TRM Labs Report: USDC Carried 99.6% of $52.7M in x402 Payments, AI Agents a Fraction

Only a Sliver of x402 Volume Looks Agentic

Most payment volume flowing through Coinbase's x402 protocol does not come from AI agents, according to new blockchain-intelligence research that dissects who is actually paying on the machine-to-machine rail. The study, published Wednesday, examined roughly $52.7 million in value across 198.9 million settlements processed by known x402 facilitators on Base, an Ethereum layer-2 network, plus Solana and Polygon, since May 2025. After stripping out self-payments and other anomalous flows, researchers put likely genuine commerce at $25.62 million — and estimated that only 0.6% to 7.5% of that remaining value appeared to originate from AI agents.

The distinction matters because an ordinary script can complete x402's payment sequence without any agent. The protocol, launched by Coinbase in 2025, embeds payment directly into web requests: the buyer receives a price, signs a payment authorization, and a facilitator verifies it, submits the blockchain transaction and covers the network fee. Scheduled jobs, load tests and self-dealing can leave identical on-chain records, which makes raw protocol totals a poor proxy for agentic commerce. The models treat a “true agent” as one that explores across multiple services, while an address repeating the same price reads as a script hitting one endpoint — a choice the researchers concede may understate activity, since many agents today are single-purpose and would fail this test.

To isolate real demand, the analysis removed self-payments, bulk flows from one or two payers, and sellers with fewer than 10 buyers; a stricter screen required facilitator-broadcast payments with varying amounts averaging under $1 sustained across months, plus public agent registration or payments to multiple sellers. Activity in late 2025 included apparent meme-token minting and payments to one AI-analysis service, before volume concentrated in a single payment contract in early 2026 and AI-service payments returned through an agent-payment router around midyear. One figure stands out: USDC settled 99.6% of all value — $52.47 million — giving the rail a stablecoin-first tokenomics profile. Industry momentum continues regardless: Binance's August Agent OS launch bundled an x402 payment layer, Coinbase's Base ran a $1 million accelerator for agent and payments startups, and Amazon announced AgentCore Payments with Coinbase and Stripe in May.

AI Capex Strains Alibaba and Amazon Cash Flow

While the agent-payment rail waits for genuine demand, the corporate AI buildout is showing the cost of front-running it. Alibaba's June-quarter results, disclosed in its August 20 earnings release, show AI cloud and computing services revenue up 45% year-on-year to 48.44 billion yuan, driven by wider adoption of AI-related products and stronger public cloud sales. But capital expenditure jumped 75% to 67.68 billion yuan in the same quarter, which the company tied to expanded AI infrastructure investment, added CPU computing capacity and rising chip component prices. Free cash flow swung deeper into deficit: a 44.67 billion yuan outflow for the quarter, versus an 18.82 billion yuan outflow a year earlier. Profitability is uneven inside the AI business itself — adjusted EBITA for AI cloud and computing rose 133% to 5.63 billion yuan, while AI application operations posted a 13.86 billion yuan adjusted EBITA loss.

Amazon's numbers follow the same pattern. AWS revenue for the second quarter of 2026 reached $42.2 billion, up 37%, with operating income of $16.6 billion, per the company's July 30 earnings announcement. Yet trailing-twelve-month free cash flow through June 30 turned to a $7.6 billion outflow from an $18.2 billion inflow the year before. Cash capital expenditure hit $53.1 billion in the quarter and $96.3 billion for the first half, and the company's 10-Q guides full-year 2026 capex to roughly $220 billion, most of it technical infrastructure supporting AWS growth. AWS's AI business has passed a $25 billion annualized revenue run-rate, though the filing flags memory chip prices, energy costs and customer demand as swing factors ahead. Unlike fast-moving futures trading, infrastructure payback is measured in quarters: outlays precede revenue, so the gap between capex and operating cash flow is the metric that will decide whether this spending compounds returns or erodes them. Readers tracking the market in real time can follow live spot and futures prices on Binance.

Agentic Commerce Still Needs Agentic Compliance

Read together, the two stories describe an AI economy whose spending is running ahead of its proof. The research document behind the x402 analysis states the rail itself works but that on-chain agent registries remain voluntary and mostly unused, calling for accurate registration, counterparty reputation signals agents can check themselves, and monitoring built for volume rather than value — in its words, “agentic commerce will need agentic compliance.” Meanwhile, the official filings from Alibaba and Amazon confirm AI demand is converting into cloud revenue but not yet into recovered capital. The gap between narrative and measurable flow — on-chain and in cash flow statements alike — is where this FOMO-tinged investment cycle will ultimately be judged.