AlphaRank
/
AlphaResearch
Crypto x AI Sector Overview AI Infrastructure
Sector Intelligence 003 · July 2026

Crypto x AI: The Sector Explained

What is being built across compute, data, models, agents, privacy, and applications, plus the evidence that shows what is actually working.

Coverage
Crypto x AI
Evidence cutoff
July 28 · 03:23 UTC
Report snapshot Charts use data available on July 28. Live prices and project metrics can change after publication.

Crypto x AI is six markets sharing one label.

  • Crypto x AI is not one market. The sector includes compute, data and intellectual property, models, agents, privacy, and applications. Each area sells a different product to a different customer.
  • The market category is useful, but messy. CoinGecko counted $22.11B of AI assets at the cutoff, about 0.98% of the full crypto market. Chainlink made up 28.1% even though it is not mainly an AI project.
  • Compute has the clearest proof of use. Providers can show available hardware, completed work, customers, and payments. The problem is that each provider reports different numbers, so clean rankings are not yet possible.
  • Agents are the fastest-moving product category. Platforms are giving software agents wallets, identities, tools, jobs, and payment rails. Creating an agent token is easy. Proving repeat agent revenue is much harder.
  • Privacy and verification solve different problems. Trusted hardware, shared secrets, encrypted computing, rerunning the work, and zero-knowledge proofs are different methods. Each protects a different part of an AI workflow.
  • Product growth does not automatically help the token. The token needs a clear job, and that job must create meaningful, repeat demand. A useful product can still have a weak token model.
  • The best proof is repeat paid work. Available capacity is only a starting point. Strong evidence means completed jobs, returning buyers, healthy economics, and a clear path from product use to token demand.

Crypto adds markets and payments around AI. It does not replace the model.

Most AI work still happens offchain. Crypto becomes useful when people or software need to buy resources from strangers, prove what happened, or share ownership.

A normal AI company can own the data center, the model, the customer account, and the payment system. A crypto AI network separates those jobs. One person can supply a GPU, another can supply data, a third can build a model, and a software agent can buy the result. The protocol connects them.

The simple test

Does the blockchain solve a real coordination problem? If not, the token may be attached to a normal AI product without adding much value.

AlphaResearch cover titled Crypto x AI: The Sector Explained with six layers: compute, data and IP, intelligence, agents, trust, and applications.
Fig. 01 · Crypto x AI: The Sector Explained. The report separates the sector into six economic jobs. One project can span several layers, but each layer needs its own evidence.
Six-column diagram showing how compute, data and IP, intelligence, agents, trust, and applications connect in the crypto AI stack. Project examples appear with their logos.
Fig. 02 · Six different markets. The taxonomy identifies what is being bought and sold. It does not imply that every project stays inside one column.
$22.11B
CoinGecko AI category · fixed snapshot
0.98%
AI category share of all crypto
62
Core projects in the report map
6
Distinct economic layers

The sector is broad, but most liquid value sits in a few assets.

The sector includes GPU markets, data networks, IP licensing, model markets, agent tools, payment systems, privacy products, and consumer apps. These are different businesses. They should not be compared as if they sell the same thing.

The map places 62 projects in one main category so the structure is easy to scan. The categories are not perfect. ASI spans agents, data, models, and compute. NEAR combines a general-purpose blockchain with confidential AI services. Bittensor combines model markets, incentives, and subnet investing.

An incomplete map of 62 crypto AI projects, shown with their logos and grouped into compute and training, data IP and context, models and intelligence, agents and commerce, privacy and verification, and applications and distribution.
Fig. 03 · An incomplete map of the crypto AI ecosystem. Projects are placed by primary product using first-party documentation. Inclusion is not endorsement, and state labels are simplified to keep the map readable.

The token market does not match the project map. CoinGecko's AI category totaled $22.11B at the cutoff, but it included broad infrastructure such as Chainlink and NEAR. Its smaller AI categories also overlap, so adding them together would count some assets more than once.

Bar chart showing a 22.11 billion dollar CoinGecko AI category, 3.14 billion dollar AI Agents category, 2.03 billion dollar AI Applications category, and 1.18 billion dollar AI Agent Launchpad category.
Fig. 04 · The size of the liquid category. CoinGecko category and global market data are fixed at July 28, 03:14 UTC. The smaller categories overlap and are not additive.

That makes classification important. Chainlink represented 28.1% of the category. NEAR and TAO were the next two assets in the snapshot. Together, the top three made up almost half of the $22.11B total.

Ranked bars with project logos for selected liquid crypto AI assets led by LINK, NEAR, TAO, RENDER, VVV, VIRTUAL, FET, KAITO, GRASS, and AKT.
Fig. 05 · Selected liquid assets. Market caps are observed CoinGecko values at 03:17 UTC. LINK is shown in gray because its inclusion demonstrates category ambiguity rather than a pure AI exposure.
How to use the map

Start with the product. Compare GPU marketplaces with other compute providers, not with agent launchpads simply because both tokens use the AI label.

Compute shows the clearest demand, but providers are hard to compare.

Compute is the easiest part of the sector to understand. A provider supplies a GPU or server. A buyer sends a job. The network runs the job and handles payment. Capacity, completed work, use, and spending can all be measured.

Each network reports a different measure. Akash reported 3.1M deployments in 2025, more than 1,000 GPUs, 1.7B daily AkashML tokens, and $5M of all-time compute spend by Q1 2026. Aethir reported more than 435K containers, 150 clients, and over $147M of annual recurring revenue. Prime Intellect reported more than 10K training runs, 6K customers, and $100M of annualized revenue.

These figures do not belong in one ranking. A container is not a GPU. A deployment is not always a paid job. Annualized revenue is not the same as audited annual revenue.

Disclosure matrix with project logos comparing capacity, jobs or runs, customers or usage, and money evidence across Akash, Aethir, Prime Intellect, Render, io.net, and Gensyn.
Fig. 06 · What compute providers disclose. Green marks confirmed product surfaces, orange marks first-party operating claims, yellow marks published states, and gray marks unknown or non-standardized data. This is a disclosure matrix, not a performance ranking.
Strongest evidence Paid spend, repeat customers, utilization, and completed jobs tied to a stable time window.
Weakest evidence Large capacity totals without a comparable series for paid use.
Training challenge Large synchronized training jobs need fast interconnects, not just scattered GPUs.
Best near-term fit Inference, batch jobs, rendering, fine-tuning, and workloads that can be divided cleanly.

Gensyn and Prime Intellect go beyond simple GPU rental. They try to coordinate training across many machines and verify the work. Gensyn makes machine-learning steps reproducible so disputed results can be checked. Prime Intellect combines a compute marketplace with an open research lab. The opportunity is larger, but the engineering challenge is also harder.

Data networks are trying to make AI inputs easier to own, license, and sell.

This part of the sector solves three different problems. Grass collects public web data through a distributed residential network. Vana lets people pool private data and allow approved buyers to query it. Story registers intellectual property and attaches rules for licenses, royalties, and disputes.

Grass has the clearest revenue claim in this group. It reported $17M of revenue in 2025 and another $17M in the first half of 2026. It also said nearly every AI customer returned. The figures are first-party and customer names are private, but they are closer to real business evidence than points or registered files.

Vana and Story focus on ownership and permission. Vana can limit who may query a dataset. Story turns intellectual property into an onchain asset with programmable terms. These systems can make rights easier to manage. They do not prove that buyers will pay enough to support every dataset or registered asset.

01 Acquire public data.

Grass turns distributed bandwidth into web-data collection for AI training and live context.

02 Query private data.

Vana lets approved buyers compute against pooled data without receiving the raw files.

03 License intellectual property.

Story attaches machine-readable terms, royalties, and dispute processes to registered IP.

Other projects cover nearby jobs. Sahara combines data labeling, model access, and agent creation. Kaito turns crypto information and social activity into market intelligence. Ocean provides data-market tools to the ASI stack. The shared goal is to reward people who contribute useful data. The shared risk is that tokenized supply grows faster than buyer demand.

Some networks sell access to models. Others create markets for answers.

Bittensor and Allora are useful examples because they do different jobs. Bittensor runs subnet markets where producers create machine intelligence and validators score it. Each subnet has its own token and market price. Allora focuses on forecasts. One group makes predictions, another estimates how accurate they will be, and the network combines both.

Gensyn treats training and evaluation as one open loop. Distributed machines train models, applications provide feedback, and evaluation markets reward better results. NEAR AI takes another route. It offers private model use and agent services through trusted hardware.

Other projects span open models, distributed training, AI-focused blockchains, and verification. The category is broad because a project can compete as a model lab, a protocol, a chain, or a verification service.

The evaluation problem

A network can reward activity without rewarding useful answers. A strong system must connect its score to a task a buyer values, then resist collusion and gaming.

Bittensor Markets and validator scoring allocate emissions across specialized subnets.
Allora Forecasts of model performance help weight and combine inferences.
Gensyn Training, verification, evaluation, and payments form one open learning loop.
NEAR AI Confidential infrastructure packages private inference and agent services.

Crypto gives AI agents wallets, identities, jobs, and payment rails.

An AI agent is software that can work toward a goal and use tools with limited human input. Crypto can give that agent a wallet, an identity, and a way to pay. This matters when an agent needs to buy an API call, hire another agent, control money, or prove which software produced a result.

Virtuals combines agent creation, tokenization, identity, jobs, and commerce. Its public site reported 45,558 unique agents, 1.48M jobs, and $2.27M of total revenue at the cutoff. These live, first-party counters are more useful than token counts alone, but their definitions have not been independently audited here.

Olas focuses on the tools needed to build and run agents. Its marketplace connects buyers with agent services. MyShell focuses on consumer creation and distribution. ElizaOS offers an open framework with plugins. Fetch provides agent infrastructure inside the wider ASI alliance.

Coinbase x402 solves a smaller but important problem: machine payments. A service asks for payment, an agent signs a stablecoin transaction, and the service returns the result. The agent can buy one API call without opening an account or waiting for a human checkout.

Six-step flow for agent identity, discovery, agreement, execution, verification, and stablecoin settlement, with logos for named project examples.
Fig. 07 · The agent commerce stack. Identity, discovery, agreement, execution, verification, and settlement are separate functions. A platform may provide one function or combine several.

Applications are where the technology becomes a product. Venice sells private generative AI and API access. Kaito sells AI-powered crypto information. Numerai uses outside models inside a hedge fund. Giza builds autonomous financial agents. These products can use crypto in the background without making the token part of the user experience.

The launchpad trap

A platform can create thousands of agent tokens before it creates one durable business. Track completed jobs, repeat buyers, paid retention, and payment volume. Token creation is supply, not demand.

Open AI systems need ways to protect data and verify results.

AI often runs outside the blockchain. That creates a trust gap. A smart contract cannot automatically know whether a remote GPU used the right model, protected the input, or returned an honest answer.

Trusted execution environments, or TEEs, isolate data and code inside protected hardware. Phala and NEAR AI use this method because it can run modern models at close to normal speed. The tradeoff is that users must still trust the hardware and its security checks.

Other methods make different tradeoffs. Multi-party computation splits a secret across several operators. Fully homomorphic encryption keeps data encrypted while it is being processed. Zero-knowledge proofs check a claim about a computation. Gensyn makes machine-learning steps reproducible so disputed work can be rerun.

Comparison table for trusted execution environments, multi-party computation, fully homomorphic encryption, deterministic replay, and zero-knowledge proofs, with logos for named examples.
Fig. 08 · Five different trust models. Each method protects a different object and accepts a different assumption. No method is universally superior.

The diligence questions are simple: What stays private? What is proven? Who could collude? What happens when a check fails? A project that says "verifiable AI" without answering those questions has not explained its security model.

A useful product does not guarantee a valuable token.

Tokens do at least five jobs in this sector. They can pay suppliers, secure work through staking, unlock a resource, govern rules, or reduce supply through burns and buybacks. The token's job matters because it determines where repeat demand could come from.

AKT connects buyers and providers inside Akash. TAO helps secure subnet competition and receives emissions. Vana data tokens can unlock dataset access. VVV can be staked for DIEM capacity, while product proceeds can support token purchases and burns. OLAS coordinates code, operators, and staking across agent economies.

None of these mechanisms creates value by itself. Buyers may be able to pay in stablecoins. Staking may be driven by emissions instead of paid work. Governance may control no cash flow. A burn may be too small to offset new supply. The link between product use and token demand must be clear and large enough to matter.

Five token value routes: pay supply, secure work, access, govern, and remove supply, each with an underwriting question and logos for named examples.
Fig. 09 · Five routes from product to token. The framework separates token utility from token value capture. Every asset requires its own mechanism and magnitude analysis.
Payment Does a buyer need the token, or can demand bypass it?
Security Is stake tied to useful paid work and credible slashing?
Access Does usage create recurring token demand rather than one-time lockup?
Supply Do burns or buybacks exceed ongoing issuance and incentives?

The sector needs to move from available capacity to repeat revenue.

Token prices can move before the business evidence improves. The goal is to identify what has already been proven and what proof should come next.

Five-level evidence ladder progressing from supply exists to work happens, buyers return, economics hold, and token captures value.
Fig. 10 · The evidence ladder. Registered resources are the starting point. Durable value requires work, retained buyers, defensible economics, and token-level capture.
Indicator What improves the evidence Review cadence
Paid work Completed jobs and spend by workload, buyer cohort, and network. Monthly
Retention Returning buyers and repeat spend after incentives are removed. Monthly or quarterly
Utilization Used GPU hours, consumed credits, or paid queries divided by available capacity. Monthly
Unit economics Revenue, supplier payout, gross margin, and coordination cost in the same period. Quarterly
Token capture Fees, required purchases, staking demand, buybacks, and burns compared with issuance. Monthly
Trust failures Disputes, slashing, attestation failures, proof cost, and time to resolution. Event-driven

The biggest gap is consistent reporting. Compute providers do not use the same units. Agent platforms mix users, agents, jobs, and tokens. Data networks mix registered files, contributors, and paid queries. Better disclosure would make it easier to judge product progress without using token price as a shortcut.

This map is broad. The evidence standards are strict.

This report uses an evidence cutoff of July 28, 2026 at 03:23:31 UTC. Market figures use CoinGecko snapshots observed between 03:14 and 03:17 UTC. The live asset cards may update after the cutoff. The article charts do not.

Methodology and limitations

The ecosystem map is a curated list, not a complete census. Each project is placed in one main category even when it spans several. Product status comes from first-party documentation and should be checked again before publication or investment.

CoinGecko decides which assets belong in each category. Its AI Agents, AI Applications, and AI Agent Launchpad categories overlap, so this report shows them separately and does not add them together. The 0.98% market share divides the $22.11B AI category by CoinGecko's $2.251T total crypto market cap.

Operating figures from Akash, Aethir, Prime Intellect, Grass, and Virtuals are first-party statements. This report did not audit them independently. Compute providers also use different units and time periods, so the report does not create a single performance ranking.

Descriptions of how protocols work come from official documentation. A documented design does not prove customer demand, decentralization, profit, or token value. When data is missing or cannot be compared, the report labels the gap instead of estimating it.

This report is for information and research. It is not investment advice or an offer to buy or sell an asset. Crypto and AI projects can change quickly, and tokens can lose substantial value. Check protocol status, market data, legal rights, and operating claims independently.

ALPHARESEARCH · SECTOR INTELLIGENCE 003 · JULY 2026
This material is for informational and research purposes only. It is not investment, legal, tax, or accounting advice; it is not an offer or recommendation to buy or sell any asset. Digital assets are volatile and may lose substantial value. Verify protocol state and market data independently.