Reading Unicorns — Monday Musings

When the model is free, where does the value go?

Three stories broke this week that look unrelated. They're not.

Moonshot open-sourced a 2.8-trillion-parameter model that matches frontier US systems at a fraction of the cost. Dwelly raised $170m to buy up UK letting agencies and gut their back offices with AI. And a UK sovereign-infrastructure startup, Valarian, keeps closing rounds on the thesis that whoever controls the deployment layer — not the model — controls the leverage.

Put those next to each other and you get an answer to the question everyone in AI is dancing around: if the model itself keeps getting commoditised, where does the value actually go?

The model is becoming table stakes, not a moat

Kimi K3 is the clearest evidence yet. Independent researchers who've actually used it are calling it the strongest open model released so far — not because Moonshot stole anything, but because building a frontier-adjacent model with far fewer resources than OpenAI or Anthropic is now a demonstrated, repeatable skill, not a one-off. Washington's response was to reach for distillation accusations and sanctions talk instead of a technical rebuttal. That's telling in itself: when the fastest available counter-argument is geopolitical rather than technical, it usually means the technical argument isn't there.

This isn't a new pattern. It's the DeepSeek moment again, with a shorter memory. And each time it happens, it gets harder to argue that model quality alone is a durable business.

If the model's free, the business wrapped around it is the asset

Which is exactly what Dwelly is betting on. It isn't a model company. It buys letting agencies — a genuinely unglamorous, regulated, people-heavy industry — and replaces the back office with AI once it owns the asset. One property manager handling 300 units instead of 100 isn't a headline number. It's a systematic re-pricing of every fragmented service business that's been valued on its existing labour constraints.

That's why the ElevenLabs, Synthesia and Legora CEOs wrote personal cheques into a lettings company rather than another lab. If you're close enough to the frontier to know model capability is becoming a commodity, the smart move isn't to build another wrapper — it's to go buy the businesses that were priced assuming AI didn't exist, and unlock the multiple expansion yourself. I'd expect this to spread well past lettings: insurance brokers, conveyancing, accountancy, anywhere the moat has always been "we know how to do the paperwork," not "we have proprietary technology."

If the model's free, control over deployment becomes the other asset

The third piece is the one that gets less attention: sovereign and physical-world infrastructure. Valarian's entire pitch is that once AI runs mission-critical workloads, someone needs to own the layer that governs how those workloads behave, where the data lives, and who can pull the plug. That thesis gets stronger, not weaker, every time a geopolitically contested model shows up cheap and capable. A commodity model you don't control the provenance of is a bigger governance problem, not a smaller one.

You can see the same logic playing out in UK robotics, just at a smaller scale. Perceptual Robotics and Kinematic Trees aren't chasing general-purpose humanoids — they're building narrow, certified, deployable autonomy for one job (turbine inspection, factory tasks) done well. That's not a lack of ambition. It's the correct strategy for a market that can't out-capitalise Figure AI or Physical Intelligence on raw model scale. You compete on deployment, certification speed, and trust in a narrow domain instead.

The pattern

Model capability is heading toward commodity pricing. That doesn't kill value in AI — it relocates it. It moves toward whoever owns the regulated, relationship-heavy business the model gets embedded into, and toward whoever controls the layer that decides where and how the model is allowed to run. Everything in between — the general-purpose chat wrapper, the "we fine-tuned GPT for your industry" pitch — is exactly where the pressure is going to land hardest.

What this means

• If you're building a thin layer on top of a foundation model with no proprietary data or distribution, this week's news should worry you more than reassure you.

• If you're an investor, the more interesting cheque right now might not be into the next model company — it might be into the acquisition of a boring, regulated, labour-constrained business, with AI as the operating lever rather than the product.

• If you're in infrastructure, defence, or physical deployment, geopolitical instability at the model layer is a demand driver for you, not a risk to hedge against.

• Watch who's writing personal cheques, not just who's leading rounds. Insiders close to model economics moving their own money into "boring" businesses is a stronger signal than any press release.

Worth thinking about

The AI story everyone's watching is who has the best model. The more useful story is who's already positioned for the world where that question stops mattering.