AGI Décentralisée Expliquée : Comment la Blockchain Alimente l'IA (2026)

Decentralized AGI is the thesis that the path to artificial general intelligence does not have to run exclusively through three or four corporate laboratories with private data centers, and that open networks coordinated by blockchains can assemble the same four ingredients - compute, data, models and capital - from a permissionless global supply base. This guide is a complete technical and market map of that thesis as it stands in September 2026.

The idea is no longer purely speculative. Over the last two years, three independent lines of research converged. First, communication-efficient training algorithms such as DiLoCo, OpenDiLoCo and DeMo cut the bandwidth requirement of distributed model training by two to three orders of magnitude, making it possible to train serious models across the public internet rather than inside a single InfiniBand fabric. Second, verifiable computation matured: trusted execution environments moved onto datacenter GPUs with low single-digit percentage overhead, while optimistic and cryptographic verification schemes made it economically rational to trust an anonymous GPU operator. Third, crypto-economic design improved enough that networks can price and reward a fuzzy, subjective good like model quality rather than a crisp, objective one like a hash.

What you will find below:

• A precise definition of decentralized AGI and, just as importantly, what it is not. Most tokens marketed under the AGI banner are not building general intelligence; they are building one layer of the stack, and the distinction matters enormously for valuation.
• The macro case for why 2026 is the year this sector stopped being a narrative trade and started producing measurable revenue, driven by the inference cost curve, energy constraints on hyperscaler expansion, and the open-weight model explosion.
• Detailed analysis of the leading networks: Bittensor (TAO), NEAR Protocol (NEAR), the Artificial Superintelligence Alliance (FET), Render (RENDER), Akash (AKT), plus the compute and agent layers built on Ethereum (ETH), Solana (SOL), Cosmos (ATOM) and Arbitrum (ARB).
• A technology deep dive on the verification problem, which is the single hardest unsolved issue in the field, covering replication consensus, optimistic fraud proofs, zero-knowledge machine learning and confidential computing.
• A market analysis that separates circulating market capitalization from actual network revenue, and an honest assessment of the gap between the two.
• A risk map covering token emission economics, subnet quality dispersion, regulatory exposure under the EU AI Act and US market structure rules, and the existential competitive risk that centralized labs simply move faster.
• A 2026-2030 outlook with explicit probability framing rather than price targets.

A note on the sentiment data accompanying this guide. Our platform currently registers maximum confidence readings on ETH, SOL, BNB, ADA, DOGE, XLM, LTC, UNI, APT, ATOM and ARB, with slightly lower but still elevated readings on XRP (92), BTC (91), NEAR (90) and TRX (87). Uniformly saturated sentiment across a broad basket of large caps is itself a signal: it typically indicates a market in a broad risk-on regime rather than one making fine-grained distinctions between assets. In that environment, thematic sectors like decentralized AI tend to outperform on the way up and drawdown harder on the way down, because positioning is crowded and the fundamental anchors are weak. Read the project analysis below with that context in mind.

Finally, a discipline note that runs through the whole document. The decentralized AI sector has an unusually high ratio of claims to verifiable facts. Wherever possible this guide distinguishes between what a network has shipped and can be measured doing, what it has demonstrated at research scale, and what remains roadmap. Figures cited are approximate and reflect conditions as of mid-2026; anything that moves daily, such as market capitalization or subnet emissions, should be verified against live data before it informs a decision.

Start with the terminology, because it is routinely abused. Artificial general intelligence refers to a system that can perform essentially any cognitive task a competent human can, transferring knowledge across domains without task-specific retraining. No such system exists in September 2026. Frontier models from the major laboratories are extraordinarily capable in language, code and increasingly in agentic tool use, but they remain narrow in ways that matter: brittle long-horizon planning, weak continual learning, no persistent self-modifying memory.

Decentralized AGI, then, is not a product you can buy today. It is an architectural bet with a specific claim: that the organizational form which produces AGI will be an open network, not a closed firm, and that blockchain rails are the coordination technology that makes such a network economically viable.

The four inputs and why they are hard to decentralize

Any path to general intelligence requires four inputs, and each has a distinct decentralization problem:

• Compute. Frontier training runs consume tens of thousands of interconnected accelerators for months. The constraint is not raw FLOPs, which are abundant in aggregate across the world, but interconnect. Gradient synchronization in standard data-parallel training moves the full parameter set between nodes on every step. At 400 Gbps inside a datacenter this is fine; at 1 Gbps across continents it is fatal. Decentralizing compute therefore requires changing the algorithm, not just the hardware sourcing.
• Data. The public web has been largely exhausted as a training corpus. The remaining frontier is proprietary, personal and real-time data: enterprise documents, health records, sensor streams, human preference feedback. This data is locked behind privacy law and commercial incentive. Decentralizing it requires mechanisms that let data contribute to a model without the contributor surrendering custody, which points to federated learning, secure enclaves and, eventually, encrypted computation.
• Model development. Research talent is the scarcest input. Centralized labs concentrate it with equity packages. Open networks must replace that with a credible ongoing income stream for contributors, which is what token emission schedules attempt to do, with mixed results.
• Capital. This is the one input crypto decentralizes natively and well. A network can raise and continuously allocate capital to contributors without a board, a term sheet or a jurisdiction.

Three distinct meanings of the phrase

When projects say decentralized AGI they usually mean one of three quite different things, and conflating them produces bad analysis.

1. Decentralized infrastructure for AI. Marketplaces for GPUs, storage, bandwidth and data. Render, Akash, io.net, Filecoin and Arweave live here. This is a real business with real revenue and essentially no AGI content. It is commodity infrastructure arbitrage: idle hardware priced below hyperscaler list rates. Valuable, measurable, and competitive on price rather than on intelligence.

2. Decentralized production of intelligence. Networks that use crypto-economic incentives to make anonymous participants produce better models, better inferences or better predictions than they would alone. Bittensor is the canonical example and remains the most intellectually ambitious design in the space. The claim here is genuinely about intelligence: a market mechanism that continuously selects for quality across a competitive population of miners. Prime Intellect and Nous Research pursue the same goal through open collaborative training rather than through a token-weighted consensus.

3. Decentralized ownership and governance of AI. The argument that whatever AGI turns out to be, its ownership should not be concentrated. NEAR articulates this most clearly with its user-owned AI framing, as does the Artificial Superintelligence Alliance. This is partly a technical program and partly a political one.

The blockchain is not the intelligence

The most common misunderstanding is that the chain itself does the thinking. It does not, and cannot: executing a transformer forward pass inside consensus would be absurdly expensive and pointlessly slow. The blockchain plays four supporting roles, all of them coordination roles: it settles payments between parties who have no relationship and no shared jurisdiction, it records commitments so that model weights, datasets and outputs have tamper-evident provenance, it runs the incentive mechanism that decides who gets paid how much for contributed work, and it provides identity and reputation so that a network can distinguish a consistently useful contributor from a freeloader without a KYC process.

Everything heavy happens off-chain. The chain is the accounting layer and the rulebook. Understanding this immediately clarifies which chains matter: the requirement is cheap, fast, high-throughput settlement with good developer tooling, which is why ETH rollups, SOL and Cosmos SDK chains dominate the sector rather than any chain claiming to compute AI on-chain.

Three structural shifts converted this sector from narrative to substance between 2024 and 2026. Each is independent of crypto market cycles, which is precisely why they matter.

The bottleneck moved from training to inference, and from chips to power

Through 2023 and 2024 the binding constraint on AI was accelerator supply. That constraint eased considerably by 2026 as successive GPU generations shipped in volume and as demand shifted composition. Today the binding constraints are electrical power and grid interconnection, and to a lesser extent memory bandwidth. Hyperscalers signing multi-gigawatt power purchase agreements and building their own generation are responding to a physical limit that capital cannot short-circuit quickly; interconnect queues in major markets run years long.

This is structurally favorable to decentralized compute for a specific reason. Distributed networks aggregate power that is already interconnected: gaming GPUs in residential buildings, underutilized racks in tier-two datacenters, stranded capacity at crypto mining sites with existing high-capacity connections, and idle enterprise hardware. They do not need new substations. When the marginal cost of centralized capacity is set by grid buildout rather than silicon, a network that monetizes existing capacity has a genuine and durable cost advantage in the workloads it can serve.

Simultaneously the workload mix shifted toward inference, and specifically toward inference-time reasoning where a model spends substantial compute per query generating and evaluating chains of thought. Inference is far more decentralization-friendly than training: each request is independent, latency tolerance varies widely by application, and there is no all-reduce. A batch job summarizing ten million documents overnight does not care whether it runs in Virginia or across four hundred consumer machines, it cares about cost per token. This is the workload that decentralized networks can serve today at genuine price advantage, and it is the fastest growing segment of AI spend.

Open-weight models closed the capability gap enough to matter

The second shift is that competent open-weight models became abundant. The efficiency shock that began in early 2025, when a series of open releases demonstrated frontier-adjacent reasoning at a fraction of the assumed training cost, permanently changed the sector's structure. By 2026 there is a deep bench of openly licensed models spanning small on-device sizes up to large mixture-of-experts architectures, plus mature distillation pipelines that compress frontier capability into models that run on a single accelerator.

This matters for decentralized AGI in a way that is easy to miss. A decentralized inference network is only useful if it has something worth serving. When the best models were closed and API-gated, decentralized serving had nothing to offer except worse models at lower prices. When open weights are within striking distance of closed frontier models on most practical tasks, the value proposition inverts: the network offers comparable capability with no vendor lock-in, no content policy imposed by a third party, no data retention, and no risk of the endpoint being deprecated. Open weights are the raw material that makes decentralized serving a real product rather than a discount bin.

Regulation started to reward auditability and provenance

The third shift is regulatory. The EU AI Act's general purpose model obligations phased in from August 2025, creating documented requirements around training data summaries, copyright compliance and systemic risk evaluation for large models. Parallel developments in the United States moved market structure for digital assets toward clarity, while sectoral rules in health and finance continued to tighten on data residency and model explainability.

Compliance is usually framed as a cost for crypto. Here it cuts the other way. Requirements to document what data trained a model, to demonstrate that contributors consented, to attribute outputs, and to prove that a given inference came from the model it claims to come from are exactly the problems that content-addressed storage, on-chain provenance records and attestation-based verification solve natively. Projects like Story Protocol on the intellectual property side and Vana on the data contribution side exist specifically to make training data attributable. A centralized lab has to build an internal audit trail and ask regulators to trust it. A network built on public commitments can simply point at them.

The countervailing force

Honesty demands the other side. The gap between frontier centralized capability and anything a decentralized network has produced remains wide. No open network has trained a model that competes with the frontier at the frontier. The largest fully decentralized training runs publicly demonstrated - roughly in the tens of billions of parameters - are one to two orders of magnitude below the largest centralized runs by compute budget. The sector's bet is that this gap is a matter of engineering and time rather than a fundamental wall. That bet is not yet settled, and anyone allocating to the sector should hold it as a hypothesis rather than a conclusion.

The decentralized AI stack has stabilized into five layers. Understanding where a project sits determines what it competes on, how defensible it is, and how its token could plausibly accrue value.

Layer 1: Physical resources

Raw GPUs, CPUs, storage and bandwidth contributed by operators ranging from individuals with a single consumer card to professional datacenters with racks of enterprise accelerators. The economics here are pure commodity supply. Networks compete on utilization rate and on the friction of onboarding supply. The key metric is not how many GPUs are registered but what fraction are actually rented and paid for, a number that historically ran far below headline capacity claims across the entire DePIN compute sector.

Layer 2: Orchestration and scheduling

Turning a heterogeneous pool into a usable cluster. This is genuinely hard engineering: matching workloads to hardware with the right memory profile, handling node churn mid-job, checkpointing so a dropped node does not kill a multi-hour run, containerizing environments so a job runs identically everywhere, and routing around network topology. Akash uses a reverse auction where providers bid on deployment requests. io.net clusters geographically dispersed GPUs into logical clusters with a Ray-based orchestration layer. Render uses a tiered node system that segments jobs by trust and hardware class.

Layer 3: Verification

The hardest layer, and the one that determines whether the whole edifice stands. The problem statement is simple: you paid an anonymous operator to run a model on your input. How do you know they ran the model you specified, on the input you gave, honestly, rather than returning cached output, a cheaper smaller model, or noise? There are four families of answer.

• Replication and consensus. Have multiple independent parties do the work and compare. Cheap to reason about, expensive in raw compute (you pay N times for one result), and vulnerable to collusion if the replicas can coordinate. Bittensor's Yuma Consensus is a sophisticated variant: validators independently score miner outputs, and a miner's reward derives from the stake-weighted consensus of those scores, with a clipping mechanism that neutralizes validators whose opinions deviate too far from the weighted median. It converts a subjective quality judgment into a Schelling point.
• Optimistic verification with fraud proofs. Assume the result is honest, post a bond, allow a challenge window during which any watcher can dispute and force a recomputation. Cheap in the happy path. Requires economically motivated watchers, meaningful bonds, and a dispute resolution path that is itself trustworthy. Gensyn's approach to training verification builds on this family, combining probabilistic proof-of-learning with graph-based pinpointing so a dispute can be narrowed to a single divergent operation rather than requiring the whole run to be redone.
• Cryptographic verification (ZKML). Produce a zero-knowledge proof that a specific computation graph was executed on specific inputs. Mathematically airtight and requires trusting nobody. The problem is overhead: proving a transformer forward pass historically cost orders of magnitude more than running it, with proving time and memory scaling badly with model size. Toolchains such as EZKL and various proving stacks have improved substantially, and ZKML is now practical for small models, classifiers and decision layers where the verified computation is worth far more than the compute it consumes. It is not yet practical for verifying large language model inference at scale.
• Hardware attestation (TEE). Run the workload inside a trusted execution environment on the accelerator itself, which produces a signed attestation of the code and data it executed. Confidential computing on datacenter GPUs is now mature, with overhead measured in low single-digit percentages for most workloads rather than the multiples imposed by cryptographic proving. The tradeoff is that you replace trust in the operator with trust in the silicon vendor and its attestation service, which is a real but very different risk. This has become the pragmatic default for production decentralized inference, and it is central to NEAR's confidential AI design.

Layer 4: Incentives and mechanism design

Given a verified output, who gets paid what? This is where blockchains are indispensable and where most of the genuine innovation sits. The design problem is that intelligence quality is subjective and multi-dimensional. You cannot write a closed-form scoring function for whether a model response is good. Networks therefore use either objective proxies (loss on a held-out set, latency, uptime), validator scoring with consensus (Bittensor), or market outcomes (a prediction network where accuracy is settled by reality). Each has failure modes: proxies get gamed through overfitting to the benchmark, validator scoring can be captured by stake concentration, and market outcomes only work where ground truth eventually arrives.

Layer 5: Applications and agents

The top of the stack is where autonomous agents transact. Two 2025-2026 standards matter here: the x402 pattern, which revives the dormant HTTP 402 status code to let a machine client pay for an API call inline with stablecoins and no account, and ERC-8004, which specifies on-chain registries for agent identity, reputation and validation so that one agent can evaluate whether to trust another without a central directory. Together these make agent-to-agent commerce a technical reality rather than a pitch deck slide. The unresolved question is demand: an economy of agents paying each other requires agents that produce enough value to justify the payments, and that remains early.

What follows is an assessment of the networks that matter, organized by what they actually do. Figures are approximate and reflect mid-2026 conditions.

Bittensor (TAO) - the intelligence market

Bittensor is the most conceptually ambitious project in the sector and the closest thing to a decentralized AGI research program with a working economy. Its architecture is a root chain plus a growing population of subnets, each a self-contained competitive market for a specific machine intelligence task. Miners in a subnet produce outputs, validators score them, and Yuma Consensus converts stake-weighted validator opinions into emissions. Subnets now number in the low hundreds and cover text inference at scale, image generation, protein folding, financial time-series prediction, web-scale data collection, decentralized pretraining and more.

The February 2025 transition to dynamic TAO was the network's most consequential change. Each subnet now has its own alpha token with a liquidity pool against TAO, and emissions are allocated according to market pricing of those alpha tokens rather than by root validator voting. This turned subnet funding into a continuous market referendum: capital flows toward subnets the market believes are producing value. It is an elegant mechanism and also a volatile one, since alpha prices are thin and reflexive.

Token economics mirror Bitcoin deliberately: a 21 million hard cap, block emissions that halve at fixed supply milestones, with the first halving already passed and subsequent halvings stretching emission far into the future. TAO has traded across a very wide band since 2024, roughly between the low hundreds and the high seven hundreds of dollars, reflecting a market that has not settled on how to value it. Bull case: Bittensor is a permissionless funding and selection mechanism for machine intelligence research with genuine product traction in several subnets. Bear case: quality dispersion across subnets is extreme, a meaningful share of emissions funds activity with no external customer, and much of the demand for TAO has historically been from participants seeking emissions rather than from end users buying inference.

NEAR Protocol (NEAR) - user-owned AI and chain abstraction

NEAR is unusual in that its founding team includes a co-author of the transformer paper, and the project has redirected substantial effort toward AI since 2024. Its thesis is user-owned AI: users should control the models and data that act on their behalf, with privacy enforced by confidential hardware rather than by policy. The practical stack combines NEAR's sharded high-throughput chain, an intents system that lets a user express a desired outcome and have solvers compete to fulfill it across chains, and confidential compute infrastructure for running models on private data with attestation.

Intents are the underrated piece. An autonomous agent that must execute across several chains has historically needed to manage keys, gas and bridges on each. NEAR's chain signatures and intent architecture collapse that into a single declarative request. This is infrastructure that agent developers actually need. NEAR carries a sentiment reading of 90 on our platform, slightly below the saturated large-cap cohort, which is consistent with an asset where the fundamental story is strong but the market is still pricing execution risk. Risk: NEAR competes for AI developer mindshare against both Ethereum's rollup ecosystem and Solana, and its AI products must win on merit rather than on chain loyalty.

Artificial Superintelligence Alliance (FET) - the merged agent stack

The ASI Alliance consolidated Fetch.ai, SingularityNET and Ocean Protocol under the FET ticker in 2024, later joined by CUDOS compute. The merger produced the sector's broadest single stack: autonomous agent frameworks and an agent marketplace from Fetch, neuro-symbolic AI research and the OpenCog lineage from SingularityNET, data exchange primitives from Ocean, and GPU capacity from CUDOS. Strength: vertical coverage from compute through data to agents, plus the longest-running AGI research pedigree in crypto. Weakness: merged organizations carry integration debt, and the combined entity has had to work to demonstrate that the whole exceeds the sum of three previously independent communities. The token is best understood as a diversified index of agent-economy exposure rather than a bet on one mechanism.

Render (RENDER) - GPU rendering expanding into AI

Render began as distributed GPU rendering for 3D artists, migrated from Ethereum to Solana, and uses a burn-and-mint equilibrium where clients burn credits and node operators are minted rewards. It has the sector's most legitimate claim to pre-existing product-market fit: real studios paid for real rendering jobs before any AI narrative existed. Its expansion into AI inference and diffusion workloads leverages the same node network. Key question: rendering and AI inference have different hardware and memory profiles, and the operators optimized for one are not automatically competitive at the other.

Akash (AKT) - the decentralized cloud

Akash is a Cosmos SDK chain running a reverse-auction marketplace for containerized compute. A user posts a deployment specification with a price ceiling; providers bid down. It is the most cloud-native design in the sector, since anything that runs in a container runs on Akash, and its GPU marketplace has been its growth engine. Strength: genuine substitutability with mainstream cloud for a real class of workloads, transparent on-chain lease data. Weakness: supply of high-end accelerators is contested against buyers who will pay more, and utilization of premium hardware is the metric to watch.

Open collaborative training: Prime Intellect and Nous Research

Neither is primarily a token story, and both matter more than most tokens. Prime Intellect demonstrated globally distributed training of a ten-billion-parameter model across continents, then followed with a distributed reinforcement learning run at larger scale, built on communication-efficient methods derived from DiLoCo. Nous Research developed the DeMo optimizer, which decouples momentum updates to slash inter-node communication, and operates a distributed pretraining network with coordination on Solana. These groups produce the empirical evidence that decentralized training works at all. If the decentralized AGI thesis is ever validated, it will be validated by results from work of this kind.

The supporting cast

• Ethereum (ETH) and Arbitrum (ARB): settlement and the natural home for agent standards including ERC-8004, plus restaking-based verification services.
• Solana (SOL): the throughput layer for high-frequency agent activity, DePIN bandwidth and data networks, and agent token launch infrastructure.
• Cosmos (ATOM): the SDK underlying Akash, Allora and several AI appchains, with IBC as the interoperability substrate.
• Filecoin and Arweave: verifiable storage for datasets and model weights, essential for provenance claims.
• Allora: a network where inference quality is judged by realized outcomes, a design that neatly sidesteps subjective scoring.

If you only understand one technical topic in this sector, make it this one. Verification and communication efficiency are the two walls between decentralized networks and serious AI workloads, and every meaningful engineering effort in the space is an attack on one of them.

Why distributed training was considered impossible

Standard data-parallel training splits a batch across N workers. Each computes gradients on its shard, then all workers perform an all-reduce to average gradients, then each applies the same update. The all-reduce moves data proportional to the model's parameter count, every single step. For a model with billions of parameters in half precision, that is gigabytes per step, thousands of times per hour. Inside a datacenter with high-bandwidth interconnect this is a manageable fraction of step time. Across the public internet at typical residential or commercial bandwidth, communication would dominate compute by orders of magnitude, and the cluster would spend nearly all of its time waiting on the network.

This is why the conventional wisdom held that serious training required colocated hardware. The conventional wisdom was correct about the naive algorithm and wrong about the problem.

How that wall was breached

Three algorithmic families changed the picture:

• Local SGD and DiLoCo. Instead of synchronizing every step, workers train independently for hundreds of local steps using an inner optimizer, then synchronize a pseudo-gradient representing their accumulated drift using an outer optimizer. This reduces synchronization frequency by two to three orders of magnitude with modest loss in final model quality. The open-source OpenDiLoCo implementation and its successors made this reproducible outside a single research lab.
• Gradient compression and momentum decoupling. The DeMo approach exploits the observation that momentum contains substantial redundancy and that only the fastest-moving components of the gradient need to be shared promptly. By decoupling and compressing what crosses the wire, bandwidth requirements fall by additional orders of magnitude, in some configurations enough to train over ordinary consumer connections.
• Asynchronous and fault-tolerant pipelines. Nodes joining and leaving mid-run, which is catastrophic for a synchronous cluster, is handled through elastic device meshes, frequent checkpointing and designs that tolerate stragglers rather than waiting for them.

The current state of the art: decentralized training of models in the tens of billions of parameters is demonstrated and reproducible. Decentralized reinforcement learning post-training, which is more communication-tolerant than pretraining because rollout generation is embarrassingly parallel, is further along still and may be where decentralized networks first achieve genuine competitive advantage. Frontier-scale pretraining across the open internet remains out of reach.

The four verification regimes, compared

Consider the decision from a network designer's perspective. You need verification that is cheap enough not to destroy your cost advantage, strong enough that fraud is unprofitable, and general enough to cover your workload mix.

Replication costs you a multiple of the base compute and gives probabilistic assurance that scales with the number of replicas and the independence of the replicators. It works well for tasks where outputs are comparable, which is why it suits scoring-based networks. It fails where outputs are legitimately non-deterministic, since two honest nodes running the same sampling-based generation produce different text, forcing you to compare on semantic quality rather than bitwise equality - which is exactly the subjective judgment problem you were trying to avoid.

Optimistic schemes cost near zero in the happy path, which is their great virtue. The economic security rests on three parameters: the bond size relative to the value of cheating, the challenge window length, and whether a rational watcher exists who profits from catching fraud. Get any of the three wrong and the scheme is theater. The subtlety for ML specifically is that recomputation for dispute resolution must be deterministic, which requires pinning floating point behavior, kernel versions and hardware classes, since the same model on two different accelerator generations can produce different bits.

ZKML is the only regime with no trust assumption at all. The overhead has fallen dramatically but remains the binding issue for large models: proving cost scales with circuit size, and a large transformer forward pass is an enormous circuit. The sensible deployment pattern in 2026 is selective: prove the small, high-stakes model that gates a financial decision, not the large model that drafts marketing copy. Expect continued progress through better proving systems, hardware acceleration and lookup-argument techniques for non-linear activations, but do not expect verified frontier inference at commodity prices this cycle.

TEE attestation has become the practical winner for production inference. Confidential computing modes on modern datacenter accelerators encrypt memory and produce a signed measurement of the executing code, with overhead low enough to be commercially irrelevant for most workloads. You get integrity and confidentiality simultaneously, which is what enterprise buyers actually want. The honest caveat is that the security model now includes the silicon vendor's attestation infrastructure and the historical record of side-channel attacks against enclave technologies. It is a trust reduction, not a trust elimination.

What this implies for project evaluation

When a network claims verifiable AI, the correct follow-up questions are: which regime, what does it cost as a percentage of the workload, what exactly is attested (the model weights, the input, the sampling parameters, all three?), and what happens economically when verification fails. Projects that cannot answer these crisply are describing an aspiration.

The decentralized AI sector's market capitalization and its revenue are two different stories, and conflating them is the most common analytical error made in this space.

Sector size and composition

Depending on how broadly you define membership, the crypto AI sector has oscillated between roughly twenty and seventy billion dollars in aggregate market capitalization since early 2024. The wide range reflects both genuine volatility and definitional drift, since index providers have periodically added agent tokens, DePIN compute and even meme assets with AI branding into the same bucket.

A more useful decomposition splits the sector into four blocks:

• Compute and infrastructure (Render, Akash, io.net, Aethir, Filecoin-adjacent storage): the largest block by count, characterized by measurable revenue, thin margins, and direct price competition with commodity cloud.
• Intelligence networks (Bittensor and its subnet alpha tokens, Allora): high conceptual value, high valuation multiples on modest external revenue, dominated by TAO.
• Platform chains with AI theses (NEAR, and the AI-focused activity on ETH, SOL, ATOM and ARB): valuations driven primarily by general chain fundamentals, with the AI narrative as an incremental multiple.
• Agent and application tokens: the most reflexive and shortest-duration block, where 2024-2025 saw explosive launch activity followed by severe attrition. Survivorship here is brutal and the base rate of any individual agent token retaining value over a full cycle is poor.

The revenue reality check

Here is the uncomfortable number that anyone allocating to this sector should internalize. Aggregate annualized external revenue across decentralized compute and inference networks - meaning payments from customers who are not themselves participants in the network's reward system - has historically been a small fraction of one percent of the sector's market capitalization. Even the strongest performers, where genuine paying demand exists, trade at revenue multiples that would be extraordinary for any conventional infrastructure business.

This is not automatically damning. Early-stage networks legitimately trade on option value, and token emissions functioning as growth subsidy is a recognized playbook that produced durable outcomes in other crypto verticals. But it does dictate the correct analytical posture: track the revenue trajectory, not the revenue level. A network whose external, non-incentivized revenue is compounding at a high rate off a small base is executing. A network whose activity metrics are flat once you strip out reward-seeking behavior is not, regardless of what its token does.

The metrics that resist gaming are: paid GPU-hours where the payer is not receiving emissions, repeat customer cohorts, gross margin per node after hardware and power, and utilization of premium hardware tiers rather than aggregate registered capacity. The metrics that are routinely gamed are: registered nodes, total addressable compute, transaction counts, and unique wallet addresses.

Reading the current sentiment picture

The sentiment data accompanying this guide shows maximum or near-maximum readings across an unusually broad set of large caps: ETH, SOL, BNB, ADA, DOGE, XLM, LTC, UNI, APT, ATOM and ARB all at 100, with XRP at 92, BTC at 91, NEAR at 90 and TRX at 87. Several observations follow.

First, uniformity is information. When sentiment saturates across assets with completely different fundamentals - a payments chain, a meme asset, a DeFi governance token and an L2 all reading identically - the readings are measuring market regime rather than asset-specific conviction. That regime is broadly risk-on.

Second, in risk-on regimes, thematic sectors with weak valuation anchors exhibit the highest beta. Decentralized AI has essentially no valuation anchor, since discounted cash flow analysis is meaningless at current revenue levels. This produces outsized upside participation and equally outsized drawdowns when the regime turns.

Third, the slightly lower readings on BTC (91) and NEAR (90) relative to the saturated cohort are worth noting for different reasons. A BTC reading below the alt cohort in a broad rally is a classic late-stage rotation signature, where capital has moved out along the risk curve. NEAR reading 90 while the broad large-cap set reads 100 suggests the market has not yet fully repriced its AI positioning, which is either an opportunity or an accurate discount for execution risk depending on your read of its product traction.

Growth trajectory scenarios

Framing the next twenty-four months as scenarios rather than forecasts:

• Infrastructure commoditization (most likely). Decentralized compute captures a meaningful but unspectacular share of the price-sensitive batch inference market. Revenue grows steadily, margins stay thin, and valuations gradually converge toward infrastructure multiples. Winners are the networks with the lowest customer acquisition cost and best orchestration engineering.
• Intelligence breakout. One or more decentralized networks produces a model or capability that is genuinely competitive on a benchmark that matters, validating the production thesis rather than just the infrastructure thesis. This would be a step-change for sector valuation and is most plausible in reinforcement learning post-training or in narrow specialized domains rather than general pretraining.
• Agent economy inflection. Machine-to-machine payments via x402-style rails and ERC-8004-style reputation reach genuine volume, making stablecoin settlement infrastructure and the chains hosting it the primary beneficiaries. In this scenario ETH, SOL and NEAR capture more value than the AI-specific tokens.
• Narrative exhaustion. Centralized capability advances fast enough and cheaply enough that the decentralized alternative is never economically compelling outside censorship-resistance niches. Sector multiples compress toward the small revenue base.

Direct comparison is complicated by the fact that projects marketed under the same banner often do not compete at all. Render and Bittensor are not substitutes; one sells GPU-hours, the other runs an incentive tournament. The useful comparisons happen within layers.

Compute marketplaces: Akash vs Render vs io.net vs Aethir

All four sell access to distributed accelerators, and they differentiate along four axes.

• Workload generality. Akash is the most general, since it orchestrates arbitrary containers and functions as a substitute for mainstream cloud compute rather than a specialist. Render is the most specialized, built around rendering pipelines with AI as an expansion. io.net sits between, purpose-built for ML clustering. Aethir focuses on enterprise-grade GPU delivery and gaming as well as AI.
• Supply quality. Enterprise-grade accelerators in professional datacenters with high-bandwidth interconnect are a fundamentally different product from consumer cards on residential connections. Networks that aggregate the former can serve training and latency-sensitive inference; networks weighted toward the latter are confined to batch and rendering work. This single distinction explains most of the variance in realized revenue per node.
• Pricing mechanism. Akash's reverse auction discovers price dynamically and transparently on-chain. Render's burn-and-mint equilibrium decouples the price a customer pays from the token's spot price, which is better for customer experience but makes token value accrual indirect. Marketplace models with fixed rate cards trade price discovery for predictability.
• Settlement layer. Cosmos SDK for Akash, Solana for Render and io.net. This matters less for end users than for the composability of the network with the rest of DeFi.

Assessment: this layer will consolidate. Commodity markets with low switching costs reward scale, and the winner will be decided by which network makes onboarding a customer's existing workload nearly frictionless, not by which has the most impressive capacity number.

Intelligence production: Bittensor vs Allora vs open collaboratives

Bittensor's edge is that it has a functioning economy with hundreds of millions of dollars in emissions actively funding thousands of independent participants to compete on machine intelligence tasks. That is a research funding mechanism operating at a scale no other decentralized system approaches. Its structural weakness is that emissions are paid regardless of whether external customers exist, which means a subnet can be well funded and commercially irrelevant simultaneously. Dynamic TAO improved this by subjecting subnets to continuous market pricing, but alpha markets are thin and can be as reflexive as they are informative.

Allora's design is philosophically cleaner for any task where reality eventually adjudicates. If the network is forecasting a price or an event, the outcome arrives and scoring is objective. This eliminates the validator-subjectivity attack surface entirely. The tradeoff is scope: it only works where ground truth is forthcoming, which excludes most generative tasks.

Prime Intellect and Nous Research represent a third model: open collaborative research without a token-weighted tournament. They produce the strongest technical results in the sector and have the clearest claim to advancing the state of the art. Their coordination problem is the opposite one - excellent research output, less developed mechanisms for sustainably compensating contributors at scale.

Platform chains: NEAR vs Ethereum vs Solana for AI

NEAR's differentiation is the tightest integration between chain-level primitives and AI requirements: intents for declarative cross-chain execution, chain signatures so an agent can control accounts on other chains from a single account, sharding for throughput, and confidential compute for private inference. It is the chain designed around the agent use case.

Ethereum's differentiation is gravity. Standards get written and adopted on Ethereum because that is where the developers, the stablecoins and the institutional infrastructure are. ERC-8004 agent registries matter precisely because they are Ethereum standards. An agent economy will settle substantial value on Ethereum and its rollups including Arbitrum regardless of where the compute happens.

Solana's differentiation is cost and throughput at the transaction layer, which is decisive for high-frequency agent activity and for DePIN networks that write many small records. Render, io.net and several data networks chose Solana for exactly this reason.

Cosmos, with ATOM at its center, remains the substrate of choice for teams that want their own sovereign chain with custom consensus and application logic, which is why Akash, Allora and several AI appchains built there.

The comparison that matters most

Set the internal rivalries aside for a moment. Every project in this guide competes primarily against centralized cloud and centralized model APIs, not against each other. A developer choosing where to run inference compares a decentralized network against a hyperscaler endpoint on four dimensions: price, latency, reliability and integration effort. Decentralized networks currently win decisively on price for batch workloads, lose on latency for interactive ones, lose on reliability SLAs, and lose significantly on integration effort. Three of those four are engineering problems with known solutions. The sector's trajectory depends on how fast they get solved.

A balanced framework for thinking about exposure to this sector, without pretending that the uncertainties are smaller than they are.

The structural opportunities

• Cost arbitrage with a durable source. The price gap between decentralized GPU-hours and hyperscaler list rates is not a temporary subsidy in every case. Part of it is genuine: aggregating already-built, already-interconnected, underutilized hardware has a real cost advantage over financing new datacenters and grid connections. Where that gap is structural rather than emission-funded, it persists.
• Regulatory tailwind on provenance. Requirements to document training data, attribute outputs and prove model identity favor architectures with native cryptographic provenance. This is a slow-moving but compounding advantage.
• Censorship resistance as a real product. A non-trivial set of applications cannot use centralized model APIs, whether for content policy reasons, jurisdictional reasons, or because the application competes with the provider. This demand is inelastic and pays a premium.
• Agent payment rails. If machine-to-machine commerce reaches meaningful volume, crypto rails are the only settlement layer that supports account-free, sub-cent, instant, programmatic payments. This may be the most underpriced opportunity in the sector, and its beneficiaries are chains and stablecoin infrastructure rather than AI-branded tokens.
• Asymmetric payoff on the production thesis. If a decentralized network ever demonstrates genuinely competitive model capability, the repricing would be violent. This is a low-probability, high-magnitude outcome, which is precisely the kind of exposure that warrants a small position rather than none.

The risks, in rough order of importance

1. Emission-funded activity masquerading as demand. The single largest analytical trap. A network paying participants in its own token generates activity that looks like traction and disappears when emissions fall. The diagnostic is to ask whether any given unit of activity would exist if emissions were zero. For most networks in this sector today, much of it would not.

2. Centralized labs simply move faster. Frontier laboratories have enormous capital, the best researchers, integrated hardware relationships and the ability to make decisions in a meeting rather than through a governance forum. If capability progress continues to require enormous colocated training runs, decentralized networks may never get closer to the frontier than they are now. The sector's entire premise is that this is not the case.

3. Token value accrual is genuinely unclear in several designs. Ask precisely: when a customer pays for inference, what happens to the token? Is it burned, staked, distributed, or merely a unit of account that the customer never touches because they paid in stablecoins? Several prominent networks have token economics where growth in usage produces little mechanical demand for the token. This is not fatal but it must be priced.

4. Quality dispersion and capital misallocation inside networks. In multi-subnet architectures, the distribution of value creation is extremely skewed. A handful of subnets produce most of the real output; many produce little of external value while still drawing emissions. An investor buying the base token is buying the average, not the best.

5. Verification failure. If a decentralized inference network is caught systematically returning cheaper or cached outputs than paid for, the reputational damage would be sector-wide and severe. Verification is not merely a technical nicety here, it is the product's integrity guarantee.

6. Regulatory exposure on two fronts. AI regulation may impose obligations that a permissionless network structurally cannot meet, such as identifying who trained a model or removing a capability on demand. Separately, securities and market-structure treatment of tokens whose value derives from network revenue remains jurisdiction-dependent and unsettled in places.

7. Hardware cycle risk. Decentralized networks are disproportionately supplied with previous-generation accelerators. When a new generation delivers a large efficiency jump, the economics of older hardware deteriorate sharply, and the networks holding that supply see margins compress.

8. Concentration and governance capture. Stake-weighted mechanisms concentrate influence in large holders. Where validators score miners, a coordinated stake majority can direct emissions toward itself. Networks vary widely in how seriously they have addressed this.

A practical evaluation checklist

Before forming a view on any project in this sector, get concrete answers to these:

• What percentage of network revenue comes from parties who receive no emissions?
• What is the verification regime and its overhead as a percentage of workload cost?
• What is the emission schedule, and what does the network look like at half the current emission rate?
• Is the token required to transact, or is it optional and bypassable with stablecoins?
• What is utilization on premium hardware tiers, not aggregate registered capacity?
• What has shipped versus what is on the roadmap, and what is the historical gap between announced and delivered timelines?
• Who are three named external customers, and what do they pay for?

Projects that answer all seven crisply are rare. That rarity is itself the most useful screening signal available.

Forecasts in this sector have an abysmal track record, so what follows is structured as directional theses with explicit reasoning and explicit ways to be wrong, rather than as predictions.

2026-2027: the consolidation and proof phase

Thesis: the compute marketplace layer consolidates from a dozen credible networks to three or four, decided by orchestration quality and enterprise integration rather than by capacity. The differentiator will be whether a customer can point existing tooling at the network and have workloads run without rewriting anything. Networks that require bespoke integration will lose to those that present a familiar interface.

Thesis: TEE-based verification becomes the production standard for decentralized inference, with ZKML confined to small, high-value, high-stakes models. The overhead economics make this close to inevitable. Watch for confidential computing to become a default advertised feature rather than a premium tier.

Thesis: decentralized reinforcement learning post-training reaches genuine competitive parity before decentralized pretraining does. RL rollout generation is embarrassingly parallel and latency-tolerant, which fits the decentralized topology far better than gradient synchronization does. If the sector produces a legitimate capability milestone in this window, it most likely comes from here.

How this is wrong: if enterprise buyers conclude that any non-SLA-backed infrastructure is unusable for production, the consolidation happens around whoever can offer conventional contractual guarantees, which favors the most centralized participants in the decentralized sector and undermines the thesis from within.

2027-2028: the agent economy test

Thesis: machine-to-machine payments become the sector's clearest commercial success, and they benefit general-purpose chains and stablecoin infrastructure more than AI-specific tokens. The technical pieces - inline HTTP payment protocols, on-chain agent identity and reputation registries, programmable spending limits - are already specified and shipping. What is missing is agents valuable enough that the payments are worth making. That is a capability question, and capability is improving fast.

Thesis: data becomes the contested resource rather than compute. As open-weight models commoditize architecture and as public web data exhausts, differentiated proprietary data becomes the durable moat. This is structurally favorable to networks that can attract data contribution under privacy guarantees, which is the intersection of confidential computing and token incentives. Expect the most interesting mechanism design work of this period to be about data, not compute.

How this is wrong: synthetic data generation may prove good enough that the proprietary data moat matters far less than expected, collapsing the thesis.

2028-2030: the structural question

By the end of the decade one of two things will be evident.

Scenario A - decentralization proves viable at scale. Communication-efficient training continues improving, verification becomes cheap and routine, and a decentralized network trains or fine-tunes a model that is competitive with contemporaneous closed models in at least a significant domain. The sector reprices dramatically upward and the composition of AI infrastructure spending shifts. In this world, the networks that survived the 2026-2027 consolidation and built durable supply relationships capture enormous value.

Scenario B - decentralization settles into a permanent niche. Frontier capability remains a function of concentrated capital and colocated compute. Decentralized networks persist and are genuinely useful, serving price-sensitive batch workloads, censorship-resistant applications, privacy-constrained sectors, and the long tail of specialized models. This is a real and durable business, perhaps a multi-billion-dollar one, but it is infrastructure rather than intelligence, and it should be valued as such.

Weighing these honestly: Scenario B is more probable on current evidence. The capability gap is wide, the trend in frontier training is toward more concentration rather than less, and no decentralized network has yet demonstrated a result that forces a reassessment. But Scenario A is far from excluded, and the algorithmic progress on communication efficiency between 2024 and 2026 was faster than most informed observers expected. The rational posture is exposure sized for Scenario B economics with option value on Scenario A.

Specific things to watch as leading indicators

• The largest parameter count trained fully decentralized, and how fast that number moves. If it is still in the tens of billions in 2028, Scenario B is confirmed.
• The ratio of external revenue to emissions across the sector. A rising ratio is the single best evidence of genuine product-market fit.
• Whether any significant enterprise publicly adopts decentralized inference for a production workload under its own name. Named references change procurement conversations.
• Verification overhead as a percentage of workload cost. If it approaches negligible, decentralized inference becomes strictly dominant on price for a much wider set of workloads.
• Whether agent payment volume on public rails grows into the billions annually. This is measurable on-chain and is the cleanest available proxy for the agent economy thesis.
• The direction of AI regulation on model provenance and data attribution. Stricter requirements favor cryptographic architectures.

This final section is a working method rather than commentary. It is the process a rigorous analyst should run before forming a view on any project in this sector.

Step 1: Classify the project into its actual layer

Before anything else, determine which of the five stack layers the project genuinely occupies: physical resources, orchestration, verification, incentives, or applications. Marketing materials will claim several. Look at the engineering: what does the codebase actually do, what does the team publish about, where does the revenue come from. A project that sells GPU-hours is an infrastructure company and should be valued against infrastructure comparables no matter how much AGI appears in its documentation. Misclassification is the origin of most bad valuation work in this sector.

Step 2: Separate emission-driven activity from organic demand

Construct the counterfactual: if token emissions went to zero tomorrow, which activity would continue? Work through each participant category. Node operators contributing hardware for emissions would leave. Customers paying market rates for a service they need would stay. Validators staking for yield would unstake. Developers building products with real users would continue.

The residual - what survives the counterfactual - is the network's actual business. Everything else is growth subsidy, which is legitimate and often correct strategy, but must be recognized as spending rather than revenue. Networks rarely publish this decomposition. You often have to construct it from on-chain data, comparing payment flows against emission distributions by address.

Step 3: Interrogate the verification claim

Ask the four questions from the technology section: which regime, what overhead, what exactly is attested, and what is the economic consequence of a detected failure. Then ask the fifth question that separates rigorous projects from marketing: has anyone ever been caught and slashed? A verification system that has never triggered in production is either perfect or inert, and inert is far more common. Look for published incident data.

Step 4: Trace token value accrual mechanically

Draw the flow of a single customer dollar through the system. Customer pays in what asset? If stablecoins, does anything convert to the native token, and is that conversion mandatory or a convenience? Does any portion get burned, and is the burn proportional to usage? Are node operators paid in the native token, and do they immediately sell it, creating structural sell pressure that scales with network growth?

This exercise produces uncomfortable conclusions surprisingly often. Several well-regarded networks have designs where usage growth creates net token supply pressure rather than demand. That does not make them bad businesses; it makes their tokens a poor expression of the business.

Step 5: Assess the team against delivery history

Compare announced timelines to actual delivery over the project's full history, not just recent quarters. Specifically look for: features announced and quietly dropped, mainnet dates that slipped repeatedly, and whether technical claims in blog posts were ever independently reproduced. In a sector where claims outrun verification, a documented history of shipping exactly what was promised is one of the strongest available signals, and it is rare enough to be genuinely differentiating.

Step 6: Size the position to the scenario, not the story

Given the analysis, determine which future scenario the project needs in order to justify its current valuation. If it requires Scenario A - decentralization proving viable at frontier scale - then the position should be sized as a low-probability option, regardless of how compelling the narrative is. If it works in Scenario B, it can carry core weighting. Most projects in this sector are priced for Scenario A while their fundamentals describe Scenario B, and recognizing that mismatch is where the analytical edge lies.

Step 7: Define what would change your mind

Write down, before taking any position, the specific observations that would invalidate the thesis: a revenue trajectory that flattens for two consecutive quarters, a verification failure incident, a key team departure, an emission schedule change, a competitor shipping the same capability at lower cost. Sectors driven by narrative are exactly the sectors where pre-committed invalidation criteria matter most, because the narrative will supply a reason to hold through any individual disappointment.

Applying the framework: a worked summary

Running the seven steps across the major networks produces a rough taxonomy. Akash and Render classify cleanly as infrastructure with real external revenue, transparent mechanics and Scenario B viability. Bittensor classifies as an incentive layer with the sector's most ambitious mechanism, meaningful but concentrated external revenue, and a valuation that requires substantial Scenario A belief. NEAR classifies as a platform chain where AI is an incremental thesis on top of general chain fundamentals, which makes it lower beta to the AI narrative in both directions. FET classifies as a diversified basket spanning three layers, which reduces single-point failure risk while diluting the clarity of any one thesis. Prime Intellect and Nous Research classify as research programs whose results should be tracked as leading indicators for the entire sector regardless of whether there is any way to take direct exposure.

None of these classifications is a recommendation. They are the output of a method, and the method is the durable thing. Projects change, mechanisms get redesigned, and the sector's composition in 2028 will look different from today. The seven steps will still apply.