All news

Scale AI's round prices training data as AI infrastructure

Racks of white ceramic ingots in a quiet foundry, a fresh cobalt ingot cooling in a mould

Scale AI doubles its valuation to $13.8 billion as model makers buy into their data supplier. Expert-made training data is becoming a strategic input.

Scale AI is now valued at $13.8 billion, about double the roughly $7 billion set in its 2021 round. The supplier of labeled and expert-written training data to the largest model developers closed a $1 billion round on 21 May that combines new shares with secondary sales by existing holders.

The round prices data as infrastructure. The scarce input for the next generation of models is no longer public web text, now a legal risk, but new data written or checked by experts: reasoning steps, code and multimodal examples. Whoever produces that data at scale sits between the labs and their next model, so value in the AI stack accrues to the data supply chain as well as to chips and models, and that chain is less visible and less crowded.

The cap table makes the case. Amazon and Meta join as new investors alongside several chip and software companies, and Nvidia returns from earlier rounds. Amazon and Meta both train their own models, so a stake in a supplier they depend on secures access and shows them where demand is heading. Scale counts Meta, Microsoft and the US Department of Defense among its customers, worked with OpenAI on early reinforcement learning from human feedback and now describes its goal as building a data foundry for AI. The secondary component carries a lesson of its own: late-stage AI companies can give early holders liquidity without an IPO.

Data work is moving up the value chain. Labeling road images for self-driving cars was a volume business, while writing step-by-step solutions in mathematics, law or medicine takes specialists and pays accordingly. That opens room for companies that build expert networks in one domain, tools that measure model quality, and platforms that turn a company's own operations into training data with clear rights.

Concentration cuts the other way. A handful of labs buy most frontier data, and any of them can bring the work in-house, so founders in this market should sign several model makers early and build evaluation tooling that makes them hard to replace. If more model makers take equity in their data suppliers, a services market turns into a strategic one. Prices for expert-written training data in reasoning and code give a direct read on how costly the next gains in model quality are, and the remaining question is whether governments, starting with defense, become a second large customer group for evaluation and testing.

Like compute before it, data production is becoming infrastructure for AI. The test in Fanaura's data-first thesis is an AI-native company that owns a hard-to-copy source of domain data or the tools to measure model quality, with customers beyond a single lab. The next gain in model quality starts with data that someone has to write first, and that writer is now a strategic supplier.

For founders

Building expert data pipelines or evaluation tools for model makers? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.