Datacurve Review: Is Datacurve Legit, and Does It Buy Code?

Is Datacurve legit? The short answer

Yes. Datacurve is a venture-backed company with named investors, a named lead investor and published research. It is a supplier to AI labs, not a buyer of company data. It pays individual software engineers to produce training data, then sells that data, along with benchmarks and environments, to the labs.

That distinction answers most "Datacurve" searches from business owners. If you are an engineer wondering whether its bounties are real, the evidence says yes. If you run a company and want to sell your codebase, Datacurve publishes no route to do that. The sections below cover both readers and set out where an owner can license code instead.

What Datacurve is: funding, products and customers

Datacurve went through Y Combinator's Winter 2024 batch and describes itself there in one line: "We generate expert quality coding data at scale for fine-tuning LLMs" (Y Combinator). In October 2025 it "announced a $15 million Series A round, led by Mark Goldberg at Chemistry with participation from employees at DeepMind, Vercel, Anthropic, and OpenAI" (TechCrunch). Maginative put total funding at $17.7 million after a $2.7 million seed (Maginative).

Its homepage calls it "the research and data collection infrastructure that will teach AI how to shoulder the work humanity has long endured" (Datacurve). Its products page lists reinforcement-learning environments, long-horizon tasks, ready-made datasets, benchmarks, supervised fine-tuning data and agent trajectories, which it describes as "Full traces of expert execution, including tool calls, checks, pivots, and recoveries" (Datacurve). Its environment domains are software engineering, data science, cyber security and machine-learning research.

Its public research includes DeepSWE, a long-horizon coding benchmark whose "reference solutions averaging 668 lines of code" are far larger than older benchmarks (Datacurve research). On customers, its lead investor says Datacurve has "signed multimillion-dollar contracts with leading AI labs" (Chemistry).

Source: TechCrunch: Datacurve raises $15 million to take on Scale AI (Oct 2025)

How Datacurve gets its data and who it pays

Datacurve sources data from people, not companies. TechCrunch reported that it "uses a "bounty hunter" system to attract skilled software engineers to complete the hardest-to-source datasets", and had paid out "over $1 million in bounties so far" (TechCrunch). The platform is Shipd, which invites engineers to "Compete on hard problems across STEM, earn bounties, and produce the data that drives LLMs forward".

Maginative reported that Datacurve "draws on a pool of over 14,000 contributors and pays for outputs, not hours" (Maginative). Its co-founder was candid about the pay in the same TechCrunch piece: "the pay will always be far lower for data work than conventional employment". Treat Shipd as paid side work, not a salary.

Does Datacurve buy source code from companies?

We found no evidence that it does. Its website, products page and research page describe what it sells to labs and how engineers contribute, and none of them describes buying code, git history or business records from a company. Its contact route is a general "Get in touch" for customers.

Maginative described a "Private Repo Taskbench" for enterprise codebases (Maginative). That points the other way: a company with private code pays to have models evaluated on it. If someone presents a Datacurve offer to buy your repositories, ask for it in writing from a datacurve.ai address and check it against the steps in how to vet an AI data buyer.

Datacurve alternatives for companies that want to license code

If your goal is to be paid for code you own, several programs do publish terms. They differ in what they price and how much context they want around the code.

• Per-token code licensing: Pangea.ai pays about $100 per million tokens of source and $40 per million of git history, non-exclusively.
• Codebase plus company records: Turing's Project Lazarus buys codebases with the tickets and documents around them.
• Workflow data including GitHub: the Mercor data program connects GitHub with Slack, email and other tools.
• Environment builders that need realistic repositories are listed in our RL environment companies tracker.
• For what makes code valuable and how to prepare it, read can a company sell its source code to AI companies.
• For the wider market, see our directory of AI data companies.

Why labs still want company code, even from suppliers like Datacurve

Datacurve's model shows what labs pay for: hard, verifiable coding work done by skilled people. Its contributors write new solutions to set problems. What they cannot easily produce is the record of a real product built over years: the commit history, the failed approaches, the tests that caught regressions, the tickets that explain why a change was made.

That record is what a company with an old codebase holds, and it is why some buyers pay for private repositories. Coding agents now work on long tasks, and Datacurve's own DeepSWE benchmark uses reference solutions averaging 668 lines of code. Environments for tasks of that size need realistic codebases to sit in. Suppliers can build them from scratch, buy them, or license them.

For an owner, the practical point is that the value of code depends on its context. A repository with history, tests, continuous integration and linked issues is worth more than the same code as a single snapshot. Before you approach any buyer, check what you own outright, what history survives, and whether client contracts limit reuse. Can a company sell its source code to AI companies walks through those checks.

Datacurve vs Mercor, Surge AI and Scale AI

TechCrunch framed Datacurve's raise as a bet against Scale AI, after Scale's founder left to run AI at Meta, and named Mercor and Surge as the other fast-growing suppliers (TechCrunch). All four sell expert data to AI labs. They differ in scale and in whether they deal with companies as sellers.

Datacurve is the smallest and the most specialised: coding first, with contributors paid per accepted output. Mercor runs a published program that pays companies for workflow data, including GitHub, alongside its expert network. Surge AI buys company archives for its RL gyms but publishes no seller program. For an owner, only the buyers with published seller terms are a direct route; the rest are customers of the data, not buyers of yours.

Source: TechCrunch: Datacurve raises $15 million to take on Scale AI (Oct 2025)

Is Datacurve worth it for engineers?

For a skilled engineer with spare hours, the bounty model can pay for hard, well-defined problems. Check three things before you invest time: what an accepted output pays, how review and rejection work, and who owns what you submit. Bounty pay rewards results, so a rejected submission earns nothing.

If you are an employee, also check your employment contract before you submit anything resembling your employer's code. Shipd problems are designed to be self-contained; your employer's repositories are not yours to license, to Datacurve or anyone else.

Next step

If your company owns code with years of history and you want to know which buyers would pay for it, start with an estimate rather than a cold offer.

Not sure whether a buyer or an offer is sound? Origin checks who the buyer is, what the license allows and how the price compares with published deals before anything is signed. Get a free estimate and a second opinion.

Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.