Decentralized AI Landscape: A Critical Survey

Date: 2026-03-12 · Scope: Comprehensive survey of decentralized AI training and inference projects, honest assessment of what has shipped vs. what remains vaporware, comparison with the l402-train Lightning-native approach

1. Introduction

As of early 2026, dozens of projects claim to be "decentralizing AI." Most are doing one of three things: (a) running inference on distributed GPUs, (b) building GPU rental marketplaces with a token, or (c) actually attempting distributed model training. Very few are doing (c), and those that are have hit hard technical walls around communication bandwidth, verification of untrusted computation, and incentive alignment.

This survey examines every major project in the space with a single question: what has actually shipped? We distinguish between working systems that have trained real models, testnet-stage protocols burning through runway, and token-driven narratives with little technical substance. The intent is to inform the design of l402-train by understanding what has worked, what has failed, and what remains unsolved.

2. Bittensor (TAO)

Architecture

Bittensor operates a network of 128+ "subnets," each an incentive-based competition marketplace for a specific AI task. Miners run AI models and submit outputs; validators score those outputs using subnet-specific incentive mechanisms; the Yuma Consensus algorithm determines how TAO emissions are distributed based on validator-submitted weights. The chain itself (Subtensor) runs Proof of Authority through trusted nodes controlled by the Opentensor Foundation.

Subnets cover text generation, image generation, text-to-speech, fine-tuning, DeFi yield optimization, zero-knowledge proof verification, and dataset curation. As of 2025, 93 subnets were active with tasks ranging from language model inference (SN1/Apex, SN4/Targon) to competitive AI gaming benchmarks.

In February 2025, the Dynamic TAO (dTAO) upgrade replaced the centralized root-network valuation system — where 64 validators controlled emission allocation — with a market-driven model where TAO holders stake directly into subnet-specific liquidity pools, receiving subnet "alpha tokens" in return.

Consensus and Verification

Yuma Consensus aggregates validator-submitted weights into miner rankings. Validators define their own scoring criteria per subnet. There is no universal verification mechanism — each subnet implements its own quality assessment, which means verification rigor varies wildly across the network.

Token Requirement

• Staking/delegation: Minimum 0.1 TAO
• Validator permit: Minimum stake-weight of 1,000 TAO on a subnet (including delegated stake); actual requirement is dynamic based on the top 64 wallets per subnet
• Miner registration: Dynamic TAO burn that fluctuates based on registration demand; this is a sunk cost

Actual Results

No frontier models have been trained on Bittensor. The network primarily incentivizes inference — miners compete to serve the best responses to validator queries. Some subnets (like SN9 for pretraining and SN25 for distributed training) have attempted training workloads, but no competitively benchmarked model has emerged from the network. The most visible outputs are inference API endpoints and curated datasets.

Honest Assessment

A peer-reviewed empirical analysis published in June 2025 (arXiv 2507.02951) examined 6.6 million events across 121,567 wallets in all 64 subnets from March 2023 to February 2025. The findings are damning:

• Rewards are driven by stake, not quality. Economic stake is the dominant predictor of rewards. Performance contributes "only modestly for validators and very weakly for miners."
• Extreme concentration. Fewer than 2% of wallets command a 51% majority in most subnets.
• Centralization persists. The Opentensor Foundation controls all Subtensor validator nodes via Proof of Authority and can censor transactions.
• Halving pressure. The December 2025 halving cut daily emissions from ~7,200 TAO to ~3,600 TAO. Without external revenue, miners face unsustainable economics.

Bittensor is best understood as a token-incentivized inference marketplace, not a training protocol. The dTAO upgrade improved governance but did not solve the fundamental problem: stake-weighted rewards create plutocratic dynamics where capital, not computation quality, determines payouts. For Bitcoin developers, the comparison to Bitcoin's proof-of-work is misleading — in Bitcoin, hash rate directly maps to block production. In Bittensor, TAO stake is a governance/reputation token that loosely correlates with AI output quality.

Status: Shipping (inference marketplace). Not training models.

Sources: Bittensor Docs, arXiv 2507.02951, Metalamp Overview, GCR Deep Dive, Taostats

3. Gensyn

Architecture

Gensyn is a decentralized protocol for coordinating ML compute across heterogeneous hardware. The design splits participants into Solvers (who execute training tasks), Verifiers (who check proofs), and Whistleblowers (who challenge incorrect verifications). Tasks are assigned, executed, and verified through an on-chain coordination layer.

Verification: Verde

Gensyn's core innovation is Verde — a verification system for machine learning based on refereed delegation. During training, Solvers checkpoint model weights at scheduled intervals and store metadata (parameters, dataset indices, random seeds) so that any optimization step can be replicated. Verifiers re-run portions of the proof and compute distance metrics against thresholds established during a profiling stage. Whistleblowers can challenge Verifier work through on-chain arbitration. This avoids re-running entire training jobs — only disputed steps are recalculated.

The approach is probabilistic rather than deterministic: ML training has inherent non-determinism from floating-point operations, so Verde uses statistical distance thresholds rather than exact replication.

Token Requirement

Yes. Gensyn launched its AI token via English Auction in December 2025 (300M tokens, 3% of supply). The token is required for network participation.

Actual Results

• Testnet: Launched March 2025 with persistent identity and on-chain coordination
• Adoption: 43,600 monthly active wallets, 10.7M signature requests on testnet
• Mainnet: As of November 2025, Gensyn stated it was "3-4 weeks from mainnet launch," followed by auditing. Current status appears to be late-stage testnet transitioning to mainnet
• Models trained: Gensyn has published some Hugging Face models under its organization, and demonstrated reinforcement learning verified on-chain, but no marquee model has emerged from the protocol

Honest Assessment

Gensyn has real technology — Verde is a genuinely novel contribution to ML verification, and $51M in funding (led by a16z Crypto) provides substantial runway. However, the project has been in development since 2022 with only a testnet shipping by 2025. The token launch preceded mainnet, which is a common pattern in crypto-AI where economic infrastructure ships before the technical product. The probabilistic verification approach is sound in theory but unproven at the scale needed for serious training runs.

The fundamental question remains: can you build a viable two-sided marketplace for ML compute when centralized alternatives (Lambda, CoreWeave, Together) offer simpler economics and guaranteed SLAs?

Status: Late testnet. Real technology, not yet proven at scale.

Sources: Gensyn Docs, Verde Paper, arXiv 2502.19405, Gensyn Litepaper, CoinDesk Opinion

4. Together AI

Architecture

Together AI started as a decentralized training research project and demonstrated genuinely novel distributed training techniques. Their key contribution was GPT-JT (6B), trained on the "Together Research Computer" — a distributed cluster of GPUs connected over commodity internet.

Their decentralized training algorithm used local SGD with randomly skipped global communications, reducing inter-GPU communication from 633 TB (traditional 4-way data parallelism) to 12.7 TB per machine for 3.53B tokens — adding only ~50% overhead to end-to-end training time. This was run over 1 Gbps connections, not datacenter interconnects.

What Actually Happened

Together shipped three concrete things:

1. GPT-JT (6B): A model competitive with InstructGPT davinci v2 on classification tasks, trained across distributed GPUs. Shipped in 2022.
2. RedPajama dataset: First v1 (1.2T tokens replicating LLaMA's training mix), then v2 (30T tokens with quality annotations). Over 500 models have been built on this data by the community.
3. Decentralized training algorithm: Published and demonstrated, proving that local SGD with sparse communication can match AdamW + All-Reduce convergence.

The Pivot

Together then pivoted to become a centralized AI cloud provider. As of 2025, the company hit $300M annualized revenue, is building its own datacenters (Maryland, Memphis), deploying NVIDIA Blackwell clusters, and acquired Refuel.ai for data transformation. The decentralized training research was the foundation, but the business is a conventional cloud inference and training API.

Honest Assessment

Together is the most instructive case study in this survey. They proved decentralized training could work at meaningful scale, then abandoned it because centralized infrastructure was a better business. The decentralized training algorithm was real and novel, but the economics of serving customers — who want reliability, SLAs, and one-click deployment — pushed them toward conventional cloud infrastructure.

Status: Pivoted to centralized cloud. Decentralized training research was real but abandoned.

Sources: Together Blog - GPT-JT, RedPajama v2, Sacra Revenue Analysis

5. Hivemind / Petals

Architecture

Hivemind is a PyTorch library from the Learning-at-Home research group for decentralized deep learning. It uses Kademlia-based distributed hash tables (DHT) — the same protocol family as BitTorrent — for peer discovery, with logarithmic search complexity scaling to tens of thousands of peers.

Key components:

• Decentralized Mixture of Experts (DMoE): A layer type where experts are hosted across different machines, with the DHT handling discovery
• Decentralized parameter averaging: Iteratively aggregates updates from workers without full network synchronization
• Fault-tolerant backpropagation: Forward/backward passes succeed even if nodes are unresponsive, treating failures as dropout

Petals builds on Hivemind specifically for large model inference and fine-tuning. It splits a model (originally BLOOM-176B) into blocks hosted across volunteers' GPUs in a BitTorrent-style pipeline.

Actual Results

• BLOOM-176B inference: Petals achieves ~1 token/second on consumer GPUs, 3-25x faster latency than offloading approaches. Clients need only 12 GB RAM (for embedding parameters) and 25 Mbps bandwidth.
• Fine-tuning: Natively exposes hidden states for parameter-efficient fine-tuning methods
• Bengali language model: Collaborative pretraining event with ~40 volunteers demonstrated the approach works in practice
• Smaller models: Several models trained collaboratively using Hivemind, though none at BLOOM's 176B scale

Honest Assessment

Hivemind/Petals is the most academically honest project in this space. It is pure open-source research with no token, no fundraising hype, and realistic claims. The DHT-based architecture genuinely works for inference of large models across consumer hardware. The limitation is economic: without a payment/incentive layer, participation depends on altruism or academic motivation. The network effect is fragile — if volunteers stop serving blocks, the model goes offline.

The BLOOM collaboration (BigScience) itself was a 1,000-researcher effort coordinated by Hugging Face, and while BLOOM-176B was a real model, it was trained on a centralized NVIDIA Jean Zay supercomputer, not through Hivemind's decentralized approach. Hivemind/Petals provides the inference infrastructure, not the training at that scale.

Status: Shipped (research-grade). Real working software, limited to inference at scale. No incentive layer.

Sources: Hivemind GitHub, Petals GitHub, Petals Paper, Distributed Inference Paper, Yandex Research Blog

6. Prime Intellect

Architecture

Prime Intellect is the standout project in actual decentralized training. Their approach is built on DiLoCo (Distributed Low-Communication Learning), originally proposed by DeepMind, which they extended into OpenDiLoCo — an open-source framework for globally distributed training with drastically reduced communication requirements.

The key insight: instead of synchronizing gradients every step (which requires datacenter-grade interconnects), DiLoCo synchronizes "pseudo-gradients" every H steps (typically H=100-500), with each node running a local optimizer in between. Combined with int8 quantization of pseudo-gradients, this achieves:

• 400x reduction in communication bandwidth vs. traditional data-parallel training
• Up to 2,000x reduction in total communication volume with 8-bit quantization
• 90-95% compute utilization across nodes on different continents

Their production framework, prime, replaced OpenDiLoCo with improved fault tolerance, bandwidth utilization, and scalability.

Actual Results

This is the strongest record in the space:

1. OpenDiLoCo (2024): Scaled DiLoCo to 1.1B parameters (3x the original), trained across two continents and three countries. Final perplexity 10.76 vs. 11.85 baseline. Published as arXiv 2407.07852.
2. INTELLECT-1 (Late 2024): First decentralized training of a 10B parameter model (LLaMA-3 architecture), trained across GPUs in the US, Europe, and Asia. Open-sourced under Apache 2.0.
3. INTELLECT-2 (May 2025): First globally distributed reinforcement learning training of a 32B parameter model, using fully asynchronous RL across permissionless compute contributors. Improved upon QwQ-32B (state-of-the-art 32B reasoning model). Built with PRIME-RL, featuring TOPLOC (verifies rollouts from untrusted inference workers) and SHARDCAST (broadcasts policy weights efficiently). Published as arXiv 2505.07291.
4. INTELLECT-3 (January 2026): 100B+ parameter Mixture-of-Experts model trained on their RL stack, achieving state-of-the-art for its size across math, code, science, and reasoning benchmarks.

Verification: TOPLOC

For INTELLECT-2, Prime Intellect introduced TOPLOC — a verification mechanism for rollouts from untrusted inference workers during RL training. This is a practical, lightweight approach to the verification problem that avoids the heavyweight proof systems used by projects like Gensyn.

Token Requirement

No native token as of early 2026. Prime Intellect operates permissionless compute pools where anyone can contribute GPUs. The economic model is based on open participation rather than token staking.

Honest Assessment

Prime Intellect is the clear leader in decentralized training by results. They have trained progressively larger models (1.1B, 10B, 32B, 100B+) across genuinely distributed infrastructure, published technical reports, and open-sourced their frameworks. The INTELLECT series represents the most significant demonstration that decentralized training can produce competitive models.

The limitation is that their verification approach (TOPLOC) is purpose-built for RL rollouts and doesn't generalize to arbitrary pretraining verification. Their economic model also remains unclear — without a payment layer, the incentive for compute contributors to participate long-term is underdeveloped.

Status: Shipping. The strongest technical results in decentralized training.

Sources: OpenDiLoCo Paper, INTELLECT-1 Blog, INTELLECT-2 Paper, INTELLECT-2 Release, Prime Intellect Blog

7. Nous Research / Psyche

Architecture

Nous Research is building Psyche, a decentralized training network coordinated on the Solana blockchain. The technical foundation is DisTrO (Distributed Training Over-the-Internet), a family of distributed optimizers that reduce inter-GPU communication by 1,000-10,000x.

DisTrO's key result: tested on 32x H100 GPUs with the LLaMA 2 1.2B architecture, it reduces per-step communication from 74.4 GB (All-Reduce) to 86.8 MB (857x reduction) while matching AdamW convergence. For fine-tuning and post-training, reduction reaches 10,000x with no loss degradation. This enables training over consumer internet connections (100 Mbps down / 10 Mbps up).

Psyche builds on DisTrO to add coordination, fault tolerance, and permissionless participation through blockchain-based consensus on Solana.

Actual Results

• DisTrO optimizer: Published and demonstrated on LLaMA 2 1.2B, showing convergence matching centralized training with 857x less communication. Published August 2024.
• 15B parameter test run: December 2024, scaled through 11,000 training steps
• 40B parameter testnet run: Launched as part of Psyche testnet, described as "powerful enough to serve as a foundation for future pursuits in open science"

Token Requirement

The NOUS token exists but Psyche is still in testnet. The full token economics for mainnet participation are not yet finalized.

Honest Assessment

Nous has strong ML credentials — they produced the Hermes fine-tune series, which are genuinely popular open-source models. DisTrO is a real optimizer with published results. However, Psyche remains in permissioned testnet as of early 2026, and the 40B training run has not produced a publicly benchmarked model that demonstrates competitive quality.

The $65M in funding (including $50M from Paradigm) provides substantial runway, and the 2026 roadmap targets scaling to 100K+ nodes. But the gap between "optimizer works on 32 H100s" and "permissionless global network trains competitive models" is enormous. The Solana coordination layer adds complexity without an obvious advantage over simpler approaches.

Status: Testnet. Real optimizer technology, network not yet permissionless.

Sources: Nous Psyche Blog, DisTrO GitHub, VentureBeat - DisTrO, VentureBeat - Training Run, SiliconANGLE - Funding

8. Compute Marketplaces

These projects provide GPU rental infrastructure rather than training coordination. They are adjacent to decentralized training but solve a different problem: matching GPU supply with demand.

io.net

• What it is: Aggregates GPU capacity from data centers, miners, and individuals into a unified cluster
• Scale: Expanded from 60,000 verified GPUs (March 2024) to 327,000 (March 2025)
• Revenue: Monthly revenue exceeded $1M by January 2025 ($64K daily average)
• Token: IO token required for transactions
• Honest take: Genuine traction and the largest GPU inventory in the decentralized space. However, aggregating GPUs is not the same as coordinating training across them. io.net is a marketplace, not a training protocol.

Akash Network

• What it is: Reverse auction marketplace — users post compute requirements, providers bid
• Scale: 736 GPUs at 70% utilization, $4.3M annual revenue
• Growth plans: Acquiring ~7,200 NVIDIA GB200 GPUs through Starcluster initiative (late 2025/2026)
• Token: AKT token for transactions and staking
• Honest take: Akash is the most mature decentralized compute marketplace with real revenue and utilization. But it is general-purpose cloud, not AI-specific. The reverse auction model is elegant for simple workloads but doesn't solve the coordination problems specific to distributed training.

Render Network

• What it is: Distributed GPU rendering network, expanding into AI inference
• Scale: Token burn increased 278.9% YoY through 2025, indicating growing usage
• Token: RENDER token for job payments
• Honest take: Originally built for 3D rendering, Render is pivoting to AI but remains primarily a rendering network. The AI inference capabilities are early-stage.

Nosana

• What it is: Solana-based GPU compute marketplace focused on AI inference
• Scale: Surpassed 50,000 GPU hosts since mainnet launch (January 2025)
• Token: NOS token
• Honest take: Rapid supply-side growth, but 50K hosts raising the question of whether demand matches supply. Inference-focused, not training.

Collective assessment: These marketplaces solve the supply-side problem (finding cheap GPUs) but not the demand-side problem (coordinating those GPUs into coherent training runs). A training protocol like l402-train would potentially use these as GPU sourcing layers.

Sources: Nansen - io.net Analysis, Akash Roadmap 2025, DePIN Scan - Nosana, Nosana 2025 Wrap, Messari - Akash Q3 2025

9. Ritual

Architecture

Ritual is building an AI inference coprocessor for smart contracts. The flagship product, Infernet, connects off-chain AI computation with on-chain smart contracts on any EVM-compatible chain. The roadmap includes Ritual Chain, a sovereign execution layer for verifiable AI inference.

Verification

Infernet provides computational proofs for inference results, allowing smart contracts to consume AI outputs with cryptographic guarantees. The approach is purpose-built for the "oracle problem" of getting AI results on-chain, not for training coordination.

Token Requirement

No token launched yet. Ritual Chain testnet is invite-only.

Actual Results

Infernet is live and integrated with multiple EVM chains. Ritual has raised $25M (Series A, June 2024) and partnered with Nillion for privacy-preserving inference. However, the product is narrowly focused on smart contract integration — it does not attempt training and is not relevant to the decentralized training problem.

Status: Shipping (inference oracle). Not a training protocol.

Sources: Ritual Blog, Ritual Docs, Gate Guide

10. Other Notable Projects

Morpheus (MOR)

Morpheus is an incentive protocol for open-source personal AI agents, live on Arbitrum. It rewards four groups: code contributors, compute providers, capital providers (stakers), and community builders. The protocol is about agent coordination, not model training. No models have been trained through the network. MOR token powers the economy.

Status: Mainnet (agent infrastructure). Not a training protocol.

Artificial Superintelligence Alliance (FET / SingularityNET / Ocean Protocol)

Three projects merged in 2024 into the ASI Alliance, consolidating tokens into FET. SingularityNET provides an AI marketplace, Fetch.ai builds autonomous agent infrastructure, and Ocean Protocol handles data tokenization. The 2026 roadmap includes ASI Chain (a modular blockchain for AI coordination) and an "Agentic Discovery Hub."

Despite grandiose naming ("Artificial Superintelligence"), no actual training infrastructure exists. These are middleware and marketplace layers with governance tokens. The merger appears primarily motivated by token economics consolidation rather than technical synergy.

Status: Shipping (marketplaces/middleware). Not training. Heavy on narrative.

AIArena

A lesser-known project deployed on Base blockchain (Sepolia testnet) that has attracted 600+ training nodes, 1,000 validators, and 63,000 delegators, producing ~19,000 AI models. Evaluations claim decentralized models "significantly outperform" baselines, though independent verification is limited.

Status: Testnet with notable participation metrics.

Sources: Morpheus, ASI Alliance, AIArena - ACM 2025

11. Comparison Table

12. What l402-train Does Differently

Every project in this survey makes at least one of these compromises:

1. Token requirement. Bittensor, Gensyn, Nous/Psyche, and every compute marketplace require buying and staking a project-specific token to participate. This creates a circular economy where the token's value depends on the network's success, which depends on adoption, which is gated by the token's cost. It also exposes participants to token price volatility that has nothing to do with their compute contribution.
2. Stake-weighted rewards. Bittensor's empirical data shows this plainly: rewards flow to capital, not computation quality. When your payout depends on how much token you hold rather than how good your gradients are, the system selects for whales, not capable miners.
3. No payment infrastructure. Prime Intellect and Hivemind have the best technical approaches but no payment layer. Participants contribute compute for free or for token emissions with unclear value. This limits participation to researchers, enthusiasts, and token speculators.
4. Centralized coordination in practice. Gensyn's testnet, Nous's Psyche, and Bittensor's Subtensor all rely on centralized entities for network operation, even when the architecture is nominally decentralized.

l402-train addresses these with Bitcoin's Lightning Network:

• No token. Payment is in bitcoin via Lightning. No project-specific token to buy, stake, or speculate on. Participants earn sats proportional to validated gradient quality.
• Hold invoice escrow. Lightning hold invoices create a trustless escrow mechanism: the coordinator locks payment when assigning work, settles on validated delivery, cancels on rejection. No on-chain transaction cost for the escrow itself. This is simpler and cheaper than smart contract escrow, DLC-based approaches, or token staking/slashing.
• Permissionless participation. Anyone with a GPU and a Lightning node can contribute. No minimum stake, no registration burn, no token purchase. The barrier to entry is compute capacity, not capital.
• Deterministic verification. Gradient validation uses deterministic replay on reference data shards — the coordinator can independently verify any submitted gradient by re-running the computation. Combined with SparseLoCo gradient compression, the verification cost is a fraction of the compute cost.
• Payment proportional to quality. Rewards are calculated from gradient utility (measured by loss reduction on validation data), not from stake weight or token holdings. Better gradients earn more sats.

The tradeoff is that l402-train starts with a single coordinator (Phase 1) and targets smaller models (0.5B-3B). This is intentional: the goal is to prove that Lightning micropayments can coordinate distributed training before scaling to federated validators (Phase 3+) and larger models.

The honest question l402-train must answer is the same one Together AI answered by pivoting: can decentralized training economics compete with centralized cloud? The hypothesis is that Lightning's near-zero transaction costs and instant settlement enable a granularity of payment (per-gradient micropayments) that token-based systems cannot match, and that this granularity is the key to aligning incentives at the individual gradient level rather than the epoch or session level.

Key Takeaways

1. Only two projects have trained competitive models through decentralized infrastructure: Prime Intellect (INTELLECT 1/2/3) and Together AI (GPT-JT, before pivoting). Everything else is either inference, marketplace, testnet, or narrative.
2. The communication bottleneck is solved. DiLoCo (Prime Intellect), DisTrO (Nous), and local SGD (Together) all independently demonstrated 100-2,000x communication reduction. This is no longer the hard problem.
3. Verification of untrusted computation remains the hard problem. Gensyn's Verde is the most sophisticated attempt, but is unproven at scale. Prime Intellect's TOPLOC works but is narrow (RL rollouts only). Bittensor punts to subnet-specific implementations. No universal, lightweight verification mechanism for distributed training has been demonstrated.
4. Token economics create misaligned incentives. The empirical evidence from Bittensor is clear: stake-weighted rewards benefit capital holders, not compute contributors. Lightning micropayments offer a fundamentally different incentive structure.
5. The market is large and underserved. With $300B+ in AI infrastructure spending in 2025 alone, even a small share of training compute shifting to decentralized infrastructure represents significant demand. The question is whether decentralized approaches can offer cost, censorship-resistance, or accessibility advantages that justify their coordination overhead.

Sources (survey-level): ACL Survey - Beyond A Single AI Cluster, Prime Intellect - State of Decentralized Training, CoinDesk - Decentralized AI Leveling the Playing Field