Channel3 moved latency-tolerant stages of its product graph pipeline to asynchronous inference on Sail, cutting AI cost per product by 10×. Token-heavy indexing work is done in the background over minutes, while the customer-facing API is still near-instant.
Building a database of every product on the internet requires a lot of inference #
Channel3 is building the product graph for agentic commerce: a structured layer that lets AI agents find and understand almost any physical product with a single API call.
Messy merchant data needs to be cleaned into data that machines can reliably use. The same product may appear across the internet with different titles, descriptions, images, identifiers, and attributes. One merchant might call a rug “washable,” another “machine cleanable,” while another buries the same information in a paragraph of product copy.
AI makes it possible to understand, reconcile, and structure that data. But applying AI across a catalog at internet scale creates a real problem: web-scale inference is expensive.
Data about products changes continuously. New products appear, existing listings change, and every data change has to move through a series of AI-powered processing steps before it becomes useful to an agent.
At Channel3's scale, that translates into trillions of tokens of inference and a workload that only grows with the catalog. The team needed a way to grow without AI costs rising at the same rate.
Moving agentic work to the background #
Channel3 and Sail lowered AI cost per product by 10× by moving agentic work to the background with Sail.
Channel3's customer-facing API has to be fast: it returns products in less than half a second.
But the work that supports their product graph has a very different latency requirement than their customers. When Channel3 adds a new product to their catalog, it doesn't have to be processed in milliseconds. A new product can join the catalog minutes or even an hour later with no impact on customers. This “lag time” created an opportunity to move the latency-tolerant workflows onto asynchronous inference, allowing for dramatically better economics.
The cost we pay for inference determines how much of the web we can afford to understand. Sail lets us use latency as a lever and run dramatically more AI for the same budget.
The Channel3 team tracks these economic gains using AI cost per product as a core efficiency metric. This way, the economic impact of the pipeline is measured independently of changes in overall catalog volume.
Most jobs are submitted to Gemma 4 31B using Sail's Flex completion window, which relaxes TTFT and TPS targets in exchange for maximum efficiency. When a job completes, its output is applied to the product and the next stage of Channel3's pipeline can continue.
Asynchronous does not have to mean slow. Channel3 measures the time between handing a request to Sail and having the completed result written back onto the product. Even on the Flex completion window, p95 latency is just 8.5 minutes.
This is fast enough to keep the pipeline churning through new and changing products continuously.
Channel3 also values Sail's reliability and deliberate approach to model rollouts. When inference sits underneath a pipeline processing products around the clock, predictable behavior matters as much as raw throughput.
Channel3 is exactly the kind of workload we built Sail for. When you're processing an entire product catalog, every request doesn't need an instant response, but you do need massive scale. We built Sail around that combination of scale, efficiency, and reliability, so teams can spend their budget understanding more products.
Since moving these inference workloads to Sail, Channel3 has reduced AI cost per product to approximately 10% of its previous baseline, while maintaining output quality.
Channel3 is growing fast, adding millions of products every month as it builds the most complete product graph for agentic commerce. As the graph grows, so will the inference needed to build and maintain it.
Sail's architecture is designed for that growth: keep the customer-facing path fast, push work that can tolerate latency into the background, and use asynchronous inference to make web-scale product understanding economically practical.
The Channel3 and Sail teams are excited to see how we can continue to innovate and bring every product on the internet within reach of AI agents.