Most investors are comfortable buying a story. AIcrypto has run almost entirely on one. AI is the future, crypto is the rails, buy the intersection. That story pushed the sector to $22.2 bill
Most investors are comfortable buying a story. AIcrypto has run almost entirely on one. AI is the future, crypto is the rails, buy the intersection. That story pushed the sector to $22.2 billion in combined market capitalization as of September 10, according to CoinGecko, led by Bittensor, NEAR Protocol and Render.
Now the sector has to answer a harder question. Who pays, and for what?
To find out, TheStreet Roundtable put that question to five founders and executives, each building a different layer of the stack AI systems actually run on. Storage, data access, verification, coordination and inference. They answered on the record, including the parts that don't flatter them. Four other big names, most of them in decentralized compute, declined or would have duplicated the five below.
Storage
Filecoin
FIL
Hosting data across independent providers, with proof of integrity
Data access
The Graph
GRT
Queries against indexed, structured blockchain data
Verification
OriginTrail
TRAC
Publishing knowledge to a tamper-evident, checkable record
Coordination
AI Rig Complex
ARC
Settling fees when an autonomous agent buys a service
Inference & settlement
NEAR Protocol
NEAR
Staked access to attested private inference; agent settlement
None of these tokens is equity. Depending on the project, it is closer to a metered access pass, posted collateral, or a claim on network usage.
Filecoin is the oldest project in this group, and its pitch starts with concentration. "Today's internet is centralized," said Marta Belcher, president and chair of the Filecoin Foundation. "The vast majority of the data behind the websites we use every day sits in data warehouses owned by just three companies. When one of them goes down, vast swaths of the web go down with it."
The concentration is measurable. Enterprise spending on cloud infrastructure hit $143 billion in the second quarter of 2026, and Amazon, Microsoft and Google took 28%, 20% and 15% of it, respectively, according to Synergy Research Group.
Filecoin's network spreads copies of data across independent storage providers, who are paid in FIL while cryptographic proofs continuously verify nothing has been altered or lost.
"Storage location is becoming one of the biggest bottlenecks for AI. The datasets are enormous, and moving them across the world to wherever your cloud provider happens to keep them is slow and expensive. That's data gravity," Belcher notes.
Past a certain size, datasets stop moving and everything else migrates to them. Belcher's answer is to distribute copies, which she says "lets you run compute close to the data instead of dragging the data to the compute," and removes the single vendor who "can raise prices, change terms, or shut them down."
The product version is Fil One, object storage compatible with Amazon's S3 API, the de facto standard. Matching that API means a company can redirect existing workloads without rewriting its software.
Asked whether decentralized storage is actually cheaper, or whether the real appeal is control, Belcher said it "can be more affordable because you're putting idle capacity to work." There is now a public price. Fil One lists storage at a flat $4.99 per terabyte per month with no egress fees, against roughly $23 per terabyte for standard S3 storage on AWS. What Filecoin still has not named is a paying customer.
Blockchains hold enormous amounts of data, and almost none of it is easy to query. Even "show me every transaction this wallet made" is painful from raw chain data. Independent operators called Indexers organize that data into queryable APIs called subgraphs. GRT pays for the queries, gets staked as collateral, and rewards the Indexers doing the work.
"As of early 2026, The Graph has served over 1.27 trillion queries for more than 75,000 projects across a 60-plus network," Hansen states.
What's changed, Hansen said, is who's asking. Alongside developers, "there have emerged two new users The Graph and many other industry participants want to serve: institutions and AI." The network recently shipped MCP servers and skills, letting AI platforms including Claude plug directly into more than 15,000 subgraphs and pull live onchain data.
The bigger bet is on autonomous agents. Hansen's checklist for an agent is a way to prove who it is, a way to see what's happening around it, and a way to pay for what it uses. He says The Graph covers all three. Agents carry identity via the ERC-8004 standard, read state through subgraphs, and settle per-query payment through x402, a machine-to-machine payment protocol, "with no API key, no account, and no human involvement."
Asked why a centralized company couldn't do this same job more cheaply, Hansen called it a complex question that applies to the whole industry, not just The Graph. His position is that users will ultimately prefer the reliability of a distributed network over a single vendor. Read that as a bet, not a moat. AWS is reliable too, and it doesn't require anyone to hold a token.
OriginTrail's question is one the other four largely sidestep. How does anyone know the information an AI is using is actually real?
"If by AI system we're limiting ourselves to large language models, then the answer is that it doesn't, and it can't. LLMs are probability engines returning the most likely next words, not the most truthful ones. Fluency and accuracy are two different things that happen to look identical on screen," Tomaž Levak, co-founder of OriginTrail, adds.
OriginTrail's Decentralized Knowledge Graph attaches provenance to information. Where a claim came from, who published it, when, and whether it's been altered. It does not try to adjudicate truth itself.
"It gives every piece of knowledge a verifiable origin, cryptographically signed and independently checkable, so an answer arrives with a trail attached," Levak says. "The AI still generates. The difference is that a human, or another system, can now audit the source instead of taking the output on faith."
TRAC is the token used to publish knowledge to that layer and to stake against the network. One structural detail sets it apart from almost everything else in this sector. Minting is permanently disabled and supply is hard-capped at 500 million tokens, so there is no emissions schedule and no unlock cliff, the scheduled date when large batches of previously locked tokens hit the market. OriginTrail says the full supply is already in circulation. CoinGecko currently lists about 447 million. Either way, the setup removes a category of risk that hangs over the rest of this list, and removes a source of growth funding along with it.
AI Rig Complex is the youngest project in this group and the most agent-first. It's an open-source framework for autonomous agents that rebalance portfolios, book services, and execute workflows by transacting with providers in a registry. ARC settles the fee each time an agent uses a service, with most of it flowing to the provider.
Founder and CEO TJ Dunham's starting point is that payment rails built for humans don't translate to software.
"Software is doing the buying now, and every payment rail we have was built to verify a human. A credit card assumes a cardholder. Software has none of that," Dunham says.
The payments themselves are tiny. "One task can fire off thousands of purchases, a fifth of a cent each, and Stripe's card minimum is fifty cents," he said. "The fee eats the payment." The deeper issue, Dunham argues, isn't the ledger. It's accountability. "Nobody can tell you who eats the loss when an agent buys the wrong thing. No ledger fixes that. A contract does." He expects the AI industry to adopt first, led by model providers and data vendors that already bill each other by the API call. Digital advertising is the dark horse, he says. "Real time bidding has been an agent economy for close to twenty years and never called itself one."
Programmatic advertising already proves that machine-speed, sub-cent transactions work at scale. It was a $162.4 billion market in the US alone in 2025, up 20.5% from a year earlier, according to the IAB. It also proves the market was built entirely without a blockchain or a token. An investor should ask ARC what the token adds that ad tech solved without one. Dunham's answer for the hardest sectors, healthcare and finance, is blunt. "Nobody there lets an agent move money until they can show a regulator exactly what it did and prove nobody edited the record afterward."
NEAR is the largest project here. Its co-founder and CEO, Illia Polosukhin, is one of the eight authors of "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture behind every modern large language model.
He's now arguing that the thing he helped start needs a different ownership structure.
"More and more of the economy is going to run through AI agents acting on your behalf: booking things, moving money, making decisions in the background. The critical question is, does that agent work for you, or for whoever built it?" Polosukhin mentions.
Today, he argues, it doesn't. "You sign away your data, your hard work is converted into training data and corporate profit, and the model is optimized for someone else's business, not your outcome." His alternative, which some institutions call sovereign AI, keeps data with the user and makes the whole stack auditable. NEAR AI runs inference inside secure enclave hardware, independently attested by Intel and NVIDIA. He named Abound, a cross-border payments platform, as a live user, and said NEAR is in discussions, not a signed deal, with the government of Bermuda.
Asked what would have to go wrong for the thesis to fail, Polosukhin named two risks. One is technical, unverifiable infrastructure handling agent credentials and capital. The other is cultural, "that users and enterprises settle for a world-historical technology they don't actually own." That second risk is the honest bear case for this entire category, from the person with the most incentive not to raise it.
This sector runs on pilots, partnerships, and press releases. The most useful discipline a TradFi investor can bring is to ask, every time, for the customer, the contract, and proof it's in production.
Three of the five answered without being pushed, and in three different currencies. The Graph answered in volume, 1.27 trillion queries, though it's unclear how much of that converts to GRT revenue versus riding on the DeFi (decentralized finance, apps that let people trade, lend, and borrow crypto without a bank) base the network was built for. OriginTrail answered in names, the British Standards Institution, Swiss Federal Railways, and security audits covering more than 25,000 factories feeding US import chains. NEAR answered in staked capital, roughly 650,000 NEAR locked up within weeks for confidential inference access, plus Abound as a named production user. Filecoin answered in architecture and, now, a published price, but still no named customer or contract, the most conspicuous gap in the group.
None of these answers is revenue. An investor should want all of it. Named counterparties, volume, locked capital, and a price per unit with a currency sign in front of it.
These tokens don't behave like equity, and reaching for a P/E ratio is a category error. A few rules of thumb instead.
Tokenomics can swamp fundamentals. Emission schedules, vesting cliffs, and burn mechanisms can overwhelm demand-side growth for quarters at a time. TRAC's capped supply is the outlier here. Most of this group carries emissions or unlock cliffs ahead. Ask about all three before assuming usage growth equals price growth.
Ask the hard question, and notice when it goes unanswered. Every project here has to explain why a centralized vendor can't do the same job cheaper. NEAR's Polosukhin named his own company's failure condition unprompted. The Graph's Hansen called the question too complex to answer. A token whose competitive case rests on user preference rather than structural advantage is priced on faith in that preference holding.
Watch which design actually routes value to the token. Staking for access and locking tokens converts usage into token demand more reliably than paying per transaction. NEAR works that way and saw hundreds of thousands of tokens locked within weeks. The Graph's massive per-query volume hasn't moved GRT the same way.
Utility tokens still trade like speculative assets, and the rules aren't written. Even genuine revenue-backed demand won't insulate a thinly traded, largely unregulated, 24/7 asset from a 60% drawdown. How U.S. regulators classify AI infrastructure tokens remains open, and it will affect both how they trade and who's allowed to hold them.
"Top AI crypto tokens" is a ranking of where money has gone. It isn't a ranking of which businesses work.
The useful exercise for a TradFi investor isn't picking off a leaderboard. It's asking two questions of any token in the category. Which layer of the AI stack does this get paid for servicing, and is that demand real yet, or still mostly a thesis? None of the five answers both cleanly. Filecoin has the most intuitive argument in the sector, a published price, and no disclosed customers. The Graph has trillions of queries served and no good answer on why a centralized rival couldn't serve them. OriginTrail has institutions in production and a standards fight ahead. AI Rig Complex has a credible machine-payments thesis and an open question about why it needs a token at all. NEAR has attested hardware and a Transformer co-author making the case, and admits the whole thing fails if buyers keep accepting the terms they're offered now. The gap between the pitch and the receipts is where the money gets made and lost, and none of it shows up on a market-cap table.
Filecoin, The Graph, OriginTrail, AI Rig Complex and NEAR Protocol each responded to written questions for this article. Bittensor, Internet Computer, Render and the Artificial Superintelligence Alliance were not interviewed. The 1.27 trillion cumulative queries, roughly 650,000 NEAR staked, 25,000-plus factories and full-circulation supply figures are the companies' own, published as provided and not independently verified. Market-cap, storage-price, cloud-share and advertising figures come from CoinGecko, vendor pricing pages, Synergy Research Group and the IAB.
This article is for informational purposes only and does not constitute investment advice or a recommendation to buy or sell any asset. Digital assets are volatile, largely unregulated, and can lose value rapidly. Readers should conduct their own research and consult a qualified financial professional before investing.
