Garry Tan says he spends $100,000–$200,000 a year on AI tokens
AI agent
A system that can take several steps to complete a task.
token
A unit used to measure AI processing.
Y Combinator
An organization that supports startup companies.
What happened
Garry Tan, CEO of Y Combinator, said he spends about $100,000 to $200,000 each year on tokens. He made the estimate during an interview with Walter Isaacson for The Builders podcast. TechCrunch reported the remark on November 2. In this context, a token is a unit used to measure an AI service’s processed text and work. It is not a physical coin. The TechCrunch report and The Builders page identify the source of the claim.
From a helper to an operator
Tan uses AI agents for research and coding. He also uses them to prepare for podcast interviews. He said an agent has read his email, Dropbox, Apple Notes, and therapy notes from twenty years. At Y Combinator, he described an internal agent called Stray Light. It holds context about founders the group backed, people who contacted the company, and applications.
Tan has also shared gstack on GitHub. It contains 23 Claude Code skills. The skills give AI different work roles, including design, engineering management, release work, documentation, and quality checks. The public repository makes this approach easier for others to inspect and try.
Why the cost matters
The number is not an industry average or a normal subscription price. It is Tan’s rough statement about his own usage. Still, it shows a shift from asking one question to connecting AI with many parts of work and life.
An agent may read information, produce an answer, and continue with another task. More work can mean more processed tokens. That can make costs rise. The story is therefore about system design, not only model prices.
Tan said costs must fall before personal agents become mainstream. He also questioned whether free agents controlled by big technology companies are the right answer. He imagines a personal AI that users can control and program. That puts price, privacy, and control in the same discussion.
Hacker News attention
The related Hacker News post received 81 points and 107 comments. Those numbers show community attention. They do not prove Tan’s spending figure or the report’s accuracy. The source’s facts and the community’s attention should remain separate.
What is confirmed
The available report confirms Tan’s own estimate, his description of broad AI use, his reference to Stray Light, and his public release of gstack. It does not establish that this spending is typical. It also does not show that every AI agent can handle the same tasks.
What is still unknown
The report does not break down the $100,000–$200,000 figure by service, model, or task. It does not say whether Tan or Y Combinator pays each cost. It also does not explain how the sensitive records are stored, protected, or reviewed. The productivity gains have not been measured in the report either.
What to watch next
The next questions are practical. Will agent prices fall? Will users get precise controls over private data? Can people stop an agent before it takes an unwanted step? Tan’s example suggests that cheaper AI will matter, but trustworthy access and clear boundaries will matter too.
Why one AI power user reports a huge yearly cost
📰 Full story: Garry Tan says he spends $100,000–$200,000 a year on AI tokens
AI agents can help with many jobs, but wider use brings bigger bills and privacy questions.
AI agent
A computer system that can do several steps for a task.
token
A unit that counts AI processing.
Y Combinator
An organization that helps new companies.
💡 The gist
- Garry Tan is CEO of Y Combinator, which supports new companies.
- He says he spends $100,000 to $200,000 yearly on AI tokens.
- His agents help with work and read personal records.
Tan discussed this on Walter Isaacson’s The Builders podcast. TechCrunch later reported his words.
Tan uses AI agents for research, coding, and podcast preparation. An AI agent can take several steps for a task. It can read information, make a plan, and help complete work.
Tan said one agent reads his email, Dropbox, Apple Notes, and therapy notes. He also described Stray Light, an internal Y Combinator agent. It uses information about founders, messages, and applications.
Tan shared 23 Claude Code skills called gstack on GitHub. The skills give AI different jobs. They cover design, engineering work, documentation, and quality checks. This shows how widely Tan wants to use agents.
A token measures how much text and work an AI service processes. More steps can use more tokens. More tokens can make the bill grow. That is why the cost depends on usage, not only on a monthly plan.
The $100,000 to $200,000 figure is not an average. It is Tan’s rough statement about his own use. The report does not show each tool’s cost. It also does not say who pays each bill.
The related Hacker News post received 81 points and 107 comments. Those numbers show community attention. They do not prove the report is correct. The Hacker News post is a separate signal.
The larger question is whether agents can become cheaper and safer. People will also need control over what agents can read. Tan wants personal AI that users can control. Future reports may show whether that idea can work for more people.
Sources: The Builders and gstack.
💬 Is a $200,000-a-year AI bill really worth it?
The main HN question was not how much was spent, but whether it produced more than cheaper methods.
- The article's reported claim puts the agent cost at up to $200,000 a year. It can read email, files, and notes, which sounds useful but creates a broad privacy risk.
- A token is a small unit of text an AI processes. If the number is calculated by multiplying tokens by list price, it may differ from the actual bill because cached inputs can be cheaper; another commenter said many tasks might fit a $100–$200 subscription plus simple programs.
- In a user's self-report, a test expected to cost about £2,500 fell to about £150 after choosing models, shortening inputs, and using a local model for one step.
- The comparison should be a roughly $30 monthly AI plan or no AI. Without measured outcomes, the $200,000 figure alone does not show value.
- One counterargument says a better investment decision could return millions. That is only a hypothesis, and commenters also worried about agents failing in inconsistent or unsafe ways.
initial digest at 107 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
AI helpers can cost lots of money
📰 Full story: Garry Tan says he spends $100,000–$200,000 a year on AI tokens
Garry Tan uses AI helpers for many kinds of work.
AI agent
A computer helper that can do steps for you.
token
A number that counts AI work.
Hacker News
A website where people discuss technology.
Garry Tan leads Y Combinator, a group that helps new companies. AI agents are computer helpers that can do steps for you. Tan says he spends $100,000 to $200,000 yearly on tokens.
Tokens count how much work the AI reads and makes. His helpers read email, Dropbox files, and old notes. Some notes are from twenty years ago.
More AI work means more tokens. More tokens can mean higher costs. Private notes also need careful protection.
Hacker News is a technology discussion site. Its related post showed 81 points and 107 comments. Those numbers show attention, not proof.
People will watch whether agents become cheaper and safer. They will also watch who controls the private information.
Sources: TechCrunch and Hacker News.
💬 Does paying a lot for AI make the work much better?
People wanted to know whether the AI really helped, not just whether it was expensive.
- The article reports a claim of up to $200,000 a year. The AI can see email and notes, which may help, but sharing secrets can be risky.
- The number may not be the real bill. In a user's self-report, a costly task estimated at about £2,500 fell to about £150 by mixing cheaper AI with an AI running on the user's own computer.
- We still do not know whether this beats a $30 monthly AI plan or no AI. Some people say one good investment decision could pay for it.
- The AI may also make unsafe mistakes or act differently each time. Spending more is not proof that the work is better.
initial digest at 107 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 The reported $200,000-a-year AI spend: can it pay off?
The figure attributed to the article is large, but HN's debate was about whether it is a real bill, what it produces, and whether it beats cheaper approaches.
initial digest at 107 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.