BREAKING: Mercor Raises… | Sourcery Newsletter Summary

1. Key Themes

Theme: Human expertise becomes the scarce input as AI improves, and training/evaluating agents becomes the new form of work

Work shifts from doing repetitive tasks to teaching agents once

"What we see emerging within enterprises is the shift away from all these repetitive tasks and how work has been done in the past, and moving towards people contributing to the economy through creating evals, updating skills files, training agents how to automate all this repetitive work."

Agents need both general and company-specific knowledge across a huge occupational surface area

"The US has more than 800 distinct occupations, each with workflows that models have yet to be evaluated against. Agents in production need 2 kinds of knowledge, how a task gets done and how a specific company does it within its own systems, policies and quality bar."

Humans are needed as long as AI has a capability gap

"As long as we don't have super intelligence, as long as there are things that humans can do that AI can't do in the economy, we need humans to measure that frontier."

Theme: Every large enterprise becomes a "Neolab," driving demand for evals and post-training

Enterprises will train their own models to protect proprietary knowledge

"Every Fortune 500 is going to look like a Neolab, where they have their core data sets, their core distribution that they care about to serve their customers, and they need to mobilize some combination of their full-time employees and contractors to curate really high-quality tasks for evaluation and training of the models that they're building."

"Allowing a frontier lab to train on a firm's internal data, using Goldman Sachs as the example, would be equivalent to sharing it with a competitor. Some Fortune 50 non-technology companies now face an existential threat as frontier labs move into their sectors, which he expects to accelerate adoption."

Eval and post-training adoption follows an inference-spend ladder

"Below $10M a year, companies pay little attention to evals. Between $10M and $100M, roughly 10% of the inference budget should go to evals that determine which models to route work to. Past $100M, as at Harvey and Cursor, post-training economics work because fixed training costs amortize across inference."

Theme: Mercor's evolution from recruiting to RL environments and data infrastructure

Not a marketplace: experts are one input to a larger dataset/environment product

"I think the largest misconception about the business is that people think about us as a marketplace like Uber, where we would just match experts with a job and then take a fee on that, when actually the experts are one component of the datasets, the RL environments that we deliver to the labs."

RL environments decompose into apps, worlds, and tasks

"An RL environment has 3 parts. Apps are clones of tools such as Google Workspace. Worlds are the content inside those apps, including messages, slides, docs and sheets. Tasks are prompts and rubrics that let a model attempt a problem and receive a grade it can learn from."

Productivity gains from agentic tooling

"Agents now make an hour of expert time produce 3 to 5X more model improvement."

Most frontier compute is now going to RL, which is where Mercor sits

"Most frontier training compute now goes to RL, where models attempt problems repeatedly and are scored against rubrics written by subject matter experts."

Theme: Hypergrowth and scale metrics

Valuation and revenue trajectory

"Revenue has grown more than 5X since September 2025."

"The new mark doubles Mercor's last priced round. Mercor raised a $350M Series C at a $10B valuation on October 27, 2025..."

Expert payouts scaling by orders of magnitude

"Now if we're paying out, call it $5 million a day, 24 months ago, we were paying out a million dollars a year."

Referral-driven network effects

"Mercor's talent platform has more than 5M people actively looking for work. 70% come through referrals, which gives the company network effects in mobilizing software engineers, doctors, lawyers and other specialists for any data type a customer needs."

Theme: App-layer survival requires owning the model and data

The model is the product

"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. Because ultimately, the model is the product, and you need to be able to have frontier capabilities to stay competitive and build a moat relative to the labs."

Two durable moats: brand and data

"Foody sees brand and data as the 2 durable moats for app layer companies."

2. Contrarian Perspectives

Data is the contrarian layer of the AI stack, and the consensus is wrong about its durability

Foody argues that investor belief in near-term superintelligence kept competition low, and he expects the market to become consensus as enterprises evaluate agents and post-train models.

"Investors have expected superintelligence within 3 years for the last 3 years, and most top Silicon Valley firms expected Mercor's market to dissipate when it raised its seed. He credits that view with keeping competition lower than in foundation models, and expects it to become consensus as enterprises evaluate agents and post-train models."

Supporting reasoning: models cannot grade themselves outside verifiable domains. "Models need outside feedback to measure their mistakes, much like a student who cannot grade their own essays, with math as an exception because answers can be checked."

Investors overrate the revenue quality of API businesses

Switching costs between model APIs are low. Mercor itself switches models daily based on internal evals.

"It feels like the switching costs for APIs are so low that I think investors overrate some of the revenue quality of API businesses."

Evidence: "Mercor switches between model APIs on a daily cadence based on its own evals." By contrast, inference providers that help customers customize models build switching costs "that cannot be ported to other infrastructure."

Rejecting the "Scale-to-Meta" structure and avoiding mimetic competition

Foody declined a Meta-style deal, and frames copying competitors as a path to second place.

"Many companies fall too much into the trap of looking at what their competitors do, which becomes mimetic. It's a really good way to be second place in a market, right? And we have no interest in being second place in a market."

Evidence: "Meta paid $14.3B for 49% of Scale AI. Alexandr Wang and the core team moved to Meta, and Scale shifted most of its business to Meta. Foody said Mercor would not pursue a similar structure." He also states recursive self-improvement will speed progress "gradually with no vertical takeoff."

3. Companies Identified

Mercor

Description: AI training data, evals, and RL environments company; evolved from AI recruiting. $750M Series D at $20B.
Why mentioned: Subject of the article/interview.
Quote: "Founded in 2023 by college dropouts, Mercor has evolved beyond AI-powered recruiting into building the data, evaluations, and reinforcement learning environments used to train AI systems on real professional work."

Deeptune

Description: Builds RL environments recreating enterprise applications (spreadsheets to Salesforce).
Why mentioned: Mercor's largest acquisition, addressing its biggest bottleneck (apps).
Quote: "Apps were Mercor's largest bottleneck. Mercor announced the Deeptune acquisition on July 9, 2026."

Sepal AI

Description: Mercor's first acquisition.
Why mentioned: Part of Mercor's M&A trajectory.
Quote: "Deeptune is Mercor's second and largest acquisition. The first was Sepal AI."

Harvey

Description: AI platform for legal work.
Why mentioned: Partner on benchmarks and post-training data; case study for app-layer companies building domain frontier models.
Quote: "Harvey's BigLaw Bench Global was designed with Mercor, and Harvey worked closely with Mercor to build and scale human expert datasets for its post-training effort."

Cognition

Description: AI coding company; early Mercor customer.
Why mentioned: Grew from a seed-stage customer into a large company; its CEO chose independence over acquisition; APEX partner.
Quote: "Foody pointed to Cognition's Scott Wu choosing independence."

Scale AI

Description: Major data labeling competitor.
Why mentioned: Competitive case study; Meta took 49%.
Quote: "Meta paid $14.3B for 49% of Scale AI... Scale shifted most of its business to Meta."

Surge AI

Description: Human data competitor.
Why mentioned: One of the 4 firms capturing most category revenue.
Quote: "more than 75% of revenue is concentrated in 4 of them, Scale AI, Surge AI, Mercor and Handshake."

Handshake

Description: Competitor expanding from expert supply into evals and RL environments.
Why mentioned: Following a similar path to Mercor.
Quote: "Handshake bought Cleanlab in January 2026 to expand from expert supply into evaluations and RL environments, the same path as Mercor and Surge."

Cleanlab

Description: Acquired by Handshake in January 2026.
Why mentioned: Evidence of competitor consolidation.
Quote: "Handshake bought Cleanlab in January 2026..."

Meta

Description: Big tech company.
Why mentioned: Paid $14.3B for 49% of Scale AI.
Quote: "Meta paid $14.3B for 49% of Scale AI."

OpenAI

Description: Frontier lab.
Why mentioned: Customer of multiple data vendors but building in-house capability.
Quote: "OpenAI, while still contracting with Surge, Mercor and Handshake, is building a significant in-house human data team."

Anthropic

Description: Frontier lab.
Why mentioned: Foody sees potential $10T+ outcome.
Quote: "Foody expects Anthropic and OpenAI could each become $10T+ companies."

Stripe (Stripe Connect)

Description: Payments infrastructure.
Why mentioned: Powers Mercor's global expert payouts.
Quote: "Mercor uses Stripe Connect to onboard and verify professionals, process global payouts and stay compliant across regions."

Wise

Description: Payments company.
Why mentioned: Fallback payout rail where Stripe is unsupported.
Quote: "In countries Stripe Connect does not support, Mercor pays through Wise."

Cursor

Description: AI coding company.
Why mentioned: Example of a company above $100M inference spend where post-training economics work.
Quote: "Past $100M, as at Harvey and Cursor, post-training economics work..."

Goldman Sachs

Description: Investment bank.
Why mentioned: Example of an enterprise that shouldn't share internal data with frontier labs.
Quote: "Allowing a frontier lab to train on a firm's internal data, using Goldman Sachs as the example, would be equivalent to sharing it with a competitor."

Investors in the round (Nvidia, Benchmark, General Catalyst, Apollo affiliates, XTX, Salesforce, Felicis, etc.)

Description: Series D participants; General Catalyst has invested in every round since the seed.
Why mentioned: Signal of institutional and strategic conviction.
Quote: "General Catalyst has invested in every round since the seed."

Andreessen Horowitz

Description: VC firm.
Why mentioned: Led Deeptune's $43M round.
Quote: "Andreessen Horowitz led Deeptune's $43M round in March, and Foody was one of the angel investors in that round."

4. People Identified

Brendan Foody

Description: Co-founder and CEO of Mercor.
Why mentioned: Interview subject; articulates the vision for Mercor.
Quote: "I think the largest misconception about the business is that people think about us as a marketplace like Uber..."

Adarsh Hiremath

Description: Co-founder; former CTO, now co-CEO.
Why mentioned: Co-leader of Mercor.
Quote: "Hiremath, formerly CTO, is now co-CEO, and Midha, formerly COO, is board chairman."

Surya Midha

Description: Co-founder; former COO, now board chairman.
Why mentioned: Founding team member.
Quote: "Mercor was founded in 2023 by Brendan Foody, Adarsh Hiremath and Surya Midha..."

Tim Lupo

Description: Founder of Deeptune.
Why mentioned: Joined Mercor in the acquisition.
Quote: "Founder Tim Lupo and the Deeptune team joined Mercor in New York."

Alfred Lin

Description: Sequoia partner.
Why mentioned: Foody's leadership influence.
Quote: "Foody agrees with Sequoia's Alfred Lin that leaders should act as shock absorbers, pressing on problems when the company is on top and rallying the team during rough patches."

Scott Wu

Description: Cognition CEO.
Why mentioned: Example of a founder choosing independence.
Quote: "Foody pointed to Cognition's Scott Wu choosing independence."

Alexandr Wang

Description: Scale AI founder.
Why mentioned: Moved to Meta with the core team in the Scale deal.
Quote: "Alexandr Wang and the core team moved to Meta..."

Jeff Bezos

Description: Amazon founder.
Why mentioned: Foody's example of long-term thinking; engaged with Foody's tax idea.
Quote: "Bezos later reposted Foody on X after Foody shared a Bezos post on the same idea. The 2 have not met in person."

Elon Musk, Sam Altman, Dario Amodei

Description: Tech leaders.
Why mentioned: Foody's named entrepreneurial influences.
Quote: "Foody named Elon Musk as the greatest entrepreneur by results and Jeff Bezos as the clearest example of long-term thinking. He also named Sam Altman and Dario Amodei."

5. Operating Insights

Start with the eval, then scale the data

Foody's advice to enterprises building AI capability is to design tasks with measurable outcomes first.

"Foody's advice to enterprises is to build tasks that produce clear eval scores, start with the eval, and then scale the data needed to climb it."

Over-invest in payments and accountability when your supply side is humans

An early payout failure at a partner taught Mercor to own the expert experience and the lab relationship directly.

"We need to be able to be accountable for the experience of the experts on our platform, and making sure that they get paid on time. And we need to over-invest in how do we have a world-class payments team that never makes mistakes, and always makes sure that people are paid out."

Manage to customer and expert outcomes, not a revenue graph

Foody changed the metrics executives focus on, and pairs this with a "shock absorber" posture when things look best.

"Foody pulled a weekly revenue graph from executive focus in favor of customer satisfaction, expert satisfaction and expert utilization."

"When everyone feels like they're on top of the world, that's when you need to be cracking down on what are all the things going wrong."

6. Overlooked Insights

Mercor's own inference spend is a leading indicator for a macro thesis

Buried in the growth discussion is a data point extrapolated into a GDP forecast.

"Mercor spends 3X as much on inference as on employee salaries. The company projects that if that ratio holds as AI spreads through the economy, it points to double-digit GDP growth over the next 5 years, nearly 3X the historical rate."

A customer-initiated partnership as an M&A signal, and several more acquisition categories are in play

Deeptune came via a customer recommendation, and further deals are signaled.

"It's sort of ironic when a customer initiates a partnership. It's a good signal that there's something there."

"Mercor is exploring several further categories for M&A."