003 Andreessen Horowitz frames Mirendil as a lab-grade research platform that could help engineers, AI researchers, and eventually scientists run frontier AI work. SV002, SV001
CV004 Kleiner Perkins says Mirendil’s loop is the product: better models do better research and better research produces better models. SV003, SV001
CV005 Carta reported median seed post-money valuation rose to $24 million in Q4 2025 from $18 million a year earlier. SV008
CV006 Carta reported median seed and Series A dilution remained around 19% to 20%, implying larger round sizes were pushing valuations higher. SV008, SV007
CV007 TechCrunch reported that AI seed startups were commonly seeing roughly $10 million rounds at $40 million to $45 million post-money valuations in 2026. SV006
CV008 TechCrunch reported investors are pricing AI seed rounds years ahead of traction and expect live product, users, and revenue much earlier than before. SV006
CV009 Crunchbase reported close to 700 seed-stage rounds of $10 million or more in 2025, putting that category on track for an all-time high. SV009
CV010 Crunchbase reported more than 42% of global seed funding and just over $15 billion had gone to AI-focused seed rounds in 2025. SV009
CV011 Crunchbase reported U.S. seed rounds of $100 million or more topped $3.6 billion in 2025, with Thinking Machines Lab as the largest driver. SV009, SV011
CV012 TechCrunch reported Thinking Machines Lab closed a $2 billion seed round at a $12 billion valuation before fully revealing its product. SV011, SV012
CV013 TechCrunch reported Safe Superintelligence raised an additional $2 billion at a $32 billion valuation while its product was still in the works. SV028, SV026
CV014 TechCrunch reported Periodic Labs raised a $300 million seed to automate scientific discovery through AI scientists and autonomous laboratories. SV013, SV014
CV015 Tech Funding News reported Periodic Labs was discussing a raise around a $7 billion valuation and had semiconductor customer traction. SV014, SV013
CV016 Lila Sciences said its $350 million Series A brought total capital raised to $550 million and would open its platform to commercial partners. SV015, SV016
CV017 Isomorphic Labs announced a Lilly collaboration with a $45 million upfront payment and potential total deal value up to $1.7 billion. SV017
CV018 Recursion’s 2025 Form 10-K reported public float of $2.03 billion as of June 30, 2025. SV018
CV019 Recursion’s 2025 Form 10-K reported 2025 total revenue of $74.681 million and R&D expense of $475.271 million. SV018
CV020 Recursion’s 2025 Form 10-K reported a 2025 net loss of $644.759 million. SV018
CV021 Schrödinger’s 2025 Form 10-K reported public float of $1.124 billion, 2025 total revenues of $255.869 million, R&D expense of $173.138 million, and net loss of $103.265 million. SV019
CV022 Gartner predicted more than 40% of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear value, or inadequate risk controls. SV020
CV023 Gartner said many agentic-AI projects were early experiments driven by hype and estimated only about 130 of thousands of agentic AI vendors were real. SV020
CV024 Crunchbase argued AI venture markets show risk-bubble characteristics because investors may accept huge systematic exposure to AI and weaker traditional risk analysis. SV010, SV021
CV025 The World Economic Forum described how an AI bubble could divert capital toward AI projects and later create job losses and risk aversion if expectations disappoint. SV022, SV021
CV026 Mordor Intelligence estimated the agentic AI market at $9.89 billion in 2026 and forecast $57.42 billion by 2031. SV023
CV027 Grand View Research estimated AI in drug discovery at $2.9 billion in 2026 and forecast $13.8 billion by 2033. SV024
CV028 Dimension Market Research estimated the autonomous chemical laboratory market at $5.7485 billion in 2026 and $19.4834 billion by 2035. SV025
CV029 TechCrunch reported Mistral was in talks to raise about €3 billion at around a €20 billion valuation in 2026. SV030, SV031
CV030 CNBC reported Mistral reached an €11.7 billion valuation in a Series C round led by ASML, more than doubling its prior valuation. SV031, SV030
CV031 TechCrunch reported xAI raised $20 billion in a Series E round and said it had about 600 million monthly active users of X and Grok. SV029, SV032
CV032 CNBC reported AI talent markets include multi-million-dollar compensation packages and that top labs may spend tens or hundreds of millions to hire engineers while models cost billions to build. SV032
CV033 Mirendil’s official site says its founding team consists of 20 researchers and engineers from Anthropic, xAI, Google DeepMind, and OpenAI. SV001, SV002, SV003
CV034 A $200 million raise at a roughly $1 billion post-money valuation implies about 20% simplified new-investor ownership before any unreported terms. SV001, SV004, SV008
CV035 Public evidence does not support Mirendil’s reported $1 billion valuation on fundamentals because no public product, revenue, customer count, benchmark, or governance package has been disclosed. SV001, SV004, SV005, SV006
CV036 The evidence supports a track or research-more recommendation rather than a buy at the reported price. SV001, SV004, SV005, SV010, SV020
CV037 The bull scenario requires private technical proof, design-partner pull, and scarcity-lab follow-on demand sufficient to support a $3 billion to $6 billion mark. SV011, SV013, SV015, SV030
CV038 The base scenario assumes research progress but limited commercialization proof, supporting a broad $0.8 billion to $2 billion next-proof-cycle range. SV001, SV004, SV006, SV008
CV039 The bear scenario assumes weak proof or market compression, supporting a $0.2 billion to $0.7 billion range and down-round or structured-bridge risk. SV010, SV020, SV021, SV022
CV040 Mirendil’s valuation stance should remain expensive until private diligence proves technical superiority, external demand, clean structure, and runway adequacy. SV005, SV006, SV008, SV010
CV041 The risk rating should be high because market opportunity, founder quality, and investor quality are offset by product, commercial, governance, and market-cycle uncertainty. SV001, SV005, SV020, SV021
CV042 Exit readiness is low today because Mirendil has not disclosed revenue, customers, product maturity, public-company controls, or strategic partnership economics. SV001, SV005, SV017, SV018, SV019
CV043 Public AI-for-science comparables suggest later markets reward proof and economics, not narrative alone, because Recursion and Schrödinger disclose revenue but still carry heavy R&D spending and losses. SV018, SV019, SV017
CV044 The largest diligence asks are technical benchmarks, customer proof, commercial model, compute and burn, cap-table structure, governance, security, and follow-on path. SV001, SV002, SV005, SV020
CV045 The recommendation should improve only if private evidence shows repeated external scientific or AI-builder use that produces valuable research outputs faster than incumbent workflows. SV002, SV003, SV017, SV025
来源
编号出版方标题引文
SO001 Mirendil Democratizing frontier AI R&D to accelerate science and technology
SO002 Andreessen Horowitz Investing in Mirendil
SO003 Kleiner Perkins Mirendil: Building the system that builds systems
SO004 SiliconANGLE Mirendil raises $200M to speed up scientific research with AI
SO005 Tech Funding News One year at Anthropic, then $200M at $1B: The researchers who just closed one of AI’s largest-ever seed rounds
SO006 Unite.AI Former Anthropic Researchers Launch Mirendil at $1 Billion Valuation With $200M Seed Round
SO007 The SaaS News Mirendil Raises $200M Seed
SO008 Andrew.ooo Mirendil $200M Seed: AI Building AI (June 2026)
SO009 Nextomoro Mirendil
SO010 Behnam Neyshabur Behnam Neyshabur
SO011 Behnam Neyshabur Behnam Neyshabur CV
SO012 Google Scholar Behnam Neyshabur - Google Scholar
SO013 arXiv Sharpness-Aware Minimization for Efficiently Improving Generalization
SO014 arXiv Solving Quantitative Reasoning Problems with Language Models
SO015 arXiv Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context
SO016 Google Scholar Harsh Mehta - Google Scholar
SO017 Happenstance Harsh Mehta
SO018 TechStartups Ex-Anthropic researchers launch Mirendil, target $175M at $1B valuation for AI-powered scientific discovery
SO019 TMCnet Insight Mirendil Secures Major Funding to Expand Scientist-Focused AI Engineering Automation
SO020 The Next Web Ex-Anthropic researchers raise $200M for self-improving AI
SO021 WhatJobs News Anthropic Veterans’ Startup Mirendil Seeks to Help Scientists Develop Their Own AI
SO022 Complete AI Training Former Anthropic researchers launch Mirendil, seek $175 million for scientific AI startup
SO023 Grokipedia Behnam Neyshabur — Grokipedia
SO024 Research.com 2026 Behnam Neyshabur: Computer Science Researcher – H-Index, Publications & Awards
SO025 StartupHub Mirendil - Funding, Investors, Team & Alternatives
SO026 The Cryptonomist No Product, No Revenue: Mirendil AI Funding Lands $200M at $1B Mirendil has not shipped a product. It has not reported revenue. It has not released technical details about its AI systems.
SO027 Crypto Briefing Mirendil secures $200M seed round led by a16z and Nvidia The article’s core claim that Mirendil closed a $200 million round with Nvidia as a lead investor is directly contradicted by the research.
SM001 Grand View Research Artificial Intelligence In Drug Discovery Market Size, Share & Trends Report, 2026-2033
SM002 Precedence Research Artificial Intelligence (AI) In Drug Discovery Market Size and Growth 2026 to 2035
SM003 Mordor Intelligence Agentic AI Market Size & Share Analysis - Growth Trends & Forecasts (2026-2031)
SM004 GMInsights AI In Drug Discovery Market Size, Share & Growth Report
SM005 Research and Markets AI in Drug Discovery Market Report 2026
SM006 Stanford HAI Economy | The 2026 AI Index Report | Stanford HAI
SM007 Nature How much of the scientific literature is generated by AI? How much of the scientific literature is generated by AI? The first studies of the size of the AI footprint in scientific journals, preprint repositories and peer-review reports give a spread of answers -- and indicate a rapidly evolving situation that it is difficult to get a handle on.
SM008 Chemical & Engineering News (ACS) NSF budget proposal slashes funding across the agency
SM009 American Institute of Physics (FYI) FY2026 National Science Foundation Budget Tracker
SM010 TechCrunch Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science
SM011 Forbes Former OpenAI Researcher To Raise $500 Million For AI Science Startup
SM012 Lila Sciences Announcing Lila's $350M Series A and Incredible Partners on Our Mission
SM013 FierceBiotech Flagship's Lila Sciences lands $235M to expand AI-powered autonomous research labs
SM014 BioSpace Undeterred by Political, Economic Headwinds, Pharma Ups R&D Investment in 2024 and Beyond
SM015 EFPIA The Pharmaceutical Industry in Figures (2025)
SM016 ZS Pharma industry outlook, trends and priorities for 2026
SM017 Gartner Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls, according to Gartner, Inc.
SM018 Epoch AI Trends in Artificial Intelligence
SM019 TechFundingNews Former OpenAI and DeepMind researchers eye $7B valuation for AI startup Periodic Labs
SM020 Congressional Research Service The National Science Foundation (NSF): FY2026 Appropriations and Funding Trends
SM021 Computing Research Association President Releases Devastating NSF Budget Request
SM022 dasroot.net The Integrity Crisis: LLM-Generated Errors and the Future of Scientific Publishing
SM023 Isomorphic Labs / PR Newswire Isomorphic Labs Announces Strategic Multi-Target Research Collaboration With Lilly
SM024 Fortune Business Insights Agentic AI Market Size, Share & Industry Analysis, 2026-2034
SM025 Dimension Market Research Autonomous Chemical Laboratory Market Size 2026-2035
SM026 Excelra The State of AI/ML in Drug Discovery 2026 -- Executive Report
SM027 McKinsey & Company The State of AI 2025: Agents, Innovation, and Transformation
SM028 Forbes AI Agents And Hype: 40% Of AI Agent Projects Will Be Canceled By 2027
SP001 Mirendil Democratizing frontier AI R&D to accelerate science and technology
SP002 Andreessen Horowitz Investing in Mirendil Mirendil is building a system that can help anyone do AI work: they train frontier models that are expert at AI R&D and build the product around it.
SP003 Kleiner Perkins Mirendil: Building the system that builds systems
SP004 TechCrunch Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science The goal of Periodic Labs is nothing less than to automate scientific discovery, creating AI scientists.
SP005 Tech Funding News Former OpenAI and DeepMind researchers seek $7B valuation to build AI scientists
SP006 Forbes Former OpenAI Researcher To Raise $500 Million For AI Science Startup
SP007 Lila Sciences LILA | Scientific Superintelligence
SP008 Lila Sciences Announcing Lila’s $350M Series A and Incredible Partners on Our Mission
SP009 Fierce Biotech Flagship's Lila Sciences lands $235M to expand AI-powered autonomous research labs
SP010 FutureHouse FutureHouse
SP011 Sakana AI The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery The AI Scientist can incorrectly implement its ideas or make unfair comparisons to baselines, leading to misleading results.
SP012 Isomorphic Labs Reimagining Drug Discovery Process with AI
SP013 Isomorphic Labs / PRNewswire Isomorphic Labs announces strategic multi-target research collaboration with Lilly Isomorphic Labs will receive an upfront cash payment of $45 million and is eligible to receive up to $1.7 billion in performance-based milestone payments.
SP014 Google DeepMind AlphaFold
SP015 Recursion Pioneering AI Drug Discovery
SP016 Recursion Investor Relations
SP017 Recursion Recursion's Drug Discovery Pipeline
SP018 Securities and Exchange Commission Recursion Pharmaceuticals submissions
SP019 Insilico Medicine Main | Insilico Medicine
SP020 Chai Discovery Chai Discovery
SP021 Cradle Cradle | Engineer better proteins, faster
SP022 NVIDIA NVIDIA AI for Healthcare and Life Sciences
SP023 Schrödinger Schrödinger - Physics-based Software Platform for Molecular Discovery & Design
SP024 Anthropic Research
SP025 Anthropic Making Claude a chemist
SP026 OpenAI Research
SP027 StartupHub Mirendil Alternatives & Competitors (2026)
SP028 Gartner Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End 2027 Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.
SP029 Excelra The State of AI/ML in Drug Discovery 2026 — Executive Report Why data moats now matter more than algorithms, with 3 of pharma’s top 4 adoption barriers being data-related as foundation models commoditize.
SI001 Mirendil Democratizing frontier AI R&D to accelerate science and technology Mirendil describes a mission to democratize frontier AI R&D, but the public page does not disclose pricing, customers, or revenue.
SI002 Andreessen Horowitz Investing in Mirendil
SI003 Kleiner Perkins Mirendil: Building the system that builds systems
SI004 SiliconANGLE Mirendil raises $200M to speed up scientific research with AI
SI005 The Cryptonomist No Product, No Revenue: Mirendil AI Funding Lands $200M at $1B Mirendil has not shipped a product. It has not reported revenue. It has not released technical details about its AI systems.
SI006 Recursion Pharmaceuticals Annual Reports
SI007 U.S. Securities and Exchange Commission Recursion Pharmaceuticals 2025 Form 10-K
SI008 U.S. Securities and Exchange Commission Schrödinger 2025 Form 10-K
SI009 Schrödinger Schrödinger Reports Fourth Quarter and Full-Year 2025 Financial Results
SI010 Carta State of Private Markets: Q1 2026
SI011 Carta Record-setting early-stage valuations
SI012 Carta State of Pre-Seed: Q1 2026
SI013 Crunchbase News Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B
SI014 TechCrunch Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science
SI015 Lila Sciences Announcing Lila’s $350M Series A and Incredible Partners on Our Mission
SI016 Fierce Biotech Flagship’s Lila Sciences lands $235M to expand AI-powered autonomous research labs
SI017 PR Newswire Isomorphic Labs Announces Strategic Multi-Target Research Collaboration with Lilly
SI018 Dimension Market Research Autonomous Chemical Laboratory Market Size 2026–2035
SI019 Excelra The State of AI/ML in Drug Discovery 2026 — Executive Report
SI020 Gartner Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
SI021 Epoch AI How Much AI Compute Do Frontier Labs Use?
SI022 arXiv The rising costs of training frontier AI models
SI023 Lambda AI Cloud Pricing | GPU Compute & AI Infrastructure
SI024 NVIDIA NVIDIA H100 GPU
SI025 Carta Data Desk by Carta: Private Market Insights
SI026 Sequoia Capital AI’s $600B Question The goal of the piece was to ask the question: “Where is all the revenue?”
SE001 Mirendil Democratizing frontier AI R&D to accelerate science and technology We train frontier models that are exceptional at it and redesign the entire lab from scratch around them to make the full loop faster, more capable, and more autonomous.
SE002 Andreessen Horowitz Investing in Mirendil The training data needs to cover the full loop of AI research, from proposing experiments, to writing and running code, interpreting results, debugging failures, improving kernels, managing compute, comparing checkpoints, and deciding what to try next.
SE003 Kleiner Perkins Mirendil: Building the system that builds systems Better models do better research. Better research produces better models. The loop is the product.
SE004 Sakana AI The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
SE005 arXiv The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
SE006 GitHub SakanaAI/AI-Scientist
SE007 arXiv MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering
SE008 OpenAI MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering
SE009 GitHub openai/mle-bench
SE010 METR Evaluating frontier AI R&D capabilities of language model agents against human experts
SE011 GitHub METR/RE-Bench
SE012 arXiv RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts
SE013 SWE-bench SWE-bench Leaderboards
SE014 GitHub SWE-bench: Can Language Models Resolve Real-world Github Issues?
SE015 Epoch AI Trends in AI supercomputers
SE016 NVIDIA Research at NVIDIA
SE017 Google DeepMind FunSearch: Making new discoveries in mathematical sciences using Large Language Models
SE018 Nature Mathematical discoveries from program search with large language models
SE019 Nature Accurate structure prediction of biomolecular interactions with AlphaFold 3
SE020 UC Berkeley RDI We Scored 100% on AI Benchmarks Without Solving a Single Problem We built an AI agent that analyzes benchmark evaluation code in depth and automatically discovers inflation of benchmark scores.
SE021 GitHub moogician/trustworthy-env
SE022 Gartner Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End 2027 Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.
SE023 arXiv MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation
SE024 GitHub snap-stanford/MLAgentBench
SE025 arXiv SciCode: A Research Coding Benchmark Curated by Scientists
SE026 GitHub scicode-bench/SciCode
SE027 Google DeepMind AlphaFold
SE028 BenchLM LLM Leaderboard 2026 — Compare 272 AI Models Across 249 Benchmarks
SE029 Kaggle Kaggle: The World’s AI Proving Ground
SU001 Mirendil Mirendil — Join us
SU002 Mirendil (via Ashby) Mirendil Jobs 14 open roles listed, every one titled Member of Technical Staff across agent harness, inference, kernels, pretraining, post-training/RL, platform, infrastructure, model evaluation, security, and design engineering.
SU003 Mirendil Democratizing frontier AI R&D to accelerate science and technology Today, any lab trying to use AI in drug discovery, chemistry, biology, or robotics must also become a frontier AI lab.
SU004 Andreessen Horowitz Investing in Mirendil
SU005 Kleiner Perkins Mirendil: Building the system that builds systems
SU006 TechCrunch Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science
SU007 PR Newswire Isomorphic Labs Announces Strategic Multi-Target Research Collaboration with Lilly
SU008 PR Newswire Isomorphic Labs Announces Strategic Multi-Target Research Collaboration with Novartis
SU009 Recursion Pharmaceuticals Partners "The collaboration with Recursion enables us to discover small molecule drug candidates targeting novel biology for the treatment of fibrotic diseases..." (Joerg Moeller, MD, Bayer AG); "...highlights the potential of technology to transform drug discovery..." (James Sabry, MD, PhD, Roche).
SU010 Recursion Pharmaceuticals Pipeline
SU011 U.S. Securities and Exchange Commission Recursion Pharmaceuticals 2025 Form 10-K
SU012 U.S. Securities and Exchange Commission Schrödinger 2025 Form 10-K
SU013 Schrödinger, Inc. Schrödinger Reports Fourth Quarter and Full Year 2025 Financial Results
SU014 Lila Sciences Lila Sciences homepage
SU015 Lila Sciences Announcing Lila's $350M Series A and Incredible Partners on Our Mission
SU016 BIO International Convention (BIO 2026) Participating Companies: Lila Sciences
SU017 Cradle Cradle homepage
SU018 Chai Discovery Chai Discovery homepage
SU019 FutureHouse FutureHouse homepage
SU020 FutureHouse About FutureHouse (FAQ) FutureHouse is funded through philanthropic partnerships and grants ... Eric and Wendy Schmidt ... OpenPhilanthropy, the National Science Foundation, and the AI Safety Institute.
SU021 Sakana AI The AI Scientist
SU022 Periodic Labs Periodic Labs homepage ...issues with heat dissipation on their chips. We're training custom agents for their engineers and researchers to make sense of their experimental data in order to iterate faster.
SU023 Observer Periodic Labs Launches With $300M to Build Real 'Science AI'
SU024 ZS Associates Pharma Industry Outlook 2026
SU025 McKinsey & Company The State of AI 2025: Agents, Innovation, and Transformation
SU026 Gartner Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
SU027 Forbes (Jason Snyder) MIT Finds 95% Of GenAI Pilots Fail Because Companies Avoid Friction A new MIT study ... reveals that billions of dollars invested in enterprise GenAI pilots are yielding no results ... 95% of GenAI pilots fail.
SU028 MIT NANDA (Project NANDA) The GenAI Divide: State of AI in Business 2025
SU029 EDUCAUSE Higher Education Community Vendor Assessment Toolkit (HECVAT)
SU030 StartupHub.ai Mirendil — Alternatives
SR001 Mirendil Democratizing frontier AI R&D to accelerate science and technology
SR002 Andreessen Horowitz Investing in Mirendil
SR003 Kleiner Perkins Mirendil: Building the system that builds systems
SR004 SiliconANGLE Mirendil raises $200M to speed up scientific research with AI
SR005 The Cryptonomist Mirendil AI Funding Raises $200M for Frontier Research Mirendil has not shipped a product. It has not reported revenue. It has not released technical details about its AI systems.
SR006 Gartner Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.
SR007 Chemical & Engineering News NSF budget proposal slashes funding across the agency The proposed budget would slash the agency’s budget by $5.1 billion, or 57%, from current funding levels.
SR008 Congressional Research Service The National Science Foundation: FY2026 Appropriations and Funding History The Trump Administration is seeking $3.9 billion in discretionary funding for NSF in FY2026, a $5.2 billion (-56.9%) decrease from the FY2025 enacted level.
SR009 European Commission AI Act High-risk AI systems are subject to strict obligations before they can be put on the market.
SR010 EU Artificial Intelligence Act EU Artificial Intelligence Act | Up-to-date developments and analyses of the EU AI Act
SR011 U.S. Food and Drug Administration Artificial Intelligence in Software as a Medical Device The FDA reviews medical devices through an appropriate premarket pathway, such as premarket clearance (510(k)), De Novo classification, or premarket approval.
SR012 U.S. Food and Drug Administration Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions The FDA is issuing this guidance to provide recommendations for predetermined change control plans (PCCPs) tailored to artificial intelligence (AI)-enabled devices.
SR013 National Institute of Standards and Technology AI Risk Management Framework The NIST AI Risk Management Framework is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems.
SR014 Foley Hoag LLP Litigating Trade Secret Claims Focused on Generative AI AI systems pose distinct challenges for pleading and responding to trade secret misappropriation claims.
SR015 Beck Reed Riden LLP Employee Departures and Trade Secret Risk in the AI Era In the AI industry, the stakes are particularly high. Trade secrets include model weights, training data, system prompts, and tuning methods.
SR016 Jones Day Protecting Trade Secrets as Generative AI Evolves If an employee inputs a company's trade secret into an AI prompt, that trade secret could be at risk of losing its trade secret protection.
SR017 JD Supra / Sheppard Mullin The AI Knows Too Much: When Employees Feed Trade Secrets into Generative AI Tools The mere act of entering trade secrets into a public generative AI tool may itself threaten their protected status.
SR018 CNBC Behind the AI talent war: Why tech giants are paying millions to top hires Meta CEO Mark Zuckerberg offered $100 million signing bonuses to top OpenAI employees.
SR019 TechCrunch It’s not your imagination: AI seed startups are commanding higher valuations Higher seed valuations mean less margin for error, less room for experimentation, less tolerance for pivots, and more scrutiny if progress doesn’t match the capital raised.
SR020 CB Insights State of Venture 2025 AI companies raised $226B in 2025, accounting for 48% of total venture funding — the largest share on record.
SR021 CB Insights State of AI Q1’26 Report Private AI companies raised $226B in Q1’26, surpassing the full-year total for 2025 in just a single quarter.
SR022 Crunchbase News The Great AI Bubble Debate Venture capitalists have chosen huge systematic risk, rather than the usual idiosyncratic risk, which jeopardizes performance if the future doesn’t match up with today’s optimistic enthusiasm.
SR023 arXiv HalluCitation Matters: Revealing the Impact of Hallucinated References with 300 Hallucinated Papers in ACL Conferences Nearly 300 papers contain at least one HalluCitation, most of which were published in 2025.
SR024 Retraction Watch One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis The analysis of articles indexed in PubMed found that about one in 277 papers published in the first seven weeks of 2026 referenced a paper that didn’t exist.
SR025 arXiv LLM hallucinations in the wild: Large-scale evidence from non-existent citations The authors estimate 146,932 hallucinated citations in 2025 alone across arXiv, bioRxiv, SSRN, and PubMed Central.
SR026 OpenAI Safety & responsibility
SR027 Anthropic Anthropic's Responsible Scaling Policy
SR028 TechCrunch Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science
SR029 Lila Sciences Announcing Lila’s $350M Series A and Incredible Partners on Our Mission
SR030 PR Newswire / Isomorphic Labs Isomorphic Labs announces strategic multi-target research collaboration with Lilly The agreement includes an upfront cash payment of $45 million to Isomorphic Labs and the potential for the company to receive up to $1.7 billion in performance-based milestone payments.
SV001 Mirendil Democratizing frontier AI R&D to accelerate science and technology We are fortunate to work with Andreessen Horowitz and Kleiner Perkins, who led our seed round of $200M, followed by an investment from NVIDIA among others.
SV002 Andreessen Horowitz Investing in Mirendil
SV003 Kleiner Perkins Mirendil: Building the system that builds systems
SV004 SiliconANGLE Mirendil raises $200M to speed up scientific research with AI
SV005 The Cryptonomist No Product, No Revenue: Mirendil AI Funding Lands $200M at $1B A brand-new AI lab with no commercial products, no disclosed revenue, and no public technical details just raised $200 million at a $1 billion valuation.
SV006 TechCrunch It’s not your imagination: AI seed startups are commanding higher valuations
SV007 Carta State of Seed Report: Winter 2025
SV008 Carta Record-setting early-stage valuations
SV009 Crunchbase News Seed Funding In 2025 Broke Records Around Big Rounds And AI, With US Far In The Lead
SV010 Crunchbase News The Great AI Bubble Debate Valuation is an opinion on the future, whereas pricing reflects the current fundraising market.
SV011 TechCrunch Mira Murati’s Thinking Machines Lab is worth $12B in seed round
SV012 Tech Funding News Mira Murati-led Thinking Machines Lab shatters records with $2B seed funding
SV013 TechCrunch Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science
SV014 Tech Funding News Former OpenAI and DeepMind researchers seek $7B valuation to build AI scientists
SV015 Lila Sciences Announcing Lila’s $350M Series A and Incredible Partners on Our Mission
SV016 Fierce Biotech Flagship’s Lila Sciences lands $235M to expand AI-powered autonomous research labs
SV017 Isomorphic Labs Isomorphic Labs announces strategic multi-target research collaboration with Lilly
SV018 Securities and Exchange Commission Recursion Pharmaceuticals 2025 Form 10-K
SV019 Securities and Exchange Commission Schrödinger 2025 Form 10-K
SV020 Gartner Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.
SV021 CNBC AI valuation fears grip global investors as tech bubble concerns grow
SV022 World Economic Forum Anatomy of an AI reckoning
SV023 Mordor Intelligence Agentic AI Market Share, Size & Growth Outlook to 2031
SV024 Grand View Research Artificial Intelligence In Drug Discovery Market Report, 2033
SV025 Dimension Market Research Autonomous Chemical Laboratory Market Size 2026–2035
SV026 StartupHub.ai Best Frontier AI Labs (2026)
SV027 ValueAddVC AI Company Valuations 2026
SV028 TechCrunch OpenAI co-founder Ilya Sutskever’s Safe Superintelligence reportedly valued at $32B
SV029 TechCrunch xAI says it raised $20B in Series E funding
SV030 TechCrunch Mistral is rumored to be raising €3B at €20B valuation
SV031 CNBC AI firm Mistral valued at $14 billion as chip giant ASML takes major stake
SV032 CNBC Behind the AI talent war: Why tech giants are paying millions to top hires