186: Anthropic招股书丨“银行也有百亿预算”，与Fusion Fund张璐聊大企业在想什么
- 01The AI Era Is Pivoting from "Build the Strongest Model" to "Embed It Cheaply Everywhere"
- 02Workflow Embedding, Not Model Performance, Drives Enterprise Retention
- 03Enterprise AI Budgets Are Larger Than the Market Realizes, and Competitive Contagion Is the Driver
- 04Two-Stage Enterprise Adoption: Automation First (3-5 Years), Optimization Later
- 05Open Source Is Eating Into Frontier Closed-Model Share in the Enterprise
- 06Cost Is the New Bottleneck: Inference Over Training, Token Bills Over Token Enthusiasm
1. Key Themes
The AI Era Is Pivoting from "Build the Strongest Model" to "Embed It Cheaply Everywhere"
Zhang Lu frames Anthropic's IPO as landing exactly at an inflection point between two AI epochs. The priority stack is changing: capability is table stakes, while cost, integration, and workflow penetration decide winners.
"Anthropic 上市正好处在一个转折点。在过去这五年吧,大家讨论最重要的一点是怎么样进行更大的资源的投入,可以去堆出来一个能力最强的 AI。但是,在接下来的这些年的话呢,我觉得更主要的不是怎么样是把 AI 的能力去嵌入到各个的工作流程和各个产业中去,模型能力不再是最重要的因素,而成本、整合等等这些才是。" — Zhang Lu [00:00:00]
(Translated: "Anthropic's IPO sits at a turning point. Over the past five years the key question was how to pour in more resources to stack up the most capable AI. Going forward the focus is embedding AI into every workflow and industry; model capability is no longer the most important factor, cost and integration are.")
Workflow Embedding, Not Model Performance, Drives Enterprise Retention
Zhang's most-impressed S-1 data point is customer depth: roughly 11,000-12,000 customers spending over $100K a year, and about 2,000 spending over $1M. Her reading is that model quality gets you the trial, but integration gets you retention.
"模型性能它只能决定说这个客户或大企业他会不会尝试这个产品。但最后能不能够长期的留下来,有一个很长的留存率是基于你是不是能够和现有的工作流程直接渗透进去整合进去。" — Zhang Lu [00:04:42]
Manchi adds the S-1's net revenue retention figure: "这个数字是 900%" [00:17:55].
Enterprise AI Budgets Are Larger Than the Market Realizes, and Competitive Contagion Is the Driver
Fortune 500 adoption is driven less by startup threat than by peer pressure: once one player in an industry moves, the rest follow fast.
"只要一个行业有一家动起来,剩下几家企业追上的速度就很快。包括美国有一家非常非常知名的一个银行,我从他们的首席 AI 官听到了他在 AI 的今年的预算是已经达到了 100 多亿美金。" — Zhang Lu [00:35:10]
She cites pharma: after Eli Lilly's multibillion-dollar Nvidia partnership and Merck's Google DeepMind deal, "60%、70% 的企业在内部的一个战略优先级里面 AI 都是第一位的" [00:34:35]. She also notes the Chief AI Officer role has proliferated over the last six months [00:33:45].
Two-Stage Enterprise Adoption: Automation First (3-5 Years), Optimization Later
Zhang says most non-tech industries won't rip out existing systems for AI-native rebuilds, so the first wave is agents plugged into current software.
"大部分的产业,大部分的企业,它并不想或并不能马上改变自己的工作流程...即使它用智能体,希望智能体在继续使用他们现在用的软件,而不是说你从头再给我自己编一套软件。" — Zhang Lu [00:36:56]
"第一步可能就需要三五年的时间,第二步可能还需要相对长一些的时间。" — Zhang Lu [00:37:41]
This matters for pacing expectations: the market is huge, but "coding-style" explosions will not repeat everywhere.
Open Source Is Eating Into Frontier Closed-Model Share in the Enterprise
Zhang cites a report showing enterprise use of closed/frontier models falling "从 57% 降到了 44%" [00:10:19], AT&T allocating 60-70% of internal AI budget to open source [00:10:50], and Nvidia (with Jensen Huang) pushing open weights and open data. Her net view is that the pie is growing faster than share is shrinking.
"整体的这个蛋糕还是越变越大的。那只是说在未来大家切分蛋糕的时候就不只是一家或者两家独大了,会有更多的开源、闭源,大企业、小企业一起来切分这个蛋糕。" — Zhang Lu [00:12:39]
Cost Is the New Bottleneck: Inference Over Training, Token Bills Over Token Enthusiasm
Zhang describes the AI cost cycle turning:
"有一个它不能说名字的大企业在大概一个月的时间之内,突然间 Anthropic 的 Claude 它的这个账单就达到了几亿美金,将近 5 亿美金的规模。这个对他们 CFO 来讲是非常震惊的一件事情。" — Zhang Lu [00:38:28]
"刚开始的时候是说鼓励大家要使用更多的 token,现在是限制你不能用太多。" — Zhang Lu [00:38:43]
She also notes compute demand is shifting: "80% 的算力消耗都是在 training。但是现在...80% 未来算力的消耗是会在 inference" [00:50:03], creating startup opportunities in inference cost reduction and local deployment.
Regulated Verticals Are Where Startups Hold Structural Advantage Over Model Labs
Enterprises in pharma, insurance, healthcare, and finance will share internal data with startups but are wary of sharing it with frontier labs, especially after Anthropic's wet lab move.
"它会让这些药厂觉得说,诶,你用了我所有的这些数据,你现在反过来跟我竞争,那我是不是还要继续让你能够拿到我的内部的数据?" — Zhang Lu [00:53:17]
"他们愿意分享给初创企业是因为他觉得他是有优势可以进行控制的。但是他们愿意把这些数据分享给 Anthropic 或 OpenAI 吗?" — Zhang Lu [00:52:52]
Local/small-model deployment is a legal requirement for some data, not a preference [00:48:08].
Coding Is the One Agent That Hits the "AGI-ish" Bar, and It Won't Be Replicated Quickly
Zhang offers a practical AGI test borrowed from Kai-Fu Lee: an agent that does over 90% of the work better than over 90% of people in its domain.
"现在看到所有的智能体里面唯一达到这个标准的,就是 AI coding。" — Zhang Lu [00:24:12]
She explains why coding exploded: tech adopts fast, usage is high-frequency, task value is high, results verify quickly, and the AI completes tasks rather than advising [00:42:52]. Legal, finance, and medical lack instant verifiability.
Compute: Demand Up, Per-Generation Asset Prices Down, and Contracts Are Cyclical
Zhang confirms mid-tier unicorns are stuck with 3-year compute contracts, but argues GPUs aren't fungible and demand remains strong.
"算力需求增长和算力资产价格下跌我觉得是可以同时发生的。" — Zhang Lu [01:03:30]
She notes shorter contracts are punitive: "签一年的话好像价格是要 double" [01:01:01].
AI Regulation for Physical AI Is the Unsolved, Urgent Problem
Zhang's top open question: with physical AI, hallucinations can injure people, so corporate CTOs are demanding regulatory tooling before deployment.
"他们一定要先买把监管层这个技术解决方案做进去之后,才可以进行大规模的 physical AI 的一个部署。" — Zhang Lu [01:08:19]
2. Contrarian Perspectives
Model Capability Is No Longer the Most Important Factor in AI
After years of an industry-wide arms race, Zhang flatly states the priority has flipped. This challenges the premise behind the multi-trillion-dollar valuations of frontier labs.
"模型能力不再是最重要的因素,而成本、整合等等这些才是。" — Zhang Lu [00:06:38]
A Weak Anthropic IPO Would Actually Be Healthy for the AI Ecosystem
Most investors fear a post-IPO drop; Zhang welcomes a pullback.
"我其实反而觉得市场如果能够更加理性的去定价,更加理性的去进行资本的投入,它对于长期的这样一个 AI 的创业生态是一个非常好的事情...如果市场的价格有回调,反而对长线是个好事情。" — Zhang Lu [00:07:38]
She also separates valuation froth from real demand: "他这个价格的虚高和 AI 的真实需求增长,他其实是两回事" [00:07:21].
Don't Chase AI-for-Drug-Discovery; AI for Clinical Trials Is the Bigger Prize
Against the headline-grabbing discovery narrative, Zhang argues the bottleneck is clinical cost, not target identification.
"本身在整个药物研发的过程中,最大的成本其实是花在临床的过程中了。那你现在反而难道要因为你发现了更多的 target,你要花更大的资金和更大的资源去做这么多的临床吗?" — Zhang Lu [00:57:26]
Her alternative: use AI to winnow hundreds of targets to the two or three most likely to succeed, targeting "至少在 50%、60% 以上" clinical success [00:58:07].
Consumer AI Will Be Won by Distribution-Rich Giants, So Startups Should Stay B2B
Zhang says startups largely cannot win consumer AI because of data and distribution advantages, citing Google.
"C 端的 AI 市场最终大头 80%、90% 还是会被大的模型公司或者说是科技公司去占领。" — Zhang Lu [01:05:45]
On Google: "你只要是有谷歌的这个浏览器,Chrome 浏览器你就可以免费使用...这个体量和这个速度,一般的初创企业根本没有办法去跟他竞争" [01:06:26]. She adds that hot consumer apps fade (Character.AI as example, per Manchi: "后来申请破产").
Thin Wrappers in Tech Are Dying; Regulation Is the Moat
Generic horizontal AI apps for the tech industry are being absorbed by models, while regulated industries retain a natural entry barrier.
"高监管它也是一个准入门槛。但如果说它是非高监管行业的话,比如说科技行业各种各样的这种 AI 应用,我觉得对于初创企业来讲的话呢,它的竞争性就会越来越差的。" — Zhang Lu [01:04:59]
3. Companies Identified
Anthropic. Frontier AI lab (Claude), focused on enterprise and coding, filing an S-1 for an IPO expected in mid-November. Mentioned for deep enterprise integration and explosive growth, and for risks around capital intensity and compute commitments.
"他已经拥有了超过我记得数字大概是 11,000 多家,可能到 12,000 家年消费超过 10 万美元以上的客户...大概有 2,000 家一年的消费额是达到了 100 万美金以上。" — Zhang Lu [00:04:06]
OpenAI. Frontier lab, consumer-leaning, reportedly delaying IPO to 2027. Mentioned as Anthropic's direct competitor and as a pioneer in AI-for-math research.
"他们两家公司...都想做第一个上市的公司...这个先发优势还是非常明确的。" — Zhang Lu [00:15:19]
Spark Capital. Venture firm (founded ~2005-06; earlier investments include Coinbase and Twitter), Anthropic's only VC board representative via partner Yasmin Razavi. Mentioned for conviction in a non-consensus bet.
"他在投 Anthropic 的时候,那个时候市场并没有形成一个共识。" — Zhang Lu [00:18:39]
Nvidia. Chip leader pushing open-source models and open weights/data to accelerate enterprise deployment, and partnering with Eli Lilly.
"你看英伟达它不仅是开放权重,它甚至开放数据,对吧?它真的是开源。就是因为他想推动企业的一个加速的部署。" — Zhang Lu [00:47:59]
AT&T. US telecom giant. Mentioned for publicly committing 60-70% of internal AI budget to open source.
"AT&T 也是美国最大的通信运营商之一。他就直接公开宣布说公司内部 AI 的这个预算 60% 到 70% 会放到开源模型上。" — Zhang Lu [00:10:50]
Eli Lilly. Pharma company with a multibillion-dollar Nvidia collaboration, cited as the move that triggered peer acceleration.
"礼来就和英伟达宣布了几十亿美金的一个合作" — Zhang Lu [00:34:54]
Merck. Pharma company with a large Google DeepMind collaboration. [00:34:54]
Google / Google DeepMind. Praised for consumer data quality and distribution (Chrome plus Gemini Flash), and as strategic early investor in Anthropic.
"我其实从几年前开始我就会公开讲说我对谷歌的一个模型能力的看好。" — Zhang Lu [01:06:06]
Amazon. Strategic Anthropic investor that wanted a diversified, competitive model market. [00:22:28]
SpaceX. Cited as a major Anthropic data-center compute supplier, as an IPO lockup-design precedent (seven staged milestones), and as an expensive compute seller.
"SpaceX 的锁定期的这样的一个包括七个不同的时间节点,最后大家会发现这是一个很聪明的设立。" — Zhang Lu [00:09:21]
CommonSpirit Health. Largest US nonprofit health system; its CTO told Zhang in May that thousands of hospitals and clinics will begin deploying AI agents.
"他们的 CTO 就跟我讲说他们上千家的这种医院啊诊所啊可能就要开始部署一些 AI 智能体了。" — Zhang Lu [00:46:40]
Novo Nordisk. Pharma partner of a Fusion Fund portfolio company building cell-therapy digital twins, targeting Parkinson's. [00:58:12]
Nebius, Together AI, CoreWeave. Compute providers, mentioned as having sold out capacity for the year.
"他们的算力基本上今年都已经被没了,都是要等明年。" — Zhang Lu [01:00:22]
Fusion Fund. Zhang's B2B-only venture fund, running a CXO network of nearly 50 Fortune 500 CTOs/Chief AI Officers since 2018. [00:33:07]
Meta. Building a personal agent with a built-in distribution advantage. [01:07:29]
Oura (Oura Ring). Smart ring company that pulled its IPO within ~24 hours, rare in Nasdaq history, attributed to Anthropic's IPO shadow. [00:02:49]
Cursor. Cited by Manchi as the 2025 start of coding-agent adoption. [00:42:03]
Character.AI. Hyped consumer AI app that later went bankrupt, used as caution on consumer AI durability. [01:07:29]
FTX / Alameda. Early Anthropic investor via SBF; stake was auctioned off in bankruptcy. [00:21:05]
Instinct. Personal-agent startup mentioned by Manchi, described by Zhang as still widely used. [01:07:20]
Cowork (Anthropic). Anthropic's enterprise product, which Zhang's firm uses internally, flagged as its path to a workflow "中台" via MCP.
"所有的工作流程它的一个入口都是从 Cloud Cowork 去走的话,我觉得这个对于他未来在企业平台层面上的一个强化...都是非常重要的。" — Zhang Lu [00:30:08]
Salesforce. Founder Marc Benioff sits on the AI for Good Commission. [00:11:06]
HP. Former employer (global CTO) of Zhang's partner who co-built the CXO network. [00:33:07]
ITU / United Nations. Co-organizers of the AI for Good Commission. [00:11:06]
4. People Identified
Zhang Lu (张璐). Founding Managing Partner of Fusion Fund, 11 years investing exclusively in B2B tech. Source of the episode's enterprise-demand read; sits on the AI for Good Commission and the CommonSpirit Health Foundation board.
"我们自己今年这一年真的是非常的累。但是一边累一边又很高兴,因为觉得是市场到了一个很成熟的阶段。" — Zhang Lu [00:41:34]
Yasmin Razavi. Spark Capital partner and the only VC on Anthropic's board; per Manchi, Dario Amodei said she strategically pushed coding as the team's focus. Backed Anthropic in May 2023 at roughly $4.5B valuation.
"Dario 自己他说过就是 Yasmin 是有在战略上去推动把 coding 这个市场作为整个团队的重点的" — Manchi [00:19:40]
Dario Amodei. Anthropic CEO, referenced for crediting Razavi's coding push. [00:19:40]
Jensen Huang (黄仁勋). Nvidia CEO, member of the AI for Good Commission; issued an open-source call to action shortly after the first meeting in July.
"那 7 月份我们第一次会议之后的话呢,你看很快英伟达包括黄仁勋就发布了倡议的这封信,然后去号召大家更多的去支持开源生态。" — Zhang Lu [00:11:31]
Marc Benioff. Salesforce founder on the AI for Good Commission. [00:11:06]
Chamath Palihapitiya. Social Capital investor, reported to have advised Anthropic to price its IPO lower for healthy market reception.
"他其实就建议说公司要不要把上市的价格定得再更低一点。" — Zhang Lu [00:08:12]
Kai-Fu Lee (开复). Source of Zhang's AGI definition: AI doing 90% of human work better than 90% of people. [00:24:12]
Sam Bankman-Fried (SBF). Early Anthropic investor via the effective-altruism circle; now serving a prison sentence. Zhang notes his Anthropic stake alone could have covered his losses had it been held.
"如果说当时没有出现暴雷,SBF 他在 Anthropic 的投资现在是完全可以覆盖掉他当时需要赔掉的那些资金的。" — Zhang Lu [00:21:38]
Rob. The other investor in the paired episode, mentioned in Manchi's wrap-up as having long-term AI compute positions and a neutral read on demand pacing. [01:10:38]
Zhang Lu's Fusion CXO co-founder (former HP global CTO, past presidential advisor). Co-built the network of ~50 Fortune 500 CTOs. [00:33:07]
5. Operating Insights
Start AI in One Team, Expand by Use Case, Then Graduate from Personal Tool to Agent Platform
Zhang's observed Fortune 500 rollout pattern is a replicable playbook: land in one department with a specific use case, expand across teams, and move from individual tools to an enterprise agent platform. Coding comes first, then maintenance, testing, and security, then product, data, sales, and support.
"他先在一个部门一个团队一个特定的业务场景先使用,那然后呢就会从事点的这个业务场景或者团队向全公司扩展,然后呢从一个个人的工具向企业级的这个智能体的平台的扩展。" — Zhang Lu [00:35:46]
Get Visibility Before Governance: Know Which AI Tools Touch Which Data
Before safety or compliance programs, leaders need an inventory.
"你作为一个公司的 CEO,你知不知道你的整个公司人员在用多少 AI 工具?那这些 AI 工具都触及到了哪些核心数据?...如果这个都没有的话,更不要去谈什么 AI 安全了。" — Zhang Lu [00:38:01]
Treat Token Spend Like Cloud Spend: Budget, Cap, and Route to Open Source
The shift from "use more tokens" to "limit tokens" means operators should set per-team consumption controls and match workloads to cheaper models. Open-source local deployment is the cost lever and a compliance lever [00:38:43].
Make Data Strategy the Core of AI Strategy
Enterprises hold the high-quality data AI labs pay for. Zhang's advice is that data cleaning and integration into a high-quality, trainable corpus is the foundation of any AI plan.
"怎么样让自己的数据我有这么多有价值的资产,但它现在还不可用。那怎么样可以把它进行整合、清洁,然后把它变成一个数据库,而且它是高质量的数据库,可以用于 AI 的训练。" — Zhang Lu [00:44:44]
Reuse the Three-Data Flywheel When Selling to Enterprises
Zhang's rationale for B2B: you get (1) high-quality training data, (2) high-quality feedback data from usage, and (3) continuously generated high-quality data from the deployed product [00:27:17]. Structured workflows give consistent, high-signal feedback that consumer products rarely match [00:26:13].
6. Overlooked Insights
Subscription Revenue Grew Faster Than API Revenue, but the Pricing Policy Distorts the Signal
Buried in Manchi's data aside: H1 2026 subscription revenue grew roughly 1,690% year on year versus 1,200% for usage-based. It may signal non-programmer knowledge-worker adoption, but Manchi flags that pre-policy subscription arbitrage inflated it, and that Anthropic now forces enterprises above 150 seats onto metered billing. This means the single best "coding-plus" indicator is currently confounded, and the real non-coding mix remains undisclosed.
"订阅的收入 26 年上半年同比增长了 1,690%,接近 1,700% 了...如果一个企业客户的规模他超过 150 人的话是不能走订阅,只能走按量的付费的。" — Manchi [00:31:03]
Private, Family-Owned Giants Are Already Deploying Billions into AI, Robotics, and Physical AI Quietly
Everyone watches public companies, but Zhang casually notes large non-listed US enterprises have been spending billions for years with fewer governance constraints. This is a hidden demand pool for AI infrastructure, automation, and physical-AI vendors that doesn't appear in public-company earnings or analyst models.
"美国包括很多大型的企业,它其实是非上市企业...他们从几年前开始就是已经有几十亿美金的投入在 AI、在自动化、在机器人、physical AI 这个层面上部署了...他们的董事会相对来说比较小,也不需要去担心股价,那手里的现金又很多。" — Zhang Lu [00:40:41]