Commoditized intelligence is the labs' moat, not their problem | implexa

back to blogThe consensus is that intelligence is commoditizing. Tokens get cheaper every quarter, open models close the gap, so the frontier labs lose pricing power and value flows to whoever sits on top: the wrappers, the routers, the app layer. OpenRouter is the purest expression of that bet. Pick any model, we route you to the cheapest one that clears the bar.

I think the causality is backwards. Commoditized intelligence is not what kills the labs. It is the weapon they use to take the entire knowledge-work economy.

Here is the move. Once intelligence is cheap, the only question that matters is who can afford to give it away. The answer is: only the company that owns the model. Claude and OpenAI can offer effectively unlimited intelligence at a flat $200/mo, run the expensive model when a task needs it, and silently reroute everything else to a small cheap model they also own. They eat the cost to own the surface where the work happens. A wrapper cannot do this. A router cannot do this. They buy tokens at retail and resell them. You cannot subsidize a market you do not manufacture.

So the thing everyone cheers, cheaper intelligence, is precisely the thing that lets the labs run the surface at a loss until nobody else can stand on it.

That crushes the middle:

OpenRouter and the wrapper economy are arbitraging a spread the labs are actively zeroing out. Squeezed from above.

"Open models will win" is real and irrelevant. The cheaper the floor, the fatter the labs' margin on the surface, because they reroute down to that floor and keep the customer. Open weights are a gift to the rerouter, not a threat to it.

Cursor and the standalone agent IDEs are a transitional species. Great products sitting in the exact spot the model owners will subsidize into the ground. Anthropic ships Claude Code, OpenAI ships Codex. The independent app layer is renting a house on land the landlord is about to build on.

There is a gentler, VC version of this, and it is the one worth arguing with. a16z frames the era as a shift from "system of record" to "system of intelligence": the database stops being the moat, the reasoning layer on top becomes the moat, and that is where the next generation of companies gets built. Directionally right, structurally wrong. There is no durable layer between the model and the customer for those companies to stand on, because the model companies are shipping that layer themselves, as platform primitives. Memory, learning, multi-agent orchestration, grading your own output against outcomes: Anthropic ships these as features, not as startup categories. Every "system of intelligence" a founder wants to build is next quarter's model release.

So the system of intelligence is not a new layer you build on the model. It is the model company. And the lock-in is harder than SaaS ever managed. Once you have paid a frontier lab millions and let its forward-deployed engineers agentify every workflow in your org, there is no specialized layer left to migrate to. You do not rip out the thing that now runs the work. Intelligence is a platform play, not a vertical one, and the platforms are already playing it.

Now the part that gets me called insane. The SaaS incumbents do not survive as software. They survive as data.

Ask any builder where they spent their day. Nobody says Salesforce, ServiceNow, Office, or the Google suite. They say Claude Code or Codex, with everything else feeding in. That is the leading indicator, and the leading edge of knowledge work, the people who define how everyone else will eventually work, has already switched. The mass follows on the usual lag.

What happens to the incumbents is not death, it is demotion. Salesforce, ServiceNow, Workday have real moats: data gravity, compliance, and a buyer who is not the user. That buys them time. It does not buy them the interface. The login dies. The database lives, as an endpoint the agent reads and writes on your behalf. You stop visiting the system of record. The agent visits it for you.

And they let it happen, because they have no better option. The dark version is more honest than the everyone-wins version: the incumbents are trapped funding their own disruptors. Every one of them is long the AI trade, on chips, on model partnerships, on their own copilots that mostly funnel attention back to the same two engines. Fighting the consolidation means shorting the trade they are all levered to. So they ship a copilot, declare victory, and become plumbing on a delay.

Why a handful of winners and not a fragmented field? This is not iOS and Android, there is no hardware lock-in here. The consolidation force is a triad only two or three companies hold at once: the capital to train a frontier model, the ability to subsidize its intelligence to roughly zero to own the surface, and the data flywheel from owning where the work actually happens. Wrappers have none of the...