How the Grass Network became the network AI builds on Today, the Grass Network can access parts of the public web that other networks cannot. Live Context Retrieval (LCR), the latest product built on it, retrieved 97% of the pages we tested, more than any comparable service, and was the only provider to retrieve every priority domain. Getting here required not only building a residential network, but making it fast and reliable enough for the hardest enterprise use cases. Progress is measured against four metrics, and each section below shows where the network became measurably better: Reach: how much of the public web the network can access Speed: how long each request takes, end to end Capacity: how much data the network can move at once Reliability: the share of web requests that succeed Over the past year, the network has more than tripled its throughput and made routing 4x faster. Building the network Many of the world’s most-visited sites are closed to automated traffic, cutting AI agents and businesses off from public web data. The Grass Network keeps that data accessible. The most reliable way for businesses and AI agents to access the public web is to route their requests through real home internet connections. This way, each page loads the way a local visitor would see it, so businesses and agents get the most accurate, up-to-date version of the site. Through the Grass App, Grass has built a network of millions of opted-in residential connections in over 190 countries. While this pool of connections continues to grow, that growth alone doesn’t capture how much the network has improved. Home connections are volatile: phones disconnect from Wi-Fi networks and laptops close mid-request. Grass addresses these reliability problems by routing around disconnects in real time to ensure data pipelines are reliably sub-second. Because Grass also optimizes the efficiency of every connection, the products built on the network don’t depend on growth to get faster. Raising throughput and success rates makes the network faster and more reliable, and it increases the value of every new high-quality node that joins. Making the network enterprise-ready Before a web request arrives at one of those opted-in residential connections, it passes through an infrastructure layer Grass designed to maximize the speed and capacity of every request. Two major architectural choices have made the largest impact: colocating the servers that run the network and making network gateways stateless. Colocation makes every request faster Each request crosses two servers in each direction: a gateway, which picks the residential node, and a bridge, which holds the connection to that node. By bringing gateway and bridge servers into the same location, Grass shortened internal routes and cut latency overhead on every request. Routing latency, the time a request spends passing through the network, has fallen more than 4x over the past year, to below one second for the vast majority of domains. Over the same year, throughput has more than tripled, from 300 Gbps to more than 1 Tbps, and it can go higher by adding servers. At the 5 Mbps Netflix recommends for 1080p video, 1 Tbps is 200,000 full HD streams at once. Stateless gateways let the network grow with demand Every request enters through a gateway, which decides which residential connection it will use. Grass redesigned these gateways so any of them can handle any request. Traffic spreads evenly across the network, so no single server becomes a bottleneck when demand spikes. Growing capacity is now as simple as adding servers. This design also makes the network more reliable: when a server fails, traffic moves elsewhere, and applications built on the network keep running without a drop in success rates. Because Grass owns every layer of the stack it is able to implement these kinds of design choices, which are the foundation that enables the network to achieve automated, internet-scale data acquisition. Proven by the hardest use cases The toughest test of this technology is collecting the enormous amount of training data that frontier AI labs require to build new models, and delivering on that work is what hardened the network. Training data: the whole public web, on any schedule For frontier AI labs, time is the scarcest resource. A new model tops the leaderboard every few months, and labs will pay to get the right data quickly to stay competitive. The web keeps changing between training runs, too: new knowledge appears every second and old pages change or disappear, so collecting training data is a job that never finishes. Grass’s training-data business runs those jobs through the LCR Contents API. Some labs order large datasets on a schedule, and the Grass Network is used to retrieve them and revisit pages as they change. Faster routing means it can re-crawl daily, hourly, or as pages appear, and more pages per server brings down the cost of every page delivered. Reach adds signal on top of speed. Because the network can access more of the public web, the raw data it collects can be distilled into dense, high-signal training corpora. That includes the long tail of the web: while the top thousand sites are covered by everyone, the next hundred million are where much of the specific, valuable information lives. Exabytes of data have been delivered to many of the world’s leading AI labs this way. That demand battle-tested the Contents API before it opened to customers, whether they’re gathering data at scale, refreshing datasets on a schedule, or grounding AI agents in live context. Building on the network: Live Context Retrieval for agents An AI model’s knowledge stops at the end of training. Live Context Retrieval (LCR) closes that gap: the LCR Contents API fetches pages through the network whenever an application calls it, so the answer reflects the page as it is at that moment. LCR is new, but it launched on a network already proven by exabyte-scale workloads from frontier labs. That matters, because serving live requests is less forgiving than building training datasets. A user is waiting on the other side of every request, and a single agent task typically touches 5 to 50 pages (more with deep reasoning), so every millisecond of routing time is multiplied. Success rates matter just as much: every blocked, incomplete, or unavailable page means a retry, a slower answer, or a wrong one. The most important test of a network is whether it delivers the page that was asked for. LCR has been benchmarked against comparable services, and it retrieved 97% of pages, 13 percentage points ahead of the next-best provider. Source: Introducing the Grass Contents API On easy pages, the top providers are close. The gap widens on the pages that matter most and are hardest to get: Hard domains. On pages that routinely turn away bots, LCR retrieved 98%, against 71% for the next-best provider. Priority domains. LCR was the only provider tested to retrieve 100% of priority domains, including e-commerce, job boards, and review sites. These are the pages agents and analysts ask for most, and the ones most likely to be blocked. That performance opens whole categories of use cases, including research agents, copilots, shopping and travel concierges, and monitoring tools. What each gain unlocks The most advanced AI labs do not value a model upgrade by its benchmark score. Instead, they project how much a model’s usage will increase in response to advances in its abilities, and work backward to decide what that increase is worth. Grass takes the same approach: each gain matters because of the use cases it unlocks.: Reach: access to the long tail of the web yields higher-signal training data that other providers cannot collect, which is critical to building frontier models. Speed: sub-second routing lets fetching sit inside an agent loop without becoming the bottleneck, which makes LCR viable for live agent products. Capacity: more than 3x the throughput and more pages per server lower the cost of every page, so labs can re-crawl on the schedule they need. Reliability: higher success rates mean fewer retries, so agents return faster and more accurate answers, directly improving the end-user experience of agentic products. What comes next Most providers struggle with hard-to-reach domains, leading to frequent failures and multi-second delays. Pushing 1 Tbps with sub-second routing removes those bottlenecks, giving teams accurate web data whether they’re pulling large training sets or running live agents. Keep up to date. Read new posts on the Grass blog, and follow @grass on X.(Historical earnings: Home Depot reported Q2 2026 (period ended 2026-08-02) basic EPS of 8.1, diluted EPS of 8.09, net income of USD 8.055 billion, and revenue of USD 89.626 billion. Consensus expectations: No upcoming consensus estimate was provided.)