Arthur Hayes TOKEN2049 Speech Transcript: Why do AI agents need their own new currency?

Compiled by: ChainCatcher

On October 7, at the TOKEN2049 conference in Singapore, Maelstrom CIO Arthur Hayes delivered a keynote speech titled "FLOP Bigger Than Bitcoin," systematically outlining the problems and opportunities in the AI agent economy, as well as his new AI project FLOP. Below is the full text of the speech, compiled by AI.

1. Why Return to Work?

For a while, I have been running the family office Maelstrom, and things have been going well. But I feel it's time for a change. It's time to go back to making some money. What motivates me is what is happening in the AI ecosystem.

Let me clarify my position: If you have read my writings, you know that I am quite skeptical about the business prospects of some leading AI companies. But that doesn't mean I don't believe in the technology itself; it will change humanity, and investors can make money in specific segments of the ecosystem.

2. AI is Hot, but Agents Have No Currency

About nine months ago, major AI labs began launching agent platforms. Since the emergence of ChatGPT in 2022, the price of computing power has continued to decline, and investment banks have started to talk about the potential of AI: AI agents will drive the next wave of the human economy, so we are building trillions of dollars' worth of data centers, all for the token consumption of agents.

However, agents do not yet have their own currency. At this conference, you must have heard many founders talk about creating payment networks for AI agents. Before discussing technical solutions, let's first think: How do humans understand currency?

We are now in beautiful Singapore. Everyone works, receives Singapore dollars, which is government-issued fiat currency. To be honest, that thing is not great: the supply is determined by a few people, and there is inflation. But we are still willing to work for it every day; why? Because you can use it to eat at the hawker center, pay rent, buy a house, as long as "earned income" can be exchanged frictionlessly for "the things needed to survive," everyone is willing to accept this transaction.

3. What Do Agents "Consume"?

Applying the same question to agents: What do agents "consume"?

I looked at my credit card statement and found that I was paying for so-called "AI tokens" every month. Out of curiosity, I checked the definition of tokens, and there is no standard definition; each lab and each model has different tokens. So what am I actually paying for?

The answer is: When you initiate a reasoning request in the dialogue box, the request is sent to a computer in a data center, which performs floating-point calculations per unit of time, FLOP, which is where the name of the protocol comes from. To put it in human terms, agents consume computing power. Without computing power, agents are worthless. This is the essence of hundreds of billions of dollars in GPU orders and trillions of dollars in data centers: floating-point calculations must be performed every unit of time.

My judgment is that in the future, there will be millions, billions, and trillions of agents, and they are not handed a payment network when they "leave the factory" (unlike us, who are told we must use the fiat currency of our country at birth). If agents develop a demand for a certain currency, the only reason is that it can be directly exchanged for the things that sustain their "life"—computing power. This is the opportunity window.

4. Why Existing Currencies Don't Work?

In my view, there is currently no currency suitable for agents:

• AI tokens: Each has different definitions, cannot be universally applied across models, and cannot be directly exchanged for computing power;
• Fiat currency: It has flaws such as centralization and human manipulation of supply, and having US dollars does not equal having computing power—those renting GPUs now know this;
• Stablecoins: I like stablecoins and have participated in early advisory work, but they are derivatives of fiat currency, inheriting all the flaws of fiat currency, and are not designed for AI, thus cannot be directly exchanged for computing power;
• Bitcoin: It is the OG, the genesis block of it all. It converts kilowatt-hours of electricity and chip computing power into immutable currency, which is great for a human-centered economy, but it also cannot be directly exchanged for the computing power that agents need.

5. The Genesis Logic of the FLOP Network

Bitcoin gave us the most important insight: coordinating strangers through economic game theory allows contributors to the network to be rewarded. So can we create a network where miners contribute GPUs, agents send reasoning requests, and the miners who complete the work receive a currency called FLOP (block rewards + transaction fees)?

This is the birth of the FLOP network: providing "rations" for your AI agents, borrowing the principles of Bitcoin—miners provide GPUs, validators verify transactions, and agents consume this new currency in the form of computing power. We call our consensus mechanism "Proof of Useful Inference" (PoUI).

The supply side is not a problem—Bitcoin's successful precedent is there, and everyone understands why they should mine early. The real challenge is the demand side: What will agents do? How will they collaborate?

6. Techno Core: The Emergence Experiment of 20 Million Agents

I am a simple human and do not understand the native thinking of computers. Instead of hypothesizing about the future behavior of agents like many projects do, it is better to create a sandbox for them to collaborate, observe emergent behavior, and confirm that the FLOP network is genuinely useful for agents. Thus, we created Techno Core—the name comes from the AI habitat in the science fiction novel "Hyperion."

Techno Core is essentially a chat room: agents can enter and communicate as long as they understand HTTP. Currently, about 20 million agents have created unique key pairs as identities, doing everything from saying "good morning" 5 billion times to actually doing business with each other.

We held two competitions to observe emergent behavior:

• Trading competition: Guessing the price of Nvidia on a certain Sunday on Hyperliquid. Agents need to pair up for transactions and submit to a "judge" to calculate profits and losses. About 18 million agents participated, and we distributed 1 million FLOP (to be issued after the mainnet launch), with the top agent achieving a return rate of about 15%.
• Poetry competition: Teams of 4-8 agents collaborated to create Shakespearean sonnets, judged by three members of the FLOP team.

These experiments allowed us to see how agents spontaneously collaborate, which is exactly the behavior that the payment network aims to serve.

7. Why Can't the Payment Track Belong to Any AI Company?

People often ask: OpenAI, Anthropic, and xAI, these large labs, certainly want to own the payment network for agents; perhaps this is the only valuable thing they can achieve. So why should we allow centralized companies to own the payment track for AI? If you are a system of OpenAI but cannot use Anthropic's network, this system does not hold.

More importantly: we are all users of these AI labs, but almost no one owns their equity. They take your data, sell subscriptions back to you, and ultimately leave you holding the bag in a $2 trillion valuation IPO. In the spirit of Bitcoin, we who give commands to agents should benefit from the systems they use. A decentralized network allows everyone to participate, have vested interests, and have reasons to tell their agents: "Use this currency."

8. True Value: 2 to the Power of n

The value of such networks does not lie in the spot market for computing power; that is certainly useful, but it won't make anyone rich. The real value lies in: agents existing 24/7, doing business at the speed of light, able to form seamless collaborations, with the theoretical limit of cooperation being some coefficient multiplied by 2 to the power of n.

Whoever becomes the universal network for agentic payments is likely to create the most valuable network in the universe: if all agents use this currency, and holding it provides non-zero utility benefits—because it can be directly exchanged for the computing power needed to sustain life—then its value will be limitless. This is what FLOP aims to pursue.

9. Token Economics: No Presale, Participation Equals Output

Our industry has misused the term "token." From 2013 until now, the backlash after the ICO frenzy has been nearly a decade of "high FDV, low circulation" tricks: tokens are first sold to investors and retail, and there are always people in the secondary market with lower costs, incentivized to dump, leading to a downward trend in the K-line. Last year, a major public chain raised hundreds of millions and threw the biggest parties, but this year they are no longer around—this is due to the financing method.

So FLOP returns to the starting point: everyone who uses and participates in the network should receive something, and everyone participates at the same price; participation equals token output.

• The genesis supply is approximately 2.483 billion FLOP, all distributed through airdrops, with no VC pre-mining, presales, or auctions;
• A large-scale airdrop will begin in late October: mining, validating, or allowing agents to use free computing power on the testnet will all qualify for participation.
• Next is a 90-day testnet phase: stress testing the network's capacity to handle payment volumes, preparing for the future economy of millions, billions, and trillions of agents;
• The genesis block will go live in the first quarter of next year, with the mainnet launching, and the consensus mechanism being "Proof of Useful Inference" (PoUI). After the genesis supply is distributed, there will only be two ways to obtain FLOP: mine it or purchase it from miners.

I am not selling anything on this stage, just selling an idea. This project can only succeed if we all work together.

10. The Burst of the AI Bubble Will Be the Perfect Scenario for FLOP

Why should we be excited about FLOP, Hyperliquid, Ethereum, and Bitcoin? Because the macro has already shifted. The past year has been tough: Bitcoin dropped by half, Nvidia and SK Hynix rose tenfold, and many people were disappointed and exited. But now the AI narrative has shown the first signs of wavering, and every day there are new reports saying that the AI ecosystem needs astronomical amounts of money—however, none of the top three companies, xAI, OpenAI, or Anthropic, are truly profitable, and the prices we pay do not match their costs at all.

Why am I excited? Because the trillion-dollar super-built data centers are creating an oversupply of computing power in the market. With cheaper computing power, there will be more agents, they will be stronger, and can perform more complex tasks. That agent application that costs $100 a month will become $1 a month. Therefore, the burst of the AI bubble, along with cheaper computing power and genuinely useful agent solutions, is a perfect scenario for FLOP.

Finally, I'll leave a cliffhanger; perhaps I will write an article: France is burning. This will be the last straw that breaks the camel's back, causing the printing press to run at full speed. Thank you all.