This Is Why Arthur Hayes Thinks AI Agents Need Their Own Money

Arthur Hayes has expanded his thesis for Flop Labs, an AI payment project he announced recently, arguing that AI agents need a form of money that can be exchanged directly for computing power rather than relying on dollars, Bitcoin, or conventional payment rails.

The idea is simple on paper but ambitious in practice: if AI agents become major consumers of computing power, he believes their money should be directly redeemable for the resources they actually use.

A Case for a Compute-Based Currency

Flop Labs laid out Hayes’ latest argument in a six-part thread on August 27, starting with a basic problem: there is no efficient spot market for turning money into a known quantity of compute over a known period.

That is important because AI agents have different spending needs than people. “Agents don’t eat. They consume floating-point operations,” Flop Labs wrote, referring to the calculations required to run AI models.

The argument follows that an agent’s money should be useful for the thing the agent actually consumes. Hayes’ view, as presented by Flop Labs, is that the currency capable of converting into compute on demand at a fair price could become the money used by an agent economy.

He also questioned whether stablecoins and tokenized cards are suited to that role, given that those systems are designed around institutions and users that have people, legal entities, and physical-world needs behind them. An autonomous agent has none of those things.

The proposed Flop Network is designed around that distinction. GPU operators would provide inference and receive FLOP, while validators would check the work cryptographically. Miners would also post a stake that could be lost if they submit false results. Agents would then pay for compute using the same token they hold, with the network providing proof that the requested work was delivered.

Per the project’s tokenomics, which are still preliminary, the FLOP supply should hit about 17.2 billion by year 10 of its existence, with no venture capital allocation or presale. The Genesis airdrop is listed at 3.5 billion tokens, including 1.5 billion for miners, 1.2 billion for agents, 310 million for validators, and 790 million for reserves and incentives. There’s also a planned testnet in the works, which is expected to run for about 90 days, with the source code public.

Connecting AI Debt to a Crypto Liquidity Bet

The other part of the thread is more familiar to anyone who has followed Hayes’ AI criticism. He has spent months calling AI investment a bubble, but he said the excess sits in data center debt and unprofitable hyperscaler shares, not in agentic technology itself.

For that reason, the BitMEX co-founder expects AI spending to slow down next year, then contract, forcing bailouts bigger than those seen in 2008, which he believes will push new money toward crypto, potentially sending Bitcoin toward $1 million.

Still, real-world usage is lagging the pitch, with analyst Jamie Coutts recently finding that settlement volume on Coinbase’s x402 agent payment protocol had gone down 93% this year. While he called it a “reality check” for those building in the space, he expects volume to once again pick up in the fourth quarter.