If you've talked to a frontier AI lab in the last six months, you've heard some version of the same complaint: the data is running out, and the data we have is legally radioactive. Redpine, a Stockholm-based AI data infrastructure startup, just raised €6.8 million to do something about it. The round, led by NordicNinja with participation from Luminar Ventures and node.vc, takes total funding to €9 million.

Founders, operators and a small group of senior people from OpenAI, Perplexity and Spotify backed the seed. That last bit isn't a coincidence.

Redpine isn't another annotation shop. It's a headless API. AI agents query the platform in real time, pay per query, and get back legally licensed datasets calibrated for the use case at hand. Pricing is token-based. Licensing is per query, not per dataset. Think of it less like Scale AI and more like a pay-per-use rights management layer for the agentic era.

The company compares the moment to music in 1999. Most AI systems are trained on scraped internet data. The legal foundation is shaky. The quality is uncertain. And the result is that no model has any real differentiation against a competitor that scraped the same internet.

Redpine's bet is that someone has to build Spotify before the next round of lawsuits.

The founders are Anders Hammarbäck (ex-McKinsey, ex-Antler) and David Österdahl (ex-Spotify, ex-iZettle), with founding data scientist Leonora Vesterbacka, a CERN PhD who ran AI R&D at Sweden's KBLab. That's a Stockholm operator team with both the institutional networks and the model expertise to convince a Spotify alum to put real money in.

The lead, NordicNinja, is the EU-Japan tech fund anchored by JBIC. They write checks where they think a Nordic company can become a global infrastructure layer. They led Pit's seed last year. They led Photoncycle's Series A in March. And now they've led Redpine.

There's a thesis forming on Sveavägen, and it's about who owns the rails AI runs on.

Annotation companies built last decade's data infrastructure. They scaled with humans, sold to model trainers, and assumed the right of access to the world's data. That model is breaking. Publishers are suing. Reddit is licensing for cash. The New York Times has lawyers permanently camped in San Francisco.

The next layer is API-native. Models call the data when they need it, pay for what they use, and never have to argue with a publisher's general counsel after the fact. Redpine is one of a small handful of teams building exactly this.

Competitors include Scale AI, Appen and Defined.ai, all of them annotation-first. Redpine wants to skip the labeling step entirely.

Look at where AI is actually going. Not bigger pretrained models. Tool-using agents. Software that browses, reads, decides, transacts. That kind of agent runs queries hundreds of times per task. Each query is a chance to need a fresh, licensed dataset that didn't exist when the model was trained.

If your business model is selling annotated training corpora, that's a great market for the past. If your business model is selling per-query licensed data on a token meter, that's the market for the next five years.

Redpine, in other words, isn't selling fuel. It's selling a fuel station network.

The capital goes to international expansion and dataset partnerships. Hammarbäck told reporters the priority is building exclusive licensing relationships with content owners that can't afford their own enterprise sales teams. Niche publishers. Specialized data providers. The long tail of premium content the big labs would rather not negotiate with one at a time.

If Redpine can become the default broker between that long tail and the agent layer, the moat is structural.

Two risks worth flagging. First, the big labs might just route around them. OpenAI has signed direct deals with Reddit, the Financial Times, Axel Springer. Anthropic is building its own. If marquee data ends up exclusively pre-licensed to model providers, Redpine's marketplace gets thinner.

Second, the legal layer is moving. Whatever the EU AI Act and the next round of US training-data litigation produce, Redpine will need to comply with it instantly. That's a regulatory ops burden you can't outsource.

Both risks are real. Both are also the reason this round priced where it did. If the company gets the model right, it's the kind of asset a Bloomberg or a Refinitiv ends up buying.

Watch the next twelve months. The data layer is where AI's real margin lives. Redpine just put a flag in it.

Source: TNW