The Secret Supplier Feeding Frontier AI

Summary

• Grass’s advantage starts with getting AI systems onto websites that block data-center traffic. Andrej Radonjic says routing through residential IPs is the best way to get broad access, with browser infrastructure handling additional defenses. Grass pays consenting participants for bandwidth; Andrej says it has delivered exabytes of data to AI labs.
• Frontier labs increasingly buy useful signal and speed alongside raw volume. A lab’s discovery process can identify a small share of a dataset—for example, 0.16%—as useful, driving more targeted purchases. With new models resetting benchmarks every few months, customers risk being “left behind”; sophisticated labs estimate data’s value through benchmark improvements, subsequent token usage and financial returns.
• Robotics creates demand for carefully selected behavioral video. Andrej updates Yan’s roughly 1 billion figure to more than 100 billion indexed videos. Yan points out that indexing does not mean downloading every video. Generative video benefits from diverse examples that can be combined, while robotics currently needs footage that teaches actions and interactions—a harder filtering problem than simply accumulating video.
• Training crawls are building the infrastructure for live inference and search. Andrej describes training-data collection as a “treasure hunt” that also maps the web. Grass has released its Content API under Live Context Retrieval; Search is planned next, while he expects training and inference data eventually to converge.
• Inference makes regional capacity and latency increasingly important. Andrej previously thought Live Context Retrieval needed ten times as many nodes, but infrastructure improvements enabled an earlier launch than the previously discussed 2027 timetable. More residential nodes are expected as regional demand grows, and the “speed of light” constrains how close routed requests can get to original latency.
• Grass’s customer business and intellectual property sit within an ownerless foundation. “Nobody owns it,” Andrej says; Wynd Labs supplies services, employs staff and holds no IP. The revenue attestation reconciles foundation purchase orders with cash receipts, including payment delays such as 60 days; it concerns the foundation rather than a separate equity vehicle.
• Yan recounts a founder’s view that superior reinvestment returns can constitute a moat, and Andrej agrees. Yan cites the AI-data market estimate of $100 billion by 2030; Andrej sees expansion opportunities in synthetic traces, reinforcement learning and other data categories. The closing moat discussion centers on getting more out of each dollar than competitors do.

Deep dive

1. Residential access is the network’s starting advantage

• Andrej’s starting point was scraping websites for arbitrage bots during his applied-mathematics PhD. The pandemic moved intelligence gathering online: governments calculating CPI could no longer easily send people into shops with clipboards. Users expected product information to be public, while websites increasingly restricted automated access.

• The blocking mechanism catches legitimate agents too. A website sees an AWS-hosted agent’s data-center IP and cannot readily distinguish it from a competitor scraping prices, so its default can be exclusion. Andrej says routing through millions of residential IPs is the best way to get broad access; some sites additionally require browser execution and different fingerprints.

• Consent is Grass’s organizing principle: participants deliberately install an application, share bandwidth and receive rewards. Andrej contrasts that arrangement with proxy software quietly inserted into televisions or software updates, which he considers immoral and commercially unsustainable. Yan discloses that Delphi Ventures invested in Grass and holds its token.

2. Training crawls build both a data factory and a web map

• A deployed model runs into the limits of its knowledge immediately: it has not learned what happened after training. Yan asks whether continuous learning would reduce training-data importance; Andrej expects training and inference data to converge over time. He says research on continuous learning is exciting and feels it will be resolved soon. Yan suggests that continuous-learning frameworks at huge scale may be available in a few years.

• Andrej describes training collection as a “treasure hunt”: crawlers seek useful material while incidentally mapping the internet. He cites billions of URLs collected within hours by relatively small clusters. That index supplies foundations for search, while Grass’s access infrastructure retrieves website content without being blacklisted or caught in honeypots.

• The physical operation includes thousands of compute servers, many petabytes of storage and exabytes delivered to AI labs. New capacity can fill within hours. For one urgent leading-lab delivery, CTO and co-founder Chris went to Ashburn, stayed awake overnight and slept on the floor while assembling servers. Storage also supports filtering and processing raw material into more useful datasets.

• Customer priorities depend on the model: video, code, rare dialects or difficult websites can each matter. Andrej emphasizes time pressure as new models surpass benchmarks every few months. Taking the best parts of a huge dataset and weighting them more heavily during training can make training more effective; Grass is working to help identify that signal.

3. Robotics makes video selection a research problem

• Yan’s reference to roughly 1 billion public-web videos is outdated: Andrej says Grass has indexed more than 100 billion. Yan points out that indexing does not mean blindly downloading every video. Video is expensive to store and process, making targeted selection essential before the material becomes a useful training product.

• Andrej explains that when researchers know the task and have relevant data, progress is easier to measure through a “loss curve.” Teaching Spanish points toward Spanish-language material. Yan offers the hypothetical of a robot learning from someone using a lawn mower—how they walk, move their hands and handle objects. Andrej gives another example: a model could watch someone pour water into a cup and learn how to do it.

• Andrej contrasts Sora, which can learn from diverse videos and combine them—for example, producing a cartoon cat in space—with robotics, where useful footage needs to be both diverse and specific to tasks. Yan imagines a robot learning from millions of examples; Andrej says gameplay might help teach environmental interaction, while cartoons or music videos currently offer little value for most robotics tasks. Determining and filtering the right behavioral footage remains active research.

4. Live retrieval turns internal infrastructure into a product

• Live Context Retrieval is the product family; the released product is the Content API. It packages infrastructure originally built for large training crawls into one call, handling residential routing, complex TLS fingerprints and server-side browsers. Andrej says residential IP networks alone are often insufficient for some tasks.

• Users provide a URL, optionally specify a country and receive content as Markdown or plain text. Yan’s example is an agent testing a stock-trading strategy using information from around the world, including product and stock details from sites that may only be accessible in certain countries, to calculate fundamentals before earnings reports.

• Search is planned next because agents also need to discover which URLs matter. Andrej puts a typical internet-enabled agent call at five to 50 pages, with deeper research potentially requiring more. The Content API suits businesses monitoring known websites; Search would let a model locate relevant sources before retrieving their contents. Audio and image grounding remain in development.

• Andrej argues that accessible ANNs, embeddings, information-retrieval tools and open-source models have made parts of search more available. His deliberately simplified example is searching 50 million documents in an AWS S3 bucket for under $100 monthly. He stops short of saying those tools are commoditized or that Grass has already built an excellent internet search engine: reliably collecting and refreshing blocked web content remains difficult.

5. Inference changes the network’s capacity requirements

• Yan cites nearly 8 million nodes as the latest public figure he knows. Andrej says inference strengthens the case for continued residential-node growth: training collection can wait when a region fills up, while an agent needing immediate information from that location has less flexibility. Capacity must therefore be available where requests arise.

• Andrej acknowledges revising his own timetable. He previously believed Grass needed ten times as many nodes before launching Live Context Retrieval; optimization revealed that fewer nodes could do more work. Yan recalls earlier discussions pointing toward 2027. Andrej expects further efficiency gains, while recognizing that scaling eventually encounters diminishing returns.

• Andrej reports a Content API success rate very close to 100%, meaning it can find a way to collect content from a public website. He considers that more meaningful than boasting about pages crawled, given how much web information is useless. Latency remains unfinished work: residential detours face the “speed of light,” and Grass wants performance as close to the original as possible.

6. Labs price capabilities and protect their research directions

• Lab demand alternates between broad discovery and specialized collection. A customer may request a large domain-specific dataset, evaluate it, then find that about 0.16% is useful for changing results and ask for more. Andrej says training-data demand has grown significantly despite changing research directions.

• Yan presses on cost versus quality; Andrej says both matter, with speed sometimes commanding a premium when delivery is needed in three hours rather than the next day. More advanced labs model the financial benefit of improving a benchmark, the additional token usage that capability could generate and an appropriate price per data unit. Useless deliveries can get vendors blacklisted or sidelined.

• Labs may have separate teams focused on text, audio or vision, with procurement operating separately. Some may already use a vendor for a particular task, but Grass can consult with them about other relevant data it has collected. Asked by Yan about defense or government applications, Andrej says customers do not discuss their work publicly because others could use that information against them; he does not describe specific Grass work with those organizations.

• Many data vendors build reinforcement-learning environments—a “jungle gym” in which models practice particular benchmarked skills. Andrej expects continued demand for them, but suppliers must track changing research priorities. His argument for Grass’s broader infrastructure is that customers still need web access when they move from one capability or data category to another.

7. The foundation holds the business while reinvestment drives the thesis

• Andrej describes the revenue attestation as an accounting firm’s reconciliation of purchase orders from AI labs and other customers against cash receipts. Timing differences can arise when invoices are paid 60 days later. The attestation concerns the Grass Foundation, an organizational distinction he says many observers missed.

• Customer relationships run through the foundation, which Andrej describes as ownerless: “Nobody owns it.” Independent directors oversee its actions. Wynd Labs supplies services and handles employment, benefits and insurance; it holds no IP, which transferred to the foundation at token launch. Wynd–foundation dealings require director approval, and Wynd’s team also holds tokens as part of the intended incentive alignment.

• Yan cites an estimate that the AI-data market will be worth $100 billion by 2030. Andrej sees opportunities for Grass across pre-training, inference, preference data, synthetic traces and reinforcement learning. Large raw datasets can support synthetic-data generation; better web access can improve learning environments. He also sees further uses for the existing consumer distribution, while acknowledging that some categories are not current priorities.

• Asked about capital, Andrej says Grass will predictably need more. Yan then recounts an unnamed AI-lab co-founder’s view that a moat can be the ability to get more out of each dollar invested than competitors. Andrej calls that an excellent moat and agrees that getting more out of an investment is an interesting idea.