To scale AI infrastructure startups, become the default first
Inference and search for agents is the most crowded, best-funded corner of AI, and the price of a token falls every quarter. The companies pulling ahead have one thing in common: they became the default — the API a developer reaches for first, and the tool an agent calls without being told to. Five of them, and how each got there — including Keenable, which launched this month with its own web index built for agents.
Jump to the six proven sales plays →The price of a token keeps falling. The value of being the default keeps rising.
Every company here sells something buyers can switch away from with a few lines of code: an API call. That should make this a race to the bottom. It hasn't been, because volume is exploding faster than prices are falling — and most of that volume flows through whoever is already wired in. Agents make the effect stronger: an agent doesn't shop around, it calls the tool it was given.
I read this market as a toll road where the toll keeps dropping. That's fine for whoever owns the busiest on-ramp, and brutal for everyone else. So the question I'd ask of any company here isn't “are you cheaper?” It's “what makes you the first call?”
Every agent call passes through a stack. Each layer has one default slot.
An agent's request moves down a short stack: something routes it, something runs the model, something fetches live information. At each layer, the agent's builder picks one default and rarely revisits it. Winning that slot is worth more than winning any single price comparison.
How AI infrastructure startups win the default: six proven plays
Each play leads with the number that proves it. None of them is “be cheaper” — price matters, but here it's the door opener, not the reason customers stay.
The plays split into two families. Fireworks and Together win by being where the customer's model runs; OpenRouter and Keenable win by being where the request passes through. The second family needs far less capital and gets a view across the whole market — which is why I think the gateways are the better product businesses, even if the clouds have the bigger revenue today.
If I had one quarter, I'd run play 5 first. Keenable's agent storefront is the most forward-looking go-to-market move in this whole group: it treats the agent as the customer, which is where the volume is heading.
Investors are paying most for the companies that own the routing, not the GPUs
Valuation divided by revenue tells you what the market believes about each moat. The ranking is the opposite of what you'd expect from size.
The pattern I see: the more a company owns the decision of where a request goes — rather than the hardware that serves it — the higher the multiple. OpenRouter is the extreme case: a small revenue base priced like a platform, because it sees every model's demand at once.
For the GPU-heavy clouds, the defensible asset is the specialized model running in production. That's why Fireworks leads with “specialized intelligence” rather than speed or price.
Keenable is building the search default for agents
Keenable came out of stealth in August with its own crawler, index and ranking — not a wrapper around someone else's search. The timing is sharp: the big search engines have been narrowing third-party access just as agents need the live web more than ever. Its founders know this territory: CEO Andrey Styskin ran search and ads at Yandex and web infrastructure at Amazon AGI, and chief scientist Matthias Petri built Alexa's web grounding. The company says it already has signed contracts with several AI labs.
What I like most is that Keenable picked a layer where the default isn't settled yet, and showed up with real infrastructure instead of a thin API. A team that has built web-scale search before, an index it owns, and products like Time Machine that nobody else offers — that's a strong hand for a seed-stage company.
The work ahead is distribution: turning lab contracts and a clever agent storefront into the search agents use by default. The playbook above shows what that looks like.