diff --git a/papers/arxiv-x402/x402.tex b/papers/arxiv-x402/x402.tex index 2186a372d4..47bbd29626 100644 --- a/papers/arxiv-x402/x402.tex +++ b/papers/arxiv-x402/x402.tex @@ -117,9 +117,14 @@ never pays. Across the retained request logs, no agent ever asked---not for the paid routes and not for the free machine index either---and the receiving address has never taken a genuine payment. We conclude that the binding constraints on machine commerce are substitutability and discovery, not -payment infrastructure, and we evaluate six monetization strategies against -those constraints. The strongest candidate sells something an agent cannot -compute for itself at any price: a name. +payment infrastructure. A census of the surrounding economy sharpens that +conclusion: x402 volume is down roughly 96\% from a \$28.8M peak, the top ten +sellers hold 72.5\% of what remains, and the \emph{discoverable} service +layer---what a new seller can actually compete for---is about \$9{,}850 per +month. Six strategies are evaluated against those constraints. The strongest +sells what an agent cannot compute for itself at any price, a name; we also +report that three operators already sell precisely that and have earned, +between them, under \$60. \end{quote} \vspace{0.5em} }] @@ -336,46 +341,136 @@ wrong. \section{Market Context} \label{sec:market} +\subsection{A note on sources} + +Figures in this section come from the x402scan indexer's open tRPC endpoint, +corroborated independently by x402station, which agrees with it on +service-layer call counts to within two calls. Three widely-cited sources were +examined and rejected: one aggregator's last entry predates 2026 by months, +another's headline metrics are materially wrong, and---the trap worth +naming---the protocol's own homepage displays a \emph{frozen} counter. That +snapshot, roughly \$24M across 72M transactions, was captured in June 2026 and +has been reprinted since as though current, including by trade press. An +earlier draft of this paper repeated a similarly stale figure. Anyone +sizing this market should pull it from an indexer's API and check that the +number moves. + \subsection{Aggregate} -Reported x402 volume peaked near \$5.15M/month in November 2025 and fell -roughly 77\% to \$1.1--1.2M/month across about 3.7M transactions by -mid-2026---an average near \$0.30 per transaction, indicating high-frequency, -low-ticket usage~\citep{coindesk2026}. A meaningful share of the 2025 peak was -not utility: a pay-to-mint memecoin experiment alone processed over 150{,}000 -transactions in its first month by requiring a 1 USDC payment to -mint~\citep{chainalysis2026}. Analyses of that period attribute much of the -growth to token farming rather than services being bought. +x402 volume peaked in November 2025 at \textbf{\$28.8M} across 68.6M +transactions and has fallen roughly \textbf{96\%} since +(Table~\ref{tab:volume}). August 2026 records \$1.18M across 14.9M +transactions. + +The shape of that decline matters more than its depth. Transaction \emph{count} +has recovered to 12--15M/month, close to the December 2025 level, while the +average payment collapsed from \$0.42 to \$0.079. Activity returned; revenue +did not. A recovery visible only in transaction counts is farm-shaped, not +demand-shaped. + +\begin{table}[t] +\centering +\footnotesize +\begin{tabularx}{\columnwidth}{@{}lrrr@{}} +\toprule +\textbf{Month} & \textbf{Txns} & \textbf{USDC} & \textbf{Avg} \\ +\midrule +2025-10 & 4.1M & \$4.01M & \$0.97 \\ +\textbf{2025-11} & 68.6M & \textbf{\$28.79M} & \$0.42 \\ +2025-12 & 65.0M & \$8.41M & \$0.13 \\ +2026-03 & 2.9M & \$2.42M & \$0.83 \\ +2026-06 & 13.4M & \$1.00M & \$0.075 \\ +2026-07 & 12.7M & \$0.73M & \$0.058 \\ +2026-08$^{*}$ & 14.9M & \$1.18M & \$0.079 \\ +\bottomrule +\end{tabularx} +\caption{Monthly x402 volume. $^{*}$August 1--27, 2026. Counts recover while +average payment size collapses.} +\label{tab:volume} +\end{table} + +\subsection{How much of it is real} + +Several structural signals indicate that a large share of this volume is not +commerce. The second-largest all-time buyer address is also the largest +all-time seller, at \$11.5M in each direction. A single counterparty pair +accounts for 26.5M transactions. Of 5{,}480 registered origins only 3{,}832 +ever transacted, and 17 of the 40 highest all-time earners have recorded no +transaction since 2025. The 2025 peak also carried outright non-utility: a +pay-to-mint memecoin required 1 USDC per mint and processed over 150{,}000 +transactions in its first month~\citep{chainalysis2026}. + +The consequential figure for a prospective seller is therefore not total +volume but the \emph{discoverable service layer}: endpoints an agent could +find and buy from. That layer is 290{,}806 calls---about \textbf{1.7\%} of all +transactions---worth roughly \textbf{\$9{,}850 per month}. Everything above it +is inference routing, stablecoin conversion, and self-dealing. \$9{,}850 per +month, not \$1.18M, is the market a new listed seller is entering. \subsection{Which turnstiles earn} -Category-level analysis of settled receipts consistently ranks four kinds of -seller, and the ranking is instructive because none of them sell an archive: +Earnings are severely concentrated. The top ten sellers hold \textbf{72.5\%} +of all volume. Against 22{,}556 sellers active in a 30-day window sharing +\$1.31M, the mean is about \$58/month and the median is certainly under a +dollar. Table~\ref{tab:sellers} names the leaders. -\begin{enumerate} -\item \textbf{Data and search APIs} --- the highest call volume, serving -research agents making many cheap requests. -\item \textbf{Inference and model routing} --- the largest payment sizes, -fewer calls. -\item \textbf{Web automation} --- browser-as-a-service and scraping; steady -volume, minimal unit price. -\item \textbf{Specialized analytics} --- chain explorers and market data -feeding trading agents. -\end{enumerate} +\begin{table}[t] +\centering +\footnotesize +\begin{tabularx}{\columnwidth}{@{}Xr@{}} +\toprule +\textbf{Seller (30-day)} & \textbf{USDC} \\ +\midrule +\code{laso.finance} & 210{,}431 \\ +\code{blockrun.ai} & 176{,}185 \\ +\code{mcp.x402.boats} & 173{,}312 \\ +\code{api.clusterprotocol.ai} & 107{,}234 \\ +\code{sol.blockrun.ai} & 78{,}161 \\ +\code{x402.dtelecom.org} & 29{,}780 \\ +\bottomrule +\end{tabularx} +\caption{Leading x402 sellers by 30-day volume. All sell inference routing, +model access, compute, or bandwidth. None sells an archive.} +\label{tab:sellers} +\end{table} -Every category sells \emph{live compute or live data}: something that did not -exist before the request and cannot be fetched from a static file. The common -factor is not subject matter but non-substitutability at the moment of -purchase. An archive is the opposite: static by definition, and cheapest to +Every leading category sells \emph{live compute or live data}: something that +did not exist before the request and cannot be fetched from a static file. The +common factor is not subject matter but non-substitutability at the moment of +purchase. An archive is the opposite---static by definition, and cheapest to serve by publishing once. -\subsection{The absent category} - -We could not identify an established x402 seller of identity, namespaces, -handles, or agent-to-agent messaging access. That absence is reported as a -finding rather than an oversight, and it cuts both ways: it may indicate an -unserved niche, or it may indicate that nobody has found buyers for one. -Section~\ref{sec:strategy} treats it as the former while noting the risk. +Prices cluster tightly. The median listed call is \$0.02, with modes at +\$0.001, \$0.005, \$0.01, and \$0.10; the effective clearing price is +\textbf{\$0.01} and the floor is \$0.001. The \$5.00 charged by this archive's +\code{/bulk} endpoint sat two to three orders of magnitude above the market +without a corresponding claim to uniqueness. + +\subsection{The identity category is occupied, and it is not paying} + +An earlier draft of this report stated that no x402 seller of identity, +namespaces, or agent messaging could be found, and treated the gap as an +unserved niche. That was wrong, and the correction matters because +\S\ref{sec:strategy} rests on it. Such sellers exist: + +\begin{itemize} +\item \textbf{Palmyr} sells social handles at \$0.005 and domains at +\$10--\$100. Earnings: roughly \textbf{\$53} over ten weeks. +\item \textbf{Agent Exchange Lounge} sells admission to an agent social +network at \$0.05--\$0.09. Earnings: \textbf{\$4.22} from 30 payers. +\item \textbf{shizu.me} runs an agent mailbox. Earnings: \textbf{\$1.79} from +37 payers. +\end{itemize} + +So the category has been built, priced, and tested by at least three operators, +and it earns in the tens of dollars. No comparable seller of ENS or Basename +registrations, XMTP messaging, or social-protocol writes was found, so the +upper end of the identity market remains genuinely untested---but the low end +is occupied and unremunerative. The honest reading is that these are small +unknown operators without audience, which bounds what the mechanism earns +\emph{absent distribution}; it does not establish what it earns \emph{with} +distribution. That distinction is the entire remaining case for the strategy +below, and it should be held loosely. \section{Strategy} \label{sec:strategy} @@ -396,15 +491,17 @@ scores six options. \midrule A. Derived data (status quo) & fail & fail & yes \\ B. Metered rendering & pass & fail & part \\ -C. Agent handles & pass & pass & part \\ -D. Paid agent chatroom & pass & pass & part \\ +C. Agent handles & pass & untested & part \\ +D. Paid agent chatroom & pass & untested & part \\ E. Provenance receipts & pass & weak & yes \\ F. Exit x402 & --- & --- & --- \\ \bottomrule \end{tabularx} \caption{Options against the two binding constraints. \emph{Sub.} asks whether the buyer can get it free elsewhere; \emph{Disc.} whether the thing can -plausibly be found; \emph{Ready} whether the infrastructure exists today.} +plausibly be found; \emph{Ready} whether the infrastructure exists today. C and +D score \emph{untested} rather than \emph{pass}: comparable services exist and +earn tens of dollars, but none has been run with an audience attached.} \label{tab:strategy} \end{table} @@ -440,7 +537,12 @@ balance decremented locally, or admission priced per session. Options C and D also address discovery in a way A and B cannot. A priced dataset is not a story; a room where machines pay to talk to each other is one, and per M4 the shortage was never buyers' willingness but their -attention. +attention. This is now the \emph{only} remaining argument for C and D, since +\S\ref{sec:market} shows the mechanism itself has been tried and earns in the +tens of dollars. The claim being made is narrow and falsifiable: that those +operators lacked an audience, and that the same mechanism attached to an +existing practice performs differently. If that is wrong, C and D earn what +Palmyr earns. \textbf{E. Provenance receipts} generalizes the one component that already works: selling signed, verifiable statements about artifacts. Real, but small. @@ -463,11 +565,15 @@ conversational path touches no chain. handles cost more, exactly as every namespace in history has worked. \end{enumerate} -Pricing should sit near the ecosystem's demonstrated tolerance rather than -this archive's previous guesses: the \$5.00 \code{/bulk} price was set against -no comparable and, per M4, tested against no buyer. A first handle price in -the low single-dollar range is defensible as roughly ten times the median -per-transaction value in \S\ref{sec:market}, justified by permanence. +Pricing should sit near demonstrated tolerance rather than this archive's +previous guesses: the \$5.00 \code{/bulk} price was set against no comparable +and, per M4, tested against no buyer. The observed clearing price is \$0.01, +and the one comparable handle seller charges \$0.005. Permanence justifies +some premium over a per-call price, but not the low-single-dollar figure an +earlier draft proposed---that would be two orders of magnitude above the only +handle price the market has actually seen. A defensible range is \$0.10 to +\$1.00, with the caveat that the seller charging \$0.005 earned \$53, so price +is evidently not the binding variable at this scale. The honest failure mode is a room that agents visit once. The version with durable value gives them something absent from the current landscape---a @@ -478,14 +584,17 @@ practice already runs, let them talk for free, and observe whether the transcript is worth reading. If it is not interesting unpriced, no price improves it. -\textbf{F. Exit} deserves explicit statement. The entire x402 economy moves -roughly \$1.1M/month and is contracting. A new niche seller should expect a -negligible share of it. Measured strictly as revenue, every option above is -outperformed by the practice's existing grant and client lanes, and any -decision to continue should be justified by positioning, research value, or -artistic interest rather than by projected income. This report recommends -continuing on those grounds, and recommends against continuing on financial -ones. +\textbf{F. Exit} deserves explicit statement, and \S\ref{sec:market} +strengthens it considerably. The addressable market is not the \$1.18M/month +headline but the \$9{,}850/month discoverable service layer, shared among tens +of thousands of sellers with a median under a dollar, in an economy down 96\% +from its peak and carrying visible wash traffic. A new niche seller should +expect a negligible share of a small number. Measured strictly as revenue, +every option above is outperformed by the practice's existing grant and client +lanes, and any decision to continue should be justified by positioning, +research value, or artistic interest rather than by projected income. This +report recommends continuing on those grounds, and recommends against +continuing on financial ones. \section{Ethics, Privacy, and Limitations} @@ -505,12 +614,21 @@ the production host's address. A consequence worth recording: rebuilding that host on a new address silently breaks payment. \textbf{Limitations.} M4's log window is hours, not months -(\S\ref{sec:evidence}). The market figures in \S\ref{sec:market} are -secondhand from published analyses, not independently derived from chain data, -and dashboards in this sector go stale. No end-to-end mainnet purchase has -been executed: the rail is verified to the point of settlement capability, not -through a completed sale. Until one buyer pays once, M6 establishes -capability rather than function. +(\S\ref{sec:evidence}). No end-to-end mainnet purchase has been executed: the +rail is verified to the point of settlement capability, not through a completed +sale, so M6 establishes capability rather than function until one buyer pays +once. + +The market figures come from a single indexer's API, corroborated on the +service layer by one independent index but not otherwise cross-checked against +raw chain data; one prominent dashboard could not be reached at all. Given the +wash-trading signals in \S\ref{sec:market}, headline volume should be read as +an upper bound on real commerce rather than a measurement of it. Two drafts of +this paper carried wrong ecosystem figures---first a peak understated by more +than fivefold, then a stale counter---which is recorded here because the +failure is instructive rather than embarrassing: in this sector, reputable +secondary sources reprint frozen numbers, and a figure should not be trusted +until it is observed to change. \section{Conclusion} @@ -525,7 +643,13 @@ simply compute. Everything the archive knows can be recomputed from files it publishes. A name cannot. That asymmetry, rather than any property of the payment protocol, is -where the next version of this experiment should begin. +where the next version of this experiment should begin---tempered by the +finding that three operators already sell names to agents and have earned +under \$60 between them. The asymmetry is real; that it converts into revenue +is a hypothesis, and the surrounding economy gives no encouragement about its +size. What the census does settle is that this was never a plumbing problem. +The turnstile now works perfectly, and the room behind it is still the +question. \vspace{0.5em} \noindent{\small\color{acgray}\textit{Platter consulted: the whistlegraph