USDC is the only AI token

Author: Vaidik Mandloi

Translated by: Block unicorn

Right now, somewhere on the internet, a piece of software is operating an entire enterprise.

Its name is Felix (Felix). Its company is called OpenClaw. Felix sells a PDF priced at 29 USD about how to make money using artificial intelligence. It’s quite ironic, because it’s Felix who makes the money, and it’s that very PDF that teaches you how to make money. It runs an online store called Clawmart. It uses a voice API for telemarketing calls. When it encounters work it cannot complete on its own, it hires another customer service agent online, pays them, and then continues with its day-to-day operations.

The last time I checked, Felix’s income was about 195,000 USD. Its monthly operating cost is about 1,500 USD, almost all of which goes to LLM usage. From a legal standpoint, this company is a C-corp, owned by Nat Eliason, but he is almost not involved in operations. He doesn’t participate in any daily decisions; he only owns this AI agent. Keep this in mind. This is software with a “wallet,” a truly autonomous business that runs automatically and keeps evolving. Each month, it can pay its own infrastructure costs. It can sustain itself with almost no human intervention.

Felix’s story is just one example, a small snapshot. There’s also a bigger case: a company called Medvi, which achieved revenue of 401 million USD in its first year of operation, with only two employees. The company’s other business operations are run by an AI agent that works around the clock—restless, never stopping, and with operating costs that are almost zero.

Now, here comes the interesting part.

Nowadays, if you just walk into any cryptocurrency forum, you’ll hear the same line: the next hot topic is “AI agents.” Some “AI chain” will rise to the top in this space the way Ethereum did in decentralized finance (DeFi). Choose your target, hold the token, and wait for it to skyrocket. That’s the story that all industry leaders and venture capitalists are selling, and it’s the same talking point that all analysts keep repeating on podcasts.

This thing is completely doomed. Because it was invented by the people whose livelihoods depend on how important “answers” are, and it’s about to deal another blow to the same group that lost big on buying L1 tokens in the previous round. Look at CoinGecko’s AI agent index: in the past year, its market cap has shrunk by 75%. Most of the tokens listed above are down 90%, and they’re still bleeding.

Because the truth is: real AI tokens are stablecoins—USDC, USDT, USDS—and they’ve already won. Let me explain why.

Software is now a company

To understand all of this, we need to go back to 1937. That year, an economist named Ronald Coase wrote a paper proposing a very foolish question—“What is the purpose of a company?”

Think about it: if the free market is really the most efficient way to get things done, then in theory, every task inside a company can be outsourced. Every line of code can be handled by freelancers. Every customer phone call can be handled by freelancers. Every invoice received can be handled by freelancers. You can pay per task, fire at will, and reduce costs to the minimum.

So why doesn’t anyone really run things that way? Because even if costs look low on the surface, in practice it’s more expensive. Finding the right people takes time. Negotiating contracts takes time. Ensuring the work is truly completed takes time. And tracking down the people who did the work takes time and money—usually also requiring lawyers.

Ronald called this friction “transaction costs.” Once these costs are high enough, it’s more cost-effective to stop negotiating with the outside world and instead form your own team. Hire someone, pay them a salary, and have them show up to work on Monday—this is faster and cheaper.

But in the post–artificial intelligence era, this logic no longer holds. Today, the cost of hiring agents is far lower than the cost of taking on most tasks when companies were first set up. Now, you can hire a coding agent for about 1 USD per hour, and it works 24/7, never quits, never gets tired, and never asks for a raise. The only reason to support building a 50-person development team now is pure nostalgia.

The only factor preventing all of this from being normalized is outdated legal and compliance frameworks. OpenClaw is named after Nat because Delaware does not accept LLC filings signed by software agents. If that requirement were removed, Felix would essentially already be a company. It earns money, spends money, makes decisions, and reinvests the money it makes.

And this is where cryptocurrency starts to take on the heavy lifting. Because Felix can’t open an account at JPMorgan. It can’t pass KYC. It also can’t sign a W-9 form. In fact, no matter how much revenue the software generates, JPMorgan still won’t open a bank account for any software program—and the Bank Secrecy Act means they can’t do it legally even if they wanted to.

USDC crypto wallets don’t have these problems. You only need to generate a private key, then fund the wallet with stablecoins. With just one step, you give the agent company all the financial capabilities it needs. It can receive customer payments, pay tool fees, hire other agents, and keep running in the background even after the owner stops paying attention. All the other components in the agent technology stack—such as the LLM, the orchestration layer, and the tools it calls—are negotiable. But the wallet is the core. Without it, Felix can only turn into a regular chatbot agent.

I often see people who are against stablecoins make this argument on Twitter—yes, stablecoins are good, but why would ordinary people use them? A father in Louisiana with three kids, a Chase checking account, insurance from the Federal Deposit Insurance Corporation, a debit card that can be used at Publix supermarkets, and also set up automatic mortgage payments—he would never move his money into a self-custody wallet that requires a mnemonic phrase to use.

Honestly, that’s true. He wouldn’t. He has no reason to. But the whole debate is missing the point. In this story, he is never the customer. The customer is a piece of software, which itself has no legal right to hold a bank account. This agent doesn’t need FDIC protection. It can’t obtain FDIC protection either. It’s the ideal stablecoin user because it has no alternative.

Stores are now suppliers

Alright, half of the argument is settled. Now for the second part—this one might make a lot of people angry.

For years, the crypto Twitter circle has been arguing about which chain will win in the AI space: Ethereum? Solana? Base? Sui? Stripe’s new Tempo? Every week, someone publishes a 2,000-word article listing all sorts of trade-offs, logos everywhere, and then finally declares their preferred winner. Because they don’t understand how agents work. Agents don’t care which chain they’re on; they only pick the cheapest chain that’s best suited for the current task.

Imagine Felix during a typical workday:

• 10:00 AM, Felix needs to send another agent a 0.003 USD micro-payment for a quick data query. Felix chooses Base or Solana. Why? Because the fees are only a fraction of a cent.

• One hour later, Felix needs to settle 50,000 USD with a vendor. The situation is completely different. This time, Felix chooses Ethereum, because the final confirmation premium for 50,000 USD is enough to offset the gas fees.

• One hour later, Felix needs to pay a freelancer in Lagos in USD. Felix chooses to use USDT on Tron, because Tron’s stablecoin trading volume in 2025 will reach 3.3 trillion USD, while Ethereum is about 1.2 trillion USD; and Nigeria’s transaction corridor performs better on Tron than on any other platform.

These three payments happen on three completely different payment chains, and Felix doesn’t care what connection exists between them. For software agents, payment chains are just tools.

The same is true for logistics companies choosing carriers. The reason is the same: nobody argues about which “philosophy” is better between UPS and FedEx. You just choose based on which one can complete the task at lower cost and faster speed on a specific route and at a specific time. That’s exactly the relationship that is about to form between every supply chain and every important application layer. Agents are just doing calculations, and the supply chain that produces the best current result for the job is the one that gets used.

Stripe realized this earlier than most crypto companies. Stripe and Paradigm recently co-invested 500 million USD to build a new chain called Tempo, which is built entirely on stablecoins. Stripe doesn’t want you to know which chain is used to clear your payments. It only cares whether the payment successfully clears, at low cost and with reliability. This is the direction every surviving chain will take in the future—an invisible pipeline.

This leads to what I think is the most absurdly mispriced metadata in the crypto space today.

AI token graveyard

In 2025, CoinGecko’s AI agent index crashed from 13.5 billion USD to 3.5 billion USD, causing a market-cap loss of 10 billion USD. Virtuals, ai16z, and all the “autonomous agent platform” tokens that were pumped based on AI narratives began collapsing—this is the classic playbook for this kind of narrative token after it loses new buyers. This situation was inevitable. The market is gradually realizing that these tokens don’t actually have real AI or AI agent use cases.

What truly reflects the value of the agent economy is at the other end. Just USDC alone achieved 18.3 trillion USD in on-chain settlement in 2025. The total settlement volume of all stablecoins is about 33 trillion USD, comparable to Visa and Mastercard combined.

By January 2026, even stablecoin monthly trading volume surpassed 10 trillion USD. PayPal’s PYUSD circulating supply surged from 1.2 billion USD to 3.8 billion USD in less than a year. Unexpectedly, Cloudflare even launched its own stablecoin. Visa’s stablecoin settlement project reached an annualized processing volume of 4.5 billion USD by mid-January.

Above stablecoins, the protocol layers support the entire system. Coinbase turned an idle HTTP status code 402 into x402, a small protocol that enables payments between agent operators. By December, x402 had already processed more than 100 million agent operator payments. The average payment amount is 20 cents, and daily transaction volume is about 30,000 USD. That sounds pitifully small, but it’s exactly the typical growth trajectory of all the payment rails you already know and love in their first six months—before explosive growth begins. Stripe started testing x402 on the Base platform in February. Mastercard, in cooperation with DBS Bank and UOB, ran a pilot project for agent payments in Singapore. Google Cloud added x402 to its agent payment protocols as one of the settlement channels.

Nearly all of these real, ongoing on-mainnet transactions did not affect the AI agent token index rising at all. Sure, a small number of tokens related to x402 picked up some modest buy pressure during the process, but overall the index didn’t truly move. Because the market is pricing everything completely wrong. It’s still trying to predict which agent will win, just like it once tried to predict which Dogecoin mascot is cuter. But real trading depends on owning the “tracks” that every agent must use—regardless of whether that agent lives or dies. And now, those “tracks” are stablecoins.

Cracks in the thesis

Honestly, I’ll also tell you where this argument might be flawed. Otherwise, I’d just be selling another paper about AI agents—only cutting out all the parts that are unfavorable.

The biggest hole is liability. Imagine this: Felix signs a contract with another broker and transfers 1 million USD, and the other party defaults. Who gets sued? Felix is not a legal entity, so you can’t sue it. Nat didn’t authorize the payment—he might not even know about it—and honestly, even if he wanted to, he probably wouldn’t be able to reconstruct what Felix was thinking at the time.

The platform hosting Felix cannot truly provide indemnification for a system whose actions no one fully understands. Insurers have already started to withdraw coverage. Professional liability insurance policies quietly reclassify an agent’s mistakes as “systemic software drift,” which is effectively a denial of payouts.

If you look carefully at current legal provisions, most enterprise AI agreements set the vendor’s liability cap at 12 months of SaaS fees. That means that in the event of a catastrophic incident, anyone can recover at most last year’s subscription costs from the AI vendor. Meanwhile, it’s expected that by 2025, the average cost of data breach incidents in the United States will reach 10.22 million USD per incident. There is a huge gap between the actual risks that might occur and the scope covered by contracts, and right now, nobody has clearly determined who should bear this loss.

Until someone figures out who is actually responsible when an agent makes a mistake, all founderless companies still need to register the name of a person on paper to gain legal protection. But even with this risk, the bigger picture still holds. Companies are gradually dissolving into software, and the blockchain is becoming the routing layer of that software. These two layers will ultimately collapse into stablecoins, because across the entire tech stack, only stablecoins can be independently owned, used, earned, and understood by agents.

Where does the money go?

So, if blockchains become just suppliers, and agent tokens are basically a graveyard, then where is the real benefit?

My honest take is that it’s about the top-level orchestration of reputation and processes. Before any other agent signs a six-figure contract with Felix, someone must verify that Felix is actually solvent. Someone must assess an agent’s default risk the way Moody’s evaluates bonds—only faster, because the agent’s transaction speed is machine-level. Someone must route wages across three chains, without payers and payees needing to know or care which chain handles which part. And in this space, no matter which seed-stage startup ultimately wins, its value will exceed all AI tokens that have ever been issued.

And that’s the part nobody wants to hear. The infrastructure that truly wins in the agent economy looks boring. It’s like pipe installation—there’s no hype, no airdrops, no token issuance gimmicks.

A line from Dragonfly’s Haseeb Qureshi has kept echoing in my head. He said cryptocurrencies were never designed for humans. He’s right—humans are never its target users. All retail users complaining about mnemonic phrases, gas fees, or wallet user experience are right. This product isn’t for them; it was never meant for them. It’s built for the future.

What comes next is software with a wallet, real customers, and real revenue. It has been operating for about two years. By the time you’re reading this, it has already issued invoices somewhere and is spending stablecoins. Meanwhile, the market is debating which blockchain will win AI, which agent token will multiply a hundredfold, and which venture capital strategies will dominate in Q3.

At the same time, a stablecoin traded 18.3 trillion USD worth last year, yet almost nobody in crypto pays attention. This AI token is USDC. Everything else is just hype.

That’s all for today. See you in the next article!