Y Combinator CEO Garry Tan Reveals He Spends Up to $200,000 Annually on AI Agent Tokens

In the rapidly evolving landscape of artificial intelligence, few figures are as deeply immersed in the practical application of agentic workflows as Y Combinator CEO Garry Tan. A vocal proponent of the "AI-first" era, Tan has previously spoken with characteristic intensity about his relationship with these digital assistants—going so far as to jokingly describe his state of constant technological engagement as a form of "cyber psychosis." Now, as the technology matures from experimental novelty to operational backbone, Tan is offering a rare look at the financial and personal investment required to live at the bleeding edge of the industry.

In an upcoming episode of Walter Isaacson’s podcast, The Builders, set to be released this Tuesday, Tan pulls back the curtain on the fiscal reality of his heavy AI usage. When asked about the scale of his reliance on these tools, Tan disclosed a figure that underscores just how significant the compute costs can be for a power user: he spends between $100,000 and $200,000 annually on tokens alone.

This staggering expenditure is not merely for hobbyist tinkering. Tan utilizes AI agents across a broad spectrum of professional and personal tasks. His commitment to the technology is further evidenced by his decision to open-source a collection of 23 specific "skills" for Claude Code, which he calls gstack and has made available on GitHub. By sharing this repository, Tan aims to provide a blueprint for others to understand the practical capabilities of agents and how to orchestrate them to perform complex, multi-step tasks.

The utility of these agents in Tan’s life is multifaceted. Professionally, he employs them to handle the rigorous preparation required for his public engagements, including researching guests and potential talking points for podcasts like Isaacson’s. However, the scope of his AI integration extends far beyond the professional sphere; it has effectively permeated his private life.

During the conversation with Isaacson, Tan candidly revealed the depth of access his agents possess. "It’s read all of my email and all of my Dropbox," he explained. "It’s read, like, all of my Apple Notes, even my therapy notes going back 20 years." For Tan, this high level of context is what allows the agent to function not just as a tool, but as a digital extension of his own cognition.

He draws a comparison to the literary figure Jiminy Cricket, describing the ideal AI agent as a constant presence on one’s shoulder. According to Tan, the power of these systems lies in their ability to know the user intimately—understanding their history, their preferences, and their goals. This context allows the agent to act autonomously, often "running ahead" to anticipate challenges, flag potential issues, and help the user "see around corners." It is this proactive, predictive capability that distinguishes modern, agentic AI from the passive software tools of the past.

As an early adopter of advanced AI models and a self-described power user, Tan offers a piece of advice that challenges the conventional wisdom regarding software development: do not treat AI agents like traditional, deterministic software. In traditional computing, repeatability and rigid safety protocols are paramount. However, Tan argues that when users become too obsessed with these metrics, they inadvertently stifle the most valuable aspects of agentic intelligence—the nuances, the reasoning, and the human-like flexibility that allow the software to handle ambiguous, real-world problems.

This philosophy is currently being put to the test at Y Combinator itself. Under Tan’s leadership, the startup accelerator has been actively integrating agentic workflows into its core operations. One such initiative involves an internal agent dubbed "Stray Light," which Tan describes as a system with its own distinct personality and a vast, company-wide "brain."

Stray Light serves as a repository for YC’s institutional knowledge, holding the total context of every founder the firm has ever backed, the entire history of its application process, and years of correspondence. By synthesizing this massive, unstructured data set, the agent provides a level of operational support that was previously impossible to achieve with traditional databases or search tools. It is, in effect, a digital institutional memory that assists the firm’s partners in decision-making and relationship management.

Despite his enthusiasm, Tan is acutely aware that the current model of AI consumption is not yet sustainable for the general public. He characterizes his own spending as a "crazy privilege," acknowledging that the high cost of tokens acts as a significant barrier to entry for most individuals and smaller organizations. For personal AI agents to achieve true mainstream adoption, Tan believes the industry must solve the fundamental problem of affordability. The current cost structure makes this level of autonomy a luxury for the few rather than a utility for the many.

Furthermore, Tan is wary of a future where these powerful, deeply integrated agents are controlled exclusively by Big Tech corporations. The potential for these companies to harvest personal data and manipulate behavior through proprietary algorithms is a concern he takes seriously. Reflecting on the impact that social media giants—specifically Meta—have had on society, Tan expressed a clear preference for a different path forward.

"This is a huge company that controls the algorithm, and we all wish that they didn’t at some level," Tan noted, highlighting the tension between the utility of centralized platforms and the desire for individual sovereignty. He envisions a future of personal AI that is controlled and programmed by the user, rather than one that serves as a walled garden for a tech conglomerate. For Tan, the ultimate goal is not just to have a smarter assistant, but to have an agent that is truly aligned with the user’s own interests, free from the opaque incentives of third-party controllers.

As the industry continues to iterate on these models, the debate over privacy, cost, and control will undoubtedly intensify. For leaders like Tan, the bet is clear: the integration of AI into the fabric of daily life is inevitable, and the focus must shift from merely building the technology to ensuring that it is accessible, controllable, and deeply integrated into the unique context of the individual.

Whether the rest of the world will follow his path of granting an AI access to two decades of personal notes and therapy sessions remains to be seen. However, as Tan’s own experience demonstrates, the pursuit of this "cyber" efficiency is already well underway for those willing to pay the price. As he prepares to discuss these developments further on The Builders, it is evident that his focus remains firmly on the horizon, where the boundaries between human intent and machine execution continue to blur.