Brendan Foody | 20VC Summary

1. Key Themes

Infrastructure Layer Will Win Over Application Layer

Foody makes a pointed case that the next 12 months will dramatically favor infrastructure companies over application layer ones. His core argument is that the model itself is becoming the product, and software abstractions built on top of APIs are increasingly replicable—even by the models themselves.

"Building defensibility in the software layer on top of the models is going to be incredibly difficult. I think over the last two years, everyone has increasingly realized that the model is the product." 00:34:12

"We're building out an eval set that measures how effectively agents can build end-to-end SaaS applications, where 2025 was the year of how do you get a model to make a PR in a code base, and 2026 is the year of how do you get the model to clone Slack end-to-end." 00:37:23

The key exception: companies with genuine network effects (Salesforce, Slack, Carta) may be defensible, but pure software moats are eroding fast.

Token Spend Will Exceed Headcount Spend Across the Enterprise

Foody offers a forward-looking view that compute costs will surpass employee salaries as a line item for most enterprises within five years. He's not projecting from theory—Mercor is already living this reality.

"Right now, we're spending more on tokens for our internal agents than we are on employee headcount. And I think most businesses are going to look like that." 00:43:10

"I would bet that in five years, the average enterprise spends more on compute than headcount." 00:47:10

This has enormous downstream implications for how enterprises will need to think about vendor selection, model evaluation infrastructure, and ROI measurement.

"Training Agents" Is the Fastest-Growing and Most Underestimated Job Category

Rather than framing AI as a pure job destroyer, Foody argues for a specific new job category—training agents—that will absorb enormous amounts of human labor. He backs this with Mercor's own data, paying out $3M/day to workers already doing this.

"All knowledge work is converging on training agents because it is structurally more efficient to do something once. Instead of having a customer support representative that is redundantly responding to hundreds of tickets, they're going to train an agent how to do that once." 00:17:52

"We're paying out over $3 million a day in the fastest job category ever created in history. And I expect that's going to continue growing exponentially from here." 00:16:52

2. Contrarian Perspectives

The API/Model Layer Will Be Commoditized—Even Anthropic and OpenAI

While consensus is building around OpenAI and Anthropic as must-own investments, Foody quietly drops a bomb: the majority of inference in five years won't use frontier models at all.

"I think that majority of inference in five years is going to be using an open source or custom fine-tuned or distilled model, not using a frontier model." 00:50:23

He still believes one of them could be worth $10T+ due to their role as "teacher models," but the inference economics will shift dramatically toward open source and distilled models.

Forward-Deployed Motion Beats Go-to-Market as the Real Enterprise Moat

Against the common belief that enterprise sales relationships and go-to-market are the primary defensibility in AI software, Foody argues it's actually the post-sales forward-deployed motion—embedding tacit organizational knowledge into agents—that creates the real lock-in.

"If you have a great forward deployed motion where you're going deep with a customer, you're training the agents based on all of this tacit knowledge within the company so that it understands how to perform effectively, that feels incredibly differentiated and hard to recreate." 00:40:01

"The Sequoia article that services are the new software resonated a lot—that these software moats are whittling away and it's the ability to layer services on top of software to meet the customer where they're at and go the last mile that is creating stronger defensibility." 00:40:28

Data Cleanliness Is a Temporary Human Problem—Models Will Self-Clean

Conventional wisdom says data structure and hygiene is the #1 barrier to enterprise AI adoption, and a major source of human employment. Foody disagrees: as reasoning improves, models will clean the data themselves. The real bottleneck is tacit knowledge.

"The caveat is that they'll be able to clean the data themselves fairly effectively as reasoning capabilities go up. The thing that humans will need to contribute to is all of the tacit knowledge within the organization that isn't written down." 00:19:01

High Margins in AI Are a Strategic Liability, Not an Asset

Most founders chase margin. Foody inverts this, suggesting that high margins signal oxygen in the market—which is an invitation to competitors.

"Pricing is not merely a question of optimizing for the next six months. It's optimizing for a structure that wins the market over the next decade...make sure that we're not leaving oxygen in the market because high margins invite competition." 00:35:27

AI Security Is at Its Absolute Beginning, Not Its Peak

In an era where hacks are normalized, Foody suggests we are just getting started—not winding down—on the security threat landscape, specifically because attackers now use agent swarms to exhaustively scan codebases at machine speed.

"It was the attacker that used a swarm of coding agents to help get access to the system as is happening in a lot of these. I think there's going to be an enormous boom in AI security engineering tools and various forms of defense." 00:09:26

3. Companies Identified

Mercor AI-powered talent network and data provider to frontier AI labs. Described as one of the fastest-growing AI companies, now doing well over $1B in revenue, profitable since nearly inception, with $500M+ in cash.

"We've expanded our relationships with all of the Frontier Labs and added $300 million in net new ARR in the last 60 days." 00:05:48

OpenAI / Anthropic Frontier model labs. Foody believes at least one will be worth $10T+, driven by their role as teacher models and their ability to distill their own small models.

"I could definitely see one of them being a $10 trillion company, maybe even significantly higher." 00:51:10

Surge (by Scale AI) A competitor to Mercor that Foody specifically calls out for excellence—particularly for staying close to research and hiring top-tier researchers.

"I admire that Edwin from Surge has done a really good job in staying super close to research...that's probably one of the largest things that differentiates both us and Surge is our ability to train models, to hire some of the best researchers in the world." 00:09:57

NVIDIA Chip manufacturer. Foody views NVIDIA as still dominant but beginning to face a multi-chip future. Even at reduced market share, it could remain the world's most valuable company.

"Even if they only have 30% or 40% market share in the largest market in the world by far, that is the world's most valuable company." 00:51:48

Salesforce / Slack / Carta Cited as rare examples of application-layer companies with genuine defensibility due to network effects.

"Salesforce has tons of companies that are building integrations on top of their platform that creates this almost marketplace and network effect around it...CART is another great example of this whole network effect." 00:38:01

4. People Identified

Edward Hu First author of the LoRA paper (low-rank adaptation), one of the most cited techniques in fine-tuning AI models. Now working with Mercor's research team.

"We've been building out an incredibly strong research team like Edward Hu, the first author on LoRA, who was previously at OpenAI, is working with us." 01:04:10

Jensen Huang (NVIDIA CEO) Called out by Foody as one of the most impressive people he's met personally.

"I really like Jensen...the jacket, his style, he's always on point. I would say Jensen is probably one of the coolest." 00:57:53

Ethan Thornton (Mock) Young founder, raised $70M at 19. Foody cites him as an example of the normalization effect—seeing someone you know personally achieve extraordinary things expands your own sense of what's possible.

"When I saw Ethan Thornton from Mock raising $70 million as a 19-year-old, I'm like, wait, Ethan is like a chill guy and a good friend. And like, maybe I could do something like that one day." 00:59:08

Victor (Benchmark) / Peter Fenton Benchmark partners who invested in Mercor's Series A at $250M post on $2.5M in revenue. Noted for high conviction and creative investor relationship-building (helicopter ride).

"Victor got super excited...he said, 'Oh, have you ever been in a helicopter?' And so Peter took us on the helicopter flight. And Benchmark really wanted to work with us." 00:30:03

Rob Walken, Ben Spector, Richard DeHaan (Prod community) MIT/Harvard nonprofit community members who supported Mercor in its earliest days with advice, customers, and working capital—taking no equity.

"Mercor wouldn't exist if it weren't for any of those individuals, I would say." 01:11:22

5. Operating Insights

Build an Eval System as the Operating System for AI Spend

Foody describes Mercor's internal infrastructure for managing token spend: a per-workflow eval system that continuously benchmarks models and determines optimal model routing. This is not theoretical—it's how they manage token spend that already exceeds headcount costs.

"Corresponding to each of these agents, we have an eval that tells us which model is best to use for this given use case. And what is the prior frontier of price performance for that specific use case? And that eval allows us to make the decisions around where should we be allocating our inference spend." 00:44:02

Takeaway for operators: As AI spend scales, the companies that build rigorous eval infrastructure per workflow will have dramatically better price/performance and be far less exposed to vendor lock-in.

Crisis Communication: Contain the Narrative Internally Before It Explodes Externally

When the Mercor hack hit, Foody's most effective move was an all-hands where leadership laid out exactly what happened—not a sanitized version, but the real picture—before the Twitter narrative hardened.

"We had an all hands with the company where we just laid out, here's exactly what's happening. Here's the trajectory of the business. And I think that was very helpful to the entire team." 00:07:36

Takeaway: In a crisis, your internal audience (team) is as important as your external one. Losing the internal narrative compounds the external damage.

Add Functional Values, Not Just Cultural Ones

After the security incident, Mercor added Security as a seventh company value—not as a communications exercise, but to structurally embed it into culture.

"We used to have six values as a company, but we added a seventh value as security to make sure it's very ingrained in the culture." 00:07:06

Takeaway: Values can be operationally reactive and specific—when a company-wide failure mode emerges, codifying it as a value is a structural fix, not just a PR fix.

6. Overlooked Insights

The "AI Project Manager" Eliminating 150 Human Coordinators Is Already Live—Not a Roadmap Item

In passing, Foody mentions that Mercor's AI project manager has already completed its first full project end-to-end—hiring experts, answering questions, building annotation tools, and managing quality. This is the automated elimination of roughly 150 people in their delivery org, and he mentions it almost as an aside.

"We have an AI project manager that just completed its first project, managing that entire thing end to end, where it's able to hire the experts. It's able to answer their questions. It's able to build the annotation tool using its coding tools within our platform and produce the end data type. And the experts all had a really good experience on the project, reporting to the AI project manager that was running it." 00:41:51

Why it matters: This is not a demo or a prototype. A company at $1B+ revenue has already replaced a 150-person coordination layer with a single AI system—and the human workers it managed reported a positive experience. This is the clearest real-world proof point yet that professional services coordination layers are imminently automatable, and investors in any services business with a project management middle layer should be stress-testing this assumption immediately.

The Power Law of Data Quality Is the Hidden Pricing Engine in AI Training

Buried in a discussion about data sourcing, Foody reveals something most people miss: within any given dataset, the top 20% of tasks drive the majority of model improvement value. This means vendors who can reliably deliver that top tier command disproportionate pricing power—not just for volume, but for quality differentiation.

"There's oftentimes this very power law nature of data that drives model improvement in that out of a data set of 10,000 tasks, the top 2,000 tasks will create majority of the value. And so it allows vendors that are extremely high quality to be super differentiated insofar as pricing power, because quality is the X factor that becomes dramatically more valuable than any other dimension." 00:25:36

Why it matters: This is a non-obvious structural advantage that most observers of the AI data market miss entirely. The market is not commoditizing on quality at the top end—it's bifurcating. Vendors who can identify and consistently produce top-tier training data have near-monopoly pricing power for that slice. This is also a strong investment signal: don't evaluate AI data companies on volume metrics alone; quality yield rate is the real moat.