The headline number is thirty-two percent.
Somewhere in the last quarter, Anthropic's economic research group put a figure on the table: that by 2030, AI-driven growth could add something in the order of a third to global output. The number has since travelled. It has moved into policy briefings, into equity research notes, into the deck templates that circulate in the weeks before a funding round closes. Thirty-two percent is not a rounding error. It is roughly the equivalent of bolting another United States and another Japan onto the world economy inside five years, without a war and without a reconstruction program.
I read projections the way I read smart contracts. I look for the omitted line item before I look at the total.
The omitted line item, in almost every version of this argument I have encountered, is settlement.
In 2025 I spent the better part of a working year analyzing automated payments, sampling more than a hundred AI-agent transactions across three networks and eleven distinct counterparty patterns. The question was never whether the agents could reason. Reasoning was improving on a curve everyone could see. The question was whether an agent could prove who it was, and whether value moved correctly when someone was actively trying to break the flow. Roughly one in five of the flows I sampled carried an identity proof that a patient counterparty could have forged at low cost. None of those failures appear in any aggregate economic statistic, because aggregate economic statistics were never built to catch them.
A projection of this kind measures the value a model assumes will be created. It does not measure the cost of making that value auditable, and it does not measure the latency between producing it and clearing it. Thirty-two percent is a production-side number. Nobody in the chain from task-level automation to national output has been asked to price the verification substrate that sits underneath it.
Protecting the ledger from the volatility of hype is not a slogan for me. It is the job description. And right now the most consequential hype in this market is not a token. It is a macro forecast.
To understand where the number comes from, and why I think it is simultaneously plausible and dangerously incomplete, it helps to look at how Anthropic builds its economic work in the first place.
The Anthropic Economic Index is, methodologically, one of the more honest pieces of research to come out of an AI laboratory. It is constructed from observed usage rather than stated preference. Rather than surveying workers about what they might delegate, it classifies what people actually ask the models to do, maps those tasks against occupational taxonomies, and derives exposure shares by role and by sector. Listening to the errors that the metrics ignore begins with crediting the metrics that were built carefully. This one was. It does not ask people to imagine their future. It reads their present.
Where the difficulty starts is the extrapolation. Moving from task exposure to national output requires a chain of inferences: exposure converts to time saved per task; time saved converts to labor reallocation; reallocation converts to measured productivity; productivity converts to gross domestic product. Each link in that chain carries a wide interval. Compounded across four links, the interval becomes very wide indeed. The thirty-two percent figure is a central estimate drawn from a distribution, not an observation pulled from a ledger. That distinction is not pedantry. It determines whether you treat the number as a fact to plan around or a hypothesis to test.
Place it beside the rest of the field and the distribution becomes visible. Goldman Sachs has published a global GDP effect in the neighborhood of seven percent spread over a decade, with a smaller annual productivity bump concentrated in advanced economies. McKinsey Global Institute has put the 2030 figure near thirteen trillion dollars of additional output. PwC has gone higher still, into the mid-teens as a share of global GDP. The IMF and the OECD have been notably more conservative, emphasizing that diffusion, not capability, is the binding constraint. Across institutions with comparable access to data, the spread spans a factor of four or five.
That spread is the most informative data point in the entire conversation. It tells you that reasonable people with good datasets disagree about the numerator, and almost all of them are agreeing on something else: that the bottleneck is absorption. Organizational capacity. Energy. Capital. Regulation. Trust. Not intelligence.
The physical constraints deserve more attention than they get in the summary versions of this argument. Data center construction is limited by grid interconnection queues measured in years, by transformer lead times, by turbine availability, by water rights in the regions where the largest clusters are being built. Compute capital expenditure has been running at a scale that would have been considered a national infrastructure program a decade ago. None of that is a reason the projection is wrong. It is a reason the projection has a schedule attached to it, and the schedule is the part that affects markets.
Policy is the second constraint, and it is moving faster than the models. Sovereign compute programs, export controls on advanced accelerators, national AI strategies, the EU AI Act's staged obligations, securities and custody guidance that keeps expanding in scope, energy disclosure regimes that now touch model training runs. Every one of these instruments assumes that the activity it regulates can be located, attributed, and measured.
That assumption is where my own work intersects this conversation, and where the macro debate has a hole in it.
All of these frameworks share a single premise: that measured output equals produced value, and that transfers of that value clear. In an economy where a growing share of transactions is initiated by software acting on behalf of humans, or by software acting on behalf of other software, that premise is doing far more work than anyone has acknowledged. Gross domestic product is a production measure assembled from surveys, filings, and administrative data. It has no atomic unit. It cannot tell you whether a specific transfer completed.
There is exactly one place in the current financial system where the transfer of value and the verification of that transfer are recorded in the same event, atomically, with a cryptographic audit trail that any third party can replay without asking permission. That place is a blockchain. Whatever one thinks of the asset prices that trade on top of it, the accounting property is real, and it is the property the macro conversation quietly assumes into existence.
This is the bridge that keeps getting skipped. If agent-mediated commerce grows at anything like the rate the thirty-two percent implies, then the rails carrying that commerce stop being a niche and become systemically important infrastructure. Those rails have latency, cost, and failure modes. No GDP model currently prices any of them.
My training is in cybersecurity, and the habit it instilled is specific: I do not ask whether a trend is real. I ask what breaks first when it scales.
The first thing that breaks is unit economics.
In 2021 I was a junior researcher at a mid-sized protocol when the NFT market came apart. I spent the following months analyzing more than fifty failing marketplace contracts to understand why liquidity had evaporated so completely. The finding was not exotic. Batch minting routines were consuming far more gas per unit than necessary, and the inefficiency was invisible in calm markets. It only became fatal when users returned and the fixed cost per action exceeded what anyone was willing to pay for the action itself. The lesson I took from that period was not about NFTs. It was about overhead at frequency.
A human buys a coffee once. A human makes perhaps a few dozen discretionary financial decisions in a day. Software does not behave that way. An agent mediating procurement, paying for inference, buying data, settling with a counterparty agent, and posting a bond for good behavior might initiate thousands of transactions in an hour. At that frequency, the fixed cost per transaction is the entire ballgame. Variable costs that look like noise at human scale become the dominant term.
Run the numbers. A minimal token transfer on an Ethereum-compatible network consumes roughly twenty-one thousand gas units before any contract logic, signature verification, or calldata is accounted for. Under volatile base fees on a congested layer one, the cost of a single action can exceed the value of the action. Rollups compress this substantially by batching, and that compression is the actual product they sell. But the compression is contingent. It depends on batching, batching depends on scheduling, and scheduling depends on whoever is assembling the blocks.
The infrastructure that makes machine-frequency spending legible has been arriving steadily: account abstraction under ERC-4337, paymasters that sponsor gas on behalf of users or agents, session keys that grant scoped and time-bounded authority, and the delegation primitives introduced under EIP-7702 that let an externally owned account behave like a smart contract for a defined period. None of these are marketing features. They are the plumbing that allows software to hold and spend value without a human approving each authorization. They are also, individually and collectively, new attack surfaces. A paymaster with a policy bug is a faucet. A session key with an over-broad scope is a standing grant. I have reviewed both.
An agent economy exists only where the marginal cost of a verified action falls below the marginal value of that action. Everything else in the narrative is downstream of that inequality. This is not a philosophical claim. It is arithmetic, and it has a testable threshold.
Identity is the second thing that breaks, and it breaks earlier than most people expect.
When I designed a verification protocol for automated payments in 2025, the design constraint was not cryptography. It was the economics of cheap identity. If an agent can mint a thousand identities at negligible cost, then every incentive system built on reputation collapses, because reputation is only meaningful when identity is scarce. A sybil that can appear as a hundred counterparties can manufacture the appearance of consensus, exhaust a bond pool, or walk away from obligations and return under a new key.
The failures I documented were not sophisticated. They were lazy. Reused attestation nonces. Signatures that verified against a key that was never bound to an operator. Self-issued credentials with no anchor. In a sample of a hundred-plus flows, the proportion carrying a proof that would not survive motivated scrutiny was around a fifth. Extrapolate that to a payment volume consistent with the thirty-two percent projection and the exposure is not a rounding error. It is a category.
The solution space is reasonably well understood, even if the implementations are immature. Zero-knowledge attestation lets an agent prove properties about itself without revealing the underlying data: that its model weights match a published commitment, that its operator is a registered entity in a named jurisdiction, that its spending authority is bounded, that it has not been revoked. I built a lightweight version of this in 2025 precisely because full disclosure was a non-starter for commercial counterparties, and because the proving cost had to stay low enough to be irrelevant at transaction scale.
Here is the trade-off that rarely makes it into the pitch decks. Verification must be cheaper than the fraud it prevents, on an expected-value basis, or it does not get used. For high-value settlement, this is trivially satisfied: a proof costing a few cents against a counterparty risk measured in thousands is obvious. For micropayments, it is not. There is a floor below which verified agent commerce cannot operate, and that floor is set by proving cost, not by model capability. The thirty-two percent projection has no line item for it.
The surrounding standards layer is consolidating as I write. HTTP-native payment negotiation protocols, payment-agnostic agent authorization frameworks, on-chain agent identity registries with reputation hooks. Each one answers a different part of the same question, and none of them are final. Standards immaturity is a real cost. It shows up as integration work, as duplicated security reviews, as counterparties that support one dialect and not another. That cost is currently being paid in engineering hours, which means it is invisible in the statistics that feed the projection.
Latency is the third thing that breaks, and it breaks in a way that has a precedent I have already documented.
In 2023 I led a forensic analysis of three production Layer 2 sequencers. I reverse-engineered their consensus and ordering mechanisms and quantified the concentration of control over block production. The report cited specific block-production latencies and identified a single-point-of-failure exposure that I estimated for the largest of the three at roughly fifteen percent. Institutional analysts picked it up because it was measurable rather than rhetorical. Decentralization either survives an audit or it does not.
That work is more relevant now, not less. Humans tolerate a twelve-second block time and a fifteen-minute confirmation because they go do something else in between. Software does not go do something else. Software retries. Retries under uncertainty are how you get duplicated payments, or a service delivered twice and billed twice, or a bond released to the wrong party. When an agent has to decide between waiting for finality and re-submitting, the answer it chooses determines whether the ledger is accurate. A centralized sequencer that can reorder or suppress transactions is not merely a philosophical problem. It is an availability and correctness problem for the machine economy specifically.
Ordering is the second-order consequence. Whoever sequences determines priority, and priority determines who extracts value from the flow. If agents become the dominant initiators of transactions, then maximal extractable value ceases to be an occasional tax on retail traders and becomes a systematic levy on machine commerce. The magnitude scales with the volume, and the volume is precisely what the projection forecasts. There is no line item for that extraction either.
Guarding the gate, not just the gold, is how I have come to think about it. Decentralization is not a political preference dressed in technical language. It is an availability property with a measurable failure budget. When the floor drops, the foundation speaks, and in an agent economy the floor drops every time a sequencer has a bad hour.
Verification economics deserves its own treatment, because the choice between proof systems is not a technical detail. It is a statement about who you trust when things go wrong.
Trusted execution environments attest quickly and cheaply, and their guarantees rest on the assumption that a small number of silicon vendors build correctly and do not collude. Zero-knowledge proofs carry a harder trust floor and cost more to generate. Optimistic systems defer verification to a challenge window and rely on watchers who must be compensated, which reintroduces a fee market with its own capture dynamics. Fraud proofs, validity proofs, bonded attestations, and staked oracles are four different answers to the same question, and they produce four different failure surfaces.
The quiet confidence of verified, not just claimed, is the property institutions are actually buying. They will not route settlement through a rail whose guarantees are asserted in documentation rather than demonstrated in a proof. My 2024 work reviewing custodial solutions after the ETF approvals taught me this in concrete terms. Auditing multi-signature implementations at three major firms, I found two using threshold signature constructions that had predated the new guidance and no longer satisfied it. The gap was not malicious. It was architectural drift: a design that was sound when it was written, sitting underneath a regulatory regime that had moved. Technical debt compounded across a compliance boundary.
What resolved it was not a legal memo. It was a roadmap that translated cryptographic requirements into language a legal team could evaluate: what the threshold means, why the assumption set changed, what the migration costs, what fails if you do nothing. Compliance is a technical feature. It is also a feature with a testing cost, and that cost belongs in any honest accounting of what AI-mediated finance will require.
The audit trail as a narrative of trust is not a metaphor. On a chain, the transaction log is the record. It is not reconstructed after the fact by an auditor sampling a population. It is the object itself. For an economy of machine-initiated payments, that property is the difference between a system that can be supervised and one that can only be believed.
Regulation is where these threads converge, and where the projection is most exposed.
Consider a single payment initiated by an autonomous agent. Who is the customer for know-your-customer purposes? The operator, the deployer, the model provider, or the principal who set the mandate? Which entity files the suspicious activity report if the agent's behavior drifts? Whose jurisdiction applies when the agent's weights are hosted in one country, its operator is incorporated in a second, its counterparty is in a third, and the settlement rail is a neutral network with validators everywhere? The EU's Markets in Crypto-Assets framework, the AI Act's obligations for general-purpose systems, securities custody rules, and the travel rule for transfers of funds each answer a fragment. None of them answer the composite.
This is not an argument for delay. It is an argument for building the answer into the architecture rather than bolting it on. Selective disclosure is the mechanism I keep returning to: the ability to prove compliance to a supervisor without publishing the commercial details of every flow. Viewing keys, zero-knowledge compliance attestations, and scoped audit credentials are the tools. They are not novel research. They are under-deployed because the demand for them arrives with the volume, and the volume is still ahead of us.
There is a harder constraint underneath all of it, and it is one I think about more than any other. If agents transact at machine frequency, state growth becomes an operational problem rather than a theoretical one. Every action writes something. Every write has to be stored, propagated, and made retrievable by whoever needs to verify it later. Archive requirements, data availability sampling, pruning policies, and the cost of permanent retrievability are all part of the verification bill, and the bill grows with the volume. Memory is the backup of the blockchain, and backups are not free.
Then there is the composition question, which is where I part company with most of the optimistic reading of the projection.
Gross domestic product is a production measure. It says nothing about who receives the output. If AI-driven growth accrues disproportionately to the owners of compute, models, energy, and application-layer equity, then output can rise while the wage share compresses. That has a direct consequence for the on-chain economy, and it is not the one currently priced.
If balances concentrate, the shape of on-chain demand changes. You get fewer, larger, more institutional flows. Those flows want regulated custody, permissioned venues, reporting hooks, and counterparties with legal personality. That is a different market from the one implied by a retail-driven agent narrative. It is also a market where the winning rails may be hybrid rather than permissionless, and where the tokens that appreciate may not be the tokens that carry the volume.
There is a second-order measurement issue buried here as well. If AI is deflationary for a broad basket of goods and services, then real output growth can exceed nominal growth, and the headline number can understate what actually happened. Conversely, if the growth is concentrated in a narrow set of capital-intensive sectors, the nominal figure can overstate the breadth of the improvement. Either way, the single number people are quoting is doing less work than it appears to.
Now the part that makes people in this industry uncomfortable.
Thirty-two percent is a model output. Models produce distributions, and institutions quote from the upper end because the upper end is what gets attention. There is a selection effect operating on which estimates become quotable, and it is not neutral. The number that travels is the one that justifies the allocation, not the one that survives the widest confidence interval.
I have watched this pattern before, at smaller scale. The phrase liquidity fragmentation circulated for years as a justification for launching new venues, new aggregators, and new tokens that solved no problem I could measure. It was a narrative with a product attached. The AI growth projection is now doing the same work at macroeconomic scale. It is being used to underwrite compute buildouts, to justify sector rotation, and, increasingly, to argue that a given token or protocol is positioned to capture the coming wave. Most of those arguments skip a step. A protocol captures value only if it charges for something scarce that the flow cannot route around. Most do not.
Here is the blind spot I would put in front of any allocator reading this. The firms positioned to capture AI-driven GDP are compute suppliers, model developers, energy producers, and application companies with conventional equity structures. A public blockchain token is not a claim on any of that unless the protocol has a fee mechanism tied to the flow. The thirty-two percent can be entirely real and still be irrelevant to the price of the asset you are holding. Those are separate questions, and they are being conflated.
There is a second, sharper possibility. AI may reduce demand for intermediation rather than increase it. If a model can hold context, evaluate counterparties, and negotiate terms directly, then a meaningful share of the value currently captured by middleware becomes a function call. The layers that depend on human friction to justify their margin are the ones exposed. Some of those layers are in this industry.
And the third blind spot is structural. If verification concentrates in a small number of attestation vendors, or if the sequencer market consolidates around two or three operators, then the thirty-two percent is delivered on infrastructure with a single point of failure. That is the fifteen percent figure I documented in 2023, scaled to the size of the global economy and charged against a projection that assumes continuous availability. Nobody has written that risk into a macro model. Somebody should.
What I would track, if I were allocating capital against this thesis rather than writing about it, are three measurements that no macro forecast contains.
The first is the realized cost of a verified agent action on a major rollup, tracked quarterly and normalized for congestion. This is the single most informative number in the entire agent economy, because it is the constraint every other claim is downstream of. When it falls below the marginal value of a routine commercial action by a comfortable margin, the agent economy stops being a projection and becomes an accounting fact. Until then, every estimate of its size is a model output, including the good ones.
The second is the concentration ratio of block production and of attestation supply. Both are measurable. Both have precedents for drifting in the wrong direction during periods of rapid growth, because growth rewards the operators with the most capacity. A margin of safety here is worth more than a marginal basis point of yield.
The third is the share of stablecoin settlement volume attributable to non-human initiators. Nobody publishes this cleanly yet. It is inferable from transaction graph structure, from gas sponsorship patterns, and from the distribution of inter-arrival times, and it is the closest thing we have to a live read on whether machine commerce is actually materializing. I would trust that series more than any forecast, because it is a measurement rather than an expectation.
The thirty-two percent may well turn out to be right. Anthropic's research group built its index honestly and its extrapolation transparently, and the direction of travel is not seriously in dispute. What I want to know is what it costs to make that growth auditable, who holds the balances when it arrives, and which gate is left standing when the volume gets there. Those are not questions the projection answers. They are the questions that decide whether the growth is durable or whether it is one more number that travelled further than its evidence did.
