Three unrelated stories landed on the same day this week. Read together, they describe one repricing.
Taste is now a product with revenue
Intelligence, the company behind Design Arena, raised 7.9 million dollars in seed funding led by Index Ventures, with Conviction, A* and Valkyrie participating. TechCrunch reported the round.
The business itself is the interesting part. Design Arena runs A versus B comparisons on AI output across websites, images and visual design, collecting human preference data from 5.3 million users. It generates 60 million dollars in annual recurring revenue, mostly from enterprise customers.
Frontier labs pay for that preference data because their automated benchmarks can be gamed and human judgment cannot be simulated cheaply. Cofounder Grace Li described the origin plainly, saying there was no substitute for human judgment when their own AI produced games that worked but were not fun.
Functional but not fun is the exact gap most AI-generated marketing output falls into. The market just put a 60 million dollar revenue line on closing it.
The structural reason this works is that automated benchmarks can be optimized against and human preference cannot, at least not cheaply. A model can learn to score well on an eval. It cannot learn to make 5.3 million people prefer something they did not prefer.
Neutrality became a procurement requirement
AWS is now letting the vibe-coding tool Superblocks run inside customers’ own private clouds, built on Aurora and integrated with Bedrock, so company data never leaves the customer’s account. TechCrunch covered the arrangement and its implications.
The architectural point is that the application layer is being separated from the model layer. Enterprises can swap between providers without rebuilding what sits on top.
The demand shift behind it is sharper than the technology. Superblocks CEO Brad Menezes described customers who sixty days earlier insisted on one named provider and now want optionality, and put it bluntly: any enterprise betting on a single model provider will see that executive fired.
The supporting number is worth holding onto. Open-source models accounted for 29 percent of traffic through Vercel’s AI gateway last month. That is not a fringe position anymore.
Control is being sold as the differentiator
Palantir reported quarterly revenue of 1.9 billion dollars, up 93 percent year over year, with 1.1 billion in profit. That profit figure exceeds the company’s total revenue in the same quarter last year.
Strip away the rhetoric from the earnings call and the commercial argument is specific. CEO Alex Karp warned that model providers may, knowingly or otherwise, capture the means of production of the partners they serve, and that enterprises adopting frontier models risk handing over the expertise those models then learn from.
Whether or not you accept the framing, it is being received. The company is positioning model-agnostic software where the customer keeps the prompts, the orchestration, and the context, and the market is paying for that position.
Note what is being sold there. Not a better model, not a cheaper one, but the guarantee that the accumulated knowledge of how you run your business stays yours after the contract ends.
That is an old enterprise software argument wearing new clothes, and it works because it has always worked. Buyers have been burned by lock-in in every previous platform cycle and they recognize the shape of it early now.
The three stories also share a shape worth naming. In each case the thing being monetized is something a model cannot produce on its own: human preference, provider independence, and institutional context.
The pattern across all three
Human judgment sells. Optionality sells. Ownership of your own context sells. The model itself is drifting toward being a component you select, the way you select a database. I argued last month that the model is not your moat, and the money is now agreeing.
For anyone buying AI capability, the practical filter this creates is short. Can you change providers in a quarter without a rebuild, or are you letting one vendor own your agents. Does your data stay in your account. Is there a human review step where taste actually matters, or did you automate the judgment along with the production.
For anyone selling it, the uncomfortable version is that your model choice is not a differentiator and has not been one for some time. What you own is the context, the workflow, and the taste applied on top.
Which means the pitch deck slide naming your model provider is doing no work. Replace it with the one showing what your system knows that a competitor’s cannot, and how that knowledge compounds with use.
The zero-sum game people expected between labs is not where the value settled. It settled one layer up, with whoever controls the context.
Pick your model like a supplier. Build your moat somewhere else.