American professional sports are full of dynasties, but truly dominant teams are surprisingly difficult to sustain.

The NFL is perhaps the clearest example. No team in the Super Bowl era has ever won three consecutive championships. Even the strongest dynasties have tended to peak, win once or twice in succession, and eventually give way to another contender.

Baseball can be similarly volatile. A team can win the World Series and find itself near the bottom of its division the following season. There are dominant stretches, but success often looks more like a window than a permanent position.

The AI market has no salary cap or draft, yet it increasingly feels like this kind of league.

That is a somewhat strange outcome.

Competitive balance in American sports is, to a large extent, engineered. Salary caps constrain how much the richest teams can spend. Draft systems give weaker teams early access to new talent. League rules are deliberately designed to make it difficult for one franchise to accumulate an overwhelming advantage indefinitely.

Frontier AI has none of these mechanisms.

There is no salary cap on researchers, no draft for new PhDs, and no rule limiting how many GPUs a company can buy. A sufficiently wealthy company can spend aggressively on compute, recruit researchers from competitors, acquire startups, and build larger infrastructure.

If anything, one might therefore expect frontier AI to look more like European football: a market in which financial strength compounds into persistent competitive advantage, with a small number of clubs maintaining a structural edge over everyone else.

Yet the actual competition has looked surprisingly fluid.

OpenAI may appear clearly ahead for a period, only for Google, Anthropic, or another competitor to close the gap in the next generation. Leadership changes across coding, reasoning, multimodality, agents, cost, and product quality. There are certainly big-market teams, but so far it has been difficult for any one of them to sustain overwhelming dominance across successive generations.

Why?

The advantage does not fully carry over

A frontier AI lab can be very roughly thought of as a combination of

talent + compute + data + ideas + execution.

All of these are difficult to obtain. But an interesting feature of the industry is that relatively few of them become permanently fixed assets of the incumbent.

Compute is enormously expensive, but it can ultimately be purchased. Researchers move between organizations. Ideas diffuse through papers, open-source projects, benchmarks, products, and people. Software implementation is becoming faster, and successful systems reveal useful information even when the underlying training recipe remains private.

Most importantly, success in one generation does not eliminate the need to execute in the next.

A company that built the best model last year still has to decide how to allocate compute this year, which directions to pursue, what data to use, how to post-train the model, how to evaluate it, and how to turn the resulting capabilities into a reliable system.

The previous championship helps. It does not automatically grant the next one.

This suggests an important distinction. Frontier AI has an extremely high barrier to entry, but that does not necessarily imply an equally strong incumbent advantage.

Only a small number of organizations can afford to compete at the frontier. But among those that can, the leader may have to rebuild a significant part of its advantage every model generation.

In sports terms, qualifying for the league is extraordinarily expensive. Once inside the league, however, last season’s trophy does not count as points in the next season.

Perhaps the important unit is the roster

AI obviously has exceptional researchers. Some individuals have had enormous influence on the direction of the field.

But frontier AI also seems less dependent on irreplaceable individuals than many professional sports.

Put the best basketball player in the world on another team, and the competitive balance of the league can change immediately. A single AI researcher, however brilliant, cannot independently build a frontier foundation model.

The value of talent in AI is highly complementary.

Researchers need compute, infrastructure, data, research engineers, evaluation systems, collaborators, and an organization capable of converting an idea into large-scale experiments. Conversely, enormous compute without the right people and research decisions is not sufficient either.

This makes the roster an interesting unit of competition.

What matters is not simply whether a company employs one superstar, but what combination of researchers, engineers, infrastructure, capital, and leadership it has assembled at a particular moment.

And unlike many traditional corporate assets, much of that roster is mobile.

Researchers leave one lab and join another. Teams form around new startups. Knowledge travels with people. A competitor with sufficient capital can assemble a credible research organization much faster than a football club can reproduce decades of academy development, fan culture, and institutional history.

The frontier is therefore expensive, but surprisingly reconfigurable.

No one knows the winning playbook yet

There is another reason sustained dominance may be difficult: the game itself is not mature.

The major AI labs certainly have different philosophies. They disagree on architecture, scaling, reasoning, agents, multimodality, safety, openness, and product strategy.

But none of them appears to possess a settled playbook for what the winning AI company ultimately looks like.

It is not even obvious what winning means.

Is the winner the company with the most capable model?

The largest consumer product?

The strongest enterprise distribution?

The best coding agent?

The lowest inference cost?

The largest developer ecosystem?

Or the company that eventually discovers the most profitable way to integrate all of these capabilities?

In a mature industry, accumulated experience can become a powerful advantage because the game remains sufficiently stable for knowledge to compound.

Frontier AI is different. The technology, the products, and even the definition of the market are changing at roughly the same time.

Experience still matters. But the target is moving.

A company can become exceptionally good at playing one version of the game just as the industry begins playing another.

No salary cap, but no guaranteed dynasty either

None of this means that frontier AI is an equal competition.

Money matters enormously. Compute matters enormously. Distribution, infrastructure, proprietary data, and existing products may eventually create much more durable advantages than we see today.

Nor does it mean that anyone can suddenly build a frontier model. The number of organizations capable of participating in this competition remains very small.

But there is something unusual about the competition among those organizations.

Their most important advantages are not entirely fixed. Talent moves. Ideas diffuse. Compute can be acquired. Technical strategies change. And the winning playbook itself is still being written.

That combination creates a peculiar kind of competitive balance.

AI has no salary cap. It has no draft. There is no league office trying to maintain parity.

And yet every new generation of models somehow feels a little like the beginning of a new season.

There are favorites. There are big-market teams. There may even be dynasties.

But dominance still has to be defended.

So far, at least, the three-peat remains surprisingly difficult.