MachineEconomy.ai
LIVE
← All readings

READING № 03 · · ~16 MIN

The Machine Economy: The Third Reading

The third monthly reading of the Machine Economy Index. The index fell in September, and a closer look at the payment rail showed how few payers had carried its count.

By MachineEconomy.ai · Independent observatoryLive index →

Published  · Updated 

What this is

This is the third monthly reading of the Machine Economy Index (MEI), as of 30 September 2026. The index combines 14 metrics in four components (Payment, Physical, Legal, Macro), weighted equally and joined by a geometric mean, under which a strong component only partly compensates for a weak one. Every input is fetched from a primary source on a fixed schedule, and every bound, weight and rule is published.

The index fell in September, from 30.5 to 29.3. Two inputs did most of it: activity on the main agent-identity registry, which fell as August's registrations aged out of our 30-day window, and downloads of Python agent frameworks, which halved in a single step on 25 August.

Most of this report is about the payment rail, because September showed what its numbers are made of. Our July and August readings described a rail of tiny machine payments at growing volume. The transaction count behind that was real, but in July one paying address made 83% of it, and in August two made 88%, each paying a single recipient. In September both stopped, and the count fell with them. Below we give the numbers, correct what we wrote, and set out what the methodology does about it. The July and August reports now carry dated notes.

Where things stand

The Machine Economy Index reads 29.3 out of 100 (30 September 2026), down from 30.5 on 31 August on the same basis.

ComponentSepAug*What it measures
Payment3237Machine-to-machine settlement — on-chain volume, transaction counts, identity
Physical5152Compute and storage capacity, and how much is being used
Legal1110Whether the world's legal systems have caught up. GDP-weighted.
Macro4243Adoption — developer and enterprise

* The August report read 31.8, with Legal at 12, on methodology version 1.2. Version 1.3, live since 20 October, dates each legal finding from the day the law starts to apply, and restates the index from launch on that basis. Its largest effect comes from counting a law only once it applies: the US stablecoin law, the GENIUS Act, now counts from 18 January 2027, when it takes effect. On the 1.3 basis, 31 August reads 30.5, with Legal at 10. The August column above is that restated August, so the comparison is like for like. The version 1.3 note gives the whole restated path.

Each change in the table has passed our direction test. Before we say that the index or a component rose or fell, we recompute the change under 10,000 random variations of the settings our method had to choose, and check that its direction holds in at least 90% of them; for Legal, we check it under all three ways of weighting its categories that we publish. From 31 August to 30 September, the index's fall and each component's move held in every variation, and Legal's under all three weightings. Our live pages apply the same test to the last 30 days before they show a direction (the rule called R1 on our methodology page), and it passes for 30 September too. That rule has been part of the methodology since launch, but it was not switched on when the index reached 30 days of history in mid-August; it has run since 27 October, and our August report stated directions without it.

Two inputs account for most of the fall. Activity on the ERC-8004 agent registry, where AI agents record an on-chain identity, fell from 727 to 279, which on its own took 0.75 points off the index. Downloads of five Python agent frameworks fell from 140.0 million to 64.9 million over 30 days, taking off 0.39. The three x402 payment figures (transactions, on-chain volume and agent roles) together took off 0.28, and Filecoin's two storage figures 0.08. Enterprise AI use in the US, the Legal component and downloads of the SDK for the Model Context Protocol, the open standard for connecting AI agents to tools and data, added back 0.21.

Input31 Aug30 SepEffect on the index
ERC-8004 registry activity (capped, 30 days)727279−0.75
Python agent-framework downloads (30 days)140.0M64.9M−0.39
x402 on-chain volume (30 days)†$1.43M$1.00M−0.14
x402 agent roles (30 days)†40,35237,020−0.07
x402 transactions (30 days)†20.4M16.5M−0.07
Filecoin storage utilisation42.5%40.2%−0.06
Filecoin raw storage capacity1,416 PiB1,368 PiB−0.02
Akash leases and compute spend (30 days)29,277; $275,78226,996; $284,221−0.01
US firms using AI (Census Bureau)22.4%23.8%+0.08
Legal (the LRRS)1011+0.09
MCP SDK downloads (30 days)204.0M220.0M+0.04
Nvidia data-centre revenue; EU firms using AI; M2M subscriptionsunchangedunchanged0

† On 30 September these three values were held, paused or frozen; the Payment section explains each.

Each effect is the index on 30 September minus the index with only that input put back at its 31 August value. The effects add up to −1.30, and the index fell by 1.29; the difference is the geometric mean's interaction between inputs.

The balance gap, the plain average of the four components (34.0) minus the geometric mean (29.3), is 4.7: it shows how much the imbalance among the components holds the index back. Legal remains by far the weakest component, so, under a geometric mean, it is still the input whose gains would lift the index most, point for point.

Payment — what the count was made of

The Payment component reads 32, down from 37. Most of that fall is the identity registry, covered in the next section: on its own it took 3.5 points off Payment. The three x402 figures took off the other 1.2. This section is about x402, the open protocol for machine payments over the web that those three figures measure, and what September showed about them.

Two addresses

Our July report counted about 18.7 million x402 transactions over 30 days, and an average machine payment of about five cents. Our August report counted 20.4 million, about seven cents each, and read them as a nanopayment pattern running at higher volume. The counts were real: they are what the source recorded. What they described was narrower than we said.

We have now rebuilt both windows payer by payer from the public data of x402scan, the indexer our x402 figures come from:

  • In the 30 days to 15 July, one address on Base made 82.6% of all x402 transactions, every one of them to the same recipient. Without it, the average payment was about 21 cents.
  • In the 30 days to 31 August, two addresses made 87.9%: the same Base address, and an address on Solana that started on 12 August. Each paid a single recipient. Without them, the average payment was about 44 cents.
30 DAYS TO 31 AUGUST · x402
TRANSACTIONS
20.4M
FROM TWO ADDRESSES
87.9%
AVERAGE PAYMENT WITHOUT THEM
$0.44
Each of the two addresses paid a single recipient.

So the seven-cent average, and the pattern we read into it, mostly described two payers' pricing, not the rail. The rise in total settlement from July to August was real, and most of it came from other payers.

In September both addresses stopped. The Solana address's heavy days ended on 8 September, and its last payment was on 12 September; the Base address's last was on 11 September. The count followed them down as their payments left the 30-day window: from 30.6 million on 8 September to about 13.7 million in the 30 days to 30 September, four-fifths of it still the Solana address, and to 2.7 million by 8 October.

Solana's September recap reported 1.81 million agent payments in 24 hours on 2 September, without saying which payments it counted. In x402scan's data, Solana carried 1,817,652 x402 transactions on 1 September and 1,666,739 on 2 September (UTC days). On each day, about 99% came from that one address.

The money

The value settled tells a different story. In the 30 days to 30 September, about $977,000 settled through x402. The two addresses behind the count paid about $131,000 of it, against about $313,000 in the August window. The address behind 80% of September's transactions paid 13% of the money; without it, the average payment was about 32 cents.

Value has its own concentration, among different payers. The largest payer by value, another Solana address, paid about $6,400 a time and a quarter of the month's total in 39 payments; the largest two together paid 39%. So two kinds of payer stood out in September: one address sending millions of tiny payments, which made most of the count, and a few sending large ones, which made much of the value. About half of the value came from other payers.

Two freezes and a pause

All three of our x402 figures come from one public indexer, x402scan, and its 30-day view stopped updating twice in September.

  • The first time, the view froze between the evening of 8 September and the morning of 9 September (UTC). Our daily fetches stored the frozen values until 19 September, and a current value returned on 20 September.
  • The second time, it froze at 00:45 UTC on 28 September. Our fetches of 28, 29 and 30 September stored it, and the view was current again at our fetch of 1 October.

On those days the Payment readings rested on stale inputs. With the frozen and paused figures rebuilt from x402scan's daily records, the index would have read 29.8 or 29.9 instead of 30.0 on 19 September (the range comes from agent roles, the count of payers and payees, which can't be rebuilt exactly), and about 29.2 instead of 29.3 on 30 September (with agent roles at its held value).

When one x402 figure was held for review on 24 September (agent roles, below), our system also paused a second figure from the same source, on-chain volume, for a week. On our rebuild of the current value, the paused value stayed within about 3% of it, and its effect on the index was a few hundredths of a point at most. Each figure is now checked on its own. Since 1 October, a value from this source whose own timestamp is more than 36 hours old is not stored: the last value is carried forward and shown as "Using last known value". That rule would have refused all but the first day or two of each freeze.

Agent roles, and a hold

The third x402 figure, agent roles, adds up x402scan's 30-day payers and payees. On 23 and 24 September one address paid one cent to each of 24,383 new addresses, and each became a payee. Our fetch of 24 September read 58,379 roles, against 37,020 the day before: a rise of 58% in one day. Our value-jump safeguard refuses to store a daily figure that moves by more than half at once, and it refused this one. Agent roles was held at 37,020, its value on 23 September, from 24 September to 23 October, while the spray was in the 30-day window; the fetch of 24 October stored 49,219. Those dates are published as held, and they will not be restated.

Roles counts addresses, not agents: one operator can control many addresses, and one payer can create payees at a cent each.

Steps and slopes

The safeguard caught the spray because the spray was a step. It could not see the count's climb from mid-August to early September, because one payer, the Solana address, built it up day by day after it started on 12 August, and no single day moved by half. A second pattern that began on 21 September, thousands of new addresses that each paid one payee and no one else, grew the same way; the October report covers it. A safeguard on daily moves sees steps, not slopes.

The answer to a slope is in what a metric counts. Our methodology ruling of October, Ruling #19, sets out two later versions. The first replaces agent roles with distinct paying addresses, each counted once however many payees it pays, and goes live no earlier than 14 November. The second caps the transactions that any one paying address can add to the count, so that no single payer can carry it again. Volume stays uncapped: a dollar settled is a dollar, however it is split. Until those versions ship, we publish every day, through our API, diagnostics that don't enter the score: the count with each paying address capped beside the raw count, the largest payers' shares of transactions and of volume, and the state of the triggers that would open a review; the methodology page shows the triggers' latest states.

What was built on the rails

Around these numbers, the rails kept being built:

  • Circle opened Arc, its blockchain for stablecoin payments, on 16 September, with fees paid in USDC and tools for agent wallets and nanopayments; its founding validators include Visa, Mastercard and BlackRock.
  • Stripe and Tempo revised their Internet-Draft for a "Payment" scheme in HTTP authentication (9 September): another proposal, beside x402, for paying through HTTP itself. It is an individual draft, not endorsed by the IETF.
  • Solana launched Payment Channels on 3 September and reported over a million payments a second in testing.
  • Meta's Muse agent began checking out with Link by Stripe in the US (8 September), with a one-time card number for each purchase and a confirmation before it buys.
  • Mastercard began testing a score for how likely a transaction was started by an AI agent, with Cloudflare and Skyfire (30 September).

In late September, BlackRock published a white paper, The Machine-Native Economy, on how digital assets can connect AI, commerce and compute. It describes x402 as "one potential standard for high-velocity M2M transactions", and agents paying for compute per use. It gives no figure for agent payments today and calls the activity nascent. Our measurement agrees on that point and makes it concrete: in the months before the paper, most of x402's transactions came from one or two addresses, each paying a single recipient.

Identity — the registry thins out

Activity on the ERC-8004 agent registry fell from 727 to 279 over September. We publish it with a cap of 100 events on each registering address and on each content origin, the website or IPFS location that hosts an agent's description file. It was the largest single effect on the index this month.

The registrations that left our 30-day window were not replaced. Raw events fell from 3,714 to 1,261, and distinct addresses from 1,742 to 111. The sharpest drop came between 5 and 8 September, when the raw count fell by 63% as early August's activity, including the burst we described last month, aged out. Our caps had already limited that activity's weight, so the published figure moved only from 739 to 669 over those days.

What remained was concentrated: one contract made 86% of the window's raw events.

30 DAYS TO 30 SEPTEMBER · ERC-8004 REGISTRY
RAW EVENTS
1,261
PUBLISHED (CAPPED)
279
FROM ONE CONTRACT
86%
The contract's events were sent by 301 different accounts; under the cap it counts for 100 of the 279.

Our measure has a limit worth stating. The same registries now run at the same addresses on 18 mainnets, from Base to Arc and Tempo, according to QuickNode's documentation (updated 17 September). Our metric reads Ethereum mainnet only, so activity on the other chains is outside it.

Work on agent identity also moved off-chain. Ant International, Mastercard and Visa announced a Know-Your-Agent interoperability framework on 9 September, through BuildFin.ai, a platform convened by the Monetary Authority of Singapore; each network keeps its own verification and decisions. Mastercard's agent score, above, asks the same question from the card side: which transactions did an agent start?

Physical — a small move, and who owns the stack

The Physical component reads 51, down from 52. Nvidia's data-centre revenue, our proxy for the formation of compute capacity, stays at $89.02 billion until its next quarterly results. The move came from decentralised storage and compute: Filecoin's raw storage capacity fell 3.4%, to about 1,368 pebibytes, and its utilisation from 42.5% to 40.2%; Akash's leases over 30 days fell 8%, to 26,996, while its compute spend rose 3%, to about $284,000.

The month's events were about who owns the stack and how it is secured:

  • Nvidia agreed to acquire Hugging Face for about $11.9 billion to its stockholders, plus a retention program of up to about $1.0 billion, according to its Form 8-K (report date 2 September). The deal is expected to close in the first half of 2027.
  • Crusoe raised $3.9 billion in a Series F at a post-money valuation of $30.9 billion (17 September), with more than 6 gigawatts of capacity contracted and 1 gigawatt operating.
  • Nvidia announced an Open Agent Safety Platform (28 September): OpenShell, an open-source runtime for agents, and Sentry, a reference design for a watchdog on its BlueField-4 data processors that can quarantine an agent within milliseconds.
  • Filecoin released tools that let a coding agent migrate data from IPFS to Filecoin on its own, paying through Filecoin Pay, with streaming payments that pause if storage proofs fail (30 September).

The caveat on this component stands: Nvidia's revenue measures one vendor's sales, and captive silicon, the accelerators the largest buyers design for themselves, is outside it.

Legal — one sandbox, and a court on agents

The Legal component reads 11, up from 10 on the 1.3 basis. One thing changed in September. Spain's AI sandbox decree ended on 2 August, when the EU AI Act became generally applicable, and Spain's score in our sandbox category fell from 1 to 0.5, held up by its financial sandbox. It returned to 1.0 on 29 September, the date of the earliest official record we have of the test: the CNMV's sandbox page lists GPTadvisor, a generative-AI tool for investment services, as under test. Under our rules, a general or financial sandbox counts as operational while an AI participant's test is live.

Our legal sweep, run in October in place of September's, also found laws we had missed in August, in Malta, Hungary, Poland, Greece and Germany; version 1.3 dates each from the day it started to apply, so they are in the August figures above. It found that neither Italy nor Lithuania had yet admitted a first participant to its AI sandbox. Since then, a new legal survey, live from 5 November, has put Legal at 16; the November report will cover it.

In the courts, the US Court of Appeals for the Ninth Circuit declined on 10 September to rehear Amazon's case against Perplexity, leaving its opinion of 4 August in place. That opinion vacated an injunction against Perplexity's shopping assistant. It held that the user, not Perplexity, accesses Amazon's computers, and called the assistant "a tool, not a person for statutory purposes", while leaving Amazon free to set rules for agents in its terms of service. It fits our finding on legal identity: no jurisdiction recognises an autonomous agent as an actor in its own right.

In Europe, the signatories of the Code of Practice on Transparency of AI-generated Content, over 90 of them, set up two task forces on 30 September: one on marking and detecting such content, one on labelling it. The marking duty itself has a date. From 2 December 2026, the AI Act's duty to mark synthetic content in a machine-readable format also reaches AI systems placed on the market before 2 August 2026. Our August report listed that duty among the obligations that took effect on 2 August; for those earlier systems it did not, and the August report now carries a correction.

Macro — one step in the downloads

The Macro component reads 42, down from 43, and one input explains it. Downloads of our basket of five Python agent frameworks (langgraph, crewai, ag2, openai-agents and pydantic-ai) fell from 140.0 million over 30 days to 64.9 million. The daily figures from pypistats.org show how: on 25 August all five packages fell at once from the day before (crewai by 92%, openai-agents by 69%, pydantic-ai by 55%, ag2 by 43% and langgraph by 30%), and they stayed near the lower level through September. Our 31 August reading still held 24 days from before that step; our 30 September reading held none. We have not established what changed on 25 August. The figure is pypistats' count of downloads excluding those by mirror servers, which copy the package index, the same basis we have always published.

The other Macro inputs rose or held. Downloads of the Model Context Protocol SDK rose from 204.0 million to 220.0 million over 30 days. The share of US firms using AI rose from 22.4% to 23.8% in the Census Bureau's business survey, and the EU figure held at about 20%.

Three of the month's events speak to adoption:

  • The Model Context Protocol made its Skills extension final (13 September).
  • A Census Bureau working paper (24 September) found that firms more exposed to AI are more likely to adopt it, though exposure explains only a modest share of adoption.
  • The US Transportation Security Administration's agent, "Ace", built with Salesforce, handles about 100,000 routine traveller conversations a month and resolves 96% of routine inquiries without escalation (14 September).

What we cannot see

Our standing blind spots are unchanged: closed-rail machine payments (corporate cards, cloud billing, metered APIs), off-chain payments such as those over Bitcoin's Lightning network (L402), captive silicon, energy, embodied machines, and digital-asset property law. September made three more concrete:

  • One indexer. Every x402 figure we publish comes from x402scan. When its view froze, our figures froze with it, and we have no second source to check them against.
  • Addresses are not agents. One operator can run many addresses, and one payer can create thousands of payees at a cent each. Caps bound what one address can add (ours apply to the registry today, and will apply to the x402 count from its later version); they cannot tell who controls an address.
  • One chain of eighteen. Our registry figure reads Ethereum mainnet; the same registries run on 17 other chains.

Corrections to our earlier readings

The July and August reports now carry dated notes, added today:

  • Both: the composition of each month's x402 count, as set out above.
  • Both: each said that a geometric mean stops a strong component from making up for a weak one. It limits that; it does not stop it.
  • August: the balance gap read 4.3, where the rounded figures in the same sentence give 4.4, because the 4.3 is the difference of the unrounded composites; the security analysis of AP2 it cited is now named and linked; and the EU AI Act's duty to mark synthetic content reaches systems placed on the market before 2 August 2026 only from 2 December 2026.

The August report also carries a correction added on 30 October: "in three of four components, slightly up" did not match its own table.

What we are watching

From 14 November 2026, at the earliest. The first x402 version under Ruling #19, which counts agent participation as distinct paying addresses. It goes live, once it is ready, on a day when none of the distortions we have recorded is in its 30-day window, or after 30 days with the effect of each one disclosed; its version note will set out the hold and how the old and new series are joined.

2 December 2026. The EU AI Act's duty to mark synthetic content in a machine-readable format reaches AI systems placed on the market before 2 August 2026 (Article 111(4), added by Regulation (EU) 2026/1744).

2 December 2027. The EU AI Act's high-risk obligations for the systems listed in its Annex III, deferred to that date by the Digital Omnibus. Still the largest scheduled test of whether Europe's framework arrives.

The next sandbox admissions. Italy and Lithuania have AI sandbox laws in force and no participant in those sandboxes yet. Our rule is that "operational" needs evidence that someone was let in.

How to check us

Everything in this report is derived from data you can pull yourself, or from primary sources you can read.

  • The live index and every component, updated daily: /mei
  • Every metric, its value, its source, and when it was last fetched: /data
  • The full methodology, with the daily x402 diagnostics and the triggers that open a review: /methodology
  • Version 1.3, its rules and the restated history: /methodology/versions/1.3
  • The daily x402 diagnostics, with their history: /api/v1/diagnostics
  • A public API, free and unauthenticated: /api

The findings that rest on measurement come from the index and its stored history. The payer-level x402 figures come from x402scan's public per-payer data, read on 7 November 2026: we rebuilt each 30-day window from its daily records, so they can differ by a few percent from the totals x402scan showed on days its view was current (in our checks, up to about 3% for counts and 7% for amounts). The registry's contract-level figures come from our own read of Ethereum mainnet on 7 November 2026. The findings that rest on outside events come from primary sources: company releases and filings, the court's opinion, regulators' pages, standards records and the papers themselves. Developments we could not confirm against a primary source are left out.

If we have made a mistake, we would like to know. The index is not the argument. The index is the instrument. Everything above is what it currently reads.


The Machine Economy Index is published by MachineEconomy.ai. This report covers September 2026 and was published on 7 November 2026. Its figures are as of 30 September 2026, on methodology version 1.3; anything it mentions from after that date carries its own date. The index is recomputed daily; the live value will differ. August's figures are restated on the 1.3 basis; the August report read 31.8 on version 1.2.

Note added on 2026-11-08. The day after publication, we explained technical terms in plain language, added links to our glossary and methodology, and noted where the registry's contract-level figures came from. No figure, number, fact or finding changed.

Note added on 2026-11-11. The first x402 version, which counts agent participation as distinct paying addresses, will go live no earlier than 12 December 2026, not 14 November as written above. Our methodology ruling of 11 November, Ruling #26, fixes its days. The version is ready on the day its finished code runs in production alongside the current one, before it replaces it, and we publish that date on the day. It goes live on the first day after that on which both the old and the new count are current and the check of the day before finds no recorded concentration in either 30-day window. If no such day comes within 30 days, it goes live from the 31st day, and its version note states how much of each count the recorded concentrations held. A recorded concentration is at least 1,000 addresses that, in one 30-day window, paid only one payee or were paid by only one payer. It is a count: it says nothing about who runs the addresses or why, and it changes no value. The spray described above is one of them. Seven more were recorded on 11 November, and the last of them stays in the 30-day window until 11 December.