Building a Niche AI Agent? The Real Money Is in the Data It Needs.

01 /The Concept

Redpine offers developers a way to "power their AI agents with premium data" – data the startup licenses officially from the original rights holders.

"Premium" here means scientific papers, research findings, textbooks, news archives, and multimedia content like films and video recordings. The target buyers are niche AI agents, enterprise AI agents, AI assistants, and AI labs building complex products.

Examples featured on Redpine's site include: charts tracking pharmaceutical company mention trends, real-time flight departure data from major airports over the past 24 hours, and sentiment shift data for S&P 500 companies over the past week.

AI agents and model-training platforms can access Redpine's data via API or MCP, paying only for what they actually consume – starting from $1–2 per thousand tokens.

Redpine was founded in 2024 in Sweden. It raised an initial $1.3M last autumn and has now closed a new $8M round backed by angels from OpenAI, Perplexity, and Spotify.

02 /Why It Matters

Redpine's core claim – "AI is only as good as the data it uses" – is simply true. AI providers crawl millions of websites and cut licensing deals even for publicly accessible content. Reddit, for instance, reportedly earns around $100M per year from data licensing.

But as Redpine points out, publicly available content represents only about 1% of all data. The most useful data is locked inside private sources: industry databases, paywalled research archives, proprietary catalogs. Some estimates put open data at closer to 4% of the total – but whether it's 1% or 4%, the gap is enormous. The data scarcity problem is real, and it specifically exists because of AI.

Before AI, the people who needed specialized data could negotiate individual access – paying subscriptions, licensing on a case-by-case basis. At the scale AI requires, that approach collapses. No AI agent can subscribe to every relevant source in the world.

One of Redpine's founders previously worked at Spotify, which is why the company frames itself as "the Spotify for data" – and the analogy holds up.

Before Spotify, no individual could buy every album ever made and then choose what to listen to on demand. Spotify solved that by shifting from "purchase" to "revenue share" – artists and labels get paid per stream rather than per disc sold. The elegance of Spotify's model is that revenue sharing is calculated against fixed monthly user payments – a structure that may eventually emerge in data licensing too.

Spotify's early job was somewhat simpler: it needed to sign a small number of major labels that controlled rights to most popular music, after which smaller labels and independents followed. Redpine's job is harder – the data market has no equivalent "big four." Fragmentation means negotiating with thousands of providers. Difficult doesn't mean impossible, though.

Which raises a strategic question worth noting: why hasn't Redpine started by systematically dominating specific niches? The playbook writes itself – accumulate critical mass of data in one vertical, watch remaining providers in that vertical join to avoid missing out, attract AI developers who want the aggregated dataset. Then move to the next niche. Repeat. It's essentially how Facebook conquered the world one university at a time.

The niches that would benefit most are already visible. A [related review](/review/v-kazhdoj-teme-pojavitsja-svoj-analog-chatgpt) covered OpenEvidence, which raised $767M – including $250M at a $12B valuation – building a "ChatGPT for doctors" powered by officially licensed data from authoritative medical sources. OpenEvidence won't be the only startup needing licensed medical data. Every player building an AI platform in healthcare will hit the same wall.

Or consider Aemon ([covered here](/review/nuzhno-prosto-nachat-jeto-sobirat)) – a Y Combinator graduate – which built an AI programmer whose edge comes from ingesting fresh scientific papers to apply new AI algorithms to existing codebases. A fair guess is that those papers are being scraped rather than licensed – which works fine while Aemon stays small and under the radar.

03 /Opportunities

Zoom out and the opportunity space is vast. In virtually every vertical, there are proprietary data sources that could dramatically improve the performance of AI platforms operating in that space.

So even without reaching for world domination, the actionable direction is clear: identify non-public data sources in a specific niche, negotiate licensing rights with those providers, and resell that access to AI developers building in that space.

At minimum, you stay a focused data broker in one niche. A more ambitious version runs the same playbook across several high-value, high-growth verticals. Or you start with the niches where you already have relationships – where you know who holds the data and have the credibility to get a deal done.

The most defensible starting position is a niche where you already have relationships – where you know who holds the data and have the credibility to get a deal done. Existing domain networks are the real asset here, not the technology.