Greg Isenberg spent 48 hours in San Francisco and wrote down 21 observations

1. Selling data is extremely profitable. He heard many stories of data brokers selling company data to labs, making around $2 million a week. He can’t tell whether this is the early days or a bubble — he’s researching which verticals are still untapped.

2. A soul-searching question. Someone at dinner asked, “When a new model drops next month, will your company get better or worse?” Half the table couldn’t answer. He thinks this is currently the single best question for evaluating an AI startup.

3. Revenue without retention. The metric everyone’s talking about now is $1B in revenue within 12 months (the new unicorn standard), but almost nobody talks about retention. He predicts we’ll see lots of AI companies with big revenue and terrible retention.

4. We’re clearly witnessing the reinvention of the computer. The model is the CPU, the harness is the operating system, and skills and plugins are the apps. Using this framework makes it much easier to spot where the gaps still are.

5. We may be only 12 months away from an agent that watches you work for a week and then takes over your job without any instructions. Screen recording + agent memory + local models — that combination can already do it today. (Creepy.)

6. Agent harnesses feel like the new generation of GPT wrappers, but the good kind — especially the vertical ones. The best teams he talked to treat models as interchangeable parts. For example, one founder demoed how he routes different models to different tasks: one for aesthetics and visuals, a cheap one for everyday grunt work, and another specialized in text.

7. More reverse engineering and cloning than he’s ever seen. People are decompiling software and rebuilding it with agents. Someone showed him a suite of roughly 10 apps, each competing with originals worth $5B+, all open-source clones he’d built himself. He and everyone else are still grappling with the question: what does it mean when software becomes this easy to copy?

8. He guesses value will shift to data. If software can be rebuilt quickly, then the thing worth buying in an old SaaS company is the years of accumulated customer data and workflows inside it. Some SaaS companies are really data companies in disguise.

9. It feels like a significant portion of legacy software won’t survive into the agent era — especially those designed around “making humans click through tons of pages.”

10. Every company he talked to wants to consolidate their internal data into one place and expose it to agents via MCP. Then when they look at the actual data, it’s a mess: five versions of the same customer, a pile of old Slack channels, PDFs nobody can find. Cleaning all that up feels like a huge consulting opportunity.

11. The companies building the building blocks will win. Many founders are working on the foundational components for agents — APIs, data layers, tools. Some companies barely have a UI because their primary user is an agent.

12. Labs aren’t short on training data, but they can’t easily get real-world examples of agents making mistakes inside enterprise environments. Whoever has those logs has something valuable.

13. Someone building a major consumer product said the next battle is memory. The model that remembers the most about you will win. Being the smartest model matters less than people think.

14. Astra is the first model that made people believe AI can do CAD, and people are starting to wonder what “vibe manufacturing” looks like. A robotics founder said: the hardest problem in AI right now is deployment, not research. Their robots work on their own, call a human when stuck, and every call becomes training data.

15. Cybersecurity companies are having their moment. Doing this work is genuinely cool right now, especially post-breach cleanup. Every agent with access to your system is a new door into your company — and many of those doors are left wide open.

16. Dead is worth more than alive. A defunct company’s five years of Slack history, emails, and code can be licensed for money. He wouldn’t be surprised if founders factor this into their shutdown timing.

17. People are starting to talk about how the “agent internet” gets monetized. It feels very early, but when the most active users on the internet become agents, what kind of startups should you be building?

18. Agent insurance is coming. Once agents start making expensive mistakes, insurers will require audits before writing policies.

19. Several people said the people they most want to hire are those who have taste and can write evals. Models are getting better at generating ideas; the hard part is judging which ideas are good.

20. The new startup trajectory: Take VC money → use investors’ money to become an influencer → company fails → founder becomes a full-time influencer and bootstraps other products with their own money.

21. Traffic is incredibly hard to get. Founders are sponsoring buses, going crazy buying billboards, and booking well-known DJs for parties — all just to be seen. It has the feel of the dot-com bubble era. Attention is the truly scarce resource. The growth lead who went viral with a UGC strategy around Wispr Flow is a great example. This is the recipe everyone wants, and probably only about 50 people in the world know how to scale these kinds of businesses efficiently in this era.

Every trip to San Francisco leaves him with plenty of stimulation coming back. Being able to experience such an exciting journey at such a golden moment in the startup world — he feels incredibly lucky. Much of the time, it almost feels surreal.

Really well written! Sharing with everyone. (He noted these came from conversations with a group of billionaires, data lab people, AI founders, and executives)

GREG ISENBERG (@gregisenberg): I spent 48 hours in SF hanging out with billionaires, data labs and AI founders and execs running $1B to $100B companies.

My observations:

1. Sooo MANY stories of data brokers making like $2M a WEEK selling company data to labs. I can’t tell if it’s the beginning or a bubble,