The author spent 48 hours in San Francisco with billionaires, AI lab insiders, and founders and executives running companies with $1 billion to $100 billion in revenue. His central observation is that data, systems for running agents in real workflows, evaluation, deployment, and distribution increasingly matter more than models alone. He also sees clear signs of bubbles and structural risks: poor retention behind fast revenue growth, easily copied software, and security vulnerabilities.

1. Exploding Data Deals and Unstable AI Revenue

His first striking observation was an overheated market for data brokers selling company data to AI labs. He heard many stories of brokers making around $2 million a week. He could not tell whether this was the beginning of a huge market or a bubble, and was actively looking for niches still available.

His most important question for an AI startup is:

"When a new model comes out next month, does your company get better or worse?"

When this came up at dinner, half the table fell silent. Companies relying on the superiority of a particular model can lose their advantage as soon as a stronger or cheaper model arrives. He considers this the best question to ask every AI startup today.

Reaching $1 billion in revenue within 12 months is being discussed almost as a new benchmark. The old symbolic goal of becoming a unicorn seems to have been replaced by far higher revenue growth. Yet he worries that almost nobody discussed customer retention. AI revenue can arrive quickly, but customers can leave just as quickly. He expects many AI companies with high revenue and terrible retention. 📈➡️📉

2. Reinventing the Computer and Agents That Do Your Work

He sees this as a period when the computer itself is being reinvented, not merely a time of new apps. His analogy is:

• AI models are CPUs.
• The harness connecting and running models is the operating system.
• Skills and plugins performing specific tasks are apps.

This framing makes unoccupied opportunities, or white space, easier to see. Models may not capture everything. The layer connecting models to actual work, controlling them, and verifying their output may become central to a new computing environment.

He also predicts that within roughly 12 months, agents may observe someone's work for a week and then perform it without further instructions. He says the necessary technical combination is already close at hand.

"Screen recording, agent memory, and local models make this possible right now."

The idea is to record which programs someone opens, what decisions they make, and the order in which they work, then have an agent reproduce that process using long-term memory and the local environment. This suggests automation is moving beyond chatbots toward learning and executing actual workflows.

3. Vertical Agent Harnesses Matter More Than General Models

He thinks agent harnesses may resemble the old "GPT wrappers," but this time can be a valuable business category in a positive sense. Vertical harnesses tailored to particular industries look especially promising.

Strong teams treat models as interchangeable components rather than committing to one. One founder described routing different jobs to different models:

• One model for aesthetic judgment and visual output
• A cheaper model for everyday, repetitive tasks
• Another model for writing and text processing

They favor prompts portable across models over fine-tuning. They A/B test by observing what users actually do with output: whether they edit, export, or share it.

He identifies four core elements of a vertical agent harness:

1. Connectors to industry data
2. Evaluations, or evals, written by people who understand the work
3. Rules defining what agents may and may not do
4. A clean interface for human review

His argument is that the model may represent only about 20% of the value. The real advantage lies in data connections, domain expertise, safeguards, and human review workflows.

Many founders are also building the APIs, data layers, and tools that serve as the agent era's basic components. Some companies barely have a user interface because their primary users are agents rather than people. He believes startups building these foundational primitives will do very well.

4. Copyable Software and the Revaluation of Data

Another observation is the sharp rise in AI-assisted software reverse engineering and copying. People are decompiling existing software and rebuilding it with agents. Someone even showed him open-source clones of roughly ten apps, each worth more than $5 billion.

"Everyone, myself included, is still thinking through what it means when software can be copied this easily."

He expects value to shift increasingly toward data. If features can be rebuilt quickly, the reason to buy an aging SaaS business may be its accumulated customer data and workflows rather than its code.

"Some SaaS companies are secretly data companies."

Declining SaaS companies might previously have been valued at something like twice annual recurring revenue, or ARR. He thinks better valuation methods may be needed to account for customer data quality, the depth of work records, and accumulated operational workflows.

Much existing software, meanwhile, may not survive the agent era. Products built around people clicking through many screens are particularly vulnerable. Once agents can pursue goals across systems, screen- and menu-centered UX may become outdated quickly.

5. Cleaning Enterprise Data and the Value of Failure Logs

He reports that almost every company wants to bring internal data together and make it accessible to agents through MCP, the Model Context Protocol. Looking at the data, however, often reveals a mess:

• Five duplicate versions of the same customer record
• Old Slack channels nobody uses
• Piles of PDFs that cannot even be searched

Cleaning and connecting this data could be a major consulting opportunity. Because agent performance depends heavily on input quality, cleanup is a prerequisite for AI adoption, not mere housekeeping. 🧹

Labs can obtain training data, but struggle to obtain examples of how agents actually fail inside companies. Organizations holding failure logs of errors, dead ends, and incorrect actions may therefore have valuable assets.

An executive at a major consumer company also said future competition would revolve around memory: how well AI remembers the user.

"The model that remembers the most about you wins. Having the smartest model matters less than people think."

As more models offer similar capabilities, continuously remembering and using a person's preferences, context, history, and relationships could become a stronger differentiator.

6. Physical AI, Security, and Insurance in the Agent Era

Physical AI featured in many conversations. He sees Astra as one of the early models that spread the idea that "AI can actually do CAD." The industry is now considering what "vibe manufacturing" might look like: a future in which natural-language intent connects AI to design and manufacturing.

A robotics founder offered a particularly striking observation:

"The hardest problems in AI are now deployment problems, not research problems."

Robots operate autonomously and call a human when stuck. Each intervention becomes more training data. Deploying into real environments, handling errors, and continuously improving now matter as much as improving performance in a lab.

Cybersecurity companies are attracting renewed interest, especially those that clean up systems and damage after breaches. Every agent with system access can be another doorway into a company, and many of those doors are not properly locked. 🔐

He also expects agent insurance to emerge. When agents begin making expensive mistakes, insurers will demand audits and verification before providing coverage. Permissions, execution logs, safety rules, and incident response could become underwriting requirements.

Several people identified people with good judgment who can write evals as their most desired hires. Models are increasingly good at generating ideas, but deciding which ideas are good remains hard. In the agent era, taste and evaluation skills that recognize and measure good results matter alongside the ability to build.

7. Failed Companies' Data, the Agent Web, and Founder Survival

He thinks some startups may now be worth more dead than alive. Even a closed company may be able to license five years of Slack conversations, emails, and code. Founders may begin considering those data assets when deciding when and how to shut down.

People are also starting to discuss monetizing the agent web. It is early, but founders should consider what to build if agents become the internet's most active users. Advertising, payments, identity, permissions, and recommendations could all change when agents search, buy, book, and consume content.

He sees a new founder trajectory:

1. A startup raises venture capital.
2. The founder uses that funding to become a creator.
3. The startup fails.
4. The founder returns as a full-time creator.
5. The founder then bootstraps multiple products using their personal reach and income.

In this pattern, a founder's personal distribution channel and audience become the foundation for new businesses even after a startup fails.

8. Attention and Distribution Are the Scarcest Resources

His final strong impression was how difficult it is to secure distribution at scale. Founders sponsor buses, buy billboards, and throw parties with famous DJs to stand out. It reminds him of the dot-com era.

"Attention is what is scarce right now!"

He mentions the growth leader who went viral by sharing Wispr Flow's UGC strategy as an example of the "secret" everyone wants. He thinks perhaps only around 50 people worldwide truly know how to scale this kind of business efficiently today.

Building technology alone is not enough. Good data, systems fitted to real work, safe deployment, rigorous evaluation, and above all distribution that reaches users must come together. Calling the trip extraordinary and fortunate, he reflects on what it feels like to build in this unusual period before heading home to get back to work. 🚀

"I feel incredibly lucky to have had such an amazing trip while building a business at such an amazing time."