Scale AI closes a $1 billion Series F at a $13.8 billion valuation

Scale AI supplies the data used to train AI models, along with the measurement and evaluation systems used to check how well those models work. In May 2024 the company announced a $1 billion financing transaction at a $13.8 billion valuation, a round it called its Series F.^[1]

Background

Building a large language model takes more than collecting text from the web. One widely used step is reinforcement learning from human feedback (RLHF): people compare or rate a model's outputs, and those judgments are used to push the model toward the answers people prefer. The method only works if someone can collect large volumes of consistent human annotation and manage it at scale.

Scale's founder Alexandr Wang wrote that the company had partnered with OpenAI on what he described as the first RLHF experiments on GPT-2, and had then scaled those techniques to InstructGPT and beyond.^[1] That history is the context in which the new round was announced: the company's business is not a model of its own, but the data work that sits underneath other companies' models.

The transaction

The company announced in May 2024 that it had closed a $1 billion financing transaction at a $13.8 billion valuation.^[1] The financing was a mix of primary and secondary. In a primary sale the money goes to the company itself; in a secondary sale the buyer purchases shares from an existing holder, so part of the total goes to shareholders rather than into the business. Splitting a round this way lets a company raise new money while also giving early backers and employees a chance to sell some of their stake.

The round was led by Accel, which was already an investor in the company.^[1]

Who put money in

Nearly all existing investors took part, among them Y Combinator, Nat Friedman, Index Ventures, Founders Fund, Coatue, Thrive Capital, Spark Capital, NVIDIA, Tiger Global, Greenoaks and Wellington Management.^[1]

The new investors included Cisco Investments, DFJ Growth, Intel Capital, ServiceNow Ventures, AMD Ventures, WCM, Amazon, Elad Gil, Meta and Qualcomm Ventures.^[1] The list mixes venture funds with companies that sell the chips, cloud capacity and enterprise software on which AI systems run. NVIDIA was already an investor; AMD Ventures, Intel Capital and Qualcomm Ventures joined as new ones, as did Amazon, Cisco Investments, ServiceNow Ventures and Meta.

What the company said it would do

Wang wrote that Scale would move into a phase of building frontier data for complex reasoning, agents and multimodality, along with measurement and evaluation systems.^[1] These are three directions in which model development has been moving. Reasoning refers to problems that take several steps to solve rather than a single reply. Agents are systems that call tools and take actions on a user's behalf instead of only producing text. Multimodality means handling text together with images, audio or other kinds of input. All three need training data that is harder to collect than plain text, and all three are hard to grade, which is why the company pairs the data work with measurement and evaluation.

Why it matters

The financing is a marker of where money in AI has been going: not only to the companies that train models and the companies that make chips, but to the layer that supplies training data and runs evaluations. The investor list reflects that position. A data company ends up with chipmakers, cloud providers and model developers all holding a stake, because each of them depends on the same inputs.

The secondary component is worth noting on its own. A round this size, at this valuation, gives existing shareholders liquidity without requiring the company to raise the entire amount as new capital.^[1] For a company whose work is tied to long research programmes at other firms, that structure keeps the cap table stable while bringing in new partners.

What the round does not settle is how the next stage of data work will be paid for and measured. On the company's own account, the answer is frontier data for reasoning, agents and multimodality, together with the evaluation systems that tell developers whether any of it worked.^[1]