Mercor以200亿美元估值融资7.5亿美元|CEO Brendan Foody
- Mercor专家的报酬在24个月内,从每年100万美元升至每天远超500万美元。 Foody表示,累计支付金额接近10亿美元;但Mercor交付的不是撮合劳动力、收取服务费,而是一整套RL环境。专家只是其中“一项要素、一道护城河”(one ingredient, one moat),产品还包括应用复刻、模拟工作内容和带评分的任务。
- Mercor的市场能否长久,取决于人类是否仍具备模型不具备的能力。 Foody称,专家借助副驾驶工具工作1小时,带来的模型进步可能达到原来的3至5倍,生产率提升反而会通过“杰文斯悖论”扩大需求。他的判断仍有前提:模型需要人类来衡量自身能力的边界。
- DeepTune补上了Mercor的主要产品短板,为专家提供他们需要填充的应用复刻。 一家客户主动促成双方合作,因为Mercor在这些应用周边的“世界”和“任务”上更强。此次收购将两种互补能力合到一起;Foody称这是“1+1=3”的组合(a one plus one equals three kind of thing)。
- Foody认为,品牌、差异化训练数据和定制模型,比软件或转售GPU更能构成护城河。 Harvey展示了垂直领域模型的打造方式,Fireworks则体现了推理环节的定制能力。他说:“转售GPU的切换成本为零。”Mercor每天根据评测结果切换API,这也解释了他为何担心投资者有时高估API收入的质量。
- 企业推理账单是Foody判断新一轮评测和后训练需求的首要指标。 推理支出达到1000万至1亿美元时,他认为约10%应投向评测;超过1亿美元后,后训练才开始足以支撑可观的固定成本。O’Shea质疑这个区间过宽,Foody则以Mercor每天近100万美元的推理支出为例,说明企业可以很快扩大规模。
- Foody的企业市场判断,建立在企业会保护自身智能资产、不把竞争数据交给前沿实验室的前提上。 Foody称Anthropic和OpenAI“完全可能成为市值超过10万亿美元的公司”,同时预测每家《财富》500强企业都会像一家“Neo lab”。他以Goldman Sachs为例:让实验室用员工、流程和文件训练模型,就像把这些竞争优势拱手交给对手。
- Foody和Adarsh曾有套现机会,但他们表示这不是自己的动力;相比之下,他们宁愿每周“工作100小时”来解决这个问题。 谈及Scale与Meta的安排,Foody说:“我不认为我们会做那样的事。”他更看重客户和专家的体验,而非每周营收曲线;他预计Mercor未来12个月内不会上市。
1. Mercor交付完整RL环境,专家只是其中一环
- Foody的万亿美元目标,起点是改变人们参与经济活动的方式:重复的客服回复、合同修改和播客剪辑,都可以变成由人教会智能体执行的工作。人的贡献将转向设计评测、更新技能文件和训练智能体。Mercor想定义这个新类别,而不只是改进招聘。
- 交付内容分为3部分:“应用”,即Google Workspace等工具的克隆版本;“世界”,即其中的消息、幻灯片、文档和表格;以及“任务”,即提示词和评分标准。模型尝试完成任务、获得评分并据此学习。专家负责填充“世界”和“任务”,构成Foody所说Mercor交付给实验室的产品中的“一项要素、一道护城河”(one ingredient, one moat)。
- Mercor通过3个平台支撑这套交付:拥有超过500万名活跃求职者的人才网络;Studio,专家可以在其中借助AI副驾驶、自动化质量检查和智能体搭建环境;以及AI项目经理。Foody将这项工作比作人类项目经理设定KPI、评估表现并提供反馈;如今Mercor的智能体已能自主运行项目、招聘人员、判断数据质量,并对客户结果负责。
- DeepTune补上了缺失的一环:Salesforce、Adobe和Google Workspace等应用的克隆版本。一家客户主动促成双方合作,因为Mercor在“世界”和“任务”上更强。DeepTune曾获得Andreessen Horowitz领投的A轮融资;这是Mercor的第2笔收购,规模远大于第1笔。Foody看重创始人Tim以及纽约团队的文化契合度,称这是“1+1=3”的组合(a one plus one equals three kind of thing)。
2. 专家供给依靠面试扩张,可靠付款留住人才
- 推荐渠道贡献了70%的候选人;Foody称,这类候选人的质量往往最高。Mercor累计完成超过700万次面试。公司在2023年4月用GPT-4搭建的首个AI面试官,延迟约为7至8秒。他对劳动力撮合的关键洞察是“可规模化的数据采集”:招聘5名斯坦福AI博士,或许就能找到优秀人选;但要在印度偏远地区找到世界级软件工程师,可能需要面试数百人甚至1,000人。Mercor也会分析推荐信息和个人项目等额外背景。
- 目前每日支付金额远超500万美元,而24个月前还只是每年100万美元;历史累计支付金额接近10亿美元。Foody还表示,自上年9月以来,公司在“中九位数规模的基数”上增长“超过5倍”。他称这段加速期“超现实”,并把人类参与模型训练形容为“历史上增长最快的职业类别”。
- Mercor与实验室直接合作,部分源于业内另一家公司的支付失灵:Mercor曾为该公司提供超过1,000名工人,后来却收到未获报酬的人员发来的求助工单,他们甚至难以养家糊口。Foody因此希望掌控专家体验,并在付款可靠性上投入更多。底层支付由Stripe Connect支持;他说,迄今为止Mercor没有遇到重大付款问题。
3. 只要人类能力仍领先,专家需求就会延续
- Foody称,AI副驾驶可能让专家投入1小时,换来原来3至5倍的模型进步。他不认为这会压缩劳动力需求,反而是“杰文斯悖论”:生产率提升会释放更多教学需求,因为“99%的知识仍在人们脑子里”。
- 对于O’Shea提出的行业持久性问题——这一类别会不会只是昙花一现?——Foody给出的答案有条件:只要人类仍能在数字或物理层面完成模型做不到的事,需求就有上限空间。这既可能是湿实验室实验,也可能是纯数字工作。Foody把模型自我评估比作让英语学生独自给所有作文打分:要衡量超出自身能力的表现并不容易。数学有可核验的答案,大多数领域则需要外部反馈。
- Foody称,过去3年里,投资者每年都预测超智能会在未来3年内出现;Mercor在种子轮阶段也曾因市场可能消失而遭到质疑。他的反驳取决于前沿能力仍在推进;他说,实验室里许多顶尖研究人员预计人类反馈在未来几十年仍然重要。怀疑论也让竞争减少,而Mercor仍能获得资金;不过他认为,这套判断“正接近成为共识”。
4. 长期经营纪律支撑继续独立
- 高标准、“穿越者”和客户至上,是Foody的3项经营原则。AI原生应用走在研究前沿之后,而更广泛的经济体又落在AI原生应用之后;与其模仿竞争对手,不如从未来倒推规划。某位高管展示每周营收曲线时,他提出反对,更看重客户满意度、专家满意度和专家利用率。“我们不想在任何市场里屈居第二。”
- Foody谈到的挫折包括安全事件、失去一名早期团队成员,以及最初约18个月服务种子阶段客户;他说,这类公司中有80%会在12个月内倒闭。“混乱才是常态。”O’Shea提到Alfred Lin关于“减震器”的比喻;Foody认同,领导者应在形势高涨时认真审视问题,在低潮时则成为团队的支持者。
- Foody称,他和Adarsh曾有机会套现,从此不再工作,但他们宁愿每周在这个问题上“工作100小时”。谈到Scale与Meta的安排——核心团队转入Meta,其余业务则主要聚焦Meta——Foody说:“我不认为我们会做那样的事。”他仍看好并购,但认为并购会同时放大错误和正面影响。
5. AI护城河取决于品牌、差异化数据和切换成本
- O’Shea提到Foody曾说AI应用层公司会消亡;Foody纠正说:“不完全是。”Harvey的护城河由品牌,以及识别客户真正关心的问题、再围绕这些问题创建基准和训练集的能力构成——“不是拿客户数据训练”。要在市场中存活,企业必须在自己的领域具备前沿能力,因为“模型就是产品”。Mercor公开发布的法律基准APEX和APEX Agents,也让Harvey对其工作产生兴趣。
- 前沿实验室拥有规模经济:分发能力和收入带来更多算力与RL环境,进而改善后续模型。Foody还补充了品牌和递归式自我改进;但后者“不会带来垂直起飞”,只会让进展随时间略微加快。
- 对推理服务商而言,“转售GPU的切换成本为零”。Foody认为,买入GPU后若其升值50%,确实能带来暂时的毛利,却不能证明商业模式具有持久性。他特别肯定Fireworks走了更有防御性的路线:帮助客户定制模型,而这些模型切换成本高,不能轻易搬到竞争对手的基础设施上。
- Foody认为,API收入质量很容易被高估。“我们每天都会切换API”,按Mercor评测的结果选择表现最好的模型。他预计推理账单庞大的成熟企业也会这么做;收入规模大,并不能消除切换成本问题。
6. 推理支出上升,企业将成为模型构建者
- Foody预测,“每家《财富》500强企业都会像一家Neo lab”,自行整理评测和训练任务。Anthropic和OpenAI“完全可能成为市值超过10万亿美元的公司”,但他不希望它们吞并整个经济,让企业只能从这些公司获取前沿智能。Goldman Sachs说明了原因:员工、流程和文件都承载着竞争优势;让前沿实验室用这些内容训练,就像把优势交给竞争对手。
- Foody以推理支出作为判断企业采用进度的经验指标。低于1,000万美元时,企业通常还不太关心评测或优化,仍处于摸索阶段;达到1,000万至1亿美元时,他建议约10%的推理预算用于评测数据集和路由决策。O’Shea质疑:“这个区间也太宽了。”Foody承认确实很宽,但指出Mercor每天的推理支出接近100万美元;有些企业会很快跨过这个区间。
- 推理支出超过1亿美元后,Harvey和Cursor等公司就需要认真考虑后训练。Foody称,企业一年可以投入数千万美元,有时达到数亿美元甚至数十亿美元,Cursor就是例子;因为训练固定成本可以摊到推理支出上。他预计企业推理账单会显著上升,采用也会按这一顺序推进。他最期待的是传统《财富》50强企业意识到其中的机遇与威胁;他预计Mercor未来12个月内不会上市。
- 实际做法是“从评测开始”。Foody称,目前前沿训练算力的大部分已用于RL:模型反复尝试问题并获得评分标准打分,而不只是阅读文档,或通过SFT学习输入输出对。领域专家必须设计任务和评分标准;先建立可信的评测,再扩大任务分布,让模型围绕评测结果持续改进。把公司文件一股脑倒进模型,并不符合他所说的学习方式。
完整逐字稿
Adarsh and I have had opportunities where we could walk away with a lot of cash, but that's not what motivates us. We'd rather work 100 hours a week to solve a really important problem in the world. Now, we're paying out about $5 million a day. Two years ago, 24 months ago, we were paying out $1 million a year.
Do you have any chips on your shoulder?
I would say I always have varying chips on my shoulder. We have no interest in being second place in a market.
Brendan, welcome to Sorcery.
Excited to be here. Thanks for having me, Molly.
We were just talking about this, but we first went on a walk probably over a year ago.
Mm-hmm.
And how much have you grown since?
I think that was in September of last year, so we've grown by more than 5X on a mid-nine-figure base.
That's crazy.
It's been pretty wild, yeah.
So how are you going to become a trillion-dollar company?
1. The Future of Human Work
Great question. I think it starts with working backward from how humans will contribute to the economy. Right now, most work is relatively repetitive. People repetitively respond to customer support tickets, redline legal contracts, or edit a podcast in all the different formats they're used to.
In the future, it will look much more like training an agent once to automate these repetitive tasks. What we see emerging within enterprises is a shift away from all these repetitive tasks and how work has been done in the past, toward people contributing to the economy by creating evals, updating skills files, and training agents to automate all this repetitive work.
If we believe that the whole way humans contribute to the economy is transforming radically, then clearly leading and defining that industry, and building all the technology around how that work happens, creates the opportunity to build a legendary, category-defining company. That's what we've always been focused on from day one: solving the problem of how humans fit into the economy as it evolves so quickly.
So let's break that down.
Mm-hmm.
Let's break down how Mercor is operationally structured. I know you want to talk about reinforcement learning environments a bit—
Mm-hmm.
—also known as RL environments. If you break down how Mercor's products work, how does it all fit together?
2. How Mercor Builds Environments
There are broadly 3 main products that we use to produce a given data type. The first is the product we built initially: our talent platform, where we have over 5 million people actively looking for work and pay out over $4 million a day. Seventy percent of the people come from referrals, so there are a lot of network effects associated with how we can mobilize software engineers, doctors, lawyers, and others for any specific type of data we want to create in the economy.
From there, we onboard those experts to Studio, our data platform, where they have an AI copilot and all the agentic QC and automated capabilities they need to create an RL environment. That means populating all the context that mirrors what they would do in their job, all the tasks, all the verifiers, and so on.
The third platform is our AI project manager, where we automate all the manual work associated with managing the experts on these projects. Similar to how project managers would set KPIs for people, say who's doing a good job or a bad job, and give them feedback, we now have agents autonomously running these projects, hiring people, determining data quality, and ultimately being accountable for customer outcomes.
So you're doing $5 million in payouts a day?
Yeah, now well over that.
How does that work?
It uses Stripe Connect as the backend.
I love how you go right into the infrastructure answer.
The team is spectacular. It's been incredible, because one of the things we realized early on was that we were working with another player in the space that was having a lot of payout issues very publicly. We were providing talent to them, and we got flooded with support tickets. We'd hired over 1,000 people for them, often in third-world countries, and they weren't getting paid and were struggling to put food on the table for their families. We were receiving many support tickets about that.
That was one of the big catalysts for realizing that we needed a direct relationship with the labs. We needed to be accountable for the experience of the experts on our platform and make sure they get paid on time. We also needed to overinvest in building a world-class payments team that never makes mistakes and always makes sure people are paid out.
To date, we've paid out just about $1 billion historically, and we're on a pretty steep ramp. We haven't had any major issues of that sort. It's been an incredible process scaling the payments team. I've written some blogs about this, because there are a lot of interesting technical problems in there.
For some people who don't understand the business model, let's zoom out to the broader umbrella of Mercor. You're not focused on GMV; it's not a marketplace. So how does the business model fit?
The largest misconception about the business is that people think of us as a marketplace like Uber, where we just match experts with a job and take a fee. Actually, the experts are one component of the data sets—the RL environments—that we deliver to the labs.
An RL environment includes 3 parts. The first part is all of the apps, which are clones of Google Workspace and other popular applications that people would want to teach agents to use within the economy. The second is the worlds, which are all the content within those apps: the messages, slides, docs, sheets, and so on. The third part is the tasks, which are prompts and rubrics, similar to how a professor would create a rubric to grade an essay. The model can attempt the problem, get a grade on how well it did, and learn from that.
The barrier to automating the entire economy is figuring out how to cover the full distribution of RL environments that mirror everything in the economy: all the different apps people use, all the different worlds, and all the different tasks. Ultimately, these RL environments are the end data we deliver to our customers. The experts who populate the worlds and the tasks are one ingredient, one moat, and have been a foundation for how we do that effectively and scalably.
3. Why Human Data Demand Expands
Obviously, a lot of this is predicated on models. How has the advancement of agents coming into the entire AI system changed and maybe increased what you're doing?
One of the big benefits is that agents have made it much more productive for someone to spend time creating data and training an agent. As an example, say an hour of someone's time would previously get X amount of model improvement. Now an hour of their time might get 3 to 5X that, because they're able to use an AI copilot to be more productive in how they build out all these data sets.
But as we've made the experts more productive, we've seen a case of Jevons paradox, where demand has just expanded exponentially. Ultimately, 99% of knowledge still lives in people's heads. If we just make it more effective and productive to have humans teach models, that increases the amount people teach models throughout the economy. It ties back to the principle I mentioned at the beginning: it doesn't make sense for people to do repetitive work. They should train the model to do it, because that's a far more scalable and productive way to contribute to the economy.
One comment on this category in general is that it's non-durable: training models using all this knowledge work and all the things that go into it is only going to last for so long. Do you see that? Obviously, you don't see that happening, but what are your arguments against it?
I think it ultimately comes down to your superintelligence timelines, because the market for humans teaching models how to do something is bounded by the amount of things humans can do that models can't do, in either the digital or physical world. As long as there are things in the economy that a human can do—whether it's joining a Google Meet and having a persuasive conversation, running a scientific experiment in a wet lab, or repairing a car—we need humans to measure that frontier.
The primary way models improve is that they need some stasis point: Where are they making a mistake? How do they measure that? And how do they use that as a reward they can learn from? It's analogous to how humans learn. Imagine you're in an English class learning how to write better essays, and the teacher just says you need to grade all your own essays and no one else can give you feedback.
You just have to grade your own essays, right? It’s pretty difficult to measure what’s beyond the frontier of your own capabilities, and that’s why we have feedback. There might be some domains where you can learn based on a clear signal. Math is a good example because there are calculators where you can check your work.
But in the vast majority of domains in the economy, the way we learn and improve is by getting feedback from our teachers, professors, managers, colleagues, et cetera. That’s very similar to how models learn and how most of the top researchers at these labs expect they’ll continue to learn for decades to come.
This episode is brought to you by Brex, my favorite. You become what you spend on, and I refuse to spend my time on work that shouldn't exist. Expense reports, receipt chasing, and manual closes, the companies building what's next from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex. Brex is the intelligent finance platform that combines cards, expenses, and banking into a single stack with agentic finance built in. AI agents that handle expenses automatically enforce policy before spend happens and close your books in minutes. That's why Sorcery runs on Brex, so I can spend time on building and not busy work. It's time to get Brex AF. Learn more at brex.com/sorcery. That's B-R-E-X.com/ S-O-U-R-C-E-R-Y. Bye. From handshake to signature faster, Harvey, the AI platform built for legal work, is trusted by eighty percent of the AM Law one hundred and over twenty percent of the Fortune five hundred. Learn more today at harvey.ai. That's harvey.ai How good is AI at real professional work? Mercour's AI Productivity Index, or APEX, tracks it with every major frontier model release. Built with partners like Harvey and Cognition, see where the frontier actually stands at mercour.com/apex.
4. Finding and Evaluating Experts
I have to ask: how do you find these experts, and how do you evaluate them?
70% of them come in through referrals. There are other channels we use, but that’s the one we’re most focused on because we find that referred people tend to be the highest quality.
Evaluating them is a really cool process, and this was actually one of our first large breakthroughs. Adarsh Sri and I built the first AI interviewer when we were in college, in April 2023, right after ChatGPT came out. The first one had about a 7- or 8-second latency. It would make an API call to GPT-4, do some preprocessing based on the person’s résumé, make another API call, ask the question, and then evaluate the interview according to a rubric or set of criteria.
The realization we had as we scaled that up is that so many of the challenges in labor matching come down to scalable data collection. An analogy is that if you’re hiring Stanford AI PhDs, you can probably meet with 5 people and one of them is going to be pretty good. If you’re hiring software engineers from a remote place in India, you sometimes need to meet with hundreds or even 1,000 people to find the really world-class engineers. But they are there, and many of them are on our team.
As we scaled up the number of interviews we did to address these differences in talent density, that became one of the major drivers of how we got spectacular people. We also analyze all the different filters people might not be considering, like who referred them and what personal projects they do on GitHub, Figma, or elsewhere. All of these things get pulled into context and analyzed. We’ve now done over 7 million interviews, and it’s been pretty crazy to see things scale up as we’ve created the fastest-growing job category in history.
It’s crazy.
Yeah.
Literally so fast.
It’s very surreal. As a frame of reference, we’re now paying out about $5 million a day. Two years ago—24 months ago—we were paying out $1 million a year.
Oh, my God.
I think many people at the company don’t fully realize that this has never happened before, and that it’s a very special opportunity. We found this giant market before anyone else, and we’ve become the category leaders in creating this really important new category.
5. Building Mercor's Winning Culture
Yeah, that’s amazing. I have so many questions about that. One of the things I’m super curious about, just to deviate for a second, is that you look at your team and you’re all in your early 20s.
Not all of us, but a lot of our—
You’re a founding team—
Early people, yeah.
Early, early 20s, right?
Mm-hmm.
You’re college dropouts. You’ve been able to grasp and really hold on to one of the fastest-growing new categories in AI and technology in years. Meanwhile, you have very large competitors getting bought out by Meta, and others popping up every other month and somehow reaching a billion dollars in revenue.
And yet, you’ve overcome challenges like the data leak and that kind of thing, but you’re still number one in the market. What is it about your culture and founding team that allows you to continue to persevere and win?
I’d say there are probably 3 values we have that are explicit company values, but they’re really operationalized and have been super impactful. The first is high standards: having a really high hiring bar and making sure we don’t compromise on it. That’s been a huge differentiator, and the quality of the people on the team is unparalleled.
The second is a value we call “time travelers,” which is about being very long-term-oriented: living in the future and asking what the world is going to look like in 5 years, then working backward from there. That’s how we found the market before all these other talent networks. We’ve always understood the research frontier really well, and how what’s happening at the frontier is a leading indicator of what happens at AI-native companies and application-layer companies, which in turn is a leading indicator of what happens in the global economy. Being thoughtful about how things systematically flow through these adoption curves has been really important.
The third thing is customer obsession. We’re an incredibly hardworking, customer-obsessed company, and we move mountains for our customers because they’re building the most important technology on Earth. We’re extremely committed to helping them do that. Those 3 things have been the most formative as we’ve developed market leadership.
Would you say you have clean fuel, or at this point do you have a little bit of redemption fuel?
What was the first one? Clean fuel?
Clean fuel.
How do you define clean fuel?
Dylan Field explained this to me before their acquisition by—I forget, was it Adobe?
Yeah.
It got scrapped. He was just running on pure clean fuel. Nothing put a chip on his shoulder. Then that happened, and he got a major chip on his shoulder and was like, “Now I’m going to fucking win,” and do it with some gumption. Do you run on clean fuel, or do you have any chips on your shoulder?
I’d say I always have varying chips on my shoulder. We’ve always had a maniacal sense of urgency. The company is super intense. If people aren’t executing, we keep a really high bar and move super quickly. We always want to be thoughtful about what competitors are doing and how we can move faster.
But I think many companies fall too much into the trap of looking at what their competitors do, which becomes mimetic. That’s a really good way to be second place in a market, and we have no interest in being second place. We want to define important categories, and I think that requires more first-principles thinking and more motivation from being time travelers—working backward from what the world will look like in 5 to 10 years, rather than focusing on competitors.
Maybe that’s related to your question about what motivates me. I think all of my co-founders and our early team are really excited about how we build a legendary company that solves the problem of how humans fit into the economy over the next decade.
What were the biggest challenges you had to overcome?
There have been so many. I think a lot of people who haven’t actually founded companies don’t know the feeling of getting punched in the face over and over again. We’ve had so many cases where we lost our seed-stage customers early on, because the initial business for roughly the first year and a half was all seed-stage companies. 80% of them go out of business within 12 months, so it was a constant tension to deal with. Some of them became large companies, like Cognition, but that was always a little bit painful.
There have been cases like the security incident. There have been cases where we lose an early member of the team. There are so many things where you just get punched in the face, and I think founders gain an appreciation that chaos is the norm. You just have to get used to chewing glass and staring into the abyss, as Elon says, because ultimately, it’s the impact on the other side of that 5- to 10-year journey that is motivating and makes it all worth it.
How do you manage that for your team? It’s really interesting. I saw you at the Upfront Summit not too long ago, and I was interviewing Alfred Lin onstage. Something he’s talked about a couple of times, even in our interviews, is that they like to be shock absorbers for their founders and their companies.
Do you like to be a shock absorber for your employees? How do you help direct them and let them know what’s going on?
I agree with this. I think what Alfred was saying is super accurate: when everyone feels like they’re on top of the world, that’s when you need to be cracking down on all the things that are going wrong. When things aren’t going well and you’re in one of the rough patches, that’s when you want to be the cheerleader telling everyone how good a job they’re doing and helping to rally the company. That is certainly super important.
The other thing that relates to this is driving the long-term orientation. Even on our executive team, someone put up a weekly revenue graph, and I was like, “No, we don’t want the company looking at this,” right? The company should be looking at how happy our customers are, how happy our experts are, how well-utilized our experts are—all these different things that actually drive creating value in the world and how that compounds.
As I’ve learned, and been fortunate to get insights from people who’ve been at a lot of the legendary companies early on, this long-term orientation is one of the most important characteristics of the people and companies that accomplish extraordinary things.
Who are some people that you look up to?
I do think Elon is the greatest entrepreneur by results. I wish he didn’t get into politics so much, but I think—
Luckily, he’s out of it.
Luckily, he’s out of it. Yeah. I heard from someone that he said the government is the one institution that wasn’t fixable, so whatever that tells us.
One thing I admire about him, beyond the repeated track record, which is obvious, is the mission orientation: enduring so much pain because he cares about building companies that shape the trajectory of a really important market and define that category. I think he’s been exceptional.
Jeff Bezos is another person we talked about. He really embodies long-term thinking more than almost anyone. Being so principled about what matters at the very beginning of Amazon, and how those principles map onto the product strategy decades later, is a lot more challenging than it looks. I’ve been amazed by his ability to create an entrepreneurial culture.
There are also a lot of founders now who are clearly extraordinary. Sam and Dario are both spectacular. I often feel bad that they both get so much hate. There’s always this constant cycle where you’re on top of the world, or you’re America’s worst enemy, and then all of a sudden you’re back. I feel bad about that. Those are definitely some of the entrepreneurs who inspire me the most.
I want to talk about Jeff Bezos. You wrote a paper—what came first?
It wasn’t a published paper.
Okay.
The joke I made on Twitter was that I remember being in a class my freshman year of college, and I had to write about how the government could pass policies that better align incentives in the economy. The thesis of my research paper was that the largest positive externality in the economy is creating jobs, and the largest way the government collects tax revenue is by taxing jobs.
For people who haven’t studied economics, if you tax the employee, that’s functionally the same as just taxing the company. It’s a direct disincentive to job creation. Maybe that makes sense for people making a billion dollars a year. I don’t make that, but I don’t care that much if I have a bigger tax bill. But the fact that we do this for the bottom half of Americans, who make under $70K a year, is deeply irrational.
The bottom half of Americans constitute 3% of federal government tax revenue, and yet it’s a giant burden on them, both economically and in terms of all the time associated with doing this. The thesis was to eliminate taxes for the bottom half of Americans, because that will incentivize more job creation, especially at the low end.
Then I saw my hero, Jeff Bezos, posting about this on X. I retweeted him, and then he reposted me, which I was on top of the world about. It’s a weird feeling: I grew up listening to all his podcasts and admiring him so much, and then finally got some engagement. I’ve never actually met him in person.
You haven’t?
I haven’t. Working on it, yeah.
Okay. Well, we’ll manifest that one.
Yeah, we’ll have to make it happen.
Well, that's exciting. AI needs more than chips. It needs power, land, and infrastructure. Zone develops next generation data center campuses, partnering with AI companies, site developers, and technology leaders to bring compute online faster and at scale. Zone is building the foundation of the AI frontier. Visit zonefrontier.com to learn more. That's zonefrontier.com to learn more. VCX by Fundrise, the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View the portfolio at getvcx.com. That's getvcx.com. Public is a multi-asset investing platform where you can automate your investing with AI agents. Create your own custom agent with a simple prompt. Agents can monitor markets, execute trading strategies, and manage your cash across stocks, bonds, options, crypto and more. Wanna buy at market close and sell the next day's open? Just tell your agent. Automate your investing at public.com. Founders, ship faster on Deal. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deel.com/sorcery. That's D-E-E-L.com/S-O-U-R-C-E-R-Y.
6. Why Mercor Acquired DeepTune
So let’s talk about the state of the business today. You recently had some big news. You acquired DeepTune. Why did you do that?
It really ties back to the critical ingredients of the RL environments we need to automate everything in the economy. The 3 parts I mentioned are, first, the apps; second, the worlds; and third, the tasks. The things we have the deepest competency in are the worlds and the tasks, because that’s what a lot of the experts do, and we have agents that help augment that and make it efficient.
Our largest bottleneck was getting the apps—clones of Salesforce, Adobe, and Google Workspace—that we could have all the experts populate with relevant content. Naturally, that’s not what we started doing; it wasn’t as much of our comparative advantage. One of the customers we work with proactively suggested that DeepTune, their top apps partner, should give us apps that we could populate as we built out these environments.
It’s sort of ironic when a customer initiates a partnership. It’s a good signal that there’s something there. We started working with them, and I got to know Tim, the founder, better. He was spectacular, and that’s obviously half of any of these things: who are the people you want to work with for the next decade?
It was clearly a one-plus-one-equals-three kind of thing. We were setting up our New York office, they were in New York, and they were a mission-aligned, culture-aligned team. We could both solve one another’s bottlenecks in scaling to solve important problems for the world. So we took the leap of faith and made it happen.
Was this your first acquisition?
This was our second, but it was much larger than our first. We also acquired a company called Siebel, founded by Robbie and Kat, who are also spectacular. But Tim had made a giant amount of progress with DeepTune and had just raised his Series A from Andreessen Horowitz, so it was a little bit further along.
Do you have an M&A plan going forward? How do you think about this strategically?
I think M&A is really important. It depends a little bit on the type of company, but it’s at least something that’s super high-leverage. You can make mistakes at large scale, but you can also achieve a high positive impact at large scale. It’s definitely something we think about, and we’re exploring a few different categories that we think could be interesting.
Have you gotten poached?
Have I gotten poached?
Yeah.
Are people trying to acquire you?
I’m sure you’ve gotten a handful.
There have been cases of that. But it really ties to what motivates me and my co-founders. Adarsh and I have had opportunities where we could walk away with a lot of cash, never work again, and get whatever we want in life, but that’s not what motivates us. We’d rather work 100 hours a week to solve a really important problem in the world, and we have the privilege of doing that in an incredible market with a spectacular team. It’s hard to know if we’ll have that privilege again. That’s really what frames it and how we think about our priorities and what we want to accomplish.
How would that work? Who would you get acquired by? What happened with the Scale acquisition? They still have outside customers.
The Scale acquisition is a little bit funky, of course, because of the M&A stuff, or the M&A restrictions. Alex and the core team went over, and a lot of the rest of the team stayed at Scale. Scale pivoted most of its business to just focus on Meta. I don’t think we would ever do something like that.
But does Meta still use other people?
Meta still uses other vendors. I shouldn’t speak too much to—
Yeah.
—what they do, but yeah.
Well, Alex is very talented.
He’s incredibly talented. I’ve only met Alex once, but I’ve been fortunate to work with a lot of the people who spend a lot of time with him. I think he was one of the most generational entrepreneurs of the last decade. It’s been great to see what he’s been able to accomplish at Meta as well.
It’s funny because we just had Scott Wu on. We just put out this episode about being independent.
Mm-hmm. I watched that. Scott is so fun to listen to.
He’s great.
Yeah.
We cut out a lot because there was a lot of laughing going on.
Yeah, yeah, yeah.
But he’s awesome. Their team is cracked.
Totally.
I don’t think people use that term anymore, but Cognition is cracked.
They’re great.
We talked about how most of the companies in coding agents have gotten acquired or are doing something like that. I asked him, “Why are you going to be independent?” He wants to build a massive company.
I think it makes sense. For all of the founders who have accomplished something extraordinary, the motivation is never economic. It wouldn’t make sense to do it for money. It’s always an emotional pull to doing something really important in the world. When I’m thinking about which companies we want to acquire, which founders we want to acquire, and which engineers and executives we want to work with, it’s the people who have that emotional pull to do something extraordinary who tend to be the most effective and accomplish the most impressive things.
Another company we were just at was Harvey.
Mm-hmm. I love those guys too.
They’re great.
Yeah.
Their office is gorgeous.
I know. I was there the other day, and it’s very beautiful.
I don’t think I’ve seen a prettier office than that.
You’ll have to check out our office soon.
Okay.
Ours is nice. It might not be the same caliber, but yeah.
Theirs is really nice.
Totally.
We had Winston on, we had Gabe on, and we had Niko on, and we talked about the labs—
Mm-hmm.
—and the recent benchmarks that they put out. You were mentioned as one of the partners that helped with that. Can you walk through how you helped and what your role in it was?
It’s similar to the process I described. The work that we do with them starts with great experts across all the granular domains they care about, and then ultimately involves helping to curate the management processes and specific data shapes those experts build to produce a really high-quality data set. We’ve put out a lot of public legal benchmarks, such as Apex and Apex Agents. They found some of that work really interesting, and of course they have a unique lens on this problem because they understand the taxonomy of what law firms care about and how to cover the distribution that the labs aren’t effective at.
I’m really excited about the work they’re doing in post-training models. My mental model for how the AI market is going to develop is that the frontier labs are a leading indicator of what AI-native companies, like app-layer companies, are going to do, which is a leading indicator for enterprises. The app-layer companies that are going to survive need to be building the frontier model in their specific domain, like Harvey building the frontier model in law. Ultimately, the model is the product, and you need frontier capabilities to stay competitive and build a moat relative to the labs. We’re super excited to partner with them on that, and there’s a lot more coming soon.
That’s really interesting. I think you also said that AI app-layer companies are going to die.
Not exactly. What I said is that there are 3 main categories of companies with meaningful revenue scale outside of data companies. The first is frontier labs, the second is inference providers, and the third is app-layer companies like Harvey. The first would be Anthropic and OpenAI; the second would be Fireworks and Base10.
My thesis is that large portions of each of those businesses get commoditized. If everyone is buying GPUs in the inference layer, then the GPUs get commoditized. If app-layer companies are building software layers, those will get commoditized as it becomes cheaper and cheaper to build software. But the 2 durable moats I’ve seen companies build are brand and data. Harvey has a world-class brand, certainly at the top in the law category and maybe more generally across application-layer startups.
The other is data moats, which we can help with meaningfully. How do they understand the flywheel—not training on customer data, but learning the distribution of what their customers care about—and use that to inspire the benchmarks they want to build and the training data sets they need to curate to be the best in the world at those? I think those are going to be the 2 primary moats to look for in application-layer companies that can build durable and potentially legendary companies.
What are the moats for the top 2 layers?
With the frontier labs, there are certainly economies of scale associated with spending more on compute and more on data. If you have the most distribution and revenue you can generate from your models, you can invest in more compute and more RL environments, which the labs are scaling up their investments in exponentially. There are also flywheels associated with a given model helping them improve their subsequent model more—that is, recursive self-improvement. It won’t be a vertical takeoff, but it will contribute and make things slightly faster over time. I’d say it’s mainly economies of scale, along with brand moats, for foundation models.
For the inference companies, the primary thing they need to get right—and Fireworks is executing really well on—is building custom models. It’s the other side of the coin from what Harvey is thinking about, because the switching costs of reselling GPUs are zero. People try to tell a nice story about how there are great margins. Of course, when you buy a bunch of GPUs and those appreciate 50%, you’ll have good margins for a period of time. The real switching costs and durable margins come from helping a customer customize a model that has high switching costs and can’t be ported over to other infrastructure. That’s how I think about the moats developing at the frontier model labs and in the inference layer, and how those companies will build durable businesses.
Which is the most vulnerable?
I think the switching costs for APIs are so low that investors overrate some of the revenue quality of API businesses. So much of this boils down to switching costs, if you make it that simple. We switch between APIs on a daily cadence, based on whichever model is performing best on our evals for how we run the business.
I think every company will do that as they mature and have scaled inference bills. Some of these companies may be really big, but I think the revenue quality is sometimes overrated.
7. Why Data Remains Mispriced
Okay. Since we're talking macro, there was another tweet you put out. By the way, that was a tweet, so we'll tag it.
Yeah.
There was another tweet you put out that AI progress requires 3 things: compute, algorithms, and data. Why do you think data is mispriced in that stack?
I think data is still the most contrarian thing. For some reason, there has been this perception by investors that we'll get superintelligence in 3 years, and AI will be better than humans at absolutely everything, so we won't need humans to create evals and train models. Investors have been saying this for the last 3 years. I remember even at our seed round, most of the top firms in Silicon Valley thought the market would dissipate within a couple of years. We've always been very fortunate to work with investors we love, but a lot of the top firms thought that.
The thing everyone has always missed is the case of Jevons paradox, or functionally the lump of labor fallacy: as we make the experts more efficient, and as they're able to improve models 3, 5, or 10 times more, that dramatically increases the demand for people to improve the models. Ultimately, as long as we don't have superintelligence, and there are things humans can do that AI can't do in the economy, we need humans to measure that frontier.
I think that has been the largest disconnect, and I honestly feel sort of grateful for it in a lot of ways. We've always had lots of access to capital, but it being contrarian reduces the amount of competition in the market. I'm glad our market is not nearly as competitive as the foundation model business. I think it's on the verge of becoming consensus, especially as people see all of the enterprises that are starting to evaluate agents they're deploying and post-train models.
So how does this play out over the next 10 years?
I think the key development is that every Fortune 500 is going to look like a Neo lab. They'll have their core data sets and the core distribution they care about to serve their customers, and they'll need to mobilize some combination of their full-time employees and contractors to curate really high-quality tasks for evaluating and training the models they're building.
Ultimately, so many processes that are currently human are going to become agentic, and this is the most efficient and scalable way to distill all of that human knowledge throughout their company into the models. That will be a giant transformation, and we're just starting the first inning of it.
Damn, I was going to ask you that. I actually asked Winston for some questions ahead of time, and he wanted me to ask about AI data sovereignty.
It's going to be so critical. There's a reason Jensen, Alex Karp, and Satya are all talking about this. I think Anthropic and OpenAI will have extraordinary businesses. Both of them could very well be $10 trillion-plus companies. But ultimately, we don't want them to subsume the global economy and be the only place you can go to access frontier intelligence.
For value to accrue to every other company in the Fortune 500 and to Americans more generally, we need to make sure that all these companies are able to train their own models and build their own intelligence that codifies their competitive advantage. I think there's sometimes this narrative from certain labs when they're fundraising that they're just going to turn on training on the usage-data flywheel of all the enterprise data, and the Fortune 500 companies are going to be fine with this. That's the way they're going to collect all the environments throughout the economy.
But clearly that's not the case. The reason a company like Goldman Sachs exists is that it has a set of competitive advantages associated with the people, processes, documents, and so on in its business. Allowing a frontier lab to train on that data would be analogous to just sharing all of it with a competitor, so they're not going to do that. I believe that instead, they're going to build their own models and deploy their own agents that codify all this knowledge.
The heuristic for what to look out for as a catalyst for this transition happening within the enterprise is how much they're spending on inference. What we basically find is that when a company is spending under $10 million on inference, they generally don't care that much about evals or how to optimize it. They're still figuring things out. Once they move into the $10 million to $100 million range on inference, they generally want to spend—
That's such a wide gap.
Yeah, it's a wide gap, but we're spending almost a million dollars a day on inference.
Mm-hmm.
Some companies scale fast. Once they're in the $10 million to $100 million range, roughly 10% of their inference budget should be going to evals: building out the data sets that tell them which models they should route things to. Once they're past $100 million in inference, like Harvey, Cursor, and all these companies, they need to be seriously thinking about how to post-train models, because the economics start to check out.
You can spend tens of millions of dollars a year, sometimes hundreds of millions or billions in the case of Cursor, training your own models because you can amortize that fixed cost across all the inference you have. I expect enterprises' inference spending will ramp materially, but they'll also go through these curves of adoption: they roll out evals in the $10 million to $100 million range, then start post-training in the $100-million-plus range as they deploy frontier intelligence throughout their businesses.
Because you have such a unique and advantageous perspective—you see literally every company, and you see how their data is being used and all this kind of stuff—what advice do you have for enterprises as they continue to build up their sovereign AI stack, taking into account everything you just mentioned about inference?
I would say the biggest thing enterprises misinterpret is how frontier models actually learn. The vast majority of the compute spent in training is no longer just reading all the documents in the company or learning from inputs and outputs in SFT. Most of the compute is now going toward RL, where the model attempts the problem many times and then gets a score on one of these rubrics. RL, in the vast majority of cases, requires subject matter experts who can create these rubrics and tasks to evaluate and train models.
Any company training its own model needs to build tasks that produce clear eval scores. Once they feel confident in the eval, they should scale up the distribution of what they care about to help climb that eval. Those are maybe the 2 pieces of advice: first, build tasks rather than pretending the model can read everything and learn; second, start with the eval, because that's the foundation for how every AI lab and successful application layer understands how well they're doing and has a baseline for improving the models they're training.
Good advice.
Thank you. Yeah.
Okay, 2 more questions. What are you most looking forward to in the next 12 months? What can we expect from Mercor? Are you going to go public?
I don't think we'll go public in the next 12 months, but it's certainly on our radar. It's funny: in 2024, I tweeted, “IPO by end of year,” and people were like, “Whoa, no way.” I thought it was clearly a joke because we had dropped out of college just 12 months earlier.
The thing I'm most looking forward to is really the stuff on enterprise. I think it's starting to happen so quickly as inference bills mature, and the really classic companies—the Fortune 50 non-technology companies—are starting to understand how important this is.
Mm.
Some of those companies are facing existential threats from the frontier models as those companies creep into their sectors. That's going to be a huge catalyst for transitioning the way intelligence gets used and deployed in the economy. I think it'll be incredible to see.
Cool. I know this is a very strange closing question, but you asked me this before we started recording.
Mm-hmm.
You were like, “How do you spend your days?” So I want to ask you, Brendan: how do you spend your days? Are you flipping donuts?
How do I spend my days?
How do you spend your days?
No longer flipping donuts.
Let’s see. I generally wake up around 7:45, and I work out most days from around 7:50 until 8:30. Then I’m in the office by 9:00. I usually have four hours of meetings a day, sometimes closer to 8 to 10 hours. I knock out lots of messages, emails, and docs in the time between.
My time is probably roughly a third hiring, a third with customers, and a third on research and product. I’m in the office until 10:00 p.m., and I rinse and repeat six days a week.
Wow.
And then Sundays I work from home.
Nice. When I talked to Winston, so many people reached out to me after listening to that episode. They said, “Oh my God, it’s really great to hear how someone going through that hypergrowth is managing their day and what they spend their time on.” There was one part where he was talking about fixing people’s calendars—
Mm-hmm.
—and deleting meetings.
Totally.
Are you trying to constantly remove distractions? How do you manage focus?
Totally. It’s challenging because I think it’s easy for everyone to want to escalate things to me. The only way the organization scales is for it not to depend only on me; people on the ground also need to have high trust in the leaders and executives in the organization. Making sure we have a really strong leadership team is probably the most important thing for that.
I feel like I could spend all day and more responding to my messages and emails. At some point, you also have to stop responding to everything. I still get to inbox zero every day because I want to make sure I get through everything. But I have to be very thoughtful about that, and about where to leverage agents to make things more efficient.
Cool. Are you guys hiring?
We are hiring a ton. It’s ironic because everyone says, “Oh, we’re going to displace all the jobs,” but all of the AI companies, us included, are hiring as quickly as possible. We’d love to work with you. Check out Mercor.com. We’re hiring experts, as well as lots of full-time employees to join our engineering, research, and operations teams.
Amazing. Great place to end it.
Yeah.
Well, thank you so much, Brendan.
Thanks for having me.