After 138 Million On-Chain Payments, What Exactly Are AI Agents Buying?

marsbitPublished on 2026-10-15Last updated on 2026-10-15

Abstract

"After 138 Million On-Chain Payments: What Are AI Agents Buying?" A little-used HTTP status code is becoming the checkout counter for the machine economy. As of September 2026, over 138 million AI Agent payment transactions have been completed via the x402 protocol on the Coinbase-developed Base chain, with a total payment volume exceeding $50 million across multiple blockchains. The vast majority of these transactions are micro-payments between $0.0001 and $0.10. The x402 protocol revives the long-dormant "402 Payment Required" HTTP status code. It enables a seamless machine-to-machine payment flow: an AI Agent receives a 402 status code with price and wallet details from an API, its wallet automatically signs and settles the transaction on-chain, granting instant access—all without human intervention or traditional accounts. This is made viable by the low transaction costs of stablecoins like USDC. Analysis reveals distinct spending patterns. Model inference (primarily to US-based providers like OpenAI) is the largest category. Data/search APIs, code audit services (heavily used by wallets from China), and financial intelligence are also significant. A geographical divide emerges: US-based agents lean towards "production" services (inference, generation), while Chinese agents favor "analytical" services (audit, intelligence), partly due to the availability of free domestic AI models. User behavior indicates a shift from speculative curiosity to utility. Wallet "stickine...

Written by: Xiaobing

A status code forgotten by the HTTP protocol for thirty years is becoming the checkout counter for the machine economy.

A set of data disclosed in Chainalysis's October 14 report on the North American crypto market: As of September 2026, AI Agent payment transactions completed via the x402 protocol on Coinbase's Base chain have cumulatively exceeded 138 million. Adding Solana and other chains, the total transaction volume of the x402 ecosystem has already surpassed 165 million, with payment amounts exceeding 50 million USD.

Among these transactions, 87% have single transaction amounts between 0.01 cents and 10 cents. The most common payment range is 1 to 10 cents.

Put another way: Approximately 5 on-chain payments by AI Agents are occurring every second, each costing possibly less than a lollipop.

The Resurrection of a Status Code

The story begins with HTTP 402.

In 1997, the HTTP/1.1 standard defined a status code "402 Payment Required", envisioning that future web services could use it to require visitors to pay before accessing content. But at the time, no payment infrastructure could support micro-payments at the webpage level. This status code was labeled "Reserved for future use" and shelved for over two decades.

In May 2025, Coinbase picked it back up. The design of the x402 protocol is very straightforward: when an AI Agent sends a request to an API, if the server requires payment, it returns a 402 status code, along with the price, payment address, and supported tokens. The Agent's wallet automatically signs, settles on-chain, and gains access. The entire process requires no account registration, no API Key, no human approval.

One HTTP request, one USDC transfer, one data exchange. Three steps complete a commercial transaction between machines.

The reason this mechanism can run is that stablecoins have pushed settlement costs low enough. Swiping a $0.03 transaction with Visa, the fee might be more expensive than the transaction itself. But settling with USDC on Base costs less than a cent. When a single transaction can be as small as one-hundredth of a cent, the entire "pay-per-use" business model becomes viable.

What Are AI Agents Buying

Chainalysis's report breaks down the consumption scenarios behind these 138 million transactions, revealing several categories:

Model Inference is the largest single category. AI Agents call large language model APIs, paying per token. One Q&A might cost $0.02, one code generation might cost $0.08. US wallets contributed 88.1% of model inference payments, consistent with the market position of leading US model service providers (OpenAI, Anthropic, Google).

Data and search APIs follow closely behind. Agents pay to obtain real-time data: price feeds, on-chain analysis, news summaries, academic paper retrieval. The geographical distribution of these payments is more even, with US wallets accounting for 23.2%.

Code auditing and security analysis is an unexpected highlight. Wallets from China contributed about 90% of this category's transaction volume, primarily driven by AgentLISA, an automated smart contract auditing service. Agents submit contract code and pay per instance with USDC to obtain audit reports.

Financial intelligence and market data is also growing rapidly. Chinese wallets account for 46.7% in this category, while the share of US wallets is surprisingly low. This might reflect a reality: Chinese AI application developers are using overseas data services to bridge information gaps.

An implicit geographical divide emerges: US Agent consumption leans towards the "production side" (inference, generation), while Chinese Agent consumption leans towards the "analysis side" (auditing, intelligence). Chainalysis attributes the extremely low share (0.1%) of China in the inference category to "AI nationalism": domestic open-source models like DeepSeek, Tongyi Qianwen are freely available, eliminating the need to call overseas commercial APIs via on-chain payments.

From Speculative Tool to Infrastructure: The Shift in Behavioral Data

More noteworthy than the transaction volume is the change in user behavior.

A significant part of x402's early growth came from PING, a meme coin that required x402 payments to mint. This wave of speculation brought in a large influx of wallets, also leading to a subsequent traffic decline. But after the tide receded, the remaining users exhibited completely different behavioral patterns.

Chainalysis tracked wallet "stickiness metrics":

The proportion of wallets that recharge again within 30 days after their first deposit increased from 70.6% in Q3 2025 to 89.3% in Q2 2026. Nearly ninety percent of wallets come back to continue using it within a month after depositing.

Even more astonishing is the change in "time to first spend". At the end of 2025, the median time from wallet creation to the first x402 payment was 73 days. By Q3 2026, this number became 18 minutes.

The proportion of wallets starting to spend within one hour after depositing soared from 4% to 67%.

These three sets of data tell the same story: x402 is transitioning from a "new toy for developers to try" to the "default payment layer in Agent workflows". Wallets are no longer exploration tools driven by human curiosity, but programmatically deployed infrastructure components. Deposit and spend, spend and refill, operating like automatic utility bill payments.

The payment amount structure is also shifting. In early 2025, transactions over $1 accounted for only 49% of total payment value; by early 2026, this proportion rose to 95%. Transactions between $0.10 and $1 plummeted from 46% to 4%. Agents are no longer just making "exploratory small calls"; they are paying for heavier, more expensive services: renting GPUs by the hour, purchasing datasets in bulk, subscribing to continuously updated financial information streams.

The Giants Are All Betting on This Track

The reason x402 deserves serious attention isn't just because on-chain transaction volume is rising, but because of the list betting on it.

In May 2026, AWS released Bedrock AgentCore Payments, enabling AI Agents running on Amazon Cloud to natively support x402 payments. Developers choose a Coinbase or Stripe wallet, deposit USDC, and the Agent can autonomously pay for API calls, MCP server access, and inter-Agent collaboration.

Coinbase and AWS jointly initiated the x402 Foundation. Cloudflare is a co-founding member and built native x402 support into Workers. Google integrated x402 into its own Agent Payments Protocol (AP2). Visa announced support for x402 via the Trusted Agent Protocol (TAP).

OpenAI and AWS jointly published a technical guide for x402 payment flows. The Solana Foundation built a similar Agent payment solution for Google Cloud services.

This list reads less like an "ecosystem partner list for a crypto project" and more like a consensus list of tech industry infrastructure players. When AWS, Google, Visa, and Stripe simultaneously appear on a protocol's supporter list, the discussion is no longer "can cryptocurrency go mainstream", but "what is the payment standard for the machine economy".

Circle's statement in the report is telling: stablecoins are providing the rails for "machine-to-machine settlement", handling transactions so small that credit card networks wouldn't even want to touch them. BlackRock's Will Su pointed out that AI Agents might create a brand-new payment demand: a continuous payment stream that is 24/7, high-frequency, and low-amount, something almost non-existent in human economic activity.

A Still Tiny but Non-Ignorable Number

It's necessary to stay grounded: a total payment amount of $50 million USD, spread over 165 million transactions, averages about $0.30 per transaction. In the context of the crypto market's daily trading volume often reaching tens or hundreds of billions, this is a statistical noise-level figure.

There are about 69,000 active AI Agents on Agentic.Market. Most of these Agents are still in the stage of "can run" but are some distance away from "large-scale commercial deployment".

But there is a noteworthy structure here: x402's growth trajectory closely resembles early DeFi. In 2019, DeFi's TVL was less than $1 billion, and no one paid it any mind. Three years later, the peak locked value exceeded $180 billion. What matters isn't the current absolute number, but the shape of the growth curve and the underlying structural drivers.

The number of AI Agents is growing exponentially, and each Agent needs to autonomously acquire and pay for external services. Traditional payment systems (credit cards, bank transfers, PayPal) are not designed for "machine-autonomous payments"; they require human identity verification, accounts, and minimum transaction amounts. Stablecoins + x402 fill a gap completely uncovered by existing financial infrastructure.

By the time you read this article, thousands of AI Agents might have silently completed another round of micropayments on Base. They don't sleep, eat, or take weekends off. They just follow the instructions of the code, one cent at a time, buying the world's data and computing power on-chain.

This might be the first large-scale use case cryptocurrency has found that doesn't rely on speculation.

Related Questions

QWhat is the HTTP 402 status code and how has it been revived by Coinbase for AI Agent payments?

AThe HTTP 402 'Payment Required' status code was defined in the HTTP/1.1 standard in 1997 but remained unused for decades due to a lack of infrastructure for web-scale micro-payments. Coinbase revived it in May 2025 as part of the x402 protocol. When an AI Agent makes an API request, if payment is required, the server responds with a 402 status code containing the price, payment address, and supported tokens. The Agent's wallet automatically signs, settles the payment on-chain, and gains access, enabling seamless machine-to-machine transactions without human intervention, API keys, or accounts.

QAccording to the Chainalysis report, what are the primary categories of goods and services that AI Agents are purchasing via x402 payments?

AThe report identifies several primary categories: 1. Model Inference: The largest single category, where Agents pay per token for LLM API calls (e.g., for Q&A, code generation). 2. Data & Search APIs: For real-time data like price feeds, on-chain analysis, and news summaries. 3. Code Audit & Security Analysis: Notably driven by Chinese wallets, using services like AgentLISA for automated smart contract audits. 4. Financial Intelligence & Market Data: A fast-growing category. The report notes a geographical divide: US-based AI Agents lean towards 'production' (inference), while Chinese-based Agents focus more on 'analysis' (audit, intelligence).

QWhat key behavioral changes in x402 wallet usage indicate it is evolving from a speculative tool to a core infrastructure component?

AKey behavioral changes include: 1. High retention: The percentage of wallets making a second top-up within 30 days of their first increased from 70.6% in Q3 2025 to 89.3% in Q2 2026. 2. Dramatically faster 'time-to-first-purchase': The median time from wallet creation to the first x402 payment dropped from 73 days at the end of 2025 to just 18 minutes by Q3 2026. 3. Shift in payment size: Transactions over $1 grew from 49% of total payment value in early 2025 to 95% in early 2026, indicating Agents are paying for more substantial services. These metrics show wallets are becoming programmatically deployed infrastructure, used immediately and continuously like a utility bill.

QWhich major technology companies are backing the x402 protocol, and what does their support signify?

AMajor backers include AWS (which integrated x402 into Bedrock AgentCore Payments), Google (integrated into its Agent Payments Protocol), Cloudflare (built-in support for Workers), Visa (support via its Trusted Agent Protocol), and Stripe (as a wallet option). OpenAI and AWS co-published technical guidelines. This list resembles a 'consensus list' of tech infrastructure giants rather than a typical crypto-ecosystem partnership. Their collective support shifts the conversation from 'crypto adoption' to defining the 'payment standard for the machine economy,' signaling strong industry belief in its potential as foundational infrastructure.

QDespite its current small scale, why is the growth of x402 payments considered significant for the future of crypto and AI?

AWhile the total $50 million payment volume is negligible compared to overall crypto markets, its significance lies in the growth trajectory and structural drivers. Similar to early DeFi, the key is the shape of the growth curve. The number of AI Agents is growing exponentially, each requiring autonomous payment for external services. Traditional payment systems (credit cards, bank transfers) are not designed for machine-initiated, high-frequency, micro-value payments. Stablecoins combined with x402 fill this completely unaddressed gap in the financial infrastructure. It represents crypto's first potential large-scale use case independent of financial speculation, enabling a new, persistent, machine-driven payment flow for data and compute power.

Related Reads

Hackers inject credit card, crypto-stealing malware into 178 major websites

Cyber attackers planted cryptocurrency and credit card-stealing malware on websites associated with Ledger, Coinpass and scores of other major businesses in an attack exploiting website optimization platform Optimizely on Monday. Dubbed "Optimjack" by researcher Amirul Iman, the campaign delivered malicious code through scripts hosted by Optimizely. The malware could replace legitimate cryptocurrency wallet addresses with attacker-controlled alternatives when visitors copied addresses, and it also captured credit card details entered into website forms, according to a separate analysis by cybersecurity firm StepSecurity. The campaign affected website scripts belonging to 178 organizations, including hardware wallet Ledger, OANDA / Coinpass and trading platforms IG, Interactive Investor and TMS Brokers, according to Report URI's investigation. Other affected brands included IKEA, Salesforce, Zoom, the BBC and CNN, websites that collectively attract millions of visitors a day. BBC was among the first websites observed to serve the Optimizely injection at around 4:21 am UTC on Monday. By 9:52 pm UTC, most of the malicious code had been removed, with Ledger's snippet cleaned at about 8:41 pm UTC. Cointelegraph reached out to Ledger, Coinpass, and a number of other major businesses reported to be impacted by the malicious injection. This is a developing story, more information will be added as it becomes available.

cointelegraph6m ago

Hackers inject credit card, crypto-stealing malware into 178 major websites

cointelegraph6m ago

Just now, Chen Danqi officially announced joining Noeri part-time, a new company betting on AI self-evolution

Princeton computer science associate professor Danqi Chen has revealed she joined AI startup Noeri on a part-time basis two months ago, expressing enthusiasm for its focus on applying AI to scientific discovery so scientists can tackle their hardest research problems. Noeri was co-founded by Stanford computer science professors Yejin Choi and Carlos Guestrin, who serve as co-CEOs. The company focuses on developing AI systems that can continuously learn and self-improve, helping enterprises build customized AI from their own data and business experience. It has raised $110 million led by Menlo Ventures, with participation from Madrona Venture Group, Kindred Ventures, Wing Venture Capital and NEA, plus strategic investors including NVIDIA, Cisco Investments and Snowflake. Choi is a leading researcher in commonsense reasoning, NLP, LLMs and AI alignment, formerly at the University of Washington and Ai2 and a senior research director at NVIDIA. Guestrin is a veteran ML scholar and entrepreneur who co-created XGBoost and co-founded Turi (acquired by Apple in 2016), later serving as Apple's senior director of ML and AI. Chen, an associate professor at Princeton, co-leads its NLP group, serves as deputy director of the Princeton Language and Intelligence initiative, and is a co-author of RoBERTa. Noeri's core idea is "Plural Intelligence" — the belief that future AI should comprise many systems of different scales and specialties, each organization building and continuously improving its own AI, rather than depending on a single corporate superintelligence. Its product acts like a persistent AI engineer: it studies a company's data and workflows and optimizes instructions and skills, tool configuration and orchestration, model selection, and even underlying model parameters. Every change comes with an explanation and requires human review, akin to code review. The system supports both closed and open-weight models and learns from each task's interactions and corrections. Chen said she looks forward to working with Choi, a longtime role model and friend, and is especially excited about plural intelligence in scientific discovery, where specialized AI agents could continuously learn around a specific research question, remember past attempts, and collaborate on hypotheses and validation.

marsbit40m ago

Just now, Chen Danqi officially announced joining Noeri part-time, a new company betting on AI self-evolution

marsbit40m ago

40 Kilometers: Shenzhen Fits an Entire Chip Supply Chain Into a One-Hour Drive

Shenzhen is compressing an entire chip supply chain into a one-hour drive, roughly 40 kilometers. Design firms cluster within 10 km of the Futian APEC venue, while fabs sit 20–40 km away in Pingshan, Pinghu, and Yanluo. The article argues this is not just local ambition but a response to a hard rule: the more mature a process, the easier design and manufacturing can be separated; the more immature, the closer they must be. TSMC's PDK model let California design and Taiwan manufacturing run for 30 years, but only because mature processes could be codified. Shenzhen's outer ring, however, hosts immature work—Huawei's "Tau Law" and stacked "logic folding" chips—where tacit knowledge lives in engineers' heads and demands face-to-face iteration. With回路B (design-process feedback) and回路C (fab-equipment feedback) cut by export controls after 2020, both loops had to return home. Shenzhen lacks an integrated giant like Samsung or Intel, so it uses city governance: a state platform invests across fabs and equipment makers, plus tape-out subsidies up to 10 million yuan. The geography forms a slope—Nanshan offices, Bantian R&D, Guanlan/Pinghu pilot lines, Pingshan fabs—connected by Metro Line 14 (40 minutes Futian to Pingshan). The lesson for planners: zone land by process maturity, not by industry category, and keep the most immature, iteration-hungry steps within an hour's drive rather than chasing a full chain.

marsbit40m ago

40 Kilometers: Shenzhen Fits an Entire Chip Supply Chain Into a One-Hour Drive

marsbit40m ago

Trading

Spot
活动图片