0G Labs' George Varghese says AI agents now use more LLM tokens than humans
The 0G Labs executive says the agent economy is at an inflection point, but agent payments are still tiny and trust remains the barrier.
- George Varghese said on Oct. 8, 2026, that AI agents now consume more large-language-model tokens than people do.
- AI agents made more than 138 million x402 payments, more than 87% of them under 10 cents, Chainalysis found.
- Fed Governor Christopher Waller said on Sept. 29 that agentic commerce is still in an early phase.
George Varghese, head of investments and partnerships at 0G Labs, says the AI agent economy has reached an inflection point, with agents now consuming more large-language-model tokens than people.
He makes the case as AI agents have made more than 138 million payments over the x402 protocol, according to Chainalysis data reported by Blockhead on Oct. 15, 2026.
Varghese set out his view on stage at TOKEN2049 in Singapore on Oct. 8, 2026, in a keynote titled "Building the Trillion-Dollar Agentic Economy". 0G Labs builds a blockchain, storage and computing network for AI applications.
Agents arrive, payments lag
Agents now transact, but in tiny amounts. More than 87% of x402 payments fell between $0.0001 and $0.10, according to Chainalysis, and most ran on Base, Coinbase's network.
The money to settle them is on-chain already. Stablecoins worth $303.1 billion circulated across blockchains on Oct. 16, according to DefiLlama.
The Federal Reserve agrees that the shift has started but is early. "Market participants broadly agree that agentic commerce is in an early phase," Governor Christopher Waller said on Sept. 29, in a speech on payments in the age of AI agents.
'An inflection point'
Varghese traced agents from chat in 2023 to co-pilots in 2024 and to multistep tasks in 2025 and 2026. "We believe we're at an inflection point in AI agent economy," he said.
"We are at a point right now where agents are actually out consuming LLM tokens compared to humans," Varghese said.
He cited OpenRouter, an AI model marketplace, as processing 126 trillion tokens a week in September 2026, up from 0.5 trillion a week in 2025. We did not independently verify those figures.
The check: consumer spending has not caught up. Varghese himself said only about 2% of U.S. households pay for AI services, his own estimate.
Small money, big forecasts
Varghese was frank about how small agent payments are. He said only $24 million in volume has gone through x402, at an average of 29 cents a payment, and put agentic payments at about 0.1% of the level forecast for 2030.
The check: Chainalysis counted more than 138 million x402 payments, mostly between one and 10 cents. Its count and Varghese's dollar figure point the same way: many payments, little money.
He said the gap is about trust. "Most of the people here in this room would not trust an agent with more than 10% of their net worth," Varghese said.
0G's pitch: a trust layer
0G's answer is verification. "We are solving the trust and verification issue that a lot of us have," Varghese said, describing a stack that lets users check which AI model served them.
He cited company-reported figures: 2.7 million live agents on 0G, a target of 10 million next year, more than $350 million raised and an $80 million ecosystem fund. We did not independently verify them.
Varghese stayed bullish on the cycle. "AI is still in a bull market, and we think it's still an early bull market for it," he said.
He also argued that costs favor open networks. "While that's occurring, the cost of inference is actually coming down," he said, adding that 0G trained a 107-billion-parameter model at 95% less cost than a conventional run, a company claim.
By the numbers
| Metric | Value | As of | Source |
|---|---|---|---|
| x402 payments by AI agents | more than 138 million | Oct. 15, 2026 | Chainalysis via Blockhead |
| Share of x402 payments from $0.0001 to $0.10 | more than 87% | Oct. 15, 2026 | Chainalysis via Blockhead |
| x402 dollar volume | $24 million (Varghese's estimate) | Oct. 8, 2026 | George Varghese |
| Stablecoins on all chains | $303.1 billion | Oct. 16, 2026 | DefiLlama |
| Live agents on 0G | 2.7 million (company-reported) | Oct. 8, 2026 | 0G Labs |
The other side: who carries the risk
The Fed frames trust as a problem of authentication, liability and fraud. Waller said fraud models built for human behavior may not translate to agents and will need recalibrating.
Card networks are also competing to be the trust layer. Mastercard launched Agent Pay for Machines on June 10, 2026, a service for payments between AI agents and machines, the company said.
That puts blockchain verification against card rules and existing fraud systems in the same market.
Waller drew a line between agent-assisted commerce, where the buyer stays in control and pays, and agent-delegated commerce, where the agent shops and pays alone. The second model, he said, needs far more trust mechanisms and guardrails.
What to watch: dollars, not counts
The first signal is dollar volume. Rising totals through x402 and similar protocols, not only payment counts, would show agents moving real money.
The second is 0G's own targets: 10 million live agents next year and new decentralized training runs open to outside GPU owners, both company plans.
The third is standards. Waller asked whether agentic commerce will settle on platform-specific or interoperable standards, a choice that will decide whether open networks get a share.
What we do not know is how much agents spend today across all payment rails, not only on-chain.
How we reported this: Varghese's remarks come from his Oct. 8, 2026, TOKEN2049 Singapore keynote, checked against the video's auto-generated captions. Payment counts come from Chainalysis as reported by Blockhead; stablecoin supply from DefiLlama as of Oct. 16; Waller's remarks from his Sept. 29 speech on the Federal Reserve website. 0G's figures are company-reported.