A month ago I spent over $3,000 USD on AI tokens.
This month I am using Codex for $300.
Same category. Same broad outcome. Completely different cost base.
That is the part of AI investing that makes my head spin.
Everyone wants to talk about moats. Who has the best model. Who has the biggest research team. Who has the most compute. Who has the most enterprise deals. Who is lining up for the monster IPO.
You see the reports around Anthropic IPO prep and the speculation around OpenAI, xAI, and the rest of the frontier labs, and it is easy to get swept into the obvious conclusion.
AI is massive. Therefore the biggest AI companies must be obvious investments.
I do not think it is that simple.
Because as an operator, I do not experience AI as a static market.
I experience it as a pendulum.
One month the winning move is to wire together APIs and pay usage. The next month the winning move is a flat subscription. Then an open-source model catches up. Then a new agent product wraps the whole workflow. Then pricing collapses again.
That is amazing for builders.
It is much harder if you are trying to value these companies like normal software businesses.
The Old Software Moat Made Sense
The old software moat was easier to understand.
Distribution. Workflow lock-in. Data. Brand. Switching costs. Network effects.
You bought the CRM. You trained the team. You moved the data in. You built the reports. You integrated the emails, the phones, the forms, the billing, the workflows.
After that, even if a competitor was a little better, you stayed where you were.
The pain of moving was bigger than the benefit.
That is a moat.
AI has some of that, but the ground is moving faster than the moat can harden.
The model layer is expensive, but for most normal business use, it is becoming more interchangeable than people want to admit.
Most companies do not need the theoretical best model on earth.
They need the model that can write the email, fix the code, review the contract, answer the customer, summarise the meeting, generate the report, and not make their team hate using it.
That means the value keeps shifting.
From model to interface.
From interface to workflow.
From workflow to distribution.
From distribution back to model when someone releases something dramatically better.
Round and round.
The Tool Provider Is Not The Religion
That is why I am careful with the hype.
Not because AI is overhyped. I think the opposite. I think AI is still under-appreciated at the operator level.
But "AI is going to change everything" and "this specific AI company deserves any valuation at any price" are not the same sentence.
As a business owner, the lesson is clear:
Use the tools aggressively.
Do not worship the tool provider.
The tool that saved you $10,000 this month might be replaced by something cheaper next month.
The product you thought was magic might become a feature.
The API you built around might get undercut by a subscription.
The expensive workflow might become a default button.
That does not mean there are no moats in AI.
It means the moat probably is not "we have a chatbot."
The moat is owning the customer.
Owning the workflow.
Owning the data loop.
Owning the distribution.
Owning the trust.
Owning the place where AI turns into actual economic value.
That is the difference.
The Operator View
From the outside, the AI market looks like a race between model companies.
From inside a business, it looks different.
I do not care which lab wins a benchmark by two percent if another tool lets me ship faster, automate more, reduce admin, or give my team leverage today.
The operator does not buy intelligence in theory.
The operator buys outcomes.
Can this make my team faster?
Can this reduce labour drag?
Can this improve decision-making?
Can this help us respond quicker?
Can this give one good person the output of three?
That is where AI becomes real.
Not in a pitch deck. Not in a valuation. Not in a leaderboard.
In the boring work.
In the inbox. In the CRM. In the sales process. In the meeting notes. In the codebase. In the quote flow. In the reports nobody wants to write but everybody needs.
That is where the economic value lands.
And it does not always land with the company that built the model.
Sometimes it lands with the company that owns the workflow.
Sometimes it lands with the company that already has the customer.
Sometimes it lands with the operator who moves first and actually implements the thing while everyone else is still arguing about which model is best.
So Is There A Moat?
Yes, but not always where the market wants it to be.
There may be moats in frontier AI.
Compute relationships matter. Research talent matters. Brand matters. Enterprise trust matters. Capital access matters.
But price matters too.
Substitution matters.
Open source matters.
Platform bundling matters.
The fact that my AI cost can move from $3,000 usage-based to $300 subscription-based in a month matters.
That is not a small detail. That is the whole point.
If the price of intelligence keeps falling, then the question is not just "who has the best intelligence?"
The question is:
Who captures the value when intelligence becomes abundant?
That is the question I keep coming back to.
The model companies may become some of the most valuable businesses in history.
They may also spend insane amounts of capital fighting a price war where yesterday's breakthrough becomes tomorrow's commodity.
Both can be true.
So when I look at the AI IPO chatter, I do not just think, "How big can this get?"
I think:
How fast can the margin disappear?
How loyal is the customer when the next model is 20% better or 80% cheaper?
Where does the value settle when everyone has intelligence on tap?
AI is not one market.
It is a force moving through every market.
And when the pendulum is swinging this fast, I would rather be the operator using it to build leverage than the investor pretending the price of intelligence has already found its floor.