Mercor CEO on Why Application Layer Companies Have No Defensibility & Token Spend Exceeds Salaries

Challenges in Building and Scaling AI Companies

• Building defensibility in the software layer on top of models is going to be incredibly difficult, and companies are facing high demand that can double overnight, but they often lack the capacity to handle it 10s.
• The model is considered the product, and services are getting automated, with companies like Mccor, valued at over $10 billion and doing over $1 billion in revenue, spending more on tokens for internal agents than on employee headcount 42s.
• The cost of hiring a high-quality AI researcher can be in the tens of millions of dollars per year, and companies like Mccor are expanding their relationships with Frontier Labs and adding hundreds of millions in net new ARR 2m6s.
• Mccor experienced a security incident, but the company handled it quickly, engaged security consulting firms, and has been performing well since, with $300 million in net new ARR added in the last 60 days 2m6s.

Leadership and Company Culture

• The CEO of Mccor, Brandon Foody, stays calm during stressful moments, such as the security incident, by having a thorough understanding of the situation, strong relationships with customers, and confidence in the company's ability to get through challenges 2m6s.
• Mccor has added a seventh value to its company culture, which is security, to ensure it is ingrained in the culture and to prevent similar incidents from happening in the future 2m6s.

Managing Public Perception and Misinformation

• Founders need to pay attention to what is being said about their company, as it can be helpful to the entire team to know what is happening and to address any misinformation, but it can also be annoying to deal with people saying things that are not true, and companies may not be able to speak out against them explicitly due to potential backlash, such as the Twitter mob, 10s.
• There are often people with economic incentives behind the scenes who will spread negative information about a company because they are aligned with a competitor, and this can be seen on Twitter where people with vested interests may make false claims, such as the claim that a company's data was being accessed by China, 2m6s.

Cybersecurity and AI Threats

• The increasing use of AI has led to a rise in cyber threats, with hackers using swarms of coding agents to access systems, and this has become a major concern for companies, which are now focusing on improving their cyber defensive capabilities, 4m37s.
• Swarms of coding agents are dangerous because they can review an entire codebase quickly and exhaustively, allowing attackers to move much faster than human attackers, and companies are now exploring ways to strengthen their cyber defensive capabilities to hedge against this type of attack, 6m15s.

Partnerships and Relationships with AI Labs

• The company has not lost its relationship with OpenAI, which is stronger than ever, but its relationship with Meta is currently paused, although all other Frontier Labs have grown their relationship with the company, 10m30s.
• The pause in the relationship with Meta is not solely due to security concerns, but there are other factors at play, 11m30s.
• The conversation starts with a discussion about Meta being a unique customer due to its scale and acquisition, and how it is working with other companies, but the specifics of the customer relationship are not disclosed 10s.

Rumors and Acquisition Speculation

• There is a mention of a rumor that the company was trying to poach Micro One team members with signing packages in the millions, but it is clarified that no such offers were made, and the rumor likely originated from an employee sending an outbound message to Micro One employees about potential hiring opportunities 2m6s.
• The topic of acquisition rumors is brought up, including a false rumor that Amazon tried to acquire the company for $13 billion, and it is stated that the company would not sell even for $30 billion, as the goal is to build a legendary company and create a new category of work 4m42s.

AI and Job Displacement

• The conversation touches on the topic of job displacement due to AI and the importance of understanding what jobs AI can and cannot do, with the company working on initiatives such as the AI productivity index or Apex to measure the tasks that can be automated 8m15s.
• The rapid progress of AI is acknowledged, with examples given of how the capabilities of coding agents and the Frontier model have increased significantly over a short period, and it is expected that this progress will continue 12m10s.
• The economy has increased productivity by 25x over the last 250 years, equivalent to automating about 96% of someone's job, and despite fears of job displacement due to technological revolutions, there are more jobs than ever before, with new problems to solve such as climate change and curing cancer 10s.
• The speed of transition to new technologies is a concern, as it took multi-decade cycles to implement and train new technologies in the past, but now technologies like Nano Banana Pro can replace certain jobs almost overnight 42s.

Emerging Job Categories in the AI Era

• The economy is becoming more effective at creating new job categories and allocating new labor, with examples such as paying out over $3 million a day in new job categories, which is expected to continue growing exponentially, and creating new roles such as training agents for deployed engineering and building data centers 2m6s.
• New job categories are being created across various fields, including AI, such as training agents, building solutions to climate change, and working on rockets to explore space, with internal projections suggesting that payouts could quadruple in the future 4m30s.
• One of the largest new job categories in the next 5 years will be training agents, as all knowledge work is converging on training agents, which is structurally more efficient than having humans perform repetitive tasks, and will lead to a paradigm shift as agents enter the workforce 6m15s.

Barriers to Enterprise AI Adoption

• Data structures and data cleanliness are significant barriers to enterprise adoption of new technologies, with some experts predicting that data cleaner will be one of the most important jobs of the next 5 years, although it is not the only barrier to adoption 8m40s.
• Models need access to data to perform their jobs effectively, but they will be able to clean the data themselves as their reasoning capabilities improve, and humans will need to contribute the tacet knowledge within an organization that isn't written down 10s.
• The new job of employees will be to codify all the knowledge that lives in people's heads and train agents so that they can perform tasks effectively across every function in the organization 42s.
• As reasoning capabilities improve, models will be able to structure data more efficiently, such as reading through messages in Slack and structuring a table of customer conversations, but humans will do tasks that models can't, such as knowledge tasks 2m6s.

Data Markets and Talent Networks

• The market for data providers is large and is seeing an unbundling into verticals, with companies specializing in specific areas such as medical real-world data, but there is also value in aggregation and economies of scale 2m6s.
• Companies like Meror are finding that having a large talent network and tooling that can be applied across different verticals is more efficient than working with many niche vendors, and this is how most labs are scaling out their data 4m30s.
• There will likely be a period of consolidation in the market as it corrects and companies that are profitable and have cash on hand, like Meror, will be well-positioned to acquire other companies and consolidate market share 6m40s.

Financial Performance and Growth

• Meror has been profitable since shortly after its seed round, having only burnt half a million dollars, and currently has over 500 million in cash, making it a significant asset for the company 8m20s.
• The business has been profitable since its early days and has more cash than it has ever raised, with the company trying to redeploy capital as fast as possible to invest in growth, but the business has grown so quickly that it has not been able to redeploy capital commensurately 10s.
• The revenue is dramatically higher than what has been posted publicly, with a gross margin of between 30 and 40%, and the company generates revenue by delivering tasks to customers, which includes finding and hiring experts, building a platform for them to work on, and managing the experts to automate coordination and produce high-quality data 2m6s.

Business Model and Service Evolution

• The company's business model is powered by a talent network, similar to how Uber is powered by a driver network, but the end product is different, with the company providing end-to-end services, including automated quality checks, to deliver tasks to customers 2m6s.
• The company has evolved from providing raw data to large models to providing fully integrated, end-to-end services, with a focus on quality and vertical integration, allowing it to inform the upstream signal with downstream data and create a power law nature of data that drives model improvement 4m30s.
• High-value tasks are those that correspond closely to economic value, such as software engineering, finance, medicine, law, and consulting, with a focus on long-horizon tasks that require high-quality vendors to be differentiated in terms of pricing power 6m40s.
• The company is moving away from simple tasks, such as preparing financial models, and towards more complex tasks that require multiple stakeholders and deliverables, such as preparing entire slide decks with financial models and analysis, which will be the capabilities that people will use in models in 6 to 12 months 8m50s.

Model Capabilities and Talent Mobilization

• The current state of model capabilities is such that there are no significant gaps in terms of data availability, as experts can be mobilized to access almost any domain, with the main challenge being finding people who are both experts in their field and power users of AI tools like chat GBT or Claude to help models learn from mistakes 10s.
• From a lab perspective, the goal is to automate everything that can be done in a workspace like Google Workspace, which requires covering the full distribution of context, including messages, tasks, and outputs, and mobilizing hundreds of thousands of people to build out this distribution across various job categories and domains 2m6s.

Revenue and Fundraising History

• The revenue model is such that 30 to 40% of revenues come from a specific source, and the company has raised funds through several rounds, including a seed round in September 2023 with a $23 million post-money valuation, and a series A round with a $250 million post-money valuation 4m30s.
• The series A round valuation felt like a bargain at the time, given the projected revenue growth, with the founder projecting $50 million in revenue run rate by the end of the year and $500 million by the end of the next year 6m40s.
• The company's fundraising process has been unconventional, with investors like Victor and Felisa's reaching out through unusual means, such as a helicopter flight and an invitation to go racing Ferraris, rather than traditional investor meetings 8m50s.
• The company experienced significant growth, with a revenue run rate of $20 million, and received a term sheet at a $2 billion valuation, which was considered high at the time but turned out to be an incredible investment 10s.
• The company continued to grow 50% month over month for over six months, and by September or October 2025, the revenue run rate had increased to $400 million, leading to another investment at a $10 billion valuation 2m6s.
• The series B investment was considered the most uncomfortable in terms of valuation, as it was priced at 100 times revenue, but it has since proven to be a great investment 4m42s.
• The company is currently considering its next round of investment, with offers at significantly higher valuations, but is taking its time to find the right partner due to its profitable state 6m15s.

Future of AI Infrastructure and Application Layers

• The next 12 months are expected to be better for infrastructure companies upstream of anthropic and open AI, such as those building modes and compounding network effects, rather than application layer companies downstream of them 8m10s.
• Application layer companies are believed to have limited defensibility due to their proximity to foundation model companies, making it difficult for them to build sustainable and profitable businesses, whereas infrastructure companies have more opportunities to build meaningful modes and achieve high margins 10m0s.
• The company Nebius increased its pricing by 30% across the board, which is expected to have no impact on demand, and this scenario is likely applicable to other companies as well, such as the one being discussed, which has the demand to double overnight but lacks the capacity to do so 10s.
• The discussion touches on the idea of pricing elasticity tests and the potential to double capacity or increase prices by 30% without significantly impacting demand, but also emphasizes the importance of optimizing pricing for long-term market structure and sustainability 42s.
• There is a mention of investment in application layer companies, such as Lorra, and the competitive landscape, including the potential threat from companies like Anthropic, which could potentially build out separate product teams to compete with specialized companies 2m6s.
• The concept of defensibility in application layer companies is debated, with one perspective arguing that deep, specialized products can provide defensibility, while another perspective suggests that models can be trained to outperform traditional solutions and that software layers can be recreated quickly, potentially disrupting companies that rely on them 2m6s.

Network Effects and Competitive Advantage

• The idea that software-as-a-service (SaaS) may become less relevant in the future is discussed, with the possibility that large companies may build their own customized solutions using models and agents, but it is also noted that companies with strong network effects, such as Salesforce and Slack, may maintain a significant moat 4m30s.
• The importance of network effects in creating a moat for companies like Salesforce, Slack, and Cart is highlighted, as these effects can make it more difficult for competitors to enter the market and can provide a sustainable advantage 6m0s.
• Companies with network effects can generate more value by iterating faster and leveraging these effects to create more value for customers, ultimately building more valuable products and increasing revenue, which is a key factor in determining whether a company will become worthless or gain dramatic value 10s.
• The absence of network effects makes it difficult for companies to maintain a defensible moat, as the software associated with their products is not unique and can be easily replicated, leading to significant struggles 10s.

Go-to-Market and Customer Retention Strategies

• The idea that the go-to-market strategy is the product is partially agreed upon, but it is argued that the forward-deployed motion, which includes post-sales efforts such as customer support and adoption, is more important than the pre-sales go-to-market strategy 42s.
• Having a great forward-deployed motion, where a company goes deep with a customer and provides training and support, can create strong defensibility and make it difficult for competitors to replicate, as seen in companies like Anthropic 1m6s.
• The ability to layer services on top of software to meet customer needs and provide a unique experience is creating stronger defensibility, and companies like Sequoia are investing in this area, as discussed in their article "Services are the new software" 2m6s.

AI-Enabled Services and Market Opportunities

• The concept of AI-enabled services is seen as a promising category, but it is essential to ensure that companies are actually leveraging AI to gain a significant competitive advantage, rather than just building services 4m10s.
• The use of AI in services can automate tasks and transform the economy, as seen in the example of Mercur, where an AI project manager completed its first project, managing the entire process end-to-end and providing a unique experience for experts 6m0s.
• Tokens play a crucial role in powering the agents used in these services, and companies are investing in this area to drive innovation and growth 8m0s.

Token Economics and Model Layer Commoditization

• The trend of increasing token consumption is expected to continue significantly before leveling off, with token costs rising for everyone, and companies may spend more on tokens for internal agents than on employee headcount, as is the case with the company, where token spend exceeds salaries 10s.
• The company manages its token spend by having AI project managers and various agents, such as interview question agents and accounting automation agents, with corresponding evaluations to determine the best model to use for each use case and the price-performance frontier 2m6s.
• The evaluations allow the company to make decisions on allocating inference spend and choosing providers, and it is believed that this system will develop across every Fortune 500 company, enabling perfect competition at the model layer with zero switching costs 2m6s.
• The commoditization of the model layer is expected to occur, with enterprise clients able to efficiently package workflows, but it is noted that the API layer will get commoditized, allowing companies to build stickiness and workflows on top of APIs 4m30s.
• The reason for the API layer becoming commoditized is that the switching costs are zero, making it easy to compare and swap models, with new frontier models emerging every two months, and decisions being made based on evaluations and scores 6m40s.

Evaluation Systems and Model Optimization

• The company has been growing quickly with the enterprise, helping them to populate the system of record and build evaluations for each use case, and it is expected that this trend will continue as companies develop their own systems of record for evaluating and specifying agent behavior 4m30s.
• The average enterprise is expected to spend more on compute than headcount in 5 years, as models become increasingly capable and provide enormous ROI, with the cost of inference and compute exceeding human intelligence at an exponential rate 42s.
• Having an evaluation for a specific workflow, such as code generation, can be a 10x lever on the price performance of a model, allowing companies like Salesforce to distill models and use open-source models that perform as well or better at a lower cost 2m6s.

Model Evaluation and Benchmarking Challenges

• The current academic benchmarks for evaluating models are not practical and are disconnected from the outcomes that enterprises care about, leading to a shift in focus towards building models that can perform end-to-end workflows and coordinate with multiple colleagues 4m6s.
• The majority of startups, especially on the West Coast, use frontier models to push their capabilities and then use open-source models to achieve similar results at a lower cost, which poses a challenge to companies like OpenAI and Anthropic 6m6s.
• Despite the potential challenge from open-source models, OpenAI and Anthropic are considered incredible investments, with a predicted increase in demand for their services in the next 5 years, although the majority of inference is expected to be done using open-source or custom fine-tuned models 8m6s.

Future Valuations and Market Dynamics

• The possibility of one company reaching a valuation of $10 trillion in five years is considered, with the opportunity associated with being a frontier model being so large that it could eat up much of the other demand within the economy, allowing for the creation of smaller models 10s.
• The idea of spending more on compute than on salaries is discussed, with Nvidia being a phenomenal business that will continue to execute well, but its potential monopoly in the market may be threatened by a move towards a multi-chip future and companies building in-house chips 2m6s.
• The concentration of value towards the top companies, with 84% of the year-to-date rally driven by the top 10 names, is a concern, but it is also seen as a natural dynamic associated with capital allocation, where giving compute to companies with marginal demand can create more value 4m30s.

Taxation and Economic Policy Considerations

• The issue of increasing inequality is raised, with the suggestion that eliminating income tax for the bottom half of Americans could help, as the current system disincentivizes jobs, and instead, taxes could be focused on things that don't negatively impact incentives in the economy, such as capital gains 6m40s.
• The idea of positive externalities in the economy is discussed, with jobs being the largest positive externality, yet the current system of income tax and payroll tax disincentivizes job creation, and a new paradigm of taxation is proposed to address this issue 8m10s.
• Taxing capital gains, especially short-term capital gains, is considered to be structurally better than taxing income, but it may disincentivize investors from taking risks, as they are already investing in innovation and taking high risks in an aggregated way in a portfolio 10s.
• Increasing the tax on capital gains could lead to investors moving to other geographies, resulting in a loss of tax revenue, and a sensitivity analysis is needed to determine the impact of increased taxation on people's decision to leave and reduce overall government revenue 2m6s.
• An alternative approach to taxation is to tax consumption of items that have negative externalities, such as carbon, rather than taxing the bottom half of Americans, as this could be a more effective way to generate revenue without damaging incentives or causing people to flee the country 4m30s.

Taxation of Financial Instruments and Prediction Markets

• Taxing prediction marketplaces, which can be considered a form of gambling, may be a viable option, as they do provide some value in allowing people to make effective predictions and hedge investments, but it is likely okay to tax them 6m40s.
• The discussion also touches on the idea that taxing the bottom 50% of Americans is not an effective way to generate revenue, as it only accounts for 3% of government revenue, and alternative methods of taxation could be more beneficial 5m50s.

Influential Figures and Personal Reflections

• The conversation also mentions various individuals, including Jeff Bezos, Jensen, Satia, Dario, and Sam, with Jensen being described as one of the coolest people met, known for his style and being "always on point" 8m30s.
• Meeting people who have accomplished extraordinary things can be inspiring and make one feel like they can achieve similar things, as they often realize that these individuals are normal people who have worked hard to accomplish their goals 10s.
• The dispersion effect of seeing friends achieve great things can be motivating, but it's also noted that this effect may be lacking in Europe, which has struggled to compete with the US in providing leading models to the world 2m6s.

Global AI Competition and Talent Migration

• Europe's inability to compete with the US in the model race is attributed to strong network effects around talent, with many European researchers choosing to work at top US labs like OpenAI and DeepMind, which then compounds to give these labs more capital, compute, and impact 2m6s.
• It's suggested that Europe may need to accept that it has lost the model race and instead focus on other areas where it can be dominant, such as energy provision, and consider having post-training capabilities to add value to foundation models 4m30s.
• The idea of needing sovereign models to keep data from being sent to the US or China is discussed, but it's noted that while there may be some value in localization, labs like OpenAI can simply hire people in other countries to teach their models to be better at local laws and regulations 6m20s.

Talent and Compensation in the AI Industry

• The high levels of compensation offered by companies like Anthropic and OpenAI make it difficult for others to compete, with the amount of money being spent on talent and model development being astonishing 8m40s.
• The current market for founding companies is extremely hot, with some employees having founded companies worth over $100 million, making it difficult for other companies to compete, especially when offering salaries and benefits, 10s.
• The challenge of building a company is not fully understood by many people, and the probability of success is low, with the company having been fortunate and lucky along the way, 42s.
• The competition for top talent is fierce, with some companies offering $20 million in cash per year, making it hard for other companies to compete, and this trend is expected to continue, but may balance out as more people gain knowledge and skills, 2m6s.
• The hardest role to hire for currently is researchers, due to a market with 10 times more demand than supply, and the cost of hiring a high-quality AI researcher can be in the tens of millions of stock per year, 2m6s.

Operational and Organizational Challenges

• Running a company has become easier with the addition of supporting functions like finance, legal, and HR, which brings stability and allows for focus on building great products, research, and customer relationships, 4m0s.
• The role of HR is seen as important for scaling culture, but can also be a pain and slow down the company, with the caveat that it is necessary for companies that are growing rapidly, 6m0s.
• To maintain a high-performing team, it is essential to keep a high talent bar, ensure people are committed to the company's mission, and have managers communicate effectively with their teams, which can be challenging with a young team and first-time managers 10s.

Work Culture and Company Values

• The company does not mandate specific working hours, and while the leadership team works extremely hard, they also prioritize having a sustainable work environment that allows employees to spend time with their families 42s.
• Building a legendary company requires immense dedication to the mission, but it is also important to ensure that the work environment is sustainable for the best people in the world to do their life's work 2m6s.

Future Plans and Public Market Readiness

• The company plans to go public in the next few years, as all legendary companies eventually do, but they are not rushing to do so and want to properly actualize their enterprise side before going public 42s.
• The team is grateful for their progress and feels extremely fortunate to have such a talented team, and they often reflect on how things could have gone differently 42s.

Reflections on Success and Gratitude

• In the last 12 months, there has been a change in perspective on the foundation model labs, with increased conviction that they will be the most valuable companies in the world due to their revenue ramp 10s.
• The company admires Jeff Bezos for his discipline in maintaining Amazon's culture and values, and they would like to have him as an investor 42s.
• The competitor that is most respected is Edwin from Serge, who has done a good job of staying close to research, which is also a priority for the company 42s.
• The ability to train models and hire top researchers is a key differentiator for certain companies, with about half of data providers being transactional talent marketplaces 10s.
• The CEO is currently looking to hire a strong head of people to handle HR-related tasks that are being escalated to them, indicating a need for more support in this area 42s.
• The CEO recalls a kindness done for them by the prod community, specifically individuals like Rob Walkin, Ben Spectre, and Richard Dhan, who provided valuable advice, funding, and support without taking any equity in the company, and notes that their company, Merur, would not exist without their help 2m6s.