All in AI、All in Web4.0 原来就是开中转站卖Token

Author: Riches

0x0 Preface

In early May 2026, three things happened simultaneously: Sun Yuchen posted about B.AI on X, the Trump family announced WorldClaw at Consensus, and Fu Sheng launched EasyRouter.io shouting “20% off for the entire venue.” These three events together, unlikely to be believed if told, actually happened.

A banana seller, a meme coin seller, and a browser seller, all started AI relay stations in the same week.

It’s a bit like three people from different provinces going to the same supermarket on the same day to buy the same bottled water—you can’t help but think, there might be something special about this water.

0x1 What is an API relay station, in plain language

First, let’s clarify this.

Companies like OpenAI, Anthropic, Google package their large models into APIs, charge by token, and sell to developers. One token is roughly half an English word or a third of a Chinese character. When you ask the model a question and get an answer back, it consumes thousands of tokens, costing you less than a cent or a few cents.

The problem is, these official APIs have barriers: require credit cards, overseas phone numbers, access to international networks, and rate limits. Domestic users find it troublesome, and large enterprises want discounts for heavy usage.

So, relay stations emerged. The logic is simple— I go to the official source in bulk for API quotas, you buy from me, and I earn the spread and service fees. In industry terms, this is called “Token import,” in simpler terms, AI resellers.

But this business itself isn’t new, nor technically complex. What’s new is that it’s dressed in a new coat—called Web4.0, AI Agent economy, blockchain infrastructure. With this coat, it becomes easy to sell tokens—not API tokens, but cryptocurrency tokens.

It’s a very old business, dressed in very new clothes.

0x2 Sun Yuchen’s B.AI: Every call is an on-chain transaction

Let’s start with Sun Yuchen, because he was the earliest and most direct.

On May 1, 2026, Sun Yuchen posted on X:

The strongest AI relay station in history, B.AI, is here. One API Key = Claude + GPT + Gemini + all domestic large models. Blockchain login, anonymous payment, tamper-proof, lowest prices across the network.

This statement is very precise. It integrates mainstream models like GPT-5.5, Claude Opus 4.7, claiming zero difference from official prices, with reportedly over 1 million users. Just from the feature description, it looks like a normal API aggregation service.

But the payment method reveals everything—no USD, no RMB, only USDT, settled on-chain.

This detail is crucial.

TRON chain carries the world’s largest USDT circulation, and Sun Yuchen controls this chain. Every time a user calls B.AI, a transaction occurs on the TRON blockchain, with fees, on-chain traffic, and ecosystem data all flowing into this system. In his own packaging language, this is “Web3 infrastructure,” a “financial identity enabling autonomous payment for AI Agents.”

In plain language: every click on my relay station feeds my public chain.

He packages this narrative as Web4.0. Sun Yuchen’s definition of Web4.0 is: readable, writable, ownable, plus autonomous AI execution. It sounds lofty, but the actual product is: an API aggregation platform that enforces settlement with on-chain stablecoins, turning each inference into a record on TRON.

The real game plan is clear: API relay profit margin is the first layer, on-chain TRON transactions earn ecosystem fees as the second layer, and the “All in Web4.0” narrative boosts TRX and Tron ecosystem valuation as the third layer.

One fish, three ways—relay station is just the head.

0x3 Trump family’s WorldClaw: Bringing Mar-a-Lago dinner into the pricing

The Trump family’s move is more direct, even spectacular.

On May 5, 2026, WLFI (World Liberty Financial, a crypto project co-founded by the Trump family) announced WorldClaw at Consensus 2026, with its core product called WorldRouter—a platform aggregating over 300 models, claiming prices 30% lower than official, with one account accessing all models, no KYC, no overseas credit card needed.

It sounds similar to other relay stations. But once the pricing table appeared, the tone changed dramatically.

Four packages: the lowest at $9.9, standard at $99, premium at $999, flagship at $9,999. The flagship includes 1 million AI credits, and a “confidential brand, undisclosed specifications” hardware device—officially noting “images for reference only, actual product may differ,” with delivery expected in Q3 2026.

Additionally, a random selection of buyers will be invited to a private dinner at Mar-a-Lago with Donald Trump Jr.

This dinner lottery is the most honest part of the product. It directly tells you that this isn’t just API sales, but faith-based consumption centered around Trump family IP.

Payment method is only USD1. This is WLFI’s stablecoin issued in March 2025, pegged to the US dollar, backed by US Treasuries and dollar deposits, operating on Ethereum, BNB Chain, and Solana. With WorldClaw, you must buy USD1 first, then use USD1 to buy credits. If you don’t want to spend money, you can lock WLFI tokens to exchange for credits—25,000 WLFI for Pro package, 2.5 million for flagship.

The logic chain is: use API services to attract traffic → enforce settlement with USD1, increasing stablecoin usage → lock WLFI to reduce circulating supply, boosting token price → harvest via token economy. API price differences are just entry fees; the token ecosystem is the main dish.

Zach Witkoff on stage said:

“In the future, trillions of AI agents will be running, paying each other, paying enterprises, but traditional banking rails are too slow for machines, KYC is impossible for machines—they need USD1.”

There’s nothing wrong with this statement. But he uses it to sell the $9,999 package and the Mar-a-Lago dinner lottery.

0x4 Who else is on this Web3 track

Sun Yuchen and the Trump family are not the only examples. AI relay stations with tokens are already a crowded track.

Gate.io launched GateRouter, aggregating over 25 models, supporting AI agents to pay with USDT autonomously—very similar to B.AI—relay entry, on-chain settlement, and fee collection.

Heurist positions itself as a decentralized AI compute cloud, aggregating idle GPU resources worldwide, providing serverless inference APIs. It has its own ZK Layer 2 (Heurist Chain), with HEU tokens, supporting x402 and ERC-8004 payment protocols, with model and agent interactions settled on its chain. Outwardly a DePIN compute network, internally it’s still API relay plus token incentives.

Virtuals Protocol is called “AI agent’s Stripe” in the community, enabling developers to tokenize AI agents, issue and trade on-chain—your agent is a token, holding it means owning a stake. This logic advances further: API relay is just one capability of the agent, the real sale is the token’s appreciation potential.

Common features of these projects are clear: aggregate model APIs, bind on-chain payments, and ultimately, a token waiting for you.

From B.AI to WorldClaw to GateRouter to Heurist, packaging varies, but the token thread runs through all.

The problem in this track isn’t the wrong direction, but the confusion over who is building infrastructure and who is using infrastructure narratives to sell tokens. Like in the 2000s internet boom, IDC companies building fiber for broadband and portal sites raising money with “internet concept” were both called internet companies, but fundamentally different.

0x5 The three ways relay stations make money and why they wear the “Web4.0” hat

Relay stations make money in three layers, each independent, together forming a complete business.

First layer, information arbitrage. Overseas APIs face regional restrictions, price differences, access barriers. Relay stations exploit these frictions, buying wholesale at lower prices, selling at official or slightly lower retail prices, pocketing the spread. They claim “zero margin,” but in reality, there are hidden markups—industry insiders say nearly half of nine suppliers are secretly inflating charges—what you pay for might be inference from Claude, but actually run on a bean bun.

Second layer, hidden taxes in payment. Abandon credit cards, introduce on-chain stablecoins or their own tokens, making each call generate traffic within their ecosystem. They earn gas fees, transaction data, and circulation revenue.

Third layer, token economy siphoning. Issue or bind their own tokens, reduce market circulation via lock-up mechanisms, push up secondary market prices, and cash out at high token prices. WLFI’s net profit sees 75% flowing directly into entities linked to the Trump family.

Stacked together, these form the true business model of “All in AI, All in Web4.0.”

Why wear the “Web4.0” hat?

Because without this hat, tokens won’t sell at high prices.

An API relay station is just a relay, with limited profit margins, subject to regulation, and zero switching costs for users. But if it’s “the infrastructure of the next-generation internet,” “the financial backbone of AI agent economy,” or “the payment protocol layer of Web4.0,” the valuation logic changes. The narrative can attract more investors, tokens can be sold at higher prices, and user migration costs become psychologically higher.

The concept’s value lies in making you feel that missing this train means missing the entire era.

0x6 What true infrastructure looks like

Let’s talk about the real shovels, not just the sickle.

AI agent economy really needs underlying infrastructure, but there are two types: payment protocol layer and compute layer. Both are being seriously developed, just less flashy than selling tokens.

The representative of the payment protocol layer is x402. Coinbase launched an open protocol based on HTTP status code 402 in May 2025—this code has been dormant in HTTP for nearly thirty years, originally meaning “Payment Required,” never activated due to lack of programmable real-time payment mechanisms. The explosion of blockchain stablecoins and AI agents finally gave it a use case. Very simple logic: client sends request, server responds with 402, including amount and payment address, client completes on-chain payment, provides proof, server grants access—no accounts, no KYC needed. Cloudflare, Google, Stripe, AWS, Visa, Mastercard are adopting it, transferred to Linux Foundation maintenance, and by late 2025, it handled nearly a million transactions per week. It’s open source, free, with no tokens to lock.

On x402, two notable application layer projects have emerged.

Pay.sh, launched on May 5, 2026, by Solana Foundation and Google Cloud, is an AI agent payment gateway built directly on x402 and MPP protocols, with an open-source registry. It integrates Google Cloud APIs like Gemini, BigQuery, Vertex AI, and over 50 community API providers, including Dune Analytics, Helius, The Graph. AI agents use Solana wallets for identity, pay stablecoins per request, no accounts or API keys needed, settled on Solana in seconds, with providers receiving fiat. In short: Google Cloud’s enterprise APIs become a storefront for agents to pay directly with cards.

Kite AI (KITE), a Layer 1 blockchain designed for agent economy, launched mainnet in late April 2026, running on Avalanche. It issues each AI agent a “passport”—Kite Agent Passport—with verifiable encrypted identity and programmable spending rights, solving compliance issues of autonomous agents spending without human oversight. x402 is its native payment primitive. Backed by PayPal Ventures and General Catalyst with a $33 million Series A, with Coinbase Ventures participating, and pilots with PayPal and Shopify. It’s not an API relay; it’s the infrastructure layer issuing IDs for machine economy.

On the compute layer, three projects stand out.

Bittensor (TAO), the oldest, often called “Bitcoin of AI”—not without reason—total supply 21 million, halving mechanism, driven by “useful computation mining”: anyone contributing models, compute, or data to the network earns TAO rewards. The network is divided into hundreds of subnets, each focused on a specific AI task, capable of calling each other to form service chains. It’s not a relay station; it’s an experiment to use token incentives to assemble a decentralized AI brain from global compute and models. As of May 2026, TAO returned to the market with daily volume over $250k. Whether it works or not, it’s not just API resale.

Phala Network (PHA) focuses on privacy computing. Its core tech is TEE (Trusted Execution Environment)—your data enters Phala nodes, runs in hardware-isolated encrypted zones, even the node operators can’t see the data, and results are verifiable on-chain. In November 2025, it migrated from Polkadot parachains to Ethereum Layer 2, handling over a billion tokens daily in AI inference. Sensitive data like medical, financial, legal info is involved, and no one wants to send raw data to a centralized relay. Phala’s direction: you can use AI without exposing your data.

Fluence (FLT) is a decentralized compute marketplace. It aggregates top-tier global data center GPU resources into a network, allowing developers to buy compute on demand at about 80% lower than AWS or Azure. The demand for FLT tokens directly correlates with platform compute usage—actual GPU hours run determine demand, not just narrative. By late 2025, platform revenue exceeded $1 million, running on 32 regions, 71 data centers, with over 1,400 GPUs. This isn’t API resale; it’s real infrastructure building a decentralized compute foundation.

These projects differ clearly from B.AI and WorldClaw: they build foundational infrastructure for others to develop on, not just middle layers reselling others’ services. x402 itself doesn’t issue tokens; OpenRouter runs a platform with hundreds of trillions of calls without tokens. TAO, PHA, FLT have tokens, but these tokens represent real compute contribution and network security staking, not just “lock 2.5 million to upgrade to a max package.”

0x7 Final words: practical advice

AI agent economy is a real direction, machine-to-machine micro-payments are real, and the “next-generation internet payment layer” narrative is real.

The only question: who is truly building this, and who is just using the narrative to sell tokens?

The simple test: remove tokens from the product—can it still survive?

Remove USDT settlement from B.AI, it’s still an API relay, but loses the ecosystem value Sun Yuchen tied to. Remove USD1 and WLFI lock-up from WorldClaw, it’s just a normal API aggregator, and the $9,999 package has no reason to exist.

Remove FLT from Fluence, and the 1,400 GPUs still run, and the $4 million saved on cloud costs is real.

Remove TAO from Bittensor, models in subnets still infer, competition continues. Remove PHA, TEE privacy computing remains, and the fact that billions of tokens are processed daily won’t disappear. x402 originally had no tokens; Cloudflare and Google still adopted it.

This is how to distinguish infrastructure from a money-making scheme: if, after removing tokens, it still exists, it’s infrastructure.

It’s not that having tokens is bad—on-chain stablecoins and decentralized compute incentives are real needs. The problem is, payment tools and token speculation are two different things. The former is a pipeline; the latter is a water seller installing a toll gate on the pipeline, also selling pipeline shares.

When someone sells you API services, stablecoins, and lock-up tokens simultaneously, one of those is your principal.

Guess which one?