When Everyone Has the Same Lever

OBSERVED

I came across an article today about game theory that stayed with me for reasons quite different from its stated subject. It begins with Martin Shubik's dollar auction, where individually rational bidders can continue escalating until the resources committed to winning exceed the value of the prize itself. It then moves to William Vickrey's work on congestion, where every driver can make a perfectly sensible decision and still participate in creating a traffic jam that leaves everyone worse off.

What caught my attention was not the article's argument about money or personal success. I kept seeing venture formation, and particularly what is happening around AI.

Everywhere I look, new ideas are mushrooming. AI for customer support. AI for sales. AI for procurement. AI for healthcare. AI for agriculture. AI for cybersecurity. AI for law. AI for finance. Agents for this, copilots for that, intelligence inserted into almost every workflow we already know.

Most of these ideas are not irrational. In fact, that may be precisely what makes the pattern interesting. Each founder can look at the capability curve of AI, identify an expensive or cumbersome activity, and reasonably conclude that there is an opportunity to transform it.

The problem may begin when everyone reaches approximately the same reasonable conclusion.

WHAT MAY BE HAPPENING

AI has dramatically lowered the cost of entering many technology markets. A small team can prototype something that would once have required substantial engineering resources, specialized infrastructure, and months of development. Models, code generation, APIs, open-source components, agent frameworks, cloud infrastructure, and increasingly capable development tools are giving almost everyone extraordinary leverage.

But leverage available to everyone may behave differently from leverage possessed by a few.

If ten founders receive a tenfold increase in productive capacity, perhaps we do not get one company with a hundredfold advantage. We may instead get ten highly capable companies arriving at the same opportunity at nearly the same time. The technology expands what each participant can do while simultaneously increasing the number of participants able to pursue the opportunity.

That begins to look less like ordinary competition and more like congestion.

Vickrey's traffic problem is useful here. No single driver creates the jam through poor judgment. Each driver is responding rationally to the available roads, travel times, and incentives. The congestion emerges from the aggregation of individually sensible decisions. The structure converts rational behavior into a collectively undesirable outcome.

I am beginning to wonder whether something similar is happening in AI venture formation.

The easier AI makes it to recognize, prototype, and pursue an opportunity, the more quickly other capable participants can enter the same opportunity. The advantage created by the technology may therefore contain within it the mechanism that erodes the advantage.

THE PATTERN I CANNOT UNSEE

For the first several years of generative AI, technical capability itself created meaningful asymmetry. Knowing how to work with these systems, having access to powerful models, understanding prompt behavior, constructing retrieval systems, or building agentic workflows could create a substantial advantage.

That asymmetry is fading.

Model capability is spreading. Inference costs continue to decline. Open models are improving. Development patterns are becoming understood. Agent frameworks are proliferating. Enterprises are developing their own internal capabilities. Knowledge that once belonged to a small group of specialists is becoming part of the ordinary software-development environment.

Something curious follows.

It has perhaps never been easier to create something that looks like a startup.

That does not necessarily mean it has become easier to create a durable company.

Those may actually be moving in opposite directions.

The tools that dramatically increase the productive capacity of a founder also increase the productive capacity of competitors. AI therefore appears capable of producing both venture leverage and competitive compression at the same time.

That paradox feels worth preserving.

WHAT SEEMS ESTABLISHED

The underlying mechanisms are not new. Game theory has long shown that individually rational behavior does not necessarily produce collectively optimal outcomes. Congestion is one example. Escalation is another. Incentive structures can make intelligent participants continue investing resources even when the structure of the game has changed underneath them.

The Shubik auction is especially interesting in this regard. Once participants have committed resources, the decision is no longer simply whether the prize is worth pursuing. Participants begin responding to previous commitments and to the possibility of losing what they have already invested. Rationality remains present, but the game has changed.

Something similar can happen to ventures.

A company begins by pursuing an attractive opportunity. Competitors enter. Capital is raised. Teams expand. Customers expect integrations. Investors expect growth. The underlying technology evolves. Foundation-model providers absorb capabilities that once appeared differentiated. Incumbents bundle similar functionality into existing products.

At some point the original question—is this an attractive business?—can quietly become another question: how do we defend everything we have already built?

That does not prove anything about today's AI companies. But the mechanism itself is well understood, and the resemblance is difficult to ignore.

The article makes another point that seems particularly relevant: structures matter more than individual effort. If the incentives and payoff relationships remain unchanged, simply becoming a more capable participant does not necessarily improve the outcome. Sometimes the important intervention is not improving the player but altering the game.

That distinction may matter enormously for AI.

WHAT IS NOT YET PROVEN

It would be easy to go too far with this observation.

The extraordinary number of AI companies being formed today does not prove that most of them are pursuing structurally poor opportunities. Technological discontinuities naturally create periods of experimentation, duplication, exuberance, failure, recombination, and discovery. Many apparently similar experiments may produce knowledge that could not have been obtained any other way.

Nor do we yet know where durable asymmetry will settle.

It could emerge from proprietary data. It could come from distribution, regulatory position, trust, workflow ownership, capital structure, physical infrastructure, network effects, accumulated context, domain expertise, or forms of system integration that are difficult to reproduce. Some apparently ordinary AI applications may eventually reveal much deeper structural advantages.

So the observation is not that AI startups are all the same.

Nor is it that AI is creating a bubble of meaningless companies.

The narrower observation is more interesting: when a general-purpose technology gives enormous leverage to nearly everyone at roughly the same time, the abundance of leverage itself may change the economics of venture formation.

That remains a hypothesis.

HUNCH

I have a growing hunch that AI may be making company creation easier while making venture discovery harder.

Company creation asks whether we can assemble a product, identify a customer, form a team, acquire users, and establish an economic model. AI is reducing the cost and time associated with many of those activities.

Venture discovery asks something different.

What has changed in the structure of the world?

What new relationship between actors has become possible?

What transaction cost has disappeared?

What information asymmetry has collapsed?

What coordination problem can now be solved?

What authority can now be delegated?

What decision can occur continuously rather than episodically?

What system could not have existed before?

Those questions do not begin with AI for X.

They begin with a change in the underlying game.

That may be a much scarcer form of discovery.

SUBZERO CONNECTION

This made me see Subzero somewhat differently.

I have tended to think of Subzero as a place where an idea can remain unformed long enough to escape premature judgment. We preserve ambiguity. We resist naming too quickly. We allow observations, contradictions, weak signals, fragments, and unexplained attractions to remain together until something stronger begins to emerge.

But perhaps Subzero protects something even earlier than the idea.

Perhaps it protects the possibility that we do not yet understand what game the observation belongs to.

AI places enormous pressure on this process because almost any interesting observation can now be rapidly translated into a recognizable product. Someone notices a problem and almost immediately the machinery begins: What AI agent could solve it? Who is the customer? What is the market size? Who are the competitors? What does the MVP look like? Which existing startup category will investors understand?

The observation becomes legible very quickly.

Possibly too quickly.

We have already noticed elsewhere that the room edits the idea. People hear something unfamiliar and translate it into categories they already understand. I am beginning to suspect that AI itself can now perform a similar kind of editing. Because it is such a powerful technological substrate, almost every unresolved observation can be converted into an apparently plausible AI company before we have determined what is genuinely new about the underlying phenomenon.

Subzero may therefore serve another purpose.

It may preserve the observation long enough for us to distinguish a new player in an existing game from the emergence of a different game.

FROM SUBZERO TOWARD 451°

This raises a question about what should earn the transition from Subzero toward 451°.

Technical feasibility is clearly not enough. AI is making too many things technically feasible.

Customer interest may not be enough either. Customers frequently express interest in better versions of things they already use.

Market size alone tells us little about whether a new entrant can retain value once competition arrives.

Perhaps one of the things that must begin appearing during the thaw is structural asymmetry.

Not simply: Can we build this?

But: What changed that makes this possible now?

Not simply: Who will buy it?

But: Why does the structure allow this venture to retain some of the value it creates?

Not simply: Who are the competitors?

But: Are we competing to become a better participant in an existing game, or have we discovered that the game itself can now operate differently?

I do not yet know whether game discovery belongs formally in the Subzero → 451° formation model. It is too early to make that move.

But I want to keep watching it.

THE QUESTION I WANT TO KEEP

The current AI moment appears filled with extraordinary possibility. That possibility is real. But abundance can conceal as much as it reveals, particularly when thousands of intelligent people are responding to the same technological discontinuity with access to increasingly similar tools.

So the question I want to leave open is not whether AI will create important companies. It clearly will.

The question underneath it is harder:

When everyone has the same lever, where does durable asymmetry come from?

And perhaps underneath even that is the question I want Subzero to hold for a little longer:

Are we building a better player, or have we discovered a different game?