Intelligence Is Cheap Now
Your moat was never the AI
I want to start with a claim that sounds wrong, and then spend the rest of this post defending it: Using AI is not a competitive advantage.
Every founder and CEO out there right now is measuring their AI adoption, tracking developer productivity, debating which model to use. I get it…I’m doing the same thing inside my own company. But if every company has access to the same models, the same APIs, the same falling token costs, then access to AI is table stakes. Not a moat. It’s electricity in 1995.
Important? Yes. Differentiating? No.
Intelligence is now the commodity, but what surrounds it isn’t.
Distribution isn’t. Domain knowledge isn’t. Customer trust isn’t. The ability to aim intelligence precisely at the right problem in a specific market — that definitely isn’t. And for startups, this changes everything about how we should be building, hiring, and thinking about the next few years.
This starts now.
The trap most teams are falling into
A 2025 randomized controlled trial found that experienced developers working in codebases they already knew were 19% slower when using AI tools. Not faster. Slower. The tools work — the problem is teams bolted them onto old workflows and expected magic.
There’s a subtler version of this trap at the company strategy level: treating AI as a productivity layer on top of the old model. Orgs still optimizing headcount and measuring output-per-engineer while sprinkling AI tools on top are missing the structural shift entirely. They’re just making the old model slightly faster. That’s it. Bolting AI onto a traditional engineering org doesn’t get you the real advantage — it gets you incremental improvement while your competitors rethink the whole game.
The teams that will look back at this moment and feel like they got it right are the ones who asked a harder question: what actually changes about how we build and compete if intelligence is cheap and everywhere?
What doesn’t work
Let me be specific about the mistakes — ones I’ve had to actively talk myself out of.
Competing on token volume
If your strategy is “we’ll throw more AI compute at the problem than the competition,” you’re cooked. Large enterprises will outspend you, negotiate better per-token rates, and lock in consumption contracts you can’t touch. Any advantage built purely on access to inference is borrowed time.
Building horizontal when you’re small
Trying to build a general-purpose AI product without a specific market wedge means you’re picking a fight with OpenAI, Anthropic, Google, and every well-capitalized AI-native company simultaneously. The generic intelligence layer is already a commodity. A startup trying to win there is playing the wrong game entirely.
Over-relying on a single model provider
Cursor is the cautionary tale. They scaled past $200M ARR and reportedly found themselves spending 100% of revenue on Anthropic — every dollar in, straight back out the door. Growth was making the problem worse, not better. They had to completely restructure their pricing mid-flight, which blew up in their face with users. Building your entire unit economics around one provider’s current pricing, without architectural flexibility, isn’t a business model — it’s a bet that someone else’s margins will stay favorable forever.
Assuming cheap tokens mean you can ignore token economics
Costs are falling, but consumption is exploding. This is Jevons Paradox playing out in real time. Startups that don’t build token management discipline early end up with infrastructure costs that have absolutely nothing to do with their revenue. Perplexity spending 164% of revenue on compute is the extreme case, but the pattern is everywhere.
Valuing generic code production as a core asset
If your primary differentiation is the ability to write a lot of application code quickly, that moat is evaporating. The value of generic code production is deflating at roughly the same rate as token costs. Which means fast.
The through-line
Every mistake above is actually the same mistake: confusing access to the commodity with a competitive advantage.
Intelligence is the commodity. I want that to land, so I’ll say it one more time as plainly as I can: the thing everyone is racing to adopt is the thing that will differentiate absolutely no one.
What won’t be commoditized:
distribution
domain knowledge
customer trust
brand, and
the ability to aim intelligence at exactly the right problem in a specific market.
That’s it. That’s the whole list. And it’s a short one.
Startups that organize around the commodity lose. The ones that organize around what surrounds it have a real shot.
The question — which I’ll get into in my next post — is what it actually looks like to build a company around that idea.
For vertical software startups especially, I think the answer is way more specific, and way more interesting, than most people realize.