
Multicoin Invests in Grass: Monetizing AI Training Data While Securing the Search Gateway of the Inference Era
Authors: Spencer Applebaum, Shayon Sengupta
Translated by: TechFlow
TechFlow Summary: While most DePIN projects are still grappling with business models, Grass has already achieved profitability through selling pretraining data and is expanding into the real-time retrieval market of the inference era. Multicoin believes the market severely underestimates the scale of this direction; if every action by AI agents requires web searches, it will be a market several orders of magnitude larger than training data.

Today, we announce an investment in Grass, the reading layer for machine intelligence, across our hedge fund and venture fund.
Andrej Radonjic recently published his plan for the Grass network to generate “rich intelligence.” We believe Grass has earned the right to pursue this ambition, and it stands as one of the clearest examples of DePIN operating at commercial scale.
Over the past eighteen months, Grass has deployed finely tuned crypto incentives to recruit contributors to share underutilized bandwidth from their internet connections. This contributor network serves as the infrastructure for querying and parsing the public internet, enabling Grass to rapidly feed data pipelines for frontier labs. These efforts require harvesting vast amounts of information, making aggregated states available for model training, and reliably delivering them to customers whose needs vary with each new training run.
We believe the market views Grass solely as an agent network serving AI labs with pretraining data, failing to account for the massive expansion into runtime data retrieval. The Grass team appears uniquely positioned to meet the demand for fresh information whenever a model searches the web to answer questions or agents retrieve context before acting—a market potentially several orders of magnitude larger than pretraining data, since these requests recur throughout a model’s entire deployment (whereas pretraining datasets are one-off deliveries for discrete runs).
Agent systems must search and retrieve information for every action. The cost and latency requirements to execute this effectively would practically necessitate rebuilding a web index—one of the most capital-intensive endeavors historically undertaken only by players like Google and Microsoft. As the internet recursively grows through agent actions, real-time retrieval is becoming increasingly critical infrastructure. We believe Grass holds structural cost and access advantages in building and providing this service, backed by relationships with labs that will consume it, along with the technical capability to make it useful for machine intelligence.
Past: Pretraining Data
The majority of the public internet is fundamentally inaccessible to traditional AI labs during pretraining. Websites rate-limit automated traffic, serve different content across geographic regions, and increasingly require pages to be rendered through standard browsers. Grass solves this by routing requests to millions of residential connections, whose owners opt into the network and are compensated for their contributions.
Unlike many agent networks of previous eras, Grass does not need to manufacture hardware, procure residential connections through wholesale ISP deals, or covertly compromise consumer devices. Users worldwide are already paying for these connections, with most capacity remaining idle. Tokens helped Grass recruit contributors and establish geographic coverage before customer demand could justify it, rapidly transforming this aggregated resource into a pretraining resource reliable enough for AI labs.
Grass must deliver massive volumes of data on tight schedules, formatted so labs can directly plug it into their training pipelines. Each delivery requires close coordination with lab teams, giving Grass deeper insight into their next requirements.
In our 2023 article on DePIN network design, we described “threshold scale” as the inflection point where a network’s distributed supply becomes commercially viable. Tokens were particularly powerful prior to this threshold, as they incentivize people to contribute to infrastructure that lacked sufficient coverage to attract customers. The true test comes after that milestone: whether a network can sell reliable services and ultimately pay for the supply it utilizes.
Grass is one of the few networks we’ve encountered that has passed this test. Over 6 million contributors have earned USD-denominated rewards, funded entirely by revenue generated directly from their contributions. Grass reported $17 million in revenue for 2025, with another $17 million in the first half of 2026, and nearly every AI client engages in repeat business. The company states it is profitable and expects training data alone to generate $75 million in revenue in 2026.
Present: Real-Time Context Retrieval
The current training data business is substantial, and we believe it will continue to grow. Frontier labs require massive volumes of new information to train ever-larger language models. However, the nature of this product is inherently “point-in-time.”
If you ask a model what a company announced this morning, whether a flight is delayed, or what a website currently displays, the information stored in its weights will fall short. It must search sources, open relevant pages, and read them in real time. An agent handling more complex tasks may need to repeat this process multiple times before it can provide an answer.
We believe the end market for runtime search and retrieval could be several orders of magnitude larger than the pretraining web data market. Training procurement occurs around model runs, but real-time information can be consumed by every model, application, and agent as long as tasks require awareness of a changing world. At historical publicly traded search API prices, even a fraction of request volume could support hundreds of millions in annual revenue. 1
Meanwhile, the internet itself is poised to grow larger. More information is being published in formats difficult for traditional search engines to process, including videos, images, spreadsheets, PDFs, and dynamic pages. Enabling machines to discover and parse this expanding body of information represents a massive opportunity, and is far from a trivial engineering challenge, as it requires continuous global-scale crawling, storage, processing, ranking, and low-latency delivery.

Grass is advancing toward that market from a position rarely held by new entrants or existing players. Its proposed Contents API will allow models to open and render public pages via the residential network, while its planned Search API and web index will help models determine which pages should be opened in the first place.
Both products build on mechanisms Grass already operates for training clients: routing requests to residential nodes based on geography and reputation, measuring response quality, and transforming harvested content into datasets ready for frontier training runs. Grass has already collected over 1 billion public web videos, a portion of which is used alongside another Multicoin portfolio company, Inference.net, to train video annotation models. Search ranking and delivering results within inference latencies are further engineering challenges, but the collection and processing systems are already running at commercial scale.
Building a web index at this scale demands significant capital and engineering resources. That has always been the price for constructing deep moats, high-margin, pricing-power infrastructure. We believe extracting a single dollar of gross margin from this opportunity would harm the business’s long-term value; pursuing this market delivers a net positive effect on long-term token value accumulation that far outweighs the suboptimal capital allocation strategies advocated by countless Twitter analysts. Therefore, we believe GRASS is at an inflection point, poised for significant revaluation driven by this new business line.
The compounding fundamental effect here is driven by Grass’s relationships with various labs. Every large-scale training data delivery gives Grass reason to expand coverage and deepen its understanding of what customers need. Better coverage improves the index; a better index enhances real-time retrieval; successful retrieval gives those labs further reason to build on Grass. Given time, this infrastructure will provide scarce data for pretraining, post-training, and inference regardless of how model architectures evolve.

Future: Abundant Intelligence
The internet agents need to read is growing faster than the web index originally built for search engines. Agents will create and modify information at machine speed, meaning answers will increasingly reside in videos, documents, and dynamic applications—formats that cannot be understood through simple web scraping. Pretraining datasets will become increasingly reliant on contextual and domain-specific knowledge.
Building and servicing the inputs these models require demands precise coordination of millions of discrete resources. To our knowledge, Grass has solved many of the puzzles here: which connections can reach specific sources, whether responses are complete, how often a source should be revisited, and how to return results within the cost or latency constraints of real-time models. The core team possesses specialized expertise in these workflows, and we expect them to continuously serve customers well as their needs shift and scale across different forms.
For years, we have argued that tokens can help infrastructure networks reach critical mass. Grass is demonstrating what becomes possible once that threshold is crossed. Training revenue funds crawling and indexing, while relationships with labs now direct that investment toward increasingly larger contracts. Andrej and other members of the Grass team are among the most focused operators we’ve encountered in resource-aggregating crypto networks, with boundless ambitions: to keep the expanding multimodal internet readable by machine intelligence in real time.
Footnote 1: OpenAI last disclosed data in July 2025 showing over 2.5 billion daily ChatGPT messages. Since then, weekly active users have exceeded 1 billion, implying roughly 3.5 billion daily messages. Currently, 22% of ChatGPT turns invoke external tools. Assuming only 10% trigger a paid search, that generates approximately 128 billion annual requests. At Brave’s price of $5 per 1,000 requests, this equals roughly $640 million. Agents will significantly amplify this number: a typical Gemini Deep Research task runs about 80 searches. Thus, if Codex and ChatGPT Work’s ~10 million weekly active users each run 5 tasks weekly with 20 searches per task, that adds 52 billion requests (~$260 million). Converting just 1% of daily messages into research-grade agent tasks would push the volume beyond 1 trillion requests annually, or roughly $5 billion.
Disclosure: Unless otherwise noted, the opinions expressed herein are solely those of the author and do not represent the views of Multicoin Capital Management, LLC or its affiliates (collectively, “Multicoin”). Certain information contained herein may be sourced from third parties, including portfolio companies invested in by funds managed by Multicoin. Multicoin considers the information provided to be reliable, but makes no representation regarding its ongoing accuracy or applicability to specific situations. This article may contain links to third-party websites (“external sites”). The mere existence of any such link does not constitute an endorsement of those sites, their content, or their operators. These links are provided solely for your convenience and do not imply our endorsement of the content of these external sites. External site content is developed and provided by others, and Multicoin assumes no responsibility for any of it. The charts provided herein are for reference only and should not be relied upon for any investment decision-making. Any forecasts, estimates, projections, targets, outlooks, and/or opinions expressed in this blog may change at any time without notice and may differ from or conflict with opinions expressed by others.
This content is for informational purposes only and should not be relied upon as a basis for investment decisions. It is not, nor should it be construed as, complete. The contents of this article should not be interpreted as legal, business, or tax advice. You should consult your own advisors regarding these matters. References to any securities or digital assets are for illustrative purposes only and do not constitute investment recommendations or offers to provide investment advisory services. Any investments or portfolio companies mentioned, referenced, or described herein do not represent all investments in any investment vehicle managed by Multicoin, and there can be no assurance that these investments will be profitable or that other future investments will have similar characteristics or results. A list of investments made by Multicoin-managed venture funds can be found here: https://multicoin.capital/portfolio/. Unannounced investments are excluded from this list due to the need to coordinate the timing and nature of public disclosures with founding teams or issuers. Additionally, for strategic reasons, the Multicoin Capital hedge fund does not disclose its positions in publicly traded digital assets.
This blog does not constitute investment advice, nor does it constitute an offer to sell or a solicitation of an offer to purchase any limited partner interests in any investment vehicle managed by Multicoin. Offers to invest or invitations to invest in any Multicoin investment vehicle are made solely pursuant to offering memoranda, limited partnership agreements, and subscription documents, and only such documents should be relied upon when making investment decisions.
Past performance does not guarantee future results. There can be no assurance that the investment objectives of any Multicoin investment vehicle will be realized, and investment results may vary significantly from year to year or even month to month. Accordingly, investors may lose all or a substantial part of their investment. The investments or products mentioned in this blog may not be suitable for you or any other party. Valuations provided are based on detailed assumptions as of the date of publication, which assumptions may no longer be applicable following publication. Our target prices or valuations, as well as any base case or bull case scenarios underlying those target prices or valuations, may not be realized.
Multicoin has established, maintained, and enforces written policies and procedures designed to reasonably identify and effectively manage conflicts of interest related to investment activities. For additional important disclosures, please refer to the disclosures and terms of use provided at https://multicoin.capital/disclosures and https://multicoin.capital/terms.















