In 1995, Netscape held roughly 90 percent of the browser market. The company had introduced millions of people to the World Wide Web, defined what a browser should look like, and convinced an entire generation of consumers and enterprises that the internet was worth their attention. Its IPO in August of that year is still cited as the moment the modern technology economy began.

By 2003, Netscape was effectively dead.

The company that defined the category did not capture it. Microsoft bundled Internet Explorer into Windows and absorbed the market through distribution. The browser, as a standalone business, ceased to exist. Netscape's contribution was not the product that won. It was the conviction that the category mattered.

The same pattern played out in search. AltaVista, Lycos, Excite, Infoseek, and Yahoo collectively spent the late 1990s teaching the world how to find information online. They burned through capital, defended early-mover positions, and competed on directory size and brand reach. Then in 1998, two graduate students published a paper on a ranking algorithm called PageRank and launched a company most observers initially dismissed as too late to matter.

Google did not create the search category. It walked into a market that had already been validated, educated, and behaviourally trained by its predecessors, and offered a fundamentally better answer to a question users were already asking.

The pioneers carried the cost of category creation. Google captured the value.

The mechanic underneath

This is not a coincidence of timing. It is a recurring structural pattern in how technology categories mature, and it deserves more attention from boards than it tends to receive.

Pioneers in any new category face an unavoidable tax. They have to convince the market that the category exists. They absorb the early regulatory friction. They train the first wave of users on behaviours those users have never performed before. They build infrastructure, often in directions that later prove to be wrong, because no one yet knows what the right architecture looks like.

By the time the category is broadly accepted, the pioneer has spent most of its capital and most of its strategic optionality on the work of legitimisation. The company is shaped by the assumptions of the early market. It cannot easily reshape itself for the mature one.

The winner enters later, often with a narrower thesis, a cleaner architecture, and the benefit of watching the pioneer's mistakes from the sidelines. The winner does not have to convince anyone the category matters. That work has already been done. The winner only has to be materially better at the use case the market has now settled on.

This is why first-mover advantage, as a concept, is widely overstated. In categories defined by network effects or distribution lock-in, early dominance can compound. In categories defined by underlying technical architecture, early dominance is frequently a liability. The pioneer's product is built on the assumptions of a market that no longer exists by the time the market is large enough to matter.

The current cycle

We are now in the middle of an almost identical cycle, and the parallels are close enough to be uncomfortable.

OpenAI, with the launch of ChatGPT in late 2022, is performing the role Netscape performed in 1994. It has introduced hundreds of millions of users to a new behaviour. It has absorbed the early regulatory shock. It has set the consumer expectation for what a conversation with an AI feels like. It has, almost single-handedly, created the category.

Around it, a generation of frontier model labs are doing the work AltaVista and Yahoo did in the late 1990s. Anthropic, Google DeepMind, Mistral, xAI, and others are competing on capability, safety, alignment, and distribution. The benchmarks change weekly. The leaderboards reshuffle. Each lab argues, with some justification, that its model is the most capable for a particular class of task.

Boards and executives observing this from the outside are being asked, implicitly, to pick a winner. The framing is almost always the same. Which model should we standardise on. Which lab will dominate. Which API should our enterprise architecture be built against.

This is the wrong question.

If the historical pattern holds, the company that captures the value of this cycle will not be one of the companies currently competing to be the best chatbot. The chatbot is the Netscape browser. It is the artefact that proves the category. It is not the artefact that captures it.

The winner of this cycle will be the company that builds the equivalent of PageRank for the AI era. A piece of architecture that takes the now-established behaviour for granted and offers something fundamentally better in a direction the pioneers were not structured to pursue.

We do not yet know what that architecture will be. We have strong hints. The most credible candidates are not chat interfaces. They are systems that assume AI works, and build the layer above it. Autonomous software that operates without prompting. Agentic platforms that coordinate work across tools. Runtime environments where business logic is generated, deployed, and executed without a human intermediary at every step.

The chatbot proved the category. The category itself is something larger.

What this means for boards

The strategic implication for executives and boards is not that they should bet against the current frontier labs. Several of them will remain enormously valuable companies. The implication is narrower and more useful.

Adoption is not advantage.

A great many organisations are now reporting AI adoption metrics to their boards. Number of seats deployed. Number of workflows touched. Number of employees trained. These metrics are real, and they are not unimportant, but they describe participation in the category, not position within it. Every competitor in your industry is reporting the same numbers.

The strategic question is not whether your organisation has adopted AI. Within eighteen months, that question will be as meaningless as asking in 2005 whether your organisation had adopted the internet. The question is which layer of the emerging stack your business will compete on, and whether you are building, partnering, or buying at that layer with the same seriousness you would apply to any other piece of long-term infrastructure.

Three questions worth putting to a board this quarter:

First, are we mistaking adoption for advantage. If every competitor has the same models, the same copilots, and the same productivity gains, where exactly is the differentiation we are paying for.

Second, which layer of the AI stack is actually strategic for our business. The model layer is unlikely to be defensible for anyone who is not a frontier lab. The application layer is crowded and largely undifferentiated. The interesting positions are usually one layer above or below where the current attention is concentrated.

Third, what is the equivalent of PageRank in this cycle, and who is building it. This is not a rhetorical question. There is an answer, and it is being built right now, and it is almost certainly not the company currently dominating the headlines.

A closing observation

History does not repeat, but it does, with reasonable consistency, instruct.

The companies that defined the early internet were not the companies that captured it. The companies defining the early AI era are unlikely to be the companies that capture it either. This is not a criticism of the pioneers. The category they created is real, and the value of that creation is enormous. It is, however, a useful corrective to the assumption that being early to the conversation is the same as being positioned to win it.

The boards that will look prescient in five years are not the ones that picked the right model in 2026. They are the ones that asked the right question.

The right question is not which AI to adopt.

The right question is what comes next, and where in the stack it is being built.