Chutes

SN64 Chutes rents out AI model computing power by the request, the way a cloud provider does, except the machines belong to strangers who are paid in the subnet's token.

The commodity, explained from zero

The commodity is inference: asking an already trained AI model a question and getting an answer, sold by the million tokens, the word-pieces those models count, in and out. Chutes hosts open-source models, ones whose files anyone may download, and lets a developer send a request to one web address and get a reply, the same way they would with a big cloud provider. The docs put it in one line: "Chutes is a decentralized, serverless inference platform for open-source AI models." Serverless means the customer never rents or manages a machine; they pay per use.

It has a price because running these models needs expensive graphics processors. The pricing page lists per-model rates, from $0.0245 per million input tokens for a small model to $3.00 in and $15.00 out for a large one, and private GPU time from $1.80 an hour. The stated principle is "No subscription, no minimum, no markup."

Who buys today: the front page claims "Powering Trillions of Tokens per Month," undated, and shows logos for OpenRouter, Kilo, Fetch.AI, Phala and Pax Historia. No customer is quoted.

Why it is on Bittensor at all

The machines are not Chutes' own. They belong to miners, who plug their GPU clusters into the network and are paid in the subnet's token according to how much work they do. Its own comparison, in a file written for AI agents, is that rivals "all rely on centralized cloud providers" while Chutes runs on "a globally distributed GPU pool with no single-cloud dependency." Here the comparison breaks: a cloud provider owns its hardware and signs contracts for the rest, while Chutes only scores strangers and lets the token do the hiring. Whether that is cheaper for the buyer is not stated in dollars anywhere.

How the work gets done

Miners run GPU servers arranged as a Kubernetes cluster, software for running many containers across machines, and host "chutes," the platform's word for a deployable model or workload. The docs say the goal of mining "is to provide as much compute as possible, optimizing for cold start times," the seconds a model takes to wake up. Validators score each miner over a rolling seven days on four weighted measures: compute units 55 percent, successful invocations 25 percent, how many distinct chutes it keeps running 15 percent, and bounties 5 percent. The subnet's emission, its slice of newly minted TAO, is split by those scores, with a rule that a miner running many identities is punished rather than paid more.

How you would know it works

The live model catalog with prices at https://chutes.ai/models and the status page with daily uptime history at https://status.chutes.ai are both public. The endpoint on the front page, https://llm.chutes.ai/v1/chat/completions, can be called with an API key.

What is missing

There is no dated usage number; "trillions of tokens" carries no month. The miner docs disagree with each other: the docs overview still describes GraVal while the miner README says GraVal is no longer supported. No page is written for stakers. The only comparison with rivals sits in llms.txt, a file meant for machines.

Go deeper

• Front door with live endpoint: https://chutes.ai
• Pricing by model and hour: https://chutes.ai/pricing
• Miner scoring weights and anti-gaming: https://chutes.ai/docs/miner-resources/scoring

Metaphor: a cloud provider whose machines belong to strangers. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.