Chutes

Chutes is a decentralized, serverless AI compute platform built on the Bittensor network as Subnet 64 (SN64). It allows developers, startups, and businesses to deploy, run, and scale AI models in seconds without managing servers, cloud contracts, or GPU hardware. The platform supports a wide range of workloads, including large language model (LLM) inference, image generation, audio processing, video generation, and custom machine learning applications. Chutes describes itself as open-source, decentralized infrastructure for the open-source AI future, offering on-demand access to a global network of GPU resources through a standard API, a Python SDK, and a web-based user interface.

The platform operates on a pay-per-use model. Users pay only for the GPU time they actually consume, billed down to individual tokens of output, rather than committing to fixed cloud subscriptions or reserved instances. According to Chutes, the platform processes trillions of tokens per month and provides the latest state-of-the-art open-source models shortly after their release, with support for models from organizations like DeepSeek, Mistral, Qwen, Microsoft, Google (Gemma), and others listed on their platform page.

Beyond its core compute service, Chutes has expanded into several additional products. Chutes Chat and Chutes Studio provide direct access to open-source AI models through a web interface. Chutes Search (search.chutes.ai) is an AI-powered web search engine. fictio (fictio.ai) is a character roleplay platform built on top of Chutes compute. The platform also includes an agent builder for running AI agents on top of the Chutes serverless network. For early-stage companies, the Chutes Startup Accelerator offers up to $20,000 in free platform credits with no equity requirement, targeting pre-Series A startups with fewer than 50 employees.

Chutes functions as a decentralized Platform-as-a-Service (PaaS) for AI workloads. A developer-facing API and web interface handle deployment and invocation, while a backend network of GPU miners executes the actual computations. Each deployed model, called a chute, is essentially a FastAPI application running on GPU-accelerated infrastructure, with individual endpoints called cords that each represent a single function or route within the application. The platform currently supports Python only through its SDK. Developers write code using simple decorators, and Chutes handles containerization, GPU selection, scaling, and shutdown automatically. Pre-built templates are available for popular frameworks like vLLM, SGLang, TEI, and Diffusion models. The system supports hardware including NVIDIA H200, B200, and AMD MI300X cards, and all miners must run on bare-metal servers or dedicated VMs managed through Kubernetes.

Security follows the project's core philosophy: don't trust, verify. Every GPU on the network is verified through a custom C/CUDA verification process that performs matrix multiplications seeded by device information and tests that at least 95% of VRAM is available, preventing hardware spoofing. All validator-to-miner traffic is encrypted with GPU-specific keys. Additional security layers include cfsv (filesystem integrity), inspecto (bytecode tampering detection), cllmv (per-token model verification), and envdump (runtime environment snapshots). For workloads requiring the highest data protection, Chutes offers publicly available Trusted Execution Environments (TEE) powered by sek8s, a hardened Kubernetes distribution running inside Intel TDX confidential virtual machines with NVIDIA Protected PCIe encryption. Validators monitor miner performance and rank them based on compute quality, uptime, and response times, while adversarial validation cross-verifies outputs from multiple miners to ensure accuracy.