Here is the contradiction most people are still walking past: the same week the industry finally admitted agents are not "digital employees you hire," two of the sharpest infrastructure operators alive told you where the money is actually going — and it isn't into the models everyone keeps benchmarking.
Union Square Ventures put it in the most quotable form. "Part of the Rebel Alliance thesis for us is that agents are moving from personified people you hire to cloud infrastructure," Nick Grossman said, describing a world where applications draw on "networks of specialized agents coordinated through layers like orchestration, memory, routing, identity, and execution."
Read that list again. Not a single item on it is a model.
Not a single item on it is a model — USV re-described the agent not as a colleague you onboard, but as a service you build on.
That's the tell. For a year the pitch-deck fantasy was the agent as tireless knowledge worker — you hire it, it does the job of a person. USV, in effect, re-described that same agent as something closer, arguably, to S3 or a message queue: a service you build on, not a colleague you onboard.
This isn't semantics. It's a bet about where value lands, and USV is now deploying against it. "We are actively investing across this stack," the firm wrote, inviting founders building "any of the layers listed in the article — or ones we haven't thought of yet."
This has reportedly been referred to as the Emergent Agent Stack, per a TBPN show promo, and USV is writing checks across it. The abstraction became a portfolio.
The reframe matters because it only pays off under one condition: the model has to be a swappable module. If intelligence is welded to the agent, then value pools in the labs and the "coordination layer" is just glue. If it's plug-and-play, the coordination layer is the product and the labs are suppliers.
Guillermo Rauch is fighting that exact fight, and unlike most people theorizing about it, he's watching it happen at scale. Vercel sees 6 million deployments a day, half of them triggered by coding agents, with more than 1 trillion tokens flowing through its AI gateway daily.
That gateway is the whole argument in miniature. When you're the pipe a trillion tokens run through, you get to see which models people actually pick when nobody's watching the leaderboard.
And what they pick isn't what the headlines say. "We're seeing a lot of growth of Gemini, even though it's not on the news as much, because people are optimizing for production now," Rauch said. "When you're optimizing for production, you start looking at price/performance."
Open models ride the same logic. "DeepSeek and GLM-5.2 are taking off," he added. "The data doesn't lie."
Sit with what that means. The model that wins in production is not the smartest one; it's the one with the best price/performance at the moment of the call. That, arguably, is the behavior of a commodity input, not a differentiated product.
Rauch framed the coupling question directly: "Do you get all your intelligence from one place? Or do you get a module or a library or a building block from one provider, and then you build on top of it." His answer is the un-sexy one — "That's more like software engineering has always been" — and his ambition follows from it: "We're going to be the AWS of this generation, so obviously we're fighting for a world of open protocols."
Now here's the part that should reorder your priors about what a "killer app" even is.
Ask a room what agents are good for and you'll hear coding. Rauch agrees that's one. But the second, in his telling, is duller and more important: "the internal agent that helps you run the company."
The interesting thing is what makes that app hard. It isn't intelligence. "The challenge there is, how do you securely access data? How do you audit what the agent is doing? How do you get a trail of all of the tool calls and access controls that the agent had to incur in order to get a job done?"
Those are governance problems, not IQ problems. And Vercel's answer to them is a cage: Vercel Sandbox, where "you put the agent in a little cage," free to express its intelligence but bound by policy on what data it can access and what data can leave.
The concrete horror story is instructive. Rauch recalls the president of Airbus — decades of what he describes as very specific C++ aerospace code — and the risk that "someone comes in and installs the wrong developer tool and boom, all the code goes out to the cloud for training."
The value there is entirely in the plumbing. Data control. Audit trails. Access policy. The sandbox around the model is worth more than the model, because the model is interchangeable and the trust boundary is not.
Rauch's other example is quieter still: a sales rep whose bottleneck "has not been her creativity, intelligence, ability to build relationships, it's been data." She couldn't ask which five accounts added the most seats in two weeks until "a Q1 project for a new sales dashboard completed." The agent's job is to break that data monopoly — "so many of these SaaS giants build their entire kingdoms on trapping your data, and that's incompatible with agents."
Un-glamorous. Boring. And exactly where the margin is migrating.
Push the same logic down and it doesn't stop at the software layer. It runs all the way to the substrate — and the substrate is telling the same story.
The Information framed the back half of the year plainly: it "could belong to AI companies focused less on bigger models and more on cutting the cost of running them." Inference cost, not model size. Baseten is, per RaiseSummit's program, on stage in Paris talking about "Inference's Scaling Era" — which arguably reads as scaling meaning cost-down, not IQ-up.
Even the exotic frontier obeys it. On space data centers, Lightspeed's Guru Chahal says whether you can run compute in space is "a foregone conclusion," and that the real bottleneck is launch cost, not radiation or cooling. "If it costs zero to put compute up in space, space would be significantly cheaper because power is free and cooling is free."
That's the whole thesis compressed into orbit. Value doesn't accrue where you make the model bigger. It accrues where you arbitrage cost — the price/performance of a token, the dollars-per-kilogram to lift a GPU past the atmosphere.
Modular, price/performance-driven, swappable at every layer. The router that picks Gemini over Anthropic on cost is arguably the same instinct as the launch company driving down dollars-per-kilo. Both capture margin by arbitraging inputs, not by owning them.
So here is the sharp version of the bet USV and Rauch are making twice, in different words. If every layer is genuinely plug and play — model, harness, data platform, sandbox, gateway — then the labs become interchangeable suppliers, and whoever owns the coordination layer captures the spread.
But notice what that thesis actually rests on. It's not a claim about intelligence. It's a claim about who owns the router and the identity layer — the two chokepoints where "plug and play" gets decided.
Rauch is fighting for open protocols precisely because a world of open protocols is a world where the router matters and the model doesn't. USV is funding orchestration, memory, routing, identity because those are the tollbooths in that world.
The uncomfortable part is that none of this is proven. The labs are not sitting still — Rauch himself notes OpenAI now publishes websites directly, walking straight into infrastructure territory. Coupling could reassert itself. The router could turn out to be a feature, not a company.
USV is buying tickets before the show. That's what a thesis is. But the reasoning is legible, and the tell is honest: they're betting on the boring primitives — audit trails, data control, price/performance routing — because those are the parts that don't get commoditized when the model underneath you gets swapped out next quarter.
If they're right, the Rebel Alliance doesn't beat the labs. It outlasts them, one interchangeable supplier at a time.