USV's Nick Grossman: the 'Rebel Alliance' thesis — why generational companies will be built around AI labs, not just by them

USV's Rebel Alliance thesis

Nick Grossman, a partner at Union Square Ventures, argues that the AI opportunity is too large to collapse into a handful of labs — and that generational companies will be built across the entire agentic stack, not just at the model layer.

USV has been making this case publicly under the label "Rebel Alliance." The framing is deliberately provocative: OpenAI and Anthropic aren't the Death Star, Grossman says, but they do operate with a vertically integrated logic that leaves most of the stack open. The prevailing fear — that Claude becomes "the world's only employee" and all value consolidates upward — is what USV is betting against.

From agent-as-employee to agent-as-infrastructure

The sharpest part of Grossman's argument isn't about application-layer niches. It's about what agents actually are becoming. He draws a distinction between personified agents hired like workers and agents deployed as cloud infrastructure, where companies programmatically spawn hundreds or thousands of them as part of production systems. That second model creates demand for model routing, orchestration, memory, identity, and payments — a full stack of components that every enterprise building agentically will need, regardless of which foundation model they use.

Model competition reinforces this. With so many capable models available, any serious builder needs an abstraction layer to swap between them on cost and capability. That's not a threat to the Rebel Alliance — it's a structural tailwind for the middleware and tooling layer.

“There are going to be generational companies built up and down the stack, not just — the meme in the market is like maybe Claude becomes the world's only employee and all value gets sucked into those companies. Our thesis, the Rebel Alliance thesis, is that the opportunity is so big and so massive and so expansive. We're moving from agent as employee to agent as infrastructure, and I think that's a big driver of why we believe so much in the Rebel Alliance thesis.”

Suno as the template

Suno is Grossman's clearest example of how application-layer companies survive model commoditization. Google has a "make music" button inside Gemini. Suno has still carved out what could be a $20 billion business, in Grossman's estimate, precisely because the labs have little incentive to prioritize a market that size. The pattern generalizes: experience specificity matters, niches are larger than they look, and a single app or chat interface doesn't absorb every use case.

Capital structure

Grossman describes a barbell. At one end, the foundation model labs are capital-intensive industrial businesses — billion-dollar training runs, GPU commitments, long amortization curves. At the other end, application companies passing through token spend can carry very low or even negative gross margins depending on how aggressively they're subsidizing usage. The middle layer — orchestration, memory, identity, payments — looks more like a traditional software business. USV's portfolio company Pluralis Research gets a mention as an example of model training done in a capital-light way, using decentralized compute rather than a centralized cluster.

Gross margins across the application layer now range from 20% to 90%, and Grossman treats 20% as potentially fine depending on growth trajectory. That's a meaningful departure from the web-era assumption that a dollar of revenue was roughly a dollar of gross profit.

The contrarian bet

USV has been thesis-driven since its founding in 2004, when Fred Wilson and Brad Burnham made a then-contrarian call that the infrastructure build of the late '90s had created a massive opening at the application layer. Twitter, Tumblr, Etsy, SoundCloud, and Indeed came out of that era. Grossman frames the current AI moment the same way — tools on the ground, expansive rather than consolidating, wide open for builders. The risk he acknowledges openly: you can be wrong on the thesis entirely, or right on the thesis but wrong on the team or the timing. USV's answer is to run a concentrated, high-conviction book rather than a coverage strategy, and to limit downside when the call misses.

The central bet is that the agentic stack gets built out company by company, layer by layer — and that most of that value accrues outside the labs.