The most sophisticated piece of AI investment thinking I’ve seen this year has nothing to do with model architecture, inference costs, or agent orchestration. It’s a dozen venture capitalists putting $9.5 million into RASA World, a company that throws electronic dance music festivals. No GPUs. No weights. No eval suite. And it may be a sharper read on where agent systems are heading than most of the infrastructure bets sitting next to it in the same portfolios.
The stated logic is simple enough: the more life moves onto screens, the more valuable the things that can’t happen on a screen become. That sounds like a lifestyle thesis. I’d argue it’s a technical one, and it follows directly from how generative systems behave at scale.
What Falls In Price When Generation Gets Cheap
Every capability we’ve pushed into models has followed the same curve. Something that used to require scarce human attention becomes a function call, the marginal cost of producing it drops toward zero, and the price of the output collapses toward that floor. Text first. Then images, voice, code, and increasingly multi-step work that used to be a junior analyst’s whole week.
Agent architectures accelerate this because they close the loop. A model that writes a draft still needs a human to act on it. An agent that can plan, call tools, check its own output, and retry removes the human from the middle of the loop entirely. The scarce resource stops being the generation and becomes the orchestration.
Follow that logic one step further and you get to the interesting question: what does not fall in price? Not content. Not analysis. Not the vast category of knowledge work that reduces to pattern completion over text. What holds value is anything whose supply is bounded by physics rather than compute.
The Properties An Agent Can’t Reproduce
A live event has a specific set of properties that no model can synthesize, and they’re worth naming precisely, because they describe the shape of the whole category.
- Non-reproducibility. A festival set exists once. You can record it, but the recording is a different good with a different price. There is no cheap copy that substitutes for the original.
- Co-presence. Value comes partly from the other people in the room knowing you’re also in the room. That’s a mutual-verification property, and it’s expensive to fake.
- Capacity limits. A venue holds what it holds. Supply doesn’t scale with capital the way inference capacity does.
- Embodiment. Sound pressure, crowd density, physical fatigue. None of it is a token stream.
Those four properties together form something like an anti-generative moat. Not because the technology is hard, but because the good is defined by scarcity that no amount of model capability erodes.
Why This Belongs In An AI Portfolio
Read as a hedge, the RASA World position is a bet on correlation structure, not on dance music. Most of a modern venture portfolio is now exposed to a single variable: whether capability gains keep translating into capturable value in software. Specialized vertical AI, defense technology, fintech, space, sustainability, health and biotech — the current funding priorities differ in sector but share a dependency on that same assumption. Corporate venture capital has become the largest source of AI funding globally, which concentrates exposure even further.
Live events are one of the few assets whose returns might actually improve if the AI thesis plays out exactly as promised. If models absorb more of the screen-mediated economy, attention that used to go to digital goods has to land somewhere. Physical, bounded, unrepeatable experiences are a plausible destination. That’s negative correlation, and in a portfolio this concentrated, negative correlation is rare and expensive to buy.
What This Means If You Build Agents
I take two things from this for anyone designing agent systems rather than funding them.
First, the defensibility question for an agent product is rarely about model quality. It’s about whether your product sits on top of something with bounded supply — a licensed dataset, a regulated workflow, a physical asset, a relationship that requires a human signature. Pure capability layers get commoditized by the next model release. That pattern has repeated often enough to treat as a design constraint rather than a risk.
Second, proof-of-presence is becoming a real engineering problem. As generated media gets harder to distinguish from recorded media, systems that can establish that a specific human was in a specific place at a specific time acquire economic weight. Ticketing, identity, attestation, provenance. That’s infrastructure work, and it looks a lot more interesting now than it did when authenticity could be assumed.
A dozen investors paid $9.5 million for exposure to things a model can’t make. The number is small. The reasoning behind it deserves more attention than the check size suggests.
🕒 Published: