What Oxylabs sells, and why AI wants it
On 9 July 2026, Oxylabs, a web-data company founded in Lithuania in 2015, took its first outside investment: 130 million dollars from the private-equity firm Warburg Pincus, at a valuation of 3.6 billion dollars. For eleven years the company grew without external capital and now reports 350 million dollars in annual recurring revenue from more than 350,000 tech teams. The headline number is large; the reason behind it is the more useful part.
Oxylabs does not build models or apps. It sells the plumbing that lets software read the public web at scale: large proxy networks that route requests through many IP addresses, scraper APIs, headless browsers, and ready-made datasets. Its own pitch is blunt. As chief executive Vytautas Savickas put it, the next generation of AI will not run on static indexes, and the future belongs to the live infrastructure that grounds these systems in real-time knowledge.
Your agent does not read the open web
When an AI agent looks something up online, it rarely touches the open web the way a person does. At any real scale, automated web reads get blocked, rate-limited, or served different content than a human sees. To get around that, agent builders route requests through commercial proxy and scraping networks, the exact layer Oxylabs sells. The free browser tab you picture is not what your software uses.
That matters because agentic workflows multiply web reads. One person checks a supplier price once; an agent checks a thousand suppliers on a schedule. The more of your operation you hand to agents that gather live data, the more of it quietly depends on a paid access layer you did not choose and may not even see on an invoice yet.
A bootstrapped toll booth, now institutional
The signal in this deal is not the dollar figure but who paid it and why. Warburg Pincus put primary growth capital into a business that took no outside money for eleven years and already runs at 350 million dollars in recurring revenue. Institutional investors do not price a self-funded, cash-generating infrastructure firm at 3.6 billion dollars unless they think the toll it collects is durable and hard to route around.
Read plainly, the bet is that web access for machines becomes a metered utility, not a free good. That is a different claim from any single model release. It says the value is moving toward whoever controls reliable, compliant access to live web data, and that this position is defensible enough to underwrite a valuation most AI application startups never reach.
The compliance bill arrives with the convenience
Reliable web access at scale is also a legal surface, not just a technical one. Large-scale scraping runs into website terms of service, bot-blocking arms races, and data-protection law the moment the pages contain personal data. Under the GDPR, a European operator whose agents collect personal information from the web is a data controller for that processing, whoever runs the proxies. The convenience of the agent just checks online does not move that responsibility to the vendor.
There is a concentration cost too. If a handful of networks supply most reliable web access, an outage, a price rise, or a legal setback at one of them ripples into every workflow that leaned on it. A dependency you cannot see on a diagram is one you cannot plan around.
What to check before you scale agents
Treat live web access as a supply chain, because that is what it now is. Map the steps in your AI workflows that read the open web, and note which ones route through a third-party proxy or scraping vendor. Put a price on that dependency, including what happens if the vendor doubles its rate or drops a source. Then check the legal footing: is the data personal, is the scraping within terms, and who is the controller if a regulator asks.
None of this argues against using agents. It argues for knowing what they stand on. The firms that scale agentic work without nasty surprises will be the ones that treated web access as procurement, with a contract, a cost, and a compliance owner, long before it showed up as a problem.
Read next: An AI Just Ran a Whole Ransomware Attack Alone | TSMC Now Keeps 60 Cents of Every Dollar
