Databento raises $100M at $1.1B: the market data five AI labs buy
The vendor that sells tick-level prices to quant funds turns out to be infrastructure for frontier AI — and it just got priced like it, with fewer than 40 employees.
Databento raised $100 million at a $1.1 billion valuation and says it now holds enterprise contracts with five of the six largest frontier AI labs, the company announced on November 12. Fifth Down Capital, an investor since the company's earliest days, led the round. Fitch Ventures — the venture arm of the credit-ratings agency — came in as a new strategic investor, and NEA, DRW Venture Capital, Tribe Capital, Redpoint Ventures, Hedosophia, Trousdale Ventures, Cross Creek, Operator Collective, AVG and TGVP all followed on.
The product is plumbing for price discovery, and it is more physical than the pitch suggests. Databento sells real-time and historical market data over one API: tick-by-tick trades and quotes, full order-book updates, equities, futures, options and indices, plus corporate actions, a security master for mapping identifiers across vendors, and lossless packet captures of the raw feed for firms that want to replay a trading day exactly as the exchange emitted it. Until recently the buyer was a quant fund, a prop shop or a brokerage — someone who needed to know what the book looked like at 09:31:07.412 and could not afford to be told "approximately."
A market-data customer used to mean a hedge fund. That is no longer true. Databento entered 2026 with a customer base of quantitative traders, brokerages, asset managers, commodities firms and digital-asset companies, and spent the year signing frontier labs. It now describes itself as a leading market data provider for AI and agentic workflows. The labs are not named, and neither are the contract values or terms — more on why that matters below.
There are three reasons financial tick data is unusually useful to a model builder, and none of them is "train a trading bot." The first is point-in-time integrity. An agent asked what a stock did when a headline crossed needs a record that is timestamped as of that moment; if you hand it a scraped snapshot, you have handed it the future, and lookahead bias is the oldest bug in quantitative research. An eval harness inherits that bug instantly and silently.
The second is normalisation. The numbers are not the hard part — the same trade prints differently across venues, formats and symbologies, and reconciling them is weeks of work per feed. Consistent timestamping across sources is the thing Databento sells hardest, and it happens to be exactly what an agent needs before it can reason about a price at all. The third is verifiability: market outcomes are public and dated, which makes finance one of the few domains where a model's answer can be graded against ground truth instead of against another model's opinion.
The efficiency claim is the part worth arguing about. Revenue has "more than doubled" since the Series B the company closed earlier this year, and Databento says it was not out raising — its existing investors pushed the round after watching the numbers. That Series B was $97 million, and per Rebellion Research's write-up of the announcement, the two rounds together put $197 million into the company within months while most of the first tranche was still unspent.
The valuation works out to roughly $28 million per employee: Databento says it reached $1.1 billion with "just under 40" people, against an industry where reaching this scale has historically taken hundreds. The same write-up reports more than 130,000 customers. Treat that number for what it is — a self-serve signup base on free credits, not the enterprise contracts that justify the price. The distinction is the whole business: signups cost almost nothing to acquire and churn quietly, while one frontier lab contract is worth more than tens of thousands of them.
Market data is capex wearing an API's clothes. In the past four months alone Databento says it has nearly completed infrastructure buildouts at nine data centres — ASX in Sydney, HKEX in Hong Kong, JPX in Tokyo, the London Stock Exchange, Equinix LD4 in London, NY3 in New York, ZH4 in Zurich, Nasdaq Väsby in Stockholm and Aruba IT3. That list is the actual moat argument: every API call is answered by a server in a cage within metres of an exchange matching engine, and the geography takes years and permits to assemble. It is also why the incumbents held the market for so long, and why a 40-person company owning nine of those sites is a stranger story than the valuation.
The stated next goal is to build the largest commercially available historical market-data archive in the world and extend the same tick-level depth to a much broader set of markets. That is a multi-year commitment to buying, cleaning and storing data nobody has bothered to normalise yet — a bet that the archive itself becomes the product, sold to labs that want one copy of history rather than forty inconsistent ones.
Three things to hold against the announcement. First, "five of the six largest frontier AI labs" is a company disclosure with no names, no contract lengths and no revenue attached; an enterprise contract can mean a production dependency or a nine-month pilot, and readers cannot tell which. Second, "revenue more than doubled" is a growth rate without a base — a doubling from a small number is not the same business as a doubling from a large one, and Databento is private. Third, the cost of goods is not servers. Vendors license the underlying feed from the venues, and exchanges set those fees on their own schedule; a challenger whose pitch is cheaper access to the same ticks is exposed every time a venue reprices, and Bloomberg, LSEG and ICE still own the entitlements and the terminal distribution that go with them.
Why it matters beyond one funding round: the data layer is where AI's cost curve is least cooperative, and labs are discovering that the cheapest way to get a domain right is to buy the version someone else already cleaned. We saw the same logic in Amazon's World Product Graph, where an internal document conceded that product answers were often wrong because the catalogue underneath them was — and the fix was data ownership, not a better model. If five of six frontier labs genuinely route their financial reasoning through one vendor's timestamps, that vendor's archive becomes a toll booth on every financial agent built on top of it, and the interesting question stops being whether $1.1 billion is a lot and starts being who else can assemble the same history.
What to watch: whether any lab goes on the record as a customer, which would convert a disclosure into a reference account; whether Fitch's involvement turns into credit data — ratings, default histories, structured-finance performance — delivered over the same API, which is a distribution deal waiting to be announced; and whether the archive ships before exchange fee schedules move.
If you build agents that touch money, would you trust a vendor's timestamps or reconcile to the raw feed yourself? Write to us via our contact page.
