The Secret Supplier Feeding Frontier AI

Grass positions itself as data infrastructure supplying frontier AI labs with public-web information, combining millions of opted-in nodes with compute servers and storage. The host cited eight million nodes as the latest publicly shared figure; Grass says it operates thousands of compute servers and many petabytes of storage, and has delivered exabytes of data to major labs. Its network routes requests through residential IP addresses, which can help reach sites that block data-center traffic. Grass says it can also use browser infrastructure to load JavaScript and vary fingerprints when websites require it.

The company says consent is central to its model: users opt in to share bandwidth in return for rewards. Andre contrasts this approach with companies that installed software without users’ consent and sold access to residential IPs, calling those practices unethical and unsustainable, though he expects they may persist. Most of Grass’s current business is collecting web data for AI training, with growing emphasis on filtering and processing raw material into useful datasets. Lab requirements differ by model and may favor volume, quality, speed, or specialized material such as video, code, and less-common languages. Delphi Ventures disclosed that it invests in Grass and holds its token.

For training, Grass can target crawls by file type, subject, or language; deployed models may instead use inference data to retrieve current information. Andre expects training and inference data to converge as models gain continuous access to updated information, but says continuous learning remains unsolved and warns that too much context can cause hallucinations. Grass reports indexing more than 100 billion public web videos, while emphasizing that indexing does not mean every video was downloaded or scraped. Collection is targeted because video is costly to store and process. Andre sees potential for robotics models to learn actions from web videos, podcasts, and text, citing a Rota AI paper on scaling robotics behavior with web video. He says useful robotics data must be diverse yet task-relevant, and finding suitable videos at scale is difficult.

Grass developed its Content API for internal data collection and says it now lets a customer submit a website URL and receive content in a chosen country and format, including Markdown or plain text, with token reduction for LLM use. The recently released API extracts content from public-web inputs with a success rate “very close to 100%,” according to the company. A planned Search API would help agents locate relevant URLs before retrieval; the broader planned live-context suite includes search, with multimodal capabilities in the pipeline. An agentic web-access cycle may involve visiting five to 50 pages, sometimes more. Andre says search tools have become more accessible, but avoiding blocks while collecting web data remains a major barrier for new search engines; open-source tools alone do not create a state-of-the-art search product.

The network’s potential uses extend beyond training crawls to inference, preference data, synthetic traces, real-world data, and web access in reinforcement-learning environments. Grass says useful content identification and crawler deployment matter more than raw page counts, and suppliers may keep their contributions confidential because a lab’s data focus can signal capabilities to competitors. Growth has come from individuals downloading the app and referring others, but routing through millions of user-device IPs adds latency, which remains unresolved. Training crawls can tolerate delays when a region’s nodes are saturated; inference needs sufficient nodes in the relevant location. Grass expects demand to drive network expansion and says infrastructure optimizations let it launch live-context retrieval products with fewer nodes than expected, though further gains will eventually diminish.

Andre says data trends can indicate the industry’s direction six to 12 months ahead, while labs may shift between large raw-data purchases and specialized datasets. Reported useful shares have ranged from 0.16% to all of a dataset, and overall training-data volumes have grown even as individual requirements differ. Buyers weigh quality, cost, and delivery speed; urgent data may command a premium, while low-quality vendors may be blocklisted or ignored temporarily. The interviewer cites estimates that data spending could approach $100 billion by 2030. Andre describes a major technology supply crunch and says, for the first time, the cost of running the same compute workload is rising.

Grass argues that owning its infrastructure can improve pricing power and lower vendor costs compared with web-access startups that pay third parties for proxies or access. The Grass Foundation owns the application’s intellectual property, has exclusive access to the app, and can route requests through it. Grass says the ownerless Foundation faces customers and is governed by independent directors who must approve transactions; Wind Labs is a service provider handling employment and operations, without owning the intellectual property or facing customers. Grass says the IP was transferred to the Foundation around the token launch and that an accounting firm reconciled purchase orders with cash receipts from large AI labs and other customers, finding a small difference caused by delayed invoice payments, including some paid after 60 days. The company is developing specialized offerings in robotics, multilingual content, legal, finance, math, and code, generally focusing on one vertical at a time. Its ambition is to provide a medium through which AI can access information for agents, training, learning, and data synthesis; Andre says future gathering could involve satellites, while today Grass uses internet IPs as “digital satellites.”