A machine can pay. That does not tell us why.
An AI agent can discover an API, receive an HTTP payment requirement, authorize a tiny payment, retrieve the resource and continue working without a person navigating checkout. That is the core promise behind x402 and similar machine-payment systems. But the existence of a machine-readable payment rail creates a measurement problem: an on-chain settlement tells us that software completed a payment flow. It does not tell us whether an AI agent independently chose the product, whether a conventional script triggered the payment, whether a developer was testing an integration, or whether related wallets were moving money between themselves.
The headline x402 numbers measure a rail, not autonomous shopping
Machine Observer observed approximately $41.93 million in cumulative settled x402 value and 192.04 million cumulative x402 transactions on October 14, 2026 across the facilitators and chains represented by its underlying source. Those figures are useful because they establish that a substantial payment rail exists. They should not be read as $41.93 million of AI-agent shopping. The protocol can be used by agents, ordinary automation, test harnesses and human-directed software, and the blockchain record alone cannot reliably distinguish those cases.
Independent analysis suggests the agentic share is much smaller
TRM Labs approached the attribution problem by screening roughly $52.7 million across 198.9 million x402 settlement transactions observed since May 2025. It removed self-payments and other concentrated or thin-traffic flows, leaving $25.62 million it classified as commerce. TRM then applied permissive and strict behavioral tests for likely agent activity. Under those models, 0.6% to 7.5% of the screened commerce value appeared agentic. The range depends on modeling choices and may omit single-purpose agents, but it illustrates why raw x402 totals cannot serve as a direct proxy for autonomous economic demand.
What is for sale is easier to see
The supply side of the market is more legible. Machine-payable directories and Machine Observer’s source catalog include web data, blockchain intelligence, model inference, compute, communications, verification, research utilities, job data and property information. Property services alone now include U.S. listing and valuation endpoints, Singapore housing statistics, Asia-Pacific land-price data and U.K. property or licensing records. This variety suggests that early machine commerce is developing around narrow digital tasks that can be delivered instantly through an API.
The three-tenths-of-a-cent economy
Machine Observer also samples what sellers ask for machine-payable services. On October 14, 21 of 26 sampled endpoints returned an observable advertised price. The median ask was $0.003, with a $0.001 minimum and a maximum around $0.02. No payment was sent to obtain those quotes, so they are asking prices rather than settlement or clearing prices. Even so, the structure is revealing. A machine does not need to commit to a monthly software subscription if it needs one narrowly defined fact, one inference, one verification or one lookup. The economic unit can become the task itself.
Information is a natural machine commodity
Much of the visible supply consists of information or computation. That makes economic sense. Physical purchases introduce shipping, returns, taxes, identity checks and often additional human authorization. Information can be discovered, purchased and consumed within the same software workflow. An agent researching an investment, evaluating a property, checking a wallet, comparing suppliers or preparing a report can potentially purchase the missing datum and immediately incorporate it into its next decision. Early machine commerce may therefore look less like robots filling shopping carts and more like software purchasing tiny pieces of intelligence from other software.
Supply is ahead of proven demand
Machine Observer’s stored service-quality observation shows 3,496 callable services among 4,267 observed services, or 81.9%. That tells us that a meaningful amount of the technical supply is usable when probed. Payment activity also exists at substantial scale. What remains harder to establish is durable, independent, economically motivated demand from autonomous buyers. Registrations, callable endpoints, advertised prices and settlements describe different parts of the market. None of them alone tells us how many external agents repeatedly choose a service because the result is worth paying for.
So what do AI agents actually buy?
The defensible answer is narrower than the marketing story. Machines demonstrably pay for digital services, and the services built for them increasingly cluster around data, computation, AI capabilities, communications, verification and specialized information. Some of those purchases almost certainly originate from autonomous or semi-autonomous agents. Current evidence does not justify treating every x402 wallet as an AI agent or every x402 settlement as an autonomous buying decision. The important development is that software can now discover a service, evaluate a price, purchase precisely the resource it needs and continue working with very little human friction.
What Machine Observer will watch next
The next stage of measurement is demand quality. Which machine-payable services attract distinct external buyers? Which buyers return? Which categories sustain revenue rather than test activity? How concentrated is spending among a small number of addresses or sellers? Those questions move the machine economy from infrastructure counts toward something closer to market intelligence. For now, the strongest signal is simple: the rails work, the catalog of things machines can buy is expanding, and the difficult work is identifying which transactions reflect genuine autonomous demand. Machine Observer will keep watching the receipts.
What the measured record says
Machine Observer observed $41.93M in cumulative settled x402 value and 192.04M cumulative x402 transactions on October 14, 2026. These are rail-level measurements across the tracked source coverage, not a measurement of autonomous-agent purchasing.
In Machine Observer’s October 14 x402 asking-price sample, 21 of 26 endpoints advertised a price; the observed median ask was $0.003, with a $0.001 minimum and approximately $0.02 maximum. These were unpaid payment challenges, not transacted prices.
TRM Labs reported about $52.7M across 198.9M x402 settlement transactions since May 2025, screened $25.62M as commerce, and estimated that 0.6% to 7.5% of that screened value appeared agentic under its permissive and strict behavioral tests.
Graded reported 3,906 discoverable x402 services, of which 3,078 were callable through its gateway at its latest published probe.
Use the live data
This guide explains the concepts; the underlying index continues to change. Open the related live index and methodology →
Interpretation boundary
Counts and probes are reported with their own populations and timestamps. Vendor declarations remain distinct from direct observations; no individual measurement is presented as a total market size or universal trust verdict.
Read the Machine Observer methodology →
For machine readers
sector: machine-commerce
last_updated: 2026-10-15
claim_ids: MO-2026-0009, MO-2026-0010, MO-2026-0011, MO-2026-0002
structured_representation: /api/v1/stories?slug=what-do-ai-agents-actually-buy
claim_endpoint: /api/v1/claims?id=<CLAIM_ID>
historical_record: /api/v1/history/observations