The Read Layer for Machine Intelligence

Today, we are announcing an investment in Grass, the read layer for machine intelligence, from our hedge fund and venture fund.

Andrej Radonjic recently published his plan for the Grass Network to generate Abundant Intelligence. We believe Grass has earned the right to pursue that ambition, and that it is among the clearest examples of a DePIN working at commercial scale.

Over the past eighteen months, Grass has deployed finely tuned crypto incentives to recruit contributors to share underutilized bandwidth on their internet connections. This network of contributors has served as infrastructure for querying and parsing the public internet, and has allowed Grass to quickly make its way into the data pipelines of frontier labs. These engagements have required gathering enormous quantities of information, making the aggregated state usable for model training, and delivering it reliably to customers whose requirements change with every new training run.

We believe the market understands Grass as a proxy network that serves AI labs via pretraining data, but has yet to price in the enormous expansion that is inference-time data retrieval. The Grass team appears to be uniquely positioned to serve the demand for fresh information whenever a model searches the web to answer a question or an agent retrieves context before acting, a market likely several orders of magnitude larger than pretraining data because these requests recur throughout a model’s deployment (as opposed to pretraining datasets that are one-time deliveries for discrete runs).

Agent systems need to search and retrieve information every time they act. The cost and latency requirements of doing so effectively require reconstructing an index of the web, one of the most capital-intensive undertakings historically pursued only by players like Google and Microsoft. As the internet grows recursively through the actions of agents, live retrieval becomes increasingly critical infrastructure. We believe that Grass has a structural cost and access advantage in building and serving this, the relationships with the labs that will consume it, and the technical ability to make it all useful to machine intelligences.

Past: Pretraining Data

Large portions of the public internet are inherently difficult for conventional AI labs to access during pretraining. Websites rate-limit automated traffic, present different content across geographies, and increasingly require pages to be rendered through ordinary browsers. Grass addressed this by routing requests through millions of residential connections whose owners opted into the network and are compensated for their contribution.

Unlike many of the proxy networks of the prior era, Grass has not needed to manufacture hardware, acquire residential connections through wholesale deals with ISPs, or infiltrate consumer devices without their knowledge. Users across the world already pay for these connections, and much of their capacity goes unused. Tokens helped Grass recruit contributors and build geographic coverage before there was enough customer demand to pay for it, and the network quickly turned this aggregated resource into something AI labs could rely on for pretraining.

Grass must deliver enormous volumes of data on tight schedules, in formats that labs can use directly in their training pipelines. Each delivery requires close coordination with lab teams and gives Grass a better understanding of what they will need next.

In our 2023 essay on DePIN network design, we described “threshold scale” as the point at which a network’s distributed supply becomes commercially useful. Tokens are particularly powerful before that point because they can reward people for contributing to infrastructure that does not yet have enough coverage to attract customers. The real test comes afterward: whether the network can sell a reliable service and eventually pay for the supply it uses.

Grass is among the few networks we have come across that has passed that test. Over 6 million contributors received dollar-denominated rewards, funded solely through revenues generated as a direct consequence of their contributions. Grass reported $17 million of revenue in 2025 and another $17 million in the first half of 2026, with repeat business from nearly every AI customer it serves. It describes the business as profitable and has guided to $75 million of 2026 revenue from training data alone.

Present: Live Context Retrieval

The current training data business is substantial and we believe it should continue to grow. Frontier labs require vast quantities of new information for increasingly larger language models. The nature of this product, however, is inherently “point-in-time.”

If you ask a model what a company announced this morning, or whether a flight has been delayed, or what a website says right now, the information stored in its weights will not be enough. It has to search for a source, open the relevant page, and read it live. An agent working through a more complex task may repeat that process several times before it can answer.

We believe the eventual market for inference-time search and retrieval could be several orders of magnitude larger than the market for pretraining web data. Training purchases arrive around model runs, but live information can be consumed across every model, application, and agent whenever a task requires knowledge of the changing world. At historical publicly listed search API prices, even a fraction of that request volume could support billions of dollars in annual revenue.¹

At the same time, the internet itself is about to get much bigger. More information will be published in forms that traditional search engines handle poorly, including video, images, spreadsheets, PDFs, and dynamic pages. Making this expanding body of information discoverable and parsable for machines is an enormous opportunity, and one that is extremely technically nontrivial since it requires continuous crawling, storage, processing, ranking, and fast delivery at global scale.

Grass is building toward that market from a position few new or existing entrants have. Its proposed Contents API would allow a model to open and render a public page through the residential network, while its planned Search API and web index would help the model determine which page to open in the first place.

Both products build on machinery Grass already operates for training customers: routing requests through residential nodes according to geography and reputation, measuring the quality of the responses, and turning what it collects into datasets large enough for frontier training runs. Grass has collected more than a billion public web videos, a subset of which was used to train a video annotation model with another Multicoin portfolio company, Inference.net. Search ranking and serving results at inference latency are further engineering challenges, but the collection and processing systems are already running at commercial scale.

Building an internet index of this scale requires substantial investment of capital and engineering resources. This is always the cost of building deeply moated, high margin, price setter style infrastructure. We believe that every dollar of gross profit directed away from this opportunity is detrimental to the long term value of the business, and that pursuing this market is a much stronger net positive for token value accrual in the long run (in lieu of subpar capital allocation strategies suggested by countless CT analysts). As such, we think GRASS is on the precipice of a substantial rerating on the back of this new line of business.

What compounds here on a fundamental basis is Grass’s relationship with the labs. Every large training delivery gives Grass a reason to improve its coverage and a better understanding of the information its customers need. Better coverage improves the index; a better index improves live retrieval; and successful retrieval gives those same labs another reason to build on Grass. In time, this infrastructure can source scarce data for pretraining, posttraining, and inference, regardless of how model architectures evolve.

Future: Abundant Intelligence

The internet agents need to read will grow faster than the one search engines were built to index. Agents will create and revise information at machine speed, and answers will increasingly sit inside video, documents, and dynamic applications that cannot be understood by fetching the text of a webpage. Pretraining datasets will become increasingly context and domain knowledge specific.

The inputs necessary to build and serve these models require coordinating millions of discrete resources in precise fashion. It is our understanding that Grass has solved many of the hard challenges here: which connection can reach the source, whether the response is complete, how often a source should be revisited, and how to serve the result within a live model’s cost or latency requirements. The core team has specific knowledge of these workflows, and we expect them to serve their customers as their requirements change and grow across form factors over time. We have argued for years that tokens can help infrastructure networks reach threshold scale. Grass is showing what becomes possible afterward. Training revenue has financed the crawl and index, and the relationships with labs can now direct that investment toward increasingly large contracts. Andrej and the rest of the Grass team are among the most committed operators we have come across in resource aggregation cryptonetworks, and their ambitions are limitless: keeping an expanding, multimodal internet legible to machine intelligence in real time.

Footnote 1: OpenAI last disclosed more than 2.5 billion ChatGPT messages per day in July 2025, and weekly users have since passed 1 billion, implying roughly 3.5 billion messages a day. 22% of ChatGPT turns now call an external tool, assuming only 10% trigger one paid search yields about 128 billion requests a year, or about $640 million at Brave's price of $5 per 1,000. Agents multiply this significantly: a typical Gemini Deep Research task runs about 80 searches, so Codex and ChatGPT Work's roughly 10 million weekly users, running five tasks a week at 20 searches each, add 52 billion requests (about $260 million), and converting just 1% of daily messages into research-grade agent tasks would push volume past 1 trillion requests or about $5 billion a year.