Eric Schmidt's Blueprint for the Next Trillion-Dollar Company: Learn Everything, Specify Nothing

Eric Schmidt built Google into one of the most valuable companies in history. He ran it as CEO for a decade, then as chairman for another five years, during the period when it went from a promising search engine to the company that defined what a technology monopoly looks like in the internet era. When he talks about how the next trillion-dollar company gets built, the framework he offers is not abstract. He has watched it happen from the inside.

In a recent conversation, Schmidt offered a specific and somewhat counterintuitive thesis for founders building in the AI era.

"The other thing I would say to a founder, it's really important that everything you do be learned and not specified. I'm doing a couple of startups on my own. We'll see how well they do. But with them, I say, I know nothing, learn everything. So you can learn how to support your customers. You can learn what the customer wants. You can learn how to... and learning meaning in the AI sense of learning, learning it as part of either supervised or unsupervised training. And if you take a learning approach, then you build a system that if it works, it will explode. Because once the learning accelerates, you get into a quasi-monopoly position."

The philosophy Schmidt describes in the rest of that passage: run as fast as you can, get there as quickly as you can, build it around learning, and if it takes off, you will be a hero. Because once it learns, it learns how to become stronger. And two or three years in, various forms of reinforcement learning begin replicating themselves, creating accelerations further. That, he argues, is the most likely path for the next trillion-dollar company, past the equivalents of Anthropic and OpenAI.

The distinction between learning and specified

The core technical distinction Schmidt is making is one that most developers understand abstractly but fewer founders apply consistently as a company strategy.

A specified system is one where a human engineer writes explicit rules for how the system behaves. A customer support flow that routes certain keywords to certain responses. A recommendation algorithm that applies manual weights to signals. A fraud detection system based on thresholds that someone set. These systems work, but they scale with the number of engineers who maintain them, not with the amount of usage they receive. Every new edge case requires a human to write a new rule.

A learning system, by contrast, improves with usage. The more queries it receives, the more corrections it processes, the more feedback it accumulates, the better it gets. The number of engineers required does not scale proportionally with improvement. Usage itself is the training signal.

What Schmidt is advising founders to do is build the second type of system from the beginning, across every part of the company. Not just in the product but in customer support, in understanding what customers want, in every operational loop. Make learning the architecture, not a feature you add later.

Why learning creates quasi-monopoly

The quasi-monopoly mechanism is the same one that made Google dominant in search, Amazon dominant in e-commerce, and Netflix dominant in streaming. It is usually called a data flywheel, though the underlying dynamic is simpler than that term implies.

When a learning system gets better with usage, and getting better attracts more users, which generates more usage, which makes the system better, the company that starts the flywheel first has a compounding advantage that is extremely difficult to overcome. A competitor who starts two years later is not just two years behind. They are behind by two years of compounding improvement that the leading system has accumulated and that the new entrant cannot replicate without going through the same volume of real-world experience.

This is what Schmidt means by quasi-monopoly. It is not a monopoly enforced by contracts or exclusivity. It is a monopoly created by accumulated learning that becomes structurally difficult to match. Google's search results got better because more people used Google, which gave Google more data about what people actually wanted when they typed a query. The improvement made more people use Google. The flywheel ran.

In the AI era, the same dynamic applies but with higher velocity. A product that uses AI to learn what each specific customer wants, how they like to be communicated with, what their actual problem is rather than what they describe, builds a relationship with that customer that cannot be replicated by a competitor who starts fresh. The data is not just a historical record. It is the actual capability of the system.

The reinforcement learning acceleration

The most ambitious part of what Schmidt describes is the timeline he sketches for reinforcement learning. In the first phase, the system learns from human-labeled data or user behavior, supervised or unsupervised learning. In the second phase, two to three years in, reinforcement learning begins to operate, where the system learns from the outcomes of its own actions rather than from labels a human provided.

This is where the acceleration Schmidt describes comes from. Supervised learning improves with more data and better labeling, which requires human effort. Reinforcement learning improves with more actions and clearer feedback loops, which can happen autonomously once the initial system is strong enough. The learning that was gated by human effort becomes increasingly self-sustaining.

What this means for a startup is that the window for establishing the initial flywheel is limited. The companies that accumulate real-world learning in the next two to three years will be the ones who can make the transition to reinforcement-based self-improvement when that phase becomes accessible. The companies that spend those years building specified systems will face a structural disadvantage that is not purely about resources. They simply will not have the accumulated learning to run reinforcement-based training on.

What Schmidt is actually telling founders to do

Translated into practical terms, the advice is specific.

Every time you are tempted to write an explicit rule, ask whether you could collect the data that would let the system learn that rule instead. Every time you want to hardcode a behavior, ask whether you could build a feedback loop that produces the correct behavior emergently. Every time you are about to specify what a good outcome looks like, ask whether you could define an objective function and let the system discover what good looks like through experience.

This applies to customer support. Rather than writing scripts, collect data about which responses lead to satisfied customers and let the system learn the script. It applies to product recommendations. Rather than manual curation, collect outcome data and let the system learn what to surface. It applies to pricing, to content moderation, to fraud detection, to every loop that currently runs on human-written rules.

The "I know nothing, learn everything" posture Schmidt describes is not false modesty. It is a specific strategic choice to substitute data collection and feedback loops for premature specification. You build systems that know they do not know, that are designed to learn from every interaction, rather than systems that think they know and apply their assumed knowledge to each new case.

The window is not permanent

Schmidt frames this as the most likely path past the equivalents of Anthropic and OpenAI. That framing matters. He is not describing how to build another frontier AI lab. He is describing how to build a product company on top of the AI infrastructure those labs have created, one that achieves the same kind of structural durability through accumulated learning in a specific domain.

The window for this is open now, in the way that the window for building on top of the internet was open in the late 1990s, and the window for building on top of mobile was open from roughly 2008 to 2012. The companies that built learning systems in those windows accumulated advantages that persist to this day. The companies that built specified systems, however well-engineered, faced rearchitecting costs as the learning-based competitors pulled ahead.

Schmidt is telling founders that 2026 is that window for AI-era companies. Whether or not you believe the trillion-dollar framing, the core technical advice is sound and urgent. Build systems that learn. Run as fast as you can. The flywheel either starts or it does not, and starting it later is not the same as starting it now.