Covenant Left. Bittensor Has Other 20+ State-of-the-Art Subnets

The Covenant exit is currently dominating headlines. But in the turbulence, something important got lost: Templar (SN3) was one subnet out of dozens. And many of the others are quietly building products that rival or outperform centralized competitors across AI, compute, finance, pharma, weather, video, security, and more.

If your entire thesis on Bittensor lived and died with Covenant, you were never paying attention to the full picture. Here’s what that picture actually looks like.

Compute & Inference

Chutes (SN64) — Currently the largest subnet on Bittensor. Chutes provides inference at one-sixth the cost of competitors and ranks number one on OpenRouter, handling thousands of daily queries. The team operates as a group of independent corporations with no CEO, with funds locked into a smart contract.

SOTA: #1 on OpenRouter. Thousands of daily queries. 1/6th the cost of centralized inference. World’s cheapest compute/inference/AI deployment.

Targon (SN4) — Decentralized GPU compute with Trusted Execution Environment (TEE) support, built for enterprise workloads. Provides service at roughly one-sixth the cost of centralized competitors offering similar infrastructure.

Why it matters: Enterprise-grade decentralized GPU with TEE, the kind of infrastructure Fortune 500 compliance teams actually require.

AI Training

IOTA (SN9) — A “Training at Home” decentralized network that has been described as the Napster of AI training infrastructure. Where Templar proved scale, IOTA is pushing accessibility, making it possible for everyday laptop owners to contribute to meaningful training runs.

Why it matters: Napsterized AI training infrastructure. If Templar is the proof that scale works, IOTA is the proof that anyone can participate.

Quasar (SN24) — Tackling two of the hardest unsolved problems in decentralized AI simultaneously: long context and Mixture of Experts (MoE) training. Most decentralized training runs focus on dense models because they’re easier to coordinate across distributed systems. Quasar is going after MoE, a fundamentally harder architecture, while also pushing context windows toward millions of tokens through its Quasar Latent Memory system.

SOTA: Outperforms state-of-the-art on OpenAI MRCR V2. First decentralized MoE training run. Context scaling toward millions of tokens.

Coding

Ridges (SN62) — An AI-powered “vibe coding” platform that beat both Claude and Cursor on the SWE Benchmark at one-seventieth of the cost. Seventy times cheaper than leading centralized coding assistants while outperforming them on standardized software engineering tasks. For anyone tracking the economics of AI-assisted development, this alone justifies attention to the Bittensor ecosystem.

SOTA: Beat Claude & Cursor on the SWE Benchmark. 1/70th the cost. World’s best open-source coding assistant.

Finance & Prediction

Synth (SN50) — A prediction market engine and oracle that turned $3,000 into $50,000 on Polymarket (a 16x return). Synth’s weighted metamodel aggregates multiple forecasting models to produce more reliable predictions than any single model could achieve. The subnet recently introduced a system where model diversity itself becomes a competitive advantage. (Read more on TAO Daily)

SOTA: $3K → $50K on Polymarket (16x ROI). World’s best price path projection forecasting. Real money, real results.

Vanta Trading — A decentralized proprietary trading platform that’s challenging the prop trading industry. Traditional firms like FTMO and Breakout run on closed, centralized infrastructure, including two-step challenges, 80–90% reward splits, restricted trading windows, and payouts that are delayed or denied. Vanta flips every one of those: one-step evaluation, 100% reward split, 25% quarterly bonuses, and weekly reward distributions verified on-chain. Over $30M in rewards already distributed through the network.

Why it matters: 100% reward split vs. the industry standard of 80–90%. The first prop trading firm where every trade, every stat, and every payout is publicly verifiable.

Computer Vision & Deepfake Detection

Score (SN44) — Decentralized computer vision for diverse purposes. Score has achieved 78% accuracy on its benchmarks, approaching the 79% human gold standard.

SOTA: 78% accuracy vs. 79% human gold standard. World’s best decentralized computer vision. Moneyball, but on-chain.

Bitmind (SN34) — The world’s best deepfake detection system, built entirely on decentralized infrastructure. As AI-generated media becomes increasingly difficult to distinguish from real content, Bitmind sits at the intersection of two of the most important technology trends of the decade.

SOTA: World’s best deepfake detection. As synthetic media explodes, this becomes critical infrastructure for trust on the internet.

ItsAI (SN32) — The world’s best AI text detector. In recent times, where AI-generated content is flooding every platform, the ability to distinguish human-written text from machine output is becoming critical infrastructure.

SOTA: World’s best AI text detector. The gatekeeper for content authenticity in an era of AI-generated everything.

Healthcare & Pharma

Metanova (SN68) — AI-driven discovery of novel pharmaceutical molecules. The subnet hunts for new stable compounds using decentralized compute. Recently introduced a bounty-based “Blueprint” reward system to incentivize discovery of specific molecular targets. If even one compound from this pipeline reaches clinical trials, the ROI math changes dramatically for the entire ecosystem. (Read more on TAO Daily)

SOTA: World’s best novel drug discovery AI. One successful clinical compound from this pipeline could be worth more than the entire current Bittensor market cap.

Weather

Zeus (SN18) — The world’s most accurate weather forecasting system, outperforming industry benchmarks used by energy traders. Weather prediction is a multi-billion-dollar market dominated by a handful of government agencies and private firms. Zeus is challenging that structure from a decentralized base.

SOTA: World’s most accurate weather forecasting. Outperforms industry benchmarks that energy traders pay millions to access.

Infrastructure & Storage

Hippius (SN65) — Decentralized cloud storage at one-four-hundredth the cost of Filecoin. That cost differential is the kind of gap that makes decentralized storage viable for real-world workloads rather than just a proof of concept.

SOTA: 1/400th the cost of Filecoin. World’s cheapest cloud storage.

Video & 3D

Vidaio (SN85) — Decentralized AI video intelligence platform specializing in advanced video upscaling from SD to HD/4K, intelligent compression, and quality enhancement. By leveraging a network of distributed miners running cutting-edge AI models, it delivers superior visual results while dramatically reducing file sizes, often by around 80% or more, without compromising perceived quality. This makes high-resolution video processing affordable and scalable for creators, streamers, broadcasters, legacy content libraries, medical/security imaging, and even autonomous vehicle fleets facing massive storage and CDN costs.

SOTA: 4K upscaling with 80% file size reduction. World’s cheapest video compression and upscaling AI.

404-Gen (SN17) — World’s largest decentralized 3D model collection and generation network. Miners compete using AI models to create high-quality 3D assets from simple text or image prompts, with validators scoring outputs through competitive mechanisms like ELO duels. It directly democratizes 3D content creation, empowering individuals and teams to build immersive virtual worlds, games, AR/VR/XR experiences, and more at unprecedented scale and accessibility.

SOTA: World’s largest collection and generator of 3D models. First Unity plugin for live 2D-to-3D. Direct pipeline into the $200B+ gaming industry.

Bitcast (SN93) — A decentralized creator marketing platform on Bittensor that connects brands with YouTube creators (acting as miners) who produce content aligned with brand-defined briefs. It uses AI to validate campaigns and rewards creators based on verified engagement metrics, creating an efficient, transparent marketplace for video content creation at scale.

SOTA: Already profitable. AI-validated campaigns. World’s largest decentralized video generation agency.

Data & Security

Data Universe (SN13) — Serves as the world’s largest decentralized data scraper. It systematically collects, processes, and stores vast amounts of web data, providing a critical raw material pipeline of high-volume, fresh information that powers training, fine-tuning, and inference across numerous other AI subnets and applications in the network. As the foundational data layer, it enables the entire Bittensor ecosystem to access decentralized, scalable data resources that would be prohibitively expensive or centralized elsewhere.

SOTA: World’s largest decentralized data scraper. The raw material pipeline that every other AI subnet ultimately depends on.

D-Sperse (SN2) — Building the world’s leading zero-knowledge machine learning (zkML) and verifiable AI inference network on Bittensor. It provides cryptographic proofs that allow AI model outputs and computations to be independently verified without revealing underlying data or models — essentially acting as the “SSL for AI.” This foundational security and trust layer is essential for enterprise adoption, real-world AI agents, and any application where verifiable intelligence is required for high-stakes decisions and interactions.

SOTA: World’s best SSL for AI. Foundational security layer for every machine learning communication on the network.

Ready AI (SN33) — Pioneering the world’s best decentralized system for text data cleaning, labeling, annotation, and structuring. Using a network of miners and advanced LLMs, it transforms raw, unstructured data — from web crawls, conversations, social media, and more — into high-quality, richly tagged, machine-readable formats optimized for training models, building vector databases, and powering AI agents.

SOTA: World’s best decentralized text data cleaning/labeling. 660× cheaper and 86% more accurate than human annotation. Structuring the web into richly tagged, machine-readable intelligence for the agentic web.

Yanez (SN54) — Adversarial testing for fraud and KYC systems, already holding multi-year contracts with major banks in a $200 billion market. This is one of the clearest examples of a Bittensor subnet generating real-world enterprise revenue. (Read more on TAO Daily)

SOTA: Multi-year contracts with major banks. $200B+ addressable market. Real enterprise revenue already streaming in.

Gradients (SN56) — Decentralized AutoML platform that makes training and fine-tuning large language models (LLMs) and diffusion/image models radically accessible to anyone. Users simply upload their dataset, select a base model, and the network’s miners compete through intelligent AutoML tournaments to automatically discover the best configurations and deliver professional-grade results in days, not weeks.

SOTA: World’s cheapest 1-click Auto ML. Democratizing machine learning for teams that don’t have PhD-level AI engineers.

The Point

Count the subnets listed above. That’s 20+ projects across compute, training, coding, finance, pharma, weather, video, gaming, security, storage, and data infrastructure. Many are state-of-the-art in their respective domains. Several are already generating revenue. Some are challenging billion-dollar incumbents on cost by orders of magnitude.

Templar was the headline act. But the ecosystem was never a one-act show.

The Covenant exit tested whether Bittensor could survive the loss of its most visible team. The answer is that the network was always deeper than any single project. And the 20+ other subnets that are yet to make mainstream headlines are still building.

Which are your favourite subnets right now?