Targon

Targon is a decentralized cloud computing platform operating as Subnet 4 (SN4) on the Bittensor network. Built by Manifold Labs, a team based in Austin, Texas, Targon focuses on providing confidential, high-performance GPU and CPU compute for AI workloads, including inference, training, and general-purpose applications. The platform positions itself as a secure alternative to centralized cloud providers, offering what it calls a non-interruptible cloud experience at interruptible prices without compromising on security.

The core differentiator of Targon is its approach to confidential computing. All workloads on the network run inside Confidential Virtual Machines (CVMs) protected by hardware-level encryption using Intel TDX (Trust Domain Extensions), AMD SEV-SNP, and NVIDIA Confidential Computing with Protected PCIe (PPCIe). This means that even the hardware operators who provide the GPUs cannot access, inspect, or tamper with the data or models running on their machines. In March 2026, Manifold Labs and Intel jointly published a technical whitepaper detailing this architecture, titled "Decentralized Compute on Untrusted Hardware Using Intel TDX and Encrypted CVMs."

Targon currently operates over 1,000 GPUs and CPUs on its network, including NVIDIA H200s, H100s, and 4090s, with support for B200s as a verified hardware configuration. The platform claims 99% uptime and sub-50ms latency. In July 2025, Manifold Labs raised a $10.5 million Series A led by OSS Capital, with participation from Digital Currency Group (DCG) and angel investors from Google, Shopify, and the Bittensor ecosystem. The company was founded in 2023 by Robert Myers, an early Bittensor contributor and former Opentensor engineer, and James Woodman, formerly COO of the Opentensor Foundation.

Beyond GPU compute, Manifold Labs also operates Sybil, a hybrid AI search engine, and Tao.xyz, a Bittensor analytics terminal and wallet.

Targon offers two primary products: Rentals (dedicated access to GPU and CPU servers) and Serverless deployments (automatic scaling without managing infrastructure). Users interact with the platform through a Python SDK, a CLI tool, and a standard API. Pricing is usage-based and transparent, with hardware providers setting their own rates through an auction system where emissions are allocated proportionally based on public demand.

The security model is built around the Targon Virtual Machine (TVM). When a hardware provider joins the network, they receive a uniquely encrypted Ubuntu 24.04 virtual machine image that can only be decrypted after the machine passes remote attestation through Intel Trust Authority (ITA), cryptographically verifying that the hardware and software configuration are genuine. The decryption key is permanently bound to the provider's network identity, preventing the encrypted disk from being copied or migrated to different hardware. Once running, each CVM performs continuous re-attestation approximately every 72 minutes, verifying both CPU-level integrity (Intel TDX measurements) and GPU authenticity (NVIDIA attestation reports) in a single unified proof. If any verification fails, the node is immediately removed from the scheduling pool and workloads are automatically rescheduled to other verified nodes.

Supported hardware configurations include Intel TDX with NVIDIA Hopper GPUs (H100, H200), Intel TDX with NVIDIA Blackwell GPUs (B200), and AMD EPYC 4th Gen with SEV-SNP for CPU-only confidential workloads. All miners must have BIOS access and run on bare-metal hardware.