红杉的终局判断：AI 应用公司的终点是每家都变成 Neo-Lab

by Mycelium Protocol

---

信息来源：红杉美国合伙人 Sonya Huang（@sonyatweetybird）一场面向创始人的闭门分享。官方未发布完整书面演讲稿，本文基于在 AI 创投圈广泛流传的分享内容整理与分析。

Sonya Huang 简介：红杉美国成长期合伙人，专注 AI 基础设施与 AI 应用投资，是 Sequoia Training Data 播客的主持人，Sequoia AI Ascent 年度峰会的主要负责人之一。个人资料页：sequoiacap.com/people/sonya-huang

---

核心论断：两个判断，一句总结

整场分享的核心落在两个相互关联的判断上：

判断一：Sovereign AI（主权 AI）才是下一代企业护城河——不是规模，不是用户数，是对核心智能的控制权。

判断二：更颠覆性——每一家认真做事的 AI 应用公司，最终都会演变成某种形态的「Neo-Lab（新一代研究实验室）」。

一句话总结：AI 应用公司不再只是 AI 技术的消费者，正在成为 AI 新的研究发源地。

---

第一部分：赛道拐点——竞争从「应用层」转向「智能层」

生成式 AI 的前两年，AI 应用创业的比拼逻辑很单一：比界面体验、比用户增长、比渠道能力。谁的 UI 更好看，谁的增长更快，谁的分发更强，谁就赢。

但这个逻辑已经过时了。

如今竞争的主战场已从应用层全面转向智能层。逻辑彻底反转：以前 AI 是产品的附加功能，是加分项；现在产品即智能，智能即产品。用户愿意留存、付费、长期使用的核心原因，就是你独有的智能能力。

AI 的质量，直接等于产品的核心竞争力。

这意味着前两年的 AI 创业本质——「拼包装」「套 API」——已经走到了尽头。

你能套 API，竞争对手也能套，底层能力没有本质差距，最后只能卷 UI、卷运营、卷流量，陷入同质化内卷，根本没有真正的壁垒。

接下来的比拼，是能不能长出别人抄不走的专属智能。

---

第二部分：哪些公司已经跑通了 Neo-Lab 路径

红杉在分享里给出了几个具体案例——它们已经走在这条路上：

公司	赛道	Neo-Lab 体现
Harvey	法律 AI	自建法律场景评测体系，专属法律模型微调
Cursor	代码工具	深度代码理解模型，非通用基础模型能力边界
Glean	企业搜索	私有数据语义理解，企业级专属索引与检索
Ramp	金融科技	财务场景专属智能，超越通用 API 的领域判断
OpenEvidence	医疗 AI	医学文献推理，专属临床场景对齐

共同路径：它们都不是先建实验室再找业务，而是先扎进真实业务场景，在解决一个个具体问题的过程中，自然长出了自研能力。

它们自己搭建领域评测体系、做 LoRA 微调、积累专属数据集、完成场景对齐，甚至训练自有小模型，也会发布论文、公开基准测试。

但出发点从来不是刷 SOTA 榜单，而是解决自身产品的真实痛点。

研究由业务驱动，研究为业务服务——这就是 Neo-Lab 的核心本质。

---

第三部分：主权 AI 四层拆解——不是要你从零训练大模型

「主权 AI」最常见的误解是「我哪有资金和算力训大模型」。红杉在分享里直接明确了边界：

主权 AI 绝不等于从零搭建基础大模型。

它的真正定义是：掌握对产品最关键那条智能链路的控制权。你不需要全链路自研，但决定产品差异化、决定用户核心体验、构成护城河的那部分智能能力，必须牢牢在自己手里，自己说了算。

具体拆成四层：

1. 数据主权

领域私有数据完全自主可控。数据的投喂方式、使用规则、流转路径，不受第三方模型厂商约束，也无需担心核心数据外泄。

2. 模型适配主权

可基于开源基座自主完成微调、LoRA 训练、领域对齐，把通用模型改造成适配自身业务的专用智能。想改就改，想调就调，不用被动等待第三方厂商迭代版本。

3. 评测与迭代主权

拥有面向业务场景的专属评测体系，能清晰量化模型在自身任务上的优劣，可持续自主迭代优化，而不是上游大模型更什么就用什么。

4. 部署主权

推理部署的位置、方式、版本升级与回滚节奏，完全自主决定。不会被第三方 API 的限流、涨价、版本下架打得措手不及。

重要补充：主权 AI 是一个连续光谱，不是非黑即白的开关：

档位	内容	适用范围
轻量级	自有业务评测集 + 深度 Prompt 工程 + RAG	轻度主权起步
中间档（推荐大多数）	开源基座微调 + 核心场景自有模型 + 长尾用外部 API	绝大多数 AI 应用公司
重模式	全量训练基础大模型	极少数公司

红杉给创始人的忠告：别一上来就选最重的方案。从业务痛点往回倒推，找出对护城河最重要的那一小段能力，先把这部分握在手里，就足以甩开同行一大截。

---

第四部分：Neo-Lab vs 传统 AI 实验室——五个本质差异

很多人有个固有印象：做 AI 研究就得建独立研究院，招顶尖博士，刷榜发论文。Neo-Lab 完全不是这个逻辑：

维度	传统基础模型实验室	Neo-Lab
目标导向	刷通用能力、冲学术榜单	第一优先级永远是业务结果
数据来源	海量公开互联网数据	业务真实产生的私有数据、真实用户反馈
核心产出	通用大模型、学术论文	领域评测方案、适配后的业务模型、垂直解法
评判标准	指标有没有刷上去	产品指标有没有提升、用户体验有没有变好
组织模式	独立研究院，与业务部门距离远	研究团队与产品、工程团队深度绑定

一句话总结：传统实验室是「先有研究，再找场景」；Neo-Lab 是「先有场景，倒逼研究」。

---

第五部分：三个创始人最容易踩的认知误区

误区一：建 Neo-Lab 就得砸几千张 GPU、招几十位博士

不对。 很多跑通的 Neo-Lab，早期只有 2-5 人的小团队：几个工程师加 1-2 位懂微调和对齐的算法人员，先盯着业务最痛的一两个任务攻坚，根本不用上来就做全栈大模型训练。

别搞形式主义——不是挂一块「AI 研究院」的牌子，就叫 Neo-Lab 了。

误区二：开源模型随处可得，下个权重跑起来就是主权 AI

不对。 拿到权重，远远不等于拥有主权。

真正的主权，是懂怎么改它、怎么测它、怎么顺着业务迭代它，而不只是能把它跑起来。如果下了开源模型，除了部署什么都不会，遇到问题毫无办法，那和调用第三方 API 本质上没有区别，照样没有真正的主权。

误区三：调用 API 省事，能一直套壳躺赢

短期没问题，长期一定死。 一旦赛道里出现走通主权 AI 路径的对手，进化成 Neo-Lab，对方会顺着业务场景越迭代越好，和你的差距会越拉越大。只会套 API 的公司，最终一定会陷入同质化价格战，没有任何护城河可言。

---

第六部分：落地指南——四条可直接执行的路径

红杉给出的四条建议，极其务实：

① 先盘点产品，划定核心边界

把产品的 AI 能力拆解清楚：哪些是核心差异化、是护城河来源，把这部分标记为「必须掌握主权」的模块；非核心能力继续用外部 API 即可，别为了「主权」而主权。

② 先建评测体系，再谈模型微调

在动模型之前，先想清楚一件事：用什么标准，能量化出模型在你的业务里到底好不好。没有靠谱的业务评测，所有微调都是盲调。

③ 从小处切入，循序渐进

优先用 LoRA、领域后对齐这些轻量手段，别上来就做全参数训练。从轻到重，一步步提升主权程度，风险小，见效快。

④ 组织上别搞孤岛

研究人员不能关起门来做实验，必须扎进业务里，看真实用户案例，和产品、工程团队深度绑定。Neo-Lab 一旦脱离了产品，就失去了存在的意义。

---

我们的补充分析：为什么这个判断值得认真对待

从竞争结构看：API 经济的核心问题是对称性——你能用的，竞争对手也能用，差异化极难持久。一旦竞争转移到智能层，数据飞轮、专属评测体系、积累的领域对齐能力，这些都不可直接复制，壁垒真实存在。

从投资视角看：Sequoia 本身投了 Fireworks（允许公司”拥有”而非”租用”智能的推理平台），这与这场分享的逻辑高度一致。「Sovereign AI」不是口号，是红杉用真金白银在下注的方向。

从中国市场看：国内 AI 应用公司面临的竞争压力更甚——字节、腾讯、阿里都在做「全家桶」，套壳产品更难生存。对国内创业公司来说，在垂直赛道里建立专属智能能力，可能比国外创业公司更紧迫。

时间窗口：红杉的判断是「前两年是红利期，接下来是淘汰赛」。如果这个时间线大致准确，留给纯套壳公司的时间已经不多了。

---

一句话结论

前两年是「人人都能做 AI 应用」的红利期，接下来是「谁掌控智能谁活下去」的淘汰赛。两年后还能留在牌桌上的垂直 AI 公司，骨子里一定是藏在业务里的 Neo-Lab。

— 红杉美国合伙人 Sonya Huang

---

相关链接

• Sonya Huang 个人页：https://www.sequoiacap.com/people/sonya-huang/
• Sonya Huang Twitter：https://twitter.com/sonyatweetybird
• Harvey（法律 AI）：https://www.harvey.ai/
• Cursor（代码工具）：https://cursor.com/
• Glean（企业搜索）：https://glean.com/
• Ramp（金融科技）：https://ramp.com/
• OpenEvidence（医疗 AI）：https://openevidence.com/
• Fireworks AI（“主权”推理平台，红杉投资）：https://fireworks.ai/

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

Sequoia’s Endgame Call: Every AI Application Company Will Become a Neo-Lab

by Mycelium Protocol

---

Source: A closed-door founders session by Sonya Huang (@sonyatweetybird), Sequoia Capital Growth Partner. No official transcript was published; this article is based on widely circulated content from that session.

Sonya Huang leads Sequoia’s AI investments in application and infrastructure companies. She hosts the Sequoia Training Data podcast and leads the Sequoia AI Ascent event. Profile: sequoiacap.com/people/sonya-huang

---

The Two Core Theses

The entire session rests on two linked arguments:

Thesis 1: Sovereign AI — control over your own intelligence — is the next-generation moat. Not scale, not users. Intelligence sovereignty.

Thesis 2 (more disruptive): Every serious AI application company will eventually become some form of “Neo-Lab” — a new-generation research laboratory.

One-sentence summary: AI application companies are no longer just consumers of AI technology. They are becoming the new source of AI research.

---

The Shift: Competition Has Moved from Application to Intelligence

For the first two years of generative AI, the competitive logic was simple: better UI, faster growth, stronger distribution. That era is over.

The battlefield has moved from the application layer to the intelligence layer. The inversion is complete: AI used to be a product’s added feature — a bonus. Now the product is the intelligence, and the intelligence is the product. The reason users stay, pay, and keep coming back is your unique intelligence capability.

AI quality directly equals product competitive advantage.

This means the first-era playbook — API wrapping, prompt engineering, slick UI — has hit a ceiling. The next phase asks: can you build proprietary intelligence that competitors can’t copy?

---

Companies Already on the Neo-Lab Path

Sequoia cited five examples that have already made this transition:

Company	Vertical	Neo-Lab evidence
Harvey	Legal AI	Proprietary legal evaluation frameworks, domain-fine-tuned models
Cursor	Code tools	Deep code understanding models beyond general LLM capability
Glean	Enterprise search	Private data semantic understanding, enterprise-specific retrieval
Ramp	Fintech	Finance-specific intelligence beyond what general APIs provide
OpenEvidence	Medical AI	Medical literature reasoning, clinical scenario alignment

None of these built a lab first and then looked for applications. They went deep into real business problems and grew their research capacity organically from solving those problems.

They run their own domain evaluations, do LoRA fine-tuning, accumulate private datasets, align to their scenarios, sometimes train small proprietary models, and publish papers. But the origin is always a real product pain point — never a SOTA benchmark.

---

Sovereign AI: Four Layers of Control

The most common misreading of “Sovereign AI” is that it requires training foundation models from scratch. Sequoia was explicit: it doesn’t.

Sovereign AI means control over the specific intelligence capabilities that determine your product’s differentiation, user experience, and moat. You don’t need full-stack self-development. But the part that matters — the part that makes you different — must be yours.

The four layers:

1. Data sovereignty — Private domain data fully under your control. How it’s fed, how it’s used, where it flows — none of it hostage to a third-party model vendor.

2. Model adaptation sovereignty — Ability to independently fine-tune, LoRA-train, and domain-align an open-source base model to fit your business. Change it when you want, don’t wait for an upstream vendor update.

3. Evaluation and iteration sovereignty — A proprietary evaluation framework scoped to your business tasks. You can measure and continuously improve model quality on your actual problems, not just follow whatever the upstream model does.

4. Deployment sovereignty — Full control over inference location, method, version rollout and rollback. Never caught off guard by third-party API rate limits, price hikes, or version deprecations.

Critical nuance: Sovereign AI is a continuous spectrum, not a binary switch.

Level	What it involves	Who needs it
Light	Domain eval set + Prompt engineering + RAG	Starting point for most
Middle (recommended for most)	Open-source fine-tuning + proprietary models for core, API for long tail	The sweet spot for most AI companies
Heavy	Full foundation model training	Extremely few companies

Sequoia’s practical advice: don’t start with the heaviest option. Work backwards from your business pain. Find the smallest capability segment most critical to your moat. Own that first. It’s enough to separate you from competitors.

---

Neo-Lab vs. Traditional Research Lab: Five Fundamental Differences

Dimension	Traditional LLM Lab	Neo-Lab
Goal	Push general capability, chase academic leaderboards	Business outcome is always priority one
Data	Massive public internet data	Private data from real operations, real user feedback
Output	General-purpose models, papers	Domain eval frameworks, business-tuned models, vertical solutions
Success metric	Did the benchmark go up?	Did the product metric improve? Did user experience improve?
Org structure	Independent research institute, remote from business teams	Research team deeply embedded with product and engineering

One-line summary: Traditional labs say “do research, then find applications.” Neo-Labs say “find the application first, let it force the research.”

---

Three Founder Misconceptions Sequoia Called Out

Misconception 1: Neo-Lab requires thousands of GPUs and dozens of PhDs.

False. Many working Neo-Labs started with 2-5 people: a few engineers plus 1-2 fine-tuning/alignment specialists, focused on the 1-2 most painful business tasks. No need for full-stack model training from day one. Putting up an “AI Research Institute” sign doesn’t make you a Neo-Lab.

Misconception 2: Download an open-source model’s weights and run it = Sovereign AI.

False. Having the weights is not having sovereignty. Real sovereignty means knowing how to modify it, how to evaluate it, how to iterate it along your business needs — not just being able to deploy it. If you can’t diagnose failures and can’t improve it, it’s functionally the same as calling a third-party API.

Misconception 3: API-wrapping is efficient, the easy path always works.

Fine for the short term. Catastrophic for the long term. Once a competitor in your vertical achieves AI sovereignty and becomes a Neo-Lab, they’ll iterate faster along their business context and the gap will compound. API-only companies end up in commoditized price wars with no defensibility.

---

Four Immediately Actionable Paths

Sequoia’s four concrete recommendations:

1. Map your product first, define the core boundary. Break down your product’s AI capabilities. Which ones drive differentiation, which ones are your moat? Label those “must own.” Non-core capabilities can keep using external APIs. Don’t pursue sovereignty for sovereignty’s sake.

2. Build the evaluation framework before touching the model. Before fine-tuning anything, answer: what metric tells you whether the model is doing your specific job well? Without a reliable business eval, all fine-tuning is blind.

3. Start small, go gradual. Prioritize lightweight methods: LoRA, domain post-alignment. Don’t start with full-parameter training. Build sovereignty incrementally, from light to heavy. Lower risk, faster results.

4. No research silos in the org. Researchers can’t lock themselves in a lab. They must be embedded in real operations, watching real user cases, deeply integrated with product and engineering teams. A Neo-Lab that disconnects from the product loses its reason to exist.

---

Our Read: Why This Call Deserves Serious Weight

From a competitive structure view: The core problem with the API economy is symmetry — what you can use, competitors can use too. Differentiation can’t last. Once competition shifts to the intelligence layer, data flywheels, proprietary evaluation frameworks, and accumulated domain alignment become genuinely non-replicable advantages.

From an investment view: Sequoia itself invested in Fireworks AI — a platform designed to let companies “own” rather than “rent” their intelligence. This is the same thesis as the talk, backed by real capital.

From a timing view: Sequoia’s framing is clear — the first two years were the dividend era. The next phase is the elimination round. If that timeline is roughly right, there’s limited runway left for pure API-wrapper companies.

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.