The $10B Startup Running on AI Agents
A company spending over $10B reveals that its AI agent token costs now exceed total employee salaries. The company runs multiple AI agents for project management and interviews, using evaluations to optimize model selection and inference spend. This token-over-headcount trend is expected to accelerate significantly across the enterprise.
Summary
In this transcript excerpt, a founder or executive at a large-scale AI-native company describes a striking financial milestone: their spending on AI agent tokens has surpassed their total employee headcount costs. This represents a fundamental shift in how operational expenses are structured at the company.
The company has deployed AI agents across several critical workflows. One example is an AI project manager that handles operations, and another is an interview question agent that has conducted over 5 million interviews. These agents are not experimental — they are deeply embedded in core business functions at scale.
To manage the complexity of running multiple agents, the company has built an evaluation system tied to each agent. These evals determine which AI model performs best for a given use case, enabling data-driven decisions about where to allocate inference spend and which provider to use. This suggests a sophisticated, performance-driven approach to AI infrastructure rather than a one-size-fits-all model selection.
The speaker concludes with a forward-looking statement: token consumption within the enterprise is expected to grow significantly before any leveling off occurs. The implication is that this cost structure — where AI inference exceeds human labor costs — will become increasingly common across industries.
Key Insights
- The speaker states that their company's spending on AI agent tokens now exceeds what they spend on employee salaries, marking a historic inversion of traditional operational cost structures.
- The company has deployed an AI project manager agent that handles operations as a core workflow, not as a pilot or experiment.
- Their interview question agent has conducted over 5 million interviews, demonstrating massive real-world deployment of AI agents in a high-stakes business process.
- The company builds a dedicated evaluation (eval) for each AI agent to determine which model performs best for that specific use case, enabling optimized provider and inference spend decisions.
- The speaker argues that enterprise token consumption will continue to grow very significantly before any leveling off occurs, suggesting the current cost inversion is just the beginning of a larger trend.
Topics
Transcript
[0:00] You're spending more on tokens for agents than you are on headcount. >> Exactly. >> token spend on agents is more than salaries. >> That's correct. We have a variety of these key workflows throughout the company where we have an AI project manager that manages operations. We have our interview question agent that where we've done over 5 million interviews and asked other questions in the interviews. And 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 that allows us to make the decisions around where should we be [0:30] allocating our inference spend, what provider should we be using, etc. This…
Full transcript available for MurmurCast members
Sign Up to AccessMore from 20VC with Harry Stebbings
Tokens Over Humans?
A tech leader argues that by end of 2025, companies will prioritize allocating budget to AI tokens over hiring additional engineers. Drawing from his experience at Adobe, he predicts a fundamental shift from headcount-based budgets to flexible spending budgets where leaders choose between humans and AI compute.
"Loosing money is like s*x"
A fast-paced financial and tech discussion covering major funding rounds including Anthropic's $65B raise, a resurgence in public markets, and growing skepticism from Uber and Microsoft about AI productivity gains. The conversation touches on the shift from capex-light to capex-heavy business models and bold predictions about AI replacing human developers.
Why we should increase capital gains tax
The speaker argues that capital gains taxes, particularly on short-term gains, are preferable to income taxes because they do not significantly disincentivize investment behavior. They suggest a tax paradigm shift toward taxing things that don't negatively impact economic incentives. The transcript cuts off before the full argument is completed.
Why AI Won't Take Your Job
The transcript argues that AI will not eliminate jobs, drawing on 250 years of economic history to counter fears of technological unemployment. The speaker highlights that productivity has increased 25x without causing net job loss, attributing recurring fears to the 'lump of labor fallacy.' Historical technology revolutions are cited as evidence that innovation creates rather than destroys employment.
This startup spends $400k/month on Anthropic
A startup founder reveals they spend $400,000 per month on Anthropic while spending nothing on OpenAI. They argue Anthropic has been outexecuting OpenAI, particularly for enterprise and workflow automation, despite crediting OpenAI with pioneering the AI lab model. The transcript ends mid-sentence hinting at a critique about stated versus revealed preferences.