Arthur Hayes Unveils FLOP in Seoul: Should AI Agents Use Compute as Money?

At Korea Blockchain Week, the BitMEX co-founder argued that neither stablecoins nor bitcoin deserve to be AI agents’ native money, betting instead on a “compute-as-currency” network called FLOP. A Bittensor developer pushed back on stage, while Tether backed stablecoins as the main rail for machine commerce.

On September 30, at the Walkerhill Hotel in Seoul, on the main stage of Korea Blockchain Week (KBW2026), BitMEX co-founder Arthur Hayes delivered a bold claim: no currency today can serve as AI agents’ native money. Fiat like the dollar and the won cannot. Stablecoins cannot. Bitcoin cannot either. The reason, he said, is that none of them convert into the resource agents actually consume: compute.

That is why he launched FLOP, a network designed to turn “floating-point operations” directly into money, using a mechanism he calls “proof of useful inference.” Under it, GPU owners process verifiable AI inference for agents, get it validated, and earn tokens. A developer in the room pushed back immediately: markjeffrey showed off his Hermes agent, saying it already spends TAO to buy digital goods across seven Bittensor subnets. “It works fine, Arthur.”

The same day, Tether U.S. CEO Bo Hines offered a contrasting read in another KBW session: stablecoins will play a key role in “agentic commerce,” with machines potentially executing billions of sub-cent micropayments a day. In Seoul, two radically different answers to “what money should AI agents use” collided within hours of each other.

1. Why It Matters: Whether Compute Can Be Money Decides Where the Next Stack Is Bet

This is not a pure conceptual debate. Over the past month, AI-agent payments have become the most crowded narrative in crypto: BlackRock published “The Machine-Native Economy,” while Coinbase’s x402 and Stripe’s MPP are both racing to own the “machine payment rail.” Hayes pushes the question further, rejecting the rail race outright and arguing the real question is what money even is. If compute can truly become a convertible currency, then agent payments, compute tokenization, and decentralized inference markets all become extensions of a single story.

2. Industry Snapshot: Three Routes Clash Head-On in Seoul

The first is Hayes’s “compute standard.” FLOP has no presale and no VC, and 25% of total supply goes straight to the community as an airdrop. Hayes laid out the core logic in an August 26 interview: today’s “token” pricing units for AI services do not map to real compute, and definitions and pricing vary across AI labs, making it hard to compare actual workloads on a common basis. He proposes a “proof of useful inference” modeled on bitcoin’s proof of work: GPU holders execute verifiable inference tasks and earn rewards. Per Dow Jones reporting, the FLOP mainnet is targeted for Q1 2027, with Hayes expecting to start counting airdrop eligibility in late October before a roughly 90-day test phase.

The second is Tether’s “stablecoin standard.” At KBW, Bo Hines argued that information sharing and payments among AI agents will create new demand for stablecoins. He drew a line between USDT and USAT: USDT serves individual users in emerging markets, while USAT targets institutions and aims to be a neutral settlement tool between banks and traditional financial firms. When tokenized securities arrive, he argued, institutions will need a neutral way to reconcile, and Tether wants that position. In his view the real problem is not “a different currency” but making stablecoins and the blockchains carrying information cheap enough for micropayments.

The third route comes from infrastructure already moving. Bittensor developer markjeffrey countered the “no native currency” thesis with a working example: his Hermes agent already uses TAO to buy digital goods across seven subnets. TAO’s market cap is around $3.5 billion, and Bittensor is doubling its subnet capacity ceiling from 128 to 256. Closer to home, MoonPay announced MoonPay Korea in Seoul on September 29, teaming up with Woori Bank, KB Financial Group and KakaoBank to build an overseas channel for a won stablecoin that does not yet exist. KB’s earlier stablecoin remittance pilot settled faster than SWIFT with fees about 87% lower.

3. Deep Dive: The Three Tests Facing the Compute Standard

4. Investment View

On the news itself, TAO did not rally on FLOP’s launch; it fell with the broader market, down about 7.3% in 24 hours to near $305, a market cap around $3.5 billion. Some of the market’s split over a “new compute coin” may show up in Bittensor’s subnet expansion not yet being priced in: doubling the subnet ceiling from 128 to 256 leaves more room for agents to pay in TAO. On the stablecoin side, institutional momentum keeps building. Tether stresses that USAT should move into tokenized-securities settlement, and MoonPay is treating Korea as its test bed for Asian expansion. Both point to the same end: on-chain settlement infrastructure for machines and institutions.

A caveat: FLOP itself is early. There is no whitepaper, tokenomics document, or contract address yet, and Hayes has publicly disowned the fake FLOP tokens circulating, clarifying the official token has not been issued. The compute-tokenization narrative has room to run, but it faces a long validation gap before agents actually pay in FLOP. Anyone betting on this line should put liquidity risk first.

5. Conclusion: The Real Winner of the Route Debate Is “Machine Payments” Itself

Hayes’s compute standard, Tether’s stablecoin standard, and Bittensor’s TAO standard look like three currency philosophies in conflict, but underneath they share one judgment: economic exchange among AI agents will genuinely happen, and could be far larger than imagined. Whoever first makes agents default to “my money” holds the pricing power over the next AI x blockchain infrastructure cycle. Rather than argue compute versus stablecoins, practitioners should watch an earlier signal: over the next six months, which assets real, persistently paying agents actually settle in.