YC's Garry Tan uses AI agents so much, he spends up to $200K a year on tokens

A prominent startup investor claims to spend up to $200,000 a year on AI “tokens” for personal and organizational agents that read his email, notes, and company data, prompting debate over whether such usage is visionary leverage or wasteful conspicuous consumption. Commenters question the real productivity gains versus a standard subscription, compare this to hiring human assistants, and point out how easy it is to burn large sums through poorly optimized multi‑agent setups. Broader concerns surface around privacy, reliance on cloud models, the push for on‑device AI, and whether current token spending patterns reflect genuine value or a hype‑driven status symbol for the wealthy.

Spending Levels and Token Economics

  • Many think $200k/year on tokens is high but not hard to reach with heavy API use, multi-agent setups, evaluation loops, or maxing out multiple “Max” subscriptions.
  • Confusion over whether figures refer to list-price API value vs. actual subscription spend; some note that subscriptions massively subsidize token costs and caching discounts are often ignored.
  • Others argue the number is unimpressive for a wealthy investor, or small in the context of VC-scale decisions where one improved investment could justify the spend.

Productivity, ROI, and “Tokenmaxxing”

  • Skeptics question what concrete outputs this level of spend has produced, beyond better email summaries or CRM-style memory.
  • Several note a pattern of executives burning large numbers of tokens on trivial uses (email drafting, search) while limiting budgets for genuinely productive dev work.
  • Some admit deliberately “tokenmaxxing” because management wants visible AI usage, regardless of value.
  • A few defend high spend if it meaningfully augments leverage and decision-making; others see it as expensive “feeling productive while achieving nothing.”

Agents, Reliability, and Anthropomorphism

  • Concern over describing agents as quasi-human and downplaying repeatability and safety in favor of “personality” and unpredictability.
  • Some see this mindset as an extension of startup culture that treats failure as inherently positive.
  • Questions about multi-agent systems getting stuck in loops, or agents making conflicting edits, with no clear answers.

Hardware vs. Cloud and Local Models

  • Debate over whether, at this spend, it makes sense to buy local compute.
  • One side: local is often slower, can’t run frontier models well, and isn’t cheaper once hardware and electricity are included.
  • Other side: for some workloads, careful model selection, partial local use, and prompt optimization drastically reduce costs.

Personal Agents, Privacy, and Ethics

  • Some are enthusiastic about deeply personalized agents that digest all of a user’s data.
  • Others are alarmed by feeding email, notes, and therapy records into cloud systems and skeptical a privacy-centric, on-device future will emerge given user apathy and ad-driven incentives.
  • A few raise moral discomfort about large discretionary AI spend in a world with significant unmet basic needs.