The AI conversation fixates on the wrong layer. While everyone fawns over the nerd-turned-class heartthrob (his name is Claude, and he's French), the real story is unfolding somewhere decidedly less photogenic: infrastructure.

And increasingly, that infrastructure isn't truly open anymore. It's being reshaped into what you might call a VC-backed "tollbooth". It's a new hybrid where open-source credibility meets high-margin control.

To understand how this works, you need to look at the Sky Computing Lab at UC Berkeley.

The Berkeley Machine

Ion Stoica, Databricks co-founder and Berkeley professor, runs what may be the most productive infrastructure factory in modern computing. His research lab doesn't just publish papers. It spins out billion-dollar companies with regularity.

The latest evidence:

RadixArk, which originated as SGLang in 2023, started as an open-source project designed to make AI models run faster and cheaper on existing hardware. It recently spun out of Berkeley at around a $400 million valuation, with backing from Accel and Intel CEO Lip-Bu Tan himself.

Then there's vLLM, the high-throughput inference engine that became standard equipment for serious AI deployment. In the last five days (as of Jan 27), vLLM transformed into Inferact with $150 million in seed funding at a whopping $800 million valuation, co-led by Andreessen Horowitz and Lightspeed.

These aren't outliers. They sit in a league that already includes multi-billion-dollar players like Fireworks AI and Baseten. Individually, they look like success stories. Collectively, they form a pattern and reveal how “open” innovation turns proprietary once venture dollars flow in.

The New Open-Source Economics

Stage One: Technical superiority as user acquisition. Launch an open-source project that solves genuine production problems and, as a result, draws developers and enterprises in.

Stage Two: Community momentum cements market position. Usage compounds and the project becomes a core piece of infrastructure that other infrastructure, models, and apps depend on. Community-first ethos combined with bleeding-edge innovation drives virality.

Stage Three: The spin-out moment. Once adoption reaches escape velocity, convert community traction into venture capital. The pitch essentially writes itself: proven PMF, existing user base, clear monetization vector.

Stage Four: Value capture at the chokepoint. Monetize the chokepoint—hosting, managed services, enterprise tools, and proprietary optimizations built around that open core.

That’s where the game flips. The open-source core maintains credibility and adoption, but the real money comes from premium hosted services, enterprise integrations, and proprietary optimizations.

Once that happens, the once fledgling community-backed project from a lab nestled in The Bay effectively becomes a VC-fueled toll operator on the AI highway. Developers still get in for free, but scaling, monitoring, and enterprise features carry a fee. And because infrastructure is sticky and deeply embedded in production workflows, those fees behave more like rent than optional services (i.e., you can’t live without them, and they occur annoyingly often).

And unlike NYC subway turnstiles, you can't jump these tollbooths.

Why VC Changes Everything

Historically, traditional open-source development faced a persistent tension: community-driven projects could achieve technical excellence but struggled to match well-funded competitors on ecosystem development, enterprise features, and market presence.

Venture capital resolves this tension by treating the open-source phase as user acquisition and PMF validation rather than altruism. Hundreds of millions in funding enable aggressive hiring, marketing, and partnership development that purely community efforts can't match. The result is winner-take-most (if not all) dynamics, but accelerated and amplified by capital deployment.

As a result, VCs are effectively manufacturing category leaders, and in turn, impact the entire makeup of the AI industry. This leads us to my next point…

The Inference Chokepoint

Think of inference as the economic engine room of the entire AI industry. Every token generated, every query answered, every user interaction flows through this layer. Inference determines unit economics for application builders, sets floors on commercial viability, and increasingly dictates who can iterate fast enough to survive. When control concentrates in a handful of venture-backed operators, downstream innovation inherits structural dependencies. A pricing adjustment, a policy change, or a strategic pivot at the infrastructure level cascades into margin compression for everyone building on top.

History offers a preview. Kubernetes began as open infrastructure; then clouds wrapped managed services (GKE, EKS, AKS) around it, capturing the lion’s share of revenue.

AI inference is moving toward the same fate: open-source at the core, but economically gated at the edges.

The Real Power Layer

Berkeley’s ecosystem just happens to be the latest full-scale demonstration of what that model looks like: an academic lab turned de facto factory for AI infrastructure. The academic legitimacy provides cover. The open-source release provides distribution and community adoption. VCs provide competitive moats.

The result is a new category of companies that benefits from public-good perception while reaping private-market rewards, all while rewriting the rules for every player in the space.

The fireworks will keep exploding overhead. But the real show is happening at ground level, where the roads are being built, and the tolls are being set.