The Machine Economy: A First Reading
The inaugural report of the Machine Economy Index. Machines are becoming economic actors. This is the first attempt to measure how ready the world is for them — and where it isn't.
Published: 2026-07-15
What this is
Machines are beginning to transact. Software agents pay for compute, buy data, hire other agents, and settle in stablecoins — without a human in the loop for any individual transaction. The infrastructure for this is being built quickly and in public: payment protocols, identity registries, decentralised compute markets, agent frameworks.
Nobody was measuring whether the world is ready for it.
The Machine Economy Index (MEI) is an attempt to. It is a composite index, in the same family as an economic-health index, built from 14 metrics across four components — Payment, Physical, Legal, and Macro — each weighted equally, and combined with a geometric mean rather than an average. That last choice matters, and we will come back to it, because it is the reason this report says what it says.
Every input is fetched from a primary source on a fixed cadence. Every bound, anchor and weight is published. Every metric we considered and rejected is on the record, with the reason. The methodology is public in full, and this report is meant to be checked.
This is the first reading.
Where things stand
The Machine Economy Index reads 32.6 out of 100 (15 July 2026).
That is the composite. It is the least interesting number in this report.
| Component | Score | What it measures |
|---|---|---|
| Payment | 43 | Machine-to-machine settlement — on-chain volume, transaction counts, agent identity |
| Physical | 51 | Compute and storage capacity, and how much of it is being used |
| Legal | 12 | Whether the world's legal systems have caught up. GDP-weighted. |
| Macro | 42 | Adoption — developer, and enterprise |
Three of the four sit between 42 and 51. One sits at 12.
That gap is not a rounding artifact, and the index is deliberately built so that it cannot be averaged away. A simple arithmetic mean of these four components would read 37. The geometric mean reads 32.6. The difference — 4.5 points — is what we call the balance gap, and it measures exactly one thing: how uneven the machine economy's foundations are.
The choice is deliberate. Under a geometric mean, a weak component cannot be compensated for by the others. If one rail is missing, the index says so. An arithmetic mean would let the strength of compute paper over the absence of law. We think that would be a lie.
One consequence governs everything that follows: the fastest way to move this index is legal infrastructure. Because Legal is the lowest component, and because the geometric mean is most sensitive to its weakest input, a point of improvement in Legal moves the headline several times more than a point anywhere else. That is a property of the arithmetic, not a preference of ours — all four components are weighted equally, at 0.25.
And we can answer the obvious objection before it is raised: you chose equal weights — what if you are wrong? We tested it. Across a fixed grid of 33 weightings — equal weights, a 35/30/15/20 split and all its permutations, and every component shifted up or down ten points — the index ranges from 28 to 38, and Legal is the weakest component in every single one. Across ten thousand randomised weightings and goalpost choices, Legal is weakest in 100% of them. You can disagree with our weights. It does not change the finding. The index level is largely a function of one thing: how much weight you place on the law. Every low reading weights Legal heavily; every high reading weights it lightly. The finding underneath is the same.
Payment — one protocol wide
The Payment component scores 43.
Over the last thirty days, roughly $871,000 settled on-chain through the x402 protocol, across approximately 18.7 million transactions.
Do that division. The average machine payment is about five cents — a shade under, at $0.047.
That single number is the clearest signal in this report that something genuinely new is happening. Humans do not transact at five cents — the fees alone forbid it. Machines do, constantly, and the economics only work because the settlement layer is designed for it. This is the nanopayment pattern, and it is not a theoretical property of the machine economy. It is already the observed one.
But the rail is narrow. x402 is, at present, essentially the entire publicly verifiable on-chain machine-payment rail. We admit it under an explicit dominance clause: when one protocol constitutes substantially all of a category's observable activity, we track it and say so, rather than pretending to a diversity that does not exist. If other machine-payment protocols publish equivalent verifiable data, they join this category — they do not expand x402's footprint.
Alongside it, the ERC-8004 agent registry recorded roughly 1,900 identity events in thirty days. That is a small number, and it should be read as one: agent identity is early. The infrastructure exists; it is barely used.
What we do not see here matters. Closed-rail machine payments — an agent paying for cloud compute on a corporate card, or metered API billing — are almost certainly larger than the on-chain volume, and they are invisible to us. So is Lightning / L402, a real and active machine-payment protocol built on Bitcoin: because Lightning settles off-chain by design, there is no public, verifiable series to measure, and there cannot be one. We have declared both as gaps rather than guessed at them.
Physical — the rail that works
The Physical component scores 51. It is the strongest of the four, and it deserves to be said plainly: the machines are being built.
Compute capacity is forming at scale — we track it through Nvidia's data-centre segment revenue, which is the best verifiable proxy that exists, though an imperfect one (more on that below). Decentralised compute markets are transacting: the Akash network is clearing leases and settling spend continuously. Decentralised storage has real capacity and real utilisation: Filecoin holds substantial raw power, and a meaningful fraction of it is doing work rather than sitting idle.
There is a caveat we want on the record, because it is the largest single weakness in our own construct. Nvidia's data-centre revenue measures one vendor's sales. If hyperscalers shift toward captive silicon — their own accelerator designs — then true compute capacity could rise while our metric falls. We have modelled this: a 30% increase in real capacity, paired with Nvidia's merchant share sliding from 90% to 65%, would move our metric down about a point and a half. It would be wrong, and it would be wrong in the worst possible way — quietly, in the right-looking direction.
So we have restated what the metric actually measures: merchant-market compute capacity formation, not total compute capacity. Captive silicon is a declared gap. And we have pre-committed to a trigger: if credible evidence shows Nvidia's merchant share falling below 60%, we review the construct. That threshold is written down now, before we know whether it will ever fire, so that it cannot be moved later to avoid an inconvenient result.
Legal — the bottleneck, and it is not close
The Legal component scores 12.
This is the component that pulls the index down, and it is the one where the substantive findings of this report live. It is measured by the Legal Rail Readiness Score (LRRS) — a GDP-weighted coverage model across 30 jurisdictions representing about 89% of world GDP, and five categories of legal instrument. It is not a milestone tally. It asks: what proportion of the world economy operates under a legal framework that can accommodate an autonomous economic agent?
Almost none of it.
No jurisdiction on earth recognises an autonomous agent
Not one. There is no legal system anywhere that grants an autonomous software agent legal status, rights, or liability as such. The UK's Financial Markets Law Committee put the global common-law consensus about as plainly as it can be put: an artificial intelligence is "not a person or an incorporated entity in the eyes of the law", and "no matter how complex it is, it would (and should) be treated as a tool."
The consequence is not abstract. In common-law jurisdictions, where an autonomous agent transacts without a legal wrapper, the law does not shrug — it defaults. It treats the deployers and token-holders behind the software as an unincorporated general partnership, which carries unlimited joint and several personal liability for everything the code does. US courts have now applied exactly this reasoning to decentralised organisations. That is the actual legal position of a great deal of machine-economy activity today, and very few of the people building it appear to know. (Civil-law systems reach the question differently, through their own doctrines; there is no single global rule, which is itself part of the problem.)
One sovereign has done something about it, and it is a rounding error in world GDP. The Republic of the Marshall Islands' DAO Act does something no other national law does: it vests the management of a company in a smart contract. Not in a director who consults a smart contract — in the contract itself, with the governing code's identifier filed publicly as a condition of formation. It works, and it is used.
It is also a Pacific atoll of 181 square kilometres. On our own world map, it is too small to draw — it appears only as a footnote beneath the map, which is as good a summary of the state of machine legal identity as we could have designed on purpose.
The UN wrote the law. Nobody has passed it.
In July 2024, UNCITRAL — the United Nations body that writes the templates the world's commercial law is built on — adopted the Model Law on Automated Contracting. The General Assembly endorsed it that December.
It is exactly on point. Article 5 provides that a contract formed by an automated system "shall not be denied validity or enforceability on the sole ground that no natural person reviewed or intervened in any action carried out in connection with the formation of the contract." It recognises contracts formed by code. It handles the attribution of automated acts.
It is, as far as we can determine, the single most relevant piece of machine-economy law in existence.
No country has enacted it. Not one, in two years. A model law has no force until a legislature adopts it, and none has. It scores zero in our index, and the zero is correct.
The nearest thing to a machine-economy free zone requires a human in the loop
Several jurisdictions have built special legal zones for digital assets. The most advanced is Abu Dhabi's DLT Foundations regime — a purpose-built legal entity for decentralised networks, with real registrants and real capital.
Read the regulations and you find this. Every DLT Foundation must appoint a Council of two to sixteen human beings. And that Council does not merely hold a veto — the regulations provide that any decision on a delegated matter shall be subject to confirmation by the Council. Not oversight after the fact: affirmative human sign-off on every delegated decision.
That is: the closest thing on earth to a machine-economy legal zone legally requires that a human confirm what the machine does. We looked hard at it and scored it zero, because our category asks whether the law contemplates machines operating economically as machines — and this one, explicitly, does not.
We rejected six other candidates on the same test. The category reads zero worldwide, and we have published every rejection with the reason and the specific evidence that would change our mind.
And Europe just deferred its own remedy by a year
The EU AI Act required every member state to have a national AI regulatory sandbox operational by 2 August 2026. On 29 June 2026, the Council adopted the Digital Omnibus, which postponed that obligation to 2 August 2027.
The coverage of that decision treated it as compliance relief for AI vendors. In machine-economy terms it is something else: one of the largest scheduled improvements in the world's legal readiness has been pushed back a year.
Today, four EU member states run an operational AI sandbox — Spain, France, Denmark and Latvia. (Latvia's went operational in May, admitting its first cohort of participants; we name the admitting instrument in our public record, because our rule is that "operational" requires evidence that someone was actually let in, not merely that a law exists.) Four more EU states have passed a law but admitted nobody. Nineteen EU member states have no sandbox at all.
An earlier version of this index credited every EU member state with partial sandbox readiness on the strength of the AI Act's mandate. We removed that credit before publishing, and the reason is worth stating, because it explains why the Legal score is 12 and not higher: a law that obliges someone else to build a framework is not a framework. Under the old rule, a thirteen-month deferral of the exact thing being credited could not have moved the score at all. A readiness score that cannot register a one-year delay in the thing it is measuring is measuring the wrong object.
(Europe is not absent everywhere. The EU's markets-in-crypto-assets regulation, MiCA, is a real, in-force stablecoin framework, and it applies across all twenty-seven member states. It is why the stablecoin category is the most-covered of the five. But a stablecoin framework is not a sandbox, and neither is a substitute for a law that recognises the agent itself.)
Macro — the developers are ready. Their employers are not.
The Macro component scores 42, and it splits cleanly in two.
Developer adoption is high. Downloads of the Model Context Protocol SDK and of the major Python agent frameworks run into the hundreds of millions per month, combined. People are building agents at scale, right now.
Enterprise adoption is not. The US Census Bureau's business survey puts AI use at about 20.6% of firms. Eurostat's equivalent puts the EU at about 19.9%. Roughly one business in five.
The gap between those two facts is one of the more interesting things this index surfaces, and we do not attempt to explain it here. We only note it, and note that both numbers come from national statistical agencies and are among the slowest-moving inputs we track. Whatever is happening in the developer ecosystem, it has not yet arrived in the enterprise.
What we cannot see
We publish our blind spots. There are fourteen declared gaps in the current methodology, each with a statement of the source that would close it. The most consequential:
- Closed-rail machine payments. Agents paying through corporate cards, cloud billing, and metered APIs. Probably larger than everything we measure on the Payment rail. No verifiable public series exists.
- Off-chain machine payments (Lightning / L402). A real, active protocol we cannot measure, because it settles off-chain by design. A construct boundary, not an omission.
- Captive silicon. Hyperscaler-designed accelerators. Real compute capacity that our merchant-market proxy cannot see.
- Energy. The machine economy's power draw. There is no Tier-1, high-cadence source for it. That is why it is a gap and not a metric.
- Embodied machines. Robotics and autonomous physical systems. Same reason.
- Digital-asset property law. A category of legal instrument that maps to none of our five. We have declared it a boundary of version 1.0 rather than quietly widening the model to accommodate it.
We would rather tell you what we cannot measure than have you find it.
What we are watching
Two dated things, stated so that we can be judged on them.
2 August 2027. The deferred EU sandbox deadline. If the twenty-seven member states stand up operational AI sandboxes, we estimate the effect at roughly a point or so on the index — one of the largest single known future movements. If it does not happen, we will say so.
Singapore. The Monetary Authority of Singapore finalised its stablecoin framework as policy in 2023. The enabling legislation has still not passed. It scores accordingly. We are watching for the amendment.
How to check us
Everything in this report is derived from data you can pull yourself.
- 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 — every bound, every anchor, every weight, the aggregation, the LRRS coverage model: /methodology
- A public API, free and unauthenticated: /api
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. Figures in this report are as of 15 July 2026. The index is recomputed daily; the live value will differ.