Over six billion artificial intelligence requests have been organized into a publicly available dataset.
In collaboration with researchers at Harvard University, the decentralized inference platform Chutes has released what may be one of the largest model-service metadata datasets in its category. The data covers one year—from April 11, 2025 to April 12, 2026—and includes 6,102,241,375,6 requests, 9,174 models, and 310,497 anonymous users. The dataset can be downloaded via the code hosting platform and the university’s storage.
What has been made public is metadata, not conversation content. It includes timestamps, token counts, latency, and time to first response. There are no prompts and no answers. This approach—publishing only metadata—is itself a trade-off between privacy and researchability.
In this year’s traffic, input tokens total about 35.8 trillion, while output is about 2.52 trillion. One conclusion provided by the researchers is that roughly 99% of requests are duplicates within 15 minutes. For the first time, these figures allow the outside world to estimate the cost structure of inference services based on real traffic.
The implications for infrastructure are very direct. Since many requests repeat within short time windows, there is significant room to optimize caching and scheduling rather than simply piling on compute. The value brought by the high duplication rate may be greater than the impact of parameter scale.
The most expensive problem is often the one that gets asked over and over again.
#人工智能 #data
In collaboration with researchers at Harvard University, the decentralized inference platform Chutes has released what may be one of the largest model-service metadata datasets in its category. The data covers one year—from April 11, 2025 to April 12, 2026—and includes 6,102,241,375,6 requests, 9,174 models, and 310,497 anonymous users. The dataset can be downloaded via the code hosting platform and the university’s storage.
What has been made public is metadata, not conversation content. It includes timestamps, token counts, latency, and time to first response. There are no prompts and no answers. This approach—publishing only metadata—is itself a trade-off between privacy and researchability.
In this year’s traffic, input tokens total about 35.8 trillion, while output is about 2.52 trillion. One conclusion provided by the researchers is that roughly 99% of requests are duplicates within 15 minutes. For the first time, these figures allow the outside world to estimate the cost structure of inference services based on real traffic.
The implications for infrastructure are very direct. Since many requests repeat within short time windows, there is significant room to optimize caching and scheduling rather than simply piling on compute. The value brought by the high duplication rate may be greater than the impact of parameter scale.
The most expensive problem is often the one that gets asked over and over again.
#人工智能 #data
