Labelbox Raises Over $100M at a $1.25B Valuation to Scale the Infrastructure Behind Smarter AI Agents

The quietest company in the AI boom just got a $1.25 billion price tag, and you have probably never clicked on anything it built. Labelbox, the San Francisco company that started out labeling the data that trains AI models, announced on October 14 that it has raised more than $100 million in Series E financing. The round, a mix of equity and a credit facility, was led by Infinitum Partners with participation from Wipro Ventures, Andreessen Horowitz, B Capital, Tau Ventures, and Snowpoint Ventures, according to the company's announcement carried by PR Newswire.

The reported pre-money valuation, $1.25 billion, comes from coverage by Unite.AI of a company blog post by CEO and co-founder Manu Sharma. What makes the number interesting is the financial shape of the company carrying it. Labelbox says it reached profitability in 2026 while growing more than fivefold year over year, and that it now serves more than 400 customers, including over 90 percent of leading U.S. AI labs. In a market where billion-dollar valuations routinely attach to companies with barely any revenue, a profitable, fast-growing infrastructure company raising a nine-figure round is the anomaly that tells you where the AI economy is actually heading.

From data labeling to teaching machines to practice

Labelbox was founded in 2018, back when the binding constraint in AI was labeled training data. Its original business was the platform where teams annotated images, text, and video so models could learn from examples. The Series E announcement makes clear how far the company has traveled from that origin story. The new money is earmarked for two products that have almost nothing to do with labeling and everything to do with what comes after training.

The first is Horizon, built for the frontier labs training the most capable models. Horizon creates reinforcement learning environments: simulated workplaces stocked with code, dashboards, spreadsheets, and documents, where a model attempts complex, multi-step tasks, gets scored on its performance, and receives expert feedback on which correct answers are actually better. The premise is that models learn the way people do, by practicing hard things and being graded, rather than by reading more text.

The second is Recursion Managed Agents, aimed at enterprises. Instead of selling AI tools that humans operate, Labelbox sells fleets of AI agents that operate inside the applications companies already use. The agents trigger off schedules, alerts, and tickets, coordinate across tools, evaluate their output against defined standards, and feed what they learn from each run back into memory so the next run goes better. Labelbox says it runs more than 50 of these agents across its own internal operations, from engineering and research to sales and finance, and it opened the product to enterprise customers earlier this month.

The bet behind the bet

The financing tells you what investors believe the next scarcity is. The round was led by Infinitum Partners, and the detail that caught the market's attention is who sits behind that firm. According to Unite.AI's coverage of Sharma's announcement, Infinitum's John Yetimoglu was an early investor in Groq and is now involved in building large-scale compute infrastructure at GroqCloud, and Sharma wrote that access to computing power will be central to Labelbox's next chapter.

In 2026, the investors worth having are the ones who can unlock compute. Money is abundant, but processing power is the bottleneck.

Read that as a market signal. Training and evaluating frontier models, and running the simulated environments where those models practice, burns enormous amounts of processing power. A lead investor with a direct line into compute infrastructure is not just writing a check. It is writing a check plus the one resource that money alone has trouble buying. In the current AI financing market, compute access has become deal currency.

Picks and shovels in the agent era

There is a pattern in this month's funding data. The largest checks are going to companies that build the layer underneath the models and the apps: the clouds, the chips, the power, the data, and now the training environments and agent harnesses. The week ending October 10 produced a $2 billion financing for Axiom, the cloud and power infrastructure arm that Flex is spinning out as an independent public company in early 2027, an $870 million Series A for TypeSafe AI, and hundreds of millions more for machine-native compute and quantum bets, according to Intellizence's weekly funding roundup. Labelbox's round belongs in that cohort.

It also fits a second pattern: investors paying up for the companies that make AI systems reliable enough to trust with real work. Models are commodities. What enterprises will pay for is the infrastructure that takes a capable model and turns it into an employee: the sandbox where it practices, the harness where it runs, the evaluation layer that grades it. That is the stack Labelbox is selling.

What profitability changes about the story

Fivefold growth with profitability is worth pausing on. The typical late-stage AI raise right now is a company buying time until revenue arrives. Labelbox is raising from a position where revenue has already arrived and then some. That changes what the money is for: not survival, but speed. Sharma framed the plan as pushing every layer, from sandbox infrastructure and the agent harness to reinforcement learning post-training, so that frontier labs and enterprises can build systems that get better with every run.

The customer list does the rest of the persuading. More than 400 customers, over 90 percent of leading U.S. AI labs, plus hyperscalers, startups, and Fortune 500 companies, per the company's announcement. When nearly every lab building a frontier model is paying you for the environment where that model practices, you have something close to a toll booth on the road to more capable AI.

The next time an AI agent reconciles your invoices or triages your support queue, it will have practiced that job thousands of times in a simulated office built by a company whose core skill is knowing how to grade a machine. Labelbox's $100 million raise is a bet that practice, not just parameters, is what makes AI useful. In a market drowning in model announcements, the smart money is buying the gym, not the athlete.