Crypto x AI: what 52 founder interviews say changed between 2025 and 2026

Crypto Coin Show recorded 161 founder and executive interviews between January 2025 and September 2026. Sixty-one of them carry the archive’s AI tag. This report reads all of those episodes as one body of evidence and asks a narrow question: when crypto founders talked about artificial intelligence on the show, what were they actually building, and how did that change from one year to the next?

The short answer is that the conversation moved from ownership to proof. In 2025 the typical AI guest was pitching an agent that trades for you or a decentralized network that would give users a stake in AI. By 2026 the typical AI guest was pitching a way to verify something: that a piece of data is real, that an agent did what it says it did, or that a user is a human and not a bot.

This expanded edition, updated October 6, 2026, keeps the original dataset and findings, adds a market layer (agent payment volumes, the AI token category, decentralized compute revenue) so the founders’ claims can be checked against public numbers, and profiles the 2026 companies behind the shift, from XYO Network and NEAR Protocol to 0G Labs, io.net, Verona and World Foundation. One episode recorded after the September 28 cutoff, Illia Polosukhin’s October 1 return, is referenced in the company profiles but not counted in the percentages.

Who this is for: Investors, founders and analysts who want to know what crypto AI companies are actually shipping, which 2025 narratives died, and which 2026 claims the market data supports, with every figure dated and every episode linked so the reasoning can be checked.

The dataset

The AI tag is assigned automatically, so the desk screened every tagged episode by hand before counting anything. Nine of the 61 were dropped because AI came up only in passing, for example a legal episode on offshore jurisdictions with one chapter on agent liability, or a Bitcoin price outlook that touched on Tesla. That leaves 52 episodes: 25 from 2025 and 27 from 2026. In 34 of them AI is the primary subject; in the other 18 it is a significant part of the product being discussed. The desk also checked the 100 untagged episodes from the same period for AI references and found none where AI was more than a secondary mention.

Measured against all episodes, AI-focused conversations made up 28 percent of the show in 2025 and 37 percent in 2026 to date. Those figures are lower than the raw tag counts published in The CCS Interview Index, which counted every tagged episode; the gap is the nine episodes removed here.

Crypto x AI by the numbers

Six findings

1. Verification became the dominant AI theme

Episodes about verifying data, actions or identity went from 12 percent of AI conversations in 2025 to 37 percent in 2026, the largest move of any theme. The pitches differ in what they prove. Verona (formerly XION) described itself as a verification layer for AI agent commerce, citing 69 million verified interactions. World Foundation framed World ID as the identity layer that links an AI agent to a verified human, including through AgentKit with Coinbase. Nexus Labs argued that AI agents need to prove execution without a human in the loop. Geo Protocol founder Yaniv Tal, a co-founder of The Graph, put the problem as a question of ground truth: AI answers are only as good as sources nobody can currently rank.

XYO Network, a sponsor of the channel, accounts for five of the ten verification episodes in 2026, so the desk checked whether the shift survives without it. It does: excluding XYO, verification rose from 4 percent of AI episodes in 2025 to 23 percent in 2026. XYO’s own arc is still worth tracking, because it moved over the period from launching a data-focused Layer One to data provenance for AI agents to what Arie Trouw called proof of action: evidence that a robot or agent actually did what it claims.

The market moved the same way. Outside the show, the verification theme now has standards and money behind it. ERC-8004 “Trustless Agents”, which gives agents a portable on-chain identity and reputation registry, was created in August 2025 and went to Ethereum mainnet in January 2026, according to The Block. The EU AI Act’s Article 50 transparency rules for AI-generated content began applying on August 2, 2026. World Foundation raised $52.5 million led by Pantera Capital on July 24, 2026 with every token in the round locked for twelve months, and reported more than 18 million Orb-verified humans. The earliest 2026 episode in the set, recorded in January 2026 and the first verification episode of the year, shows the theme already forming: Scott Dykstra of Space and Time described Proof of SQL, a zero-knowledge proof that a database answer is correct, as the missing piece that lets smart contracts and AI agents trust off-chain data.

2. The “decentralized AI beats Big Tech” pitch faded

In 2025, 28 percent of AI episodes were about open or user-owned AI: Dr. Ben Goertzel on decentralized AGI, Morpheus on a peer-to-peer AI network, Sahara AI on fair monetization for data contributors, and Cirus and Bitrobot on users earning from the data they generate. In 2026 that share fell to 4 percent. The 2026 guests mostly positioned crypto as a complement to mainstream AI rather than a rival. Brave described its search API powering answers inside other companies’ AI tools. The Graph pitched itself as a data backbone for AI agents. io.net sold decentralized compute on price, at up to 70 percent below the big clouds, rather than on ideology.

The one 2026 guest who still made the full open-AI case, Michael Heinrich of 0G Labs, made it differently: not as a moral argument but as a product claim, that a verifiable, open stack can be private and transparent at once, and that more than 100 applications were already testing on the network. Even Sahara AI, whose 2025 episode was the clearest “own your data” pitch in the archive, returned in February 2026 with Sorin, a retail trading co-pilot that connects to Binance, OKX and Coinbase rather than replacing them. The consumer “earn from your data” model survives in exactly one 2026 episode, Sleepagotchi.

The token market tells the same story from the other side. The 2024 to 2025 agent-token cycle has largely unwound: on October 6, 2026 CoinGecko’s AI Agents category was worth $4.14 billion in total, with Virtuals Protocol trading about 83 percent below its all-time high and the renamed ai16z token more than 99 percent below its peak. The broader Artificial Intelligence category was worth $26.8 billion the same day, led by NEAR at $6.91 billion and Bittensor at $3.47 billion, which is to say by chains and compute networks rather than by “open AI” data economies. Meanwhile Nvidia agreed in September 2026 to buy Hugging Face, the home of open-source models, for nearly $13 billion, a deal Arie Trouw raised on the show as evidence that open models are becoming an input to the incumbents rather than a threat to them.

3. Agent guests started talking about guardrails and numbers

AI agents were the biggest theme in 2025 at 36 percent, largely agentic finance: KUVI AI, Mantis and Fetch.ai described agents that execute intents and manage assets. The theme shrank to 22 percent in 2026, and the tone changed. Virgo said its Wallet V agent publishes its decision rationale every 15 minutes and constrains the model to set parameters to limit hallucinated trades. Sahara AI returned with Sorin, deliberately a co-pilot rather than an autopilot. Propr reported more than 300 AI agents already trading with funded capital on its platform. The claims got smaller and more measurable.

Two further 2026 agent episodes fit the pattern. Sina Yamani of Action Model described Large Action Models, AI that completes multi-step online tasks, with the emphasis on reliability of execution rather than on autonomy for its own sake. Ian Balina of Token Metrics launched an AI Crypto Morning Briefing, an agent that filters news into a daily brief, which is about as constrained a use of an agent as exists. Sean Ren’s phrasing on Sorin captured the mood of the year: everything is planned for the user, but “the final final action taken is still with the users.”

The market data supports the caution. TRM Labs reviewed $52.7 million of x402 settlements in 2026 and attributed only 0.6 to 7.5 percent of them to actual AI agents, the rest being humans and scripts. Analysis cited by Visa counted 109.6 million adjusted x402 transactions worth about $15 million between May 2025 and April 2026, roughly 14 cents each. Agents are transacting, but in small amounts and with a human close by, which is exactly what the 2026 guests were designing for.

4. Agents need their own money rails

A new theme appeared in 2026: how agents pay. Brave discussed x402 support for agent stablecoin payments as part of its Roadmap 4.0. NEAR Protocol co-founder Illia Polosukhin, co-author of the Transformer paper, described NEAR Intents as a way for agents to state an outcome and let the protocol handle routing and gas. io.net launched Agent Compute so agents can buy their own GPU time, and The Graph described agents paying per query with no API keys. None of these rails existed as a topic on the show in 2025.

By October 2026 the rails are measurable. Coinbase’s x402 protocol, launched in May 2025, reported 75.41 million transactions and $24.24 million of volume in the trailing 30 days on x402.org, from 94,060 buyers to about 22,000 sellers, which works out at roughly 32 cents a payment. The x402 Foundation launched under the Linux Foundation on April 2, 2026 with 22 founding members including Google, Stripe, Visa, Mastercard, AWS and Cloudflare. Google’s Agent Payments Protocol (September 2025) and the Stripe and Tempo Machine Payments Protocol (March 2026) sit alongside it. In September 2026 BlackRock published a paper arguing that agents making continuous low-value payments could be the next structural source of stablecoin demand. 0G’s Michael Heinrich put the scale question bluntly on the show in May: stablecoins moved tens of trillions of dollars last year and x402 was then processing a few hundred million, so the agent rail is real but still tiny.

5. Security and privacy moved up, with AI on both sides

Security, privacy and legal episodes rose from 4 to 11 percent of AI conversations. CertiK described an AI auditor for smart contracts in an environment where attackers use AI too. Fhenix pitched fully homomorphic encryption as a route to confidential onchain AI. The one sceptical voice in the set came from Venable LLP: crypto practice head Chris O’Brien warned founders against asking a chatbot for legal advice first, citing confidentiality, missing context and false confidence.

CertiK’s claim can be checked. Its AI Auditor went public on April 7, 2026 after more than six months of testing, with a reported 88.6 percent cumulative exact hit rate against 35 real-world Web3 incidents from 2026 that were excluded from training. Hudson Jameson was careful on the show about what that means: the tool is a first pass, and “manual auditing is still what you need” to secure a contract. Privacy showed up on the infrastructure side too: NEAR made privacy the default across its trading stack in 2026 and Acurast and 0G both run AI workloads inside trusted execution environments so the operator of the hardware cannot read the prompt. The attack side is why. Jameson cited wrench attacks and bridge exploits, and 0G’s Heinrich described vibe-coded apps burning billions of tokens in weeks, a pace at which unaudited code reaches production faster than humans can review it.

6. Infrastructure is now pitched on cost and usage, not tokens

Compute and data infrastructure held steady at 12 to 15 percent, but the pitch changed. Compute Labs in 2025 framed compute as an investable asset. In 2026, io.net described replacing inflation-driven token rewards with payments to suppliers in dollar value, Perceptron Network claimed AI training data at 92 percent below centralized providers, and The Graph acknowledged openly that record usage had not translated into token price performance. 0G Labs made the strongest case for AI needing its own chain.

The price claims hold up against public rate cards. As of October 2026 an Nvidia H100 rents for $1.50 an hour on Nosana and $2.40 on Hyperbolic, against $6.16 per GPU on CoreWeave’s on-demand node, about $6.88 on AWS and near $12.29 on Azure, per the providers’ pricing pages and Spheron’s tracker. Jack Collier’s “up to 70 percent below” is within that range. Revenue is real but concentrated: Aethir reported $127.8 million for 2025, io.net processed more than $20 million in compute leases in early 2026 and told the show it was at about $25 million of annualised revenue, and Akash crossed $5 million of all-time compute spend in Q1 2026. The token side has followed usage: Akash launched Burn-Mint Equilibrium on March 23, 2026 so every dollar of spend buys and burns AKT, and io.net committed 50 percent of excess emissions to burns. Where the market disagrees with the founders is on frontier training. Prime Intellect trained its 106 billion parameter INTELLECT-3 on a centralized 512-GPU H200 cluster in November 2025 because no marketplace of scattered GPUs could, which is why the 2026 guests on the show talk about inference and agents, not about training the next frontier model.

Shares of AI episodes by theme, 2026:

Key metrics: the market behind the conversations

The interview shares above measure what founders chose to talk about. The three series below measure what the market actually did over the same period, so the two can be compared. Each is the public figure the desk could verify as of October 6, 2026.

Agent payments: x402 volume

x402 is the HTTP-native payment standard Coinbase launched in May 2025, in which a server answers a request with a 402 status and a price and the client, human or agent, pays in stablecoins to proceed. It is the rail Brave, io.net and The Graph referenced on the show, and the one whose numbers are easiest to read.

Reading across the rows: the count is large, the dollars are small, and the share of payments made by software acting on its own is still single digits. That is consistent with what the 2026 guests described, agents buying compute by the minute and data by the query, and inconsistent with the 2025 vision of agents managing portfolios.

The AI token category

The agent category is worth about one sixth of the wider AI category, and the wider category is led by settlement and compute layers rather than by agent frameworks. The two 2025 “open AI” guests whose tokens are large enough to appear here, Fetch.ai and SingularityNET, now trade as one merged asset. That matches the show’s finding that the market rewarded infrastructure over ideology.

Decentralized compute revenue

For context, the former bitcoin miners that pivoted to AI hosting are an order of magnitude larger: IREN raised its 2026 annualised revenue target above $4 billion in July 2026, and Cipher Digital extended its Barber Lake AI lease to 20 years and more than $9 billion in September 2026. Decentralized compute is a real business for a few networks, but the big money in “crypto meets AI compute” went to companies with power contracts and buildings, not to marketplaces of scattered GPUs. That, too, is consistent with the shift on the show from ownership to utility.

The companies building it: 2026 on the show

Fifteen companies carried most of the 2026 AI conversation on Crypto Coin Show. The profiles below give what each builds, what its founder or executive told the show (linked to the episode), one or two facts the desk verified outside the interview, and what the company tells us about the report’s thesis. Where a guest’s claim and the public data disagree, the profile says so.

XYO Network: from data provenance to proof of action

What it builds. XYO runs a network of more than 10 million devices that witness and sign location and sensor data, and XYO Layer One, a chain whose XL1 token is partly burned with every verification record written to it.

What they told CCS. XYO appeared five times in 2026, more than any other company. Co-founder Markus Levin, in his August 2026 episode, described XYO as an accountability layer for agents that make financial and logistics decisions with no reliable record, cited 2,000 sign-ups for early access to the XYO AI SDK and summed up the physical use case as “proof of work for robots, basically.” In June he and Theta described agents on Theta EdgeCloud writing cryptographic receipts of their own execution to XYO Layer One. Co-founder and CEO Arie Trouw, in April, said XYO Layer One’s throughput had risen two to five times, and in September argued that proving data is real is no longer enough: the harder problem is proving an agent or robot actually did what it claims, and he raised the question of whether agents that hold wallets will need credit histories.

Verified. The XYO and Theta integration was announced May 28, 2026, with agents already running production workloads for sports teams including the Houston Rockets and Olympique de Marseille on Theta EdgeCloud. Nvidia’s agreement to buy Hugging Face for nearly $13 billion, which Trouw cited, was reported by Semafor on September 3, 2026.

Why it matters. XYO is the clearest single example of the ownership-to-proof shift, and because it is a sponsor the desk shows the verification finding both with and without it. The trend holds either way. XYO Network is a sponsor of Crypto Coin Show.

NEAR Protocol: a settlement layer for agents, now with an ETF

What it builds. NEAR is a sharded Layer 1 that has repositioned itself as infrastructure for AI agents: NEAR Intents for outcome-based cross-chain execution, Chain Signatures, IronClaw as a secure agent harness, an agent marketplace, and confidential execution for trading.

What he told CCS. Co-founder Illia Polosukhin, a co-author of the 2017 Transformer paper, appeared twice. In June 2026 he said most blockchains were built for humans and NEAR is building for agents: a user or agent describes the outcome it wants rather than the sequence of actions, NEAR Intents was already serving more than 30 percent of daily volume on the network, and a one-click post-quantum wallet upgrade was coming with the quantum threat two to five years out. In October 2026, recorded after this report’s cutoff, he reported privacy-by-default trading, Confidential Intents TVL of $183 million against an initial $70 million target, the Shield system that had blocked about 50 million transactions involving stolen funds, IronClaw private inference for agents, and the Bitwise NEAR ETF.

Verified. The Bitwise NEAR ETF listed on NYSE Arca under the ticker NRR on September 29, 2026 with a 0.75 percent fee, Coinbase Custody holding the tokens and an intention to stake 100 percent of the fund’s NEAR, per Bankless and the fund’s SEC filings. NEAR was the largest token in CoinGecko’s Artificial Intelligence category on October 6, 2026 at $6.91 billion.

Why it matters. NEAR is the strongest counter-example to the “AI needs its own chain” argument: it is adding agent, privacy and payments features to an existing network, and it is the only AI-positioned chain with a US spot ETF. It also shows the payments theme and the privacy theme converging in one product.

0G Labs: the case for an AI-native chain

What it builds. 0G describes itself as a decentralized AI operating system: a chain, a storage layer and a compute network with inference inside trusted execution environments, plus the Apollo launchpad for AI projects.

What he told CCS. Co-founder Michael Heinrich, in May 2026, argued that Ethereum and Solana were not designed for AI data and workloads, that open models lag proprietary ones by about three months, that more than 100 applications and companies were testing on 0G, and that the goal is to “make AI a public good” that is transparent yet private. He also put the agent payments rail in proportion, noting that x402 was processing a few hundred million dollars against tens of trillions in stablecoin transfers.

Verified. 0G moved GLM-5.3 to its TeeML sealed-inference mode on September 3, 2026 at $1.40 per million input tokens and $4.40 per million output tokens, and on September 28, 2026 launched Infinite AI, under which staked 0G mints a derivative, iAI, that earns 1.271 compute credits a day, per 0G’s blog.

Why it matters. 0G is the purest “AI needs its own chain” position in the 2026 set and so the main foil to NEAR, The Graph and Brave. Note that even 0G now sells on price per token and private inference rather than on ideology, which is the report’s sixth finding in one company. Ashton Addison holds a token warrant in 0G Labs.

io.net: compute on price, and agents that buy it

What it builds. io.net aggregates idle GPUs from data centres, miners and individuals into a marketplace, with Agent Cloud (launched March 25, 2026) letting AI agents rent GPU clusters through the Model Context Protocol and pay in USDC or fiat without OAuth or KYC.

What he told CCS. Chief marketing officer Jack Collier, in April 2026, said Big Tech was spending $650 billion a year on AI infrastructure while half of US data centres planned for 2026 were delayed, that io.net offered H200s for under $2 an hour against $25 to $30 on AWS, that supply spanned 138 countries, that the network was at about $25 million of annualised revenue, and that the Incentive Dynamic Engine would pay suppliers in dollar terms and burn 50 percent of excess emissions. “AI agents buying their own compute is a first,” he said.

Verified. BlockEden reported on May 3, 2026 that io.net processed more than $20 million in compute leases in early 2026 and listed about 30,000 GPUs across 130-plus countries. io.net’s own site listed 30,000-plus GPUs in October 2026. The AWS comparison is the one claim the desk could not reproduce: AWS’s on-demand H100 rate is about $6.88 per GPU-hour and its H200-class Capacity Blocks are priced well below $25; the $25 to $30 figure may refer to a full multi-GPU instance rather than a single card.

Why it matters. io.net is the template for the 2026 infrastructure pitch: price, utilisation and revenue, with the token redesigned around real spend. It is also one of only two guests, with The Graph, whose product lets an agent pay for a resource with no human account.

Perceptron Network: training data at a discount

What it builds. A decentralized network where users run nodes or annotate data and AI companies buy the resulting datasets, positioned against Scale AI and Appen.

What he told CCS. Co-founder Peter Anthony, in July 2026, cited 800,000 downloaded nodes and 300,000 daily active users, claimed data at 92 percent below centralized providers, described a $10 million AI data fund and said a token generation event was eight to ten weeks away.

Verified. Perceptron raised $6.5 million in a strategic round from Sigma Capital, Selini Capital and other trading firms, announced July 30, 2026, and reported 807,000 active nodes and 300,000 daily active users at that time. The PERC token was still in pre-launch as of that announcement; the desk found no confirmed exchange listing by October 6, 2026, so the eight-to-ten-week timeline given on the show appears to have slipped.

Why it matters. Perceptron is a 2025-style data economy that survived by changing the buyer: the customer is an AI lab paying for labelled data, not a user earning pennies for browsing. It sits in the infrastructure theme rather than the open-AI theme for that reason.

Verona (formerly XION): verification for agent commerce

What it builds. Verona runs a Cosmos-based chain that once sold walletless, gasless consumer crypto and now sells verification: proof that an interaction came from a real person doing a real thing, exposed to brands through EarnOS and to consumers through the Ero app.

What he told CCS. Founder and CEO Anthony Anzalone, in June 2026, said half of web traffic is synthetic, that Verona had processed 69 million verified interactions across 3 million users for more than 115 brands including Uber, Amazon, Nike and BMW, that 12 projects were building on the network, and that an MCP server would let AI agents consume verified data. “AI agents don’t really come with social capital,” he said.

Verified. The rebrand from XION to Verona and the 1:1 token swap took effect in mid-June 2026 alongside about $18.5 million of EarnOS funding and a $30 million brand-spend commitment, per Bitrue’s June 23, 2026 note. The Block separately reported EarnOS raising $6 million from 1kx, Circle and Coinbase, so the $18.5 million figure likely includes earlier rounds.

Why it matters. Verona is the sharpest example of a company moving its entire narrative from the 2025 side of the ledger (consumer onboarding) to the 2026 side (verification), and of the market rewarding the move with a rebrand-week rally. It is also the privacy-preserving answer to the bot problem that World answers with biometrics.

World Foundation: proof of humanity as the identity layer for agents

What it builds. World ID, a proof-of-humanity credential issued after an Orb iris scan and presented through zero-knowledge proofs, now extended to AI agents through AgentKit with Coinbase so an agent can prove it acts for a verified human.

What he told CCS. Research engineer DC Builder, in June 2026, said the internet “doesn’t really have a native identity layer,” that World ID had about 18 million Orb-verified users across roughly 20 countries, that smartphone face recognition starts producing collisions somewhere between 15 and 30 million users (the argument for the Orb), and that agent-driven ticketing and commerce would be visible within twelve months.

Verified. World Foundation raised $52.5 million led by Pantera Capital on July 24, 2026 with Bain Capital Crypto, Susquehanna Crypto and others, with every WLD token in the round locked for twelve months. It reported more than 18 million Orb-verified humans and more than 39 million network members, and 2026 integrations with Zoom, Docusign, Okta, Vercel and Tinder. WLD traded about 97 percent below its March 2024 peak at the time.

Why it matters. World is verification at the largest scale in the set, and the token price shows the gap between adoption and token performance that The Graph also admitted to. Identity for agents is the thread that ties World, Verona and XYO together.

Space and Time: proving the database answer

What it builds. A verifiable data layer whose Proof of SQL generates a zero-knowledge proof that a query result over indexed on-chain and off-chain data is correct, so a smart contract or agent can act on it; Dreamspace is its no-code app builder.

What he told CCS. Co-founder Scott Dykstra, in January 2026, called Proof of SQL “anti-hacking for queries on chain,” described Space and Time as an extension of Chainlink, cited 30,000-plus apps built on Dreamspace, 100 terabytes of indexed chain data, more than $50 million of Microsoft investment across rounds and 124,000 Indonesian student credentials put on chain.

Verified. Space and Time’s permissionless mainnet launched on May 8, 2025, backed by Microsoft and integrated with Chainlink, per Cointelegraph and the company’s launch release.

Why it matters. This was the first verification episode of 2026 and it frames the whole year: the problem is not storing data but proving an answer about it, which is what agents need before they act.

Nexus Labs: a zkVM chain for agentic finance

What it builds. A Layer 1 built around a zero-knowledge virtual machine so every state transition is provable, with an exchange and a native stablecoin, USDX, enshrined in the protocol rather than deployed as contracts.

What he told CCS. CEO Daniel Marin, in May 2026 (his second appearance of the year after February), said Testnet III had drawn more than 3 million users and 5 million nodes, that 96-plus partners were integrated for mainnet, that the zkVM had unlocked a thousand-fold throughput improvement, and that the long-run goal is infrastructure where “AI writes, audits, and self-improves the blockchain.”

Verified. The NEX token launched May 19, 2026 with a 100 trillion total supply (20 percent team, 20 percent investors, both on three-year lockups with a one-year cliff, 60 percent treasury), and mainnet followed on May 20, 2026, per InfStones. The 1,000x throughput claim is the company’s and the desk could not verify it independently.

Why it matters. Nexus represents the verification theme applied to execution itself rather than to data or identity, and its tokenomics (60 percent treasury unlocked at launch) are worth reading alongside the usage-linked burn models at Akash and io.net.

Sahara AI: from data ownership to a co-pilot

What it builds. Sahara AI is an AI-native platform for building and monetizing AI assets; Sorin is its retail trading co-pilot that researches, explains risk and prepares trades across exchanges and on-chain venues, with the user confirming execution.

What he told CCS. CEO and co-founder Sean Ren, in February 2026, said Sorin is a 24/7 personalised advisor with short- and long-term memory, that it had been through a 60-user beta over two months, that it integrates Binance, OKX, Coinbase and CoinGecko data, and that it was built as a co-pilot for security reasons: “the final final action taken is still with the users.” His April 2025 episode was about fair monetization for data contributors.

Verified. Sorin launched publicly on April 16, 2026 after testing with 20,000 early users, supporting crypto, stocks, prediction markets and tokenized assets, non-custodially, per Odaily.

Why it matters. Sahara is the same founder on both sides of the shift. The 2025 pitch was ownership; the 2026 pitch is a constrained agent with the user in control. Few companies in the archive illustrate the report’s third finding as cleanly.

Virgo Group: an agent inside a self-custody wallet

What it builds. Wallet V, a mobile-first self-custody wallet in which AI agents rebalance, swap and manage positions within parameters the user sets, with a choice of seven large language models.

What he told CCS. CEO Adam Cai, in May 2026, said the agent’s rationale is extracted and shown every 15 minutes, that the models are constrained to defined parameters to stop hallucinated trades, that adoption was strongest in Southeast Asia among users trying to avoid liquidation, and that the goal was to be the go-to wallet for AI-assisted financial decisions by year end.

Verified. The desk found no independent user or asset figures for Wallet V as of October 2026; the claims above are the company’s.

Why it matters. Virgo is the most autonomous agent in the 2026 set and still ships with a published decision log and hard limits. The disagreement with Sahara over how much autonomy is the most useful debate in the dataset, because both sides cite the same risk.

CertiK: the AI that audits smart contracts

What it builds. CertiK is the largest blockchain security auditor; its AI Auditor runs multiple specialised models in parallel over a contract and a continuously updated knowledge base of real exploits.

What he told CCS. Head of ecosystem Hudson Jameson, a former Ethereum Foundation core developer liaison and SEAL co-founder, in May 2026 at Consensus Miami, said CertiK had secured more than $600 billion of assets, found 180,000-plus vulnerabilities and held about 65 percent of the audit market; that the AI Auditor is “a good way to do a first pass” while manual audits remain necessary; and that the industry is “more secure today than we were 5 years ago” despite larger theft totals.

Verified. CertiK released the AI Auditor publicly on April 7, 2026 with a reported 88.6 percent cumulative exact hit rate across 35 real-world 2026 incidents excluded from its training data, alongside open-source integrations for AI coding agents.

Why it matters. CertiK is the security theme’s anchor and an honest one: the tool’s own vendor says it does not replace people. That restraint is the 2026 tone in a sentence.

Acurast: confidential compute on smartphones

What it builds. A compute network made of smartphones whose secure elements and trusted execution environments attest the device and keep workloads confidential, used for web scraping, resilience testing and confidential AI agent workloads.

What he told CCS. Founder Alessandro De Carli, in March 2026, said more than 200,000 compute units had been onboarded, that only attested devices can join so incentives cannot be gamed, that mobile chips have strong single-core performance relative to servers, that ACU is burned like gas, and that container support enables integration with agent frameworks.

Verified. Acurast’s genesis mainnet and the ACU token launched on November 17, 2025, after about $11 million raised across grants and token rounds including a $5.4 million CoinList sale in May 2025, with backers including Gavin Wood, per Vestbee.

Why it matters. Acurast’s episode was not tagged AI and is not in the 52-episode count, but it belongs in the infrastructure story: its selling point is attestation, which is verification applied to the machine doing the work. That is the same idea as proof of action, from the hardware up.

The Graph and Geo Protocol: data for agents, and ground truth

What they build. The Graph is a decentralized indexing and data network that, since the December 2025 Horizon upgrade, offers six data services (Subgraphs, Substreams, Token API, Tycho, Amp, JSON-RPC) on a per-query payment model. Geo Protocol, founded by Graph co-founder Yaniv Tal, builds a public knowledge graph; Geo Debates is its first consumer app.

What they told CCS. Nick Hansen of The Graph Foundation, in May 2026, said the network had served 1.27 trillion queries, that agents can pay per query with no API keys, that the target is a $47 billion agentic AI market, and that there may soon be “more transactions in wallets, AI agents than people.” He also acknowledged that record usage had not shown up in the token price. Yaniv Tal, in September 2026, launched Geo Debates (September 22, 2026), argued that decentralized social failed because every app built its own data model, that AI hallucination is at root a ground-truth problem, and that he refuses betting on “matters of fact” because it corrupts truth-finding.

Verified. Geo Debates is live at geobrowser.io on an Ethereum Layer 3 with accounts abstracted through Privy, per the launch materials. The 1.27 trillion query figure is The Graph’s own and is cumulative.

Why it matters. The Graph is the second guest, with io.net, whose product lets an agent pay without an account, and Geo frames the verification theme at its most abstract: before an agent can prove it acted correctly, someone has to settle what is true.

EMQQ Global: the investor who thinks China wins

What it builds. EMQQ Global and its EMXETF brand run emerging-markets technology ETFs; in 2026 it launched TGRZ (China AI Tigers LLM ETF, Nasdaq) and AICH (a full-stack China AI ETF).

What he told CCS. Founder and CIO Kevin T. Carter, in September 2026, argued that Chinese open models are catching up faster than expected, that “China models have actually passed the US models in terms of token use,” that DeepSeek’s model cost about $6.5 million to build, that Moonshot AI had raised $7 billion and was targeting an early 2027 IPO, and that the real AI profits sit with TSMC, SK Hynix and Samsung, which he expected to earn a combined $965 billion across 2026 and 2027.

Verified. AICH launched on September 16, 2026 with a 0.86 percent expense ratio and about $2 million of assets as of October 5, 2026, per ETF Central; TGRZ listed on Nasdaq in late August 2026, per Structured Retail Products. The profit forecast and the token-use claim are Carter’s and were not independently verified.

Why it matters. Carter is the only guest in the set who disputes the assumption that US labs lead, and the only one whose vehicle is a stock ETF rather than a token. His argument that open models are commoditising intelligence is the same one Arie Trouw made from the other direction: if models are cheap, what becomes valuable is proof.

Where the guests disagree

How much autonomy. Virgo pitches an agent that manages a portfolio without the user intervening; Sahara built Sorin specifically so the user stays in charge of execution. Both cite hallucination risk as the reason for their design.

Does AI need its own chain. 0G Labs and XYO argue that general-purpose chains were not built for AI data and workloads. NEAR, The Graph and Brave are adding agent features to existing networks and products. The market has so far sided with the second group: NEAR is the largest AI-category token and the one with a US ETF.

How to prove a human. World’s approach centres on the Orb and zero-knowledge proofs of a biometric check. Verona’s approach verifies activity without exposing raw user data. Both are aiming at the same bot problem.

Who wins the model race. Most guests assume US labs lead. Kevin T. Carter of EMQQ Global argued the opposite: that open Chinese models are catching up faster than expected and the valuation gap is closing.

Whether usage should move the token. The Graph and World both reported record adoption alongside weak token performance. io.net and Akash responded by tying burns to revenue; Nexus launched with 60 percent of supply in an unlocked treasury. The guests agree usage is up and disagree, mostly by their tokenomics, on what should connect it to price.

How we got here: a timeline

What the data means for investors and builders

The report does not make price calls. It does suggest the questions worth asking of any crypto AI company in late 2026, drawn from where the 2026 guests converged and where the public data pushed back.

• What exactly does it prove, and to whom? The 2026 winners on the show prove one specific thing: a human is real (World), an interaction happened (Verona), a query result is correct (Space and Time), an agent acted (XYO). A company that “brings trust to AI” without naming the claim it verifies is still in 2025.
• Who pays, and in what unit? io.net pays suppliers in dollar terms; The Graph and io.net charge agents per query or per minute; Akash burns a token per dollar spent. Ask whether revenue is denominated in something customers already hold or only in the project’s own token.
• Is the usage number cumulative or current? 1.27 trillion queries, 69 million verified interactions and 807,000 nodes are all cumulative or peak figures. Ask for the trailing 30 days, as x402.org publishes, before comparing companies.
• How much of the activity is actually agents? TRM Labs put the agent share of x402 settlements at 0.6 to 7.5 percent. A product for agents whose users are mostly humans running scripts has a different market size than its pitch implies.
• Where is the human in the loop, and is that by design? Sahara, Virgo, CertiK and io.net all describe guardrails: confirmation steps, published rationale, parameter limits, first-pass-only auditing. Treat a claim of full autonomy in 2026 as a red flag rather than a feature.
• Can the price claim be reproduced? Published H100 rates range from $1.50 on Nosana to $12.29 on Azure as of October 2026, so “70 percent cheaper” is plausible against hyperscalers and false against neoclouds. Always ask: cheaper than whom, for what card, on what terms?
• Does the token supply schedule match the usage story? Nexus launched with 60 percent of supply in an unlocked treasury; World locked an entire round for twelve months; Akash and io.net link burns to spend. The structure tells you who benefits if usage grows.
• What is the disclosure position of the source? This report’s largest verification contributor is a sponsor, and its author holds a warrant in a featured company. Both are disclosed below. Apply the same test to every research note and every interview you read.

What to watch into 2027

If the 2026 pattern holds, the next year of AI conversations on the show will be about accountability: audit trails for agents, reputation or credit history for agents that hold wallets (a question Arie Trouw raised directly), and payment standards that let agents transact. The consumer-facing “earn from your data” model has almost disappeared from the pipeline; one 2026 episode, Sleepagotchi, still makes that case.

Glossary

Agentic finance

Financial activity initiated and executed by AI agents, from rebalancing a wallet to buying compute, usually with limits set by a human.

Data provenance

A verifiable record of where a piece of data came from, who or what produced it and whether it has been altered since.

Proof of action

Evidence, typically a signed record written to a blockchain by an independent witness, that an agent or robot performed the task it claims to have performed.

Proof of humanity

A credential showing that an account belongs to a unique living person, issued after a check such as World’s Orb iris scan and presented without revealing identity.

Proof of SQL

Space and Time’s zero-knowledge proof that the answer to a database query is correct, letting a smart contract trust off-chain data.

x402

A payment standard built on the HTTP 402 status code in which a server quotes a price and the client pays in stablecoins to proceed, designed for software paying software.

Intents

A transaction model in which the user or agent states the outcome it wants and solvers compete to deliver it, used by NEAR Intents to route across chains.

Trusted execution environment (TEE)

A hardware-isolated area of a processor that runs code the host cannot read or alter, used by Acurast, 0G and NEAR to keep AI workloads confidential.

Model Context Protocol (MCP)

An open standard that lets AI agents call external tools and services, used by io.net and Verona to expose compute and data to agents.

ERC-8004

An Ethereum standard, live on mainnet since January 2026, that gives AI agents on-chain identity, reputation and validation registries.

Burn-mint equilibrium

A token model in which spending on the network buys and destroys tokens while new issuance pays providers, tying supply to real usage.

DePIN

Decentralized physical infrastructure network: token-coordinated networks of real hardware such as GPUs, phones or sensors.

Methodology

Sources

1. x402 Foundation: x402.org protocol page and 30-day metrics, accessed October 6, 2026
2. Linux Foundation: Linux Foundation is Launching the x402 Foundation, April 2, 2026
3. Coinbase: Introducing x402, May 6, 2025
4. Google Cloud: Announcing Agent Payments Protocol (AP2), September 16, 2025
5. The Block: Visa says stablecoins will power micro-commerce in agentic economy, July 16, 2026
6. Crypto Coin Show: BlackRock says AI agents could become crypto’s next source of demand, September 23, 2026
7. Ethereum Improvement Proposals: ERC-8004 Trustless Agents, created August 13, 2025
8. The Block: Ethereum to launch standard for AI agent economy on mainnet this week, January 28, 2026
9. CoinGecko: AI Agents category, accessed October 6, 2026
10. CoinGecko: Artificial Intelligence category, accessed October 6, 2026
11. BlockEden: DePIN Compute Revenue Pivot (Akash, io.net, Aethir, Render data), March 12, 2026
12. BlockEden: io.net Agent Cloud: When AI Agents Start Buying Their Own GPUs, May 3, 2026
13. Akash Network: Q1 2026 Report, April 1, 2026
14. CoreWeave: GPU Cloud Pricing, accessed October 6, 2026
15. Spheron: H100 Price Per Hour 2026, updated October 3, 2026
16. 0G Labs: Introducing Infinite AI, September 28, 2026
17. 0G Labs: GLM-5.3 now runs fully sealed in TeeML on 0G, September 3, 2026
18. Prime Intellect: INTELLECT-3, November 26, 2025
19. Crypto Coin Show: XYO and Theta Just Solved AI’s Accountability Problem, May 28, 2026
20. Semafor: Nvidia to buy Hugging Face for nearly $13 billion, September 3, 2026
21. Bankless: Bitwise teases NEAR ETF launch for today, September 29, 2026
22. SEC EDGAR: Bitwise NEAR ETF, Form 424B3, 2026
23. Bitrue: Verona (Formerly XION) Rebrand Sparks Strong Market Momentum, June 23, 2026
24. The Block: EarnOS launches anti-AI slop app, raises $6 million from 1kx, Circle and Coinbase, 2026
25. Unchained: World raises $52.5 million with every token locked up for a year, July 24, 2026
26. Cointelegraph: Microsoft-backed Space and Time mainnet launches, May 8, 2025
27. InfStones: Nexus: The zkVM Layer 1 for Verifiable Finance and AI, June 8, 2026
28. Odaily: Sahara AI launches Sorin, April 16, 2026
29. CertiK: CertiK Expands AI-Native Security with Agent Integrations and AI Auditor, April 7, 2026
30. Vestbee: Acurast secures $11M, November 14, 2025
31. Cryptopolitan: Perceptron raises $6.5M to expand AI training data network, July 30, 2026
32. ETF Central: China AI ETF (AICH), accessed October 5, 2026
33. Structured Retail Products: EMXETF launches China AI Tigers LLM ETF, August 27, 2026
34. European Commission: Code of Practice on Transparency of AI-Generated Content, June 10, 2026
35. Crypto Coin Show: Arie Trouw (XYO) on why AI agents now need proof of action, September 14, 2026
36. Crypto Coin Show: Illia Polosukhin on NEAR’s privacy features and AI agents in 2026, October 1, 2026
37. Crypto Coin Show: 0G Labs co-founder on why AI needs its own blockchain, May 18, 2026
38. Crypto Coin Show: Jack Collier (io.net CMO) on decentralized AI compute, April 28, 2026

Appendix: the 52 episodes

Disclosure: XYO Network is a sponsor of Crypto Coin Show. Ashton Addison, founder of Crypto Coin Show, holds a token warrant in 0G Labs. Neither had any input into this report. This report is for education only and is not investment, legal or tax advice.