AI that is not centralized.

Inference verification is easy (re-run and compare). Training verification is an unsolved problem because training is non-deterministic across hardware. If Gensyn cracks this at scale, it becomes the trust layer for every distributed training network.

Arguably the most technically important project in decentralized AI. Unlike most projects here, it is not a DePIN with a token — it is a research lab producing open-source infrastructure. The INTELLECT series is doing for decentralized training what Bitcoin's whitepaper did for decentralized money: proving the concept works at meaningful scale.

The highest-risk, highest-upside project on this entire map. Unextractability creates a digital commons that is usable but not appropriable. If Protocol Learning scales, it breaks the foundation-model oligopoly at the root. No token yet, pure research — the kind of team VCs should track closely even if production is 18-24 months out.

The strongest proof that decentralized training can produce frontier-quality models. What makes Nous singular is the full-stack position: the most popular open-weight models + a working training network + the fastest-growing agent framework + a $1B Paradigm mark. DisTrO's 10,000x bandwidth reduction is a genuine breakthrough on the bottleneck every training project faces.

The strongest consumer proof point in decentralized AI — real traction, not whitepaper vapor. Venice integrates new open-source releases within hours; as open-source closes the gap with GPT-5 and Claude, its privacy moat becomes the deciding factor for a growing segment of users.

Exo makes private, local inference practical at frontier-model scale. Before RDMA over Thunderbolt 5, consumer distributed inference was a toy; after it, a cluster of Macs matches datacenter latency at a fraction of the cost. No tokenomics — pure open-source tooling, which is both its strength (no mercenary dynamics) and its limit (no built-in flywheel).

The intersection of EigenCloud's verifiable compute and the Exo-style thesis that consumer Apple Silicon is untapped. The 95% operator share is designed to bootstrap supply. Watch it as a complement to Exo — Darkbloom adds the economic and verification layer Exo deliberately leaves out.

While most projects focus on verifiable compute (proving correctness), Nillion focuses on confidential compute (ensuring nobody sees the data). If AI agents process health, financial, and identity data, confidential compute is not optional — it is infrastructure. The Uber/Hedera/Indiegogo founding team is unusually strong.

The most ambitious attempt to make AI a first-class citizen of blockchain consensus. EVM compatibility is the right go-to-market — existing protocols integrate without migrating. Watch for the first DeFi protocol that ships a production AI integration on Infernet.

PIPE is a genuinely novel contribution and MemSync gives them a consumer-facing wedge that isn't purely crypto-native. If agents start managing real money, verifiable inference becomes a regulatory requirement, not a nice-to-have.

From an AI standpoint the restaked-ETH TVL is beside the point — what matters is EigenAI's bit-exact determinism: identical inference across 10,000 runs, which is what lets a smart contract trust a model's output. Restaking just supplies the cryptoeconomic security underneath; the verifiable-cloud pivot is what makes EigenCloud relevant to AI at all.

The most practical path to making smart contracts "intelligent." The IMO mechanism creates a novel funding model for open-source AI — potentially more sustainable than corporate charity.

The most traction of any AI agent platform by a wide margin — a functioning economy. Watch SAM and the robotics integration; that is where agents cross from digital to physical.

Analogous to what React became for web frontends. The plugin model and community size create hard-to-replicate network effects. The risk: framework standardization is a thin-margin "picks and shovels" play. The opportunity: every agent platform needs a framework, and ElizaOS is already there.

OML tackles the sustainability problem of open-source AI. Meta releasing Llama is corporate charity, not a durable model; OML creates verifiable ownership and economic flows for contributors. The Princeton/IISc pedigree and $85M war chest make this one of the most credibly resourced projects here.

NEAR's founder literally co-invented the architecture behind every modern LLM. Polosukhin: "The users of blockchain will be AI agents, not humans." Post-quantum integration shows forward thinking most L1s haven't started.

Addresses the most fundamental structural problem in AI: the disconnect between data creation and value capture. A Microsoft/Amazon/MIT client list suggests real demand; 1.4M daily actives on testnet is an unusually strong pre-mainnet signal (though testnet metrics are gameable).

0G tackles the most fundamental bottleneck for on-chain AI: data throughput. If the 50 GB/s holds in production, it is a step-change in what's possible — though the figure is a controlled-testnet measurement against a general-purpose settlement layer.

The only project that has actually trained a real foundation model on user-contributed data with verifiable ownership. It operationalizes "your data, your value" rather than theorizing about it. The MIT/Fed/World Bank pedigree is unusual for crypto.

A unique position: not compute, not inference, but data collection. 2.5M residential nodes is a supply-side moat no centralized company can replicate. Legal questions around scraping at scale are unresolved — but if training data becomes a regulated commodity, Grass's ZK-verified pipeline becomes the compliance layer.

The credibility anchor: $1M+ quarterly revenue, 70% GPU utilization, 8 years of operation. The thesis was never to out-scale AWS — it was to be the long tail, the censorship-resistant option, the 60-85% cheaper alternative. On those terms, Akash is delivering.

$128M+ in real revenue and a NASDAQ company signing a $260M GPU contract are numbers that make centralized clouds pay attention. Aethir is less "decentralized" in the purist sense — enterprise SLAs require coordination — but dramatically more attractive to enterprise buyers. The Axe deal is arguably the largest single contract in DePIN history.

The largest decentralized GPU aggregator by raw node count. Ray integration is pragmatic, and focusing on AI/ML cluster orchestration (vs. general compute) is the right product decision.

The deepest real-world utility roots — serving Hollywood studios years before "DePIN" existed. The pivot into AI is strategically sound: GPU rendering and AI inference share identical hardware. The Salad integration could dramatically expand supply.

When Stripe integrates it on Base, AWS embeds it in Bedrock AgentCore, and Visa partners with Coinbase for settlement, the thesis is validated. The economics are unanswerable: sub-$0.001 L2 fees vs. $0.30 + 2.9% on cards. For an agent making 10,000 API calls in a workflow, card rails are not expensive — they are mathematically impossible.

Skyfire occupies the critical gap: x402 handles payment mechanics but doesn't answer "should this agent be trusted to transact?" Enterprise adoption requires identity and compliance — KYA is becoming the identity primitive alongside the payment rails.

The most intellectually novel project on this entire map. The cryptographic primitive is peer-reviewed and solves a problem everyone thought impossible: making mining do useful AI work with zero overhead. The Together AI partnership shows it isn't theoretical. The Princeton/Columbia/Hebrew University pedigree puts it in a different credibility class. Watch this one closely.

Rather than training one model, Allora creates a competitive marketplace where accuracy is the only thing that matters. 692M inferences demonstrate real throughput, and the self-improving dynamic is the kind of network effect that compounds.