Bank of Italy Governor Fabio Panetta said it plainly this week. Artificial intelligence will move the euro area's inflation path — but the direction depends on who captures the productivity gain.
The market read it as dovish. Euribor barely flinched. Crypto repriced nothing and tweeted everything.
That conditional clause is the most honest sentence a European central banker has produced about AI in two years. It is not a policy signal. It is an admission of measurement failure. The ECB has no real-time instrument that distinguishes AI compressing corporate marginal costs from AI expanding corporate margins. It sees negotiated wages on a quarterly lag, profit data on a three-month lag, and productivity revisions that outlive the business cycle. The ledger remembers what the market forgets. Right now the ledger holds information the Governing Council does not.
Panetta's positioning matters. He arrived at the Bank of Italy in late 2023 after four years on the ECB's Executive Board, and he spent that tenure as the institution's most dependable dove. When a dove summons the word "careful," the signal is not that easing is coming. It is that easing has become contested.
The backdrop is awkward. The deposit facility rate sits in the low-2% range after a sequence of cuts. Headline harmonised inflation is within a breath of target. Core is stuck north of 2.5%. Negotiated wage growth has decelerated from a 4%-plus peak but has not returned to anything the ECB can comfortably ignore. Euro area productivity has been roughly flat since 2019 — several points of output per hour below the US trend — while AI capital expenditure concentrates almost entirely in a handful of American hyperscalers and their supply chains.
So Panetta is being asked to set a single policy rate for an economy where the most consequential technological shock of the decade is being deployed somewhere else, financed in another currency, and measured by statistics agencies that will report on it twelve months after it has already repriced labour markets.
The crypto leg is worse. Europe's AI-transactional layer is being built on dollar rails. MiCA produced a workable licensing regime for euro-denominated stablecoins and almost nobody to issue them — euro stablecoin float remains in the low single-digit billions. Meanwhile the EU AI Act's phased implementation pushes compliance cost onto exactly the domestic firms that were supposed to capture the productivity gain. Brussels wrote the rulebook for a market it does not host.
There is a live index for the cost of AI that no central bank queries. Decentralised GPU networks — Render, Akash, io.net, Aethir — publish continuous, permissionless spot rates for H100 and A100 compute hours. These are not representative of the entire AI economy. They are, however, the only machine-readable, real-time price series for AI input costs that exists anywhere.
During an audit I ran on a GPU-rental protocol in the previous cycle, I watched per-hour rates on a single network move 30% in one week. The equivalent hyperscaler list price did not budge — because list price is a sticker, not a clearing price, and contract renegotiation lags it by six to twelve months. On-chain rates clear every block.
The implication is narrow and useful. If AI is disinflationary at the input layer, that shows up in decentralised compute prices weeks before it shows up in euro area producer prices. If it is inflationary through energy and chip scarcity, that shows up first too. The ECB publishes nothing comparable.
The distribution question — who benefits — is also already encoded. In token markets, distribution is not an estimate. It is a schedule. Every AI-adjacent token issued since 2023 carries a public vesting calendar, a public treasury address, and a public emission curve. You do not need a nowcast. You need a block explorer.
Here is what that ledger shows across the 2024-2025 AI token cohort: insider allocations clustered between 40% and 60% of supply, with cliffs clustered at twelve and thirty-six months. Power lies in the code, not the community. That is precisely the capital-capture scenario Panetta is worried about — the one that produces profit expansion without wage pass-through, and therefore the demand-deficient, disinflationary outcome rather than the wage-price spiral. The ECB is asking the question at the macro level. It has been answered at the firm level, on-chain, for two years.
Now follow the money that AI actually moves. Agent-to-agent payment rails are being standardised in stablecoins. x402-style HTTP payment headers, AP2 agent mandates, and machine-payable USDC on Base and Solana all settle in dollars. There is no euro-native equivalent deployed at scale.
This matters for Panetta more than any wage print. If AI's transactional layer denominates in dollar stablecoins, the euro area's AI-driven demand shock leaks out of the ECB's jurisdiction on the first hop. Total stablecoin float is a live, public proxy for offshore dollar demand — a channel the ECB can neither observe nor influence with a Frankfurt policy rate. A rate cut does not reach it. Whoever owns the settlement rail owns both the seigniorage and the data. Brussels owns neither.
Crypto's own response function has changed, and the coverage of Panetta's remarks missed it. Through 2024, digital assets traded as a high-beta proxy for the expected policy path. Through 2025, spot ETF creation and redemption flows, the basis trade, and custody mechanics have made crypto increasingly trade its own flow. My institutional framework last year flagged the early signs of this decoupling from tech equities — distinct regulatory regimes, distinct custody plumbing, distinct marginal buyer.
The consequence is unglamorous. Panetta's "careful, data-dependent" language probably moves crypto less than it moves Italian bank equities. The Crypto Briefing framing — that this is a liquidity signal — projects a 2021 mental model onto a 2025 market.
Finally, the yield. Where capital is concentrated, the gain lands. Restaking and liquid-staking venues concentrate collateral and MEV capture into a small number of operator sets; the yield accrues to whoever already holds the tokens and the validators. That is the same structural asymmetry Panetta named, expressed in code rather than in national accounts. Different mechanism. Identical distribution.
Set the lags side by side. On-chain compute prices clear in twelve seconds. Agent payment volumes are queryable in real time. Token distribution is knowable before issuance. Against that: euro area negotiated wages publish quarterly with a thirty-day delay, national accounts publish with a two-to-three month lag and revise for years, and productivity nowcasts are fitted to a post-2019 sample that contains no comparable technology shock. The ECB describes itself as data-dependent. Data without a ledger is just a story. It is dependent on data that arrives after the position is already set.
The prevailing crypto interpretation is that AI uncertainty keeps the ECB dovish, dovish keeps liquidity loose, and loose liquidity is structurally bullish. That is a category error, and it is the same error recycled in every crypto-media summary of every central bank speech since 2020.
Panetta's uncertainty is not dovishness. It is paralysis dressed as prudence. When a policymaker says the inflation direction depends on a distributional outcome he cannot measure, he is not signalling accommodation. He is telling you the reaction function is broken. In that regime, the sensible positioning is not leverage — it is volatility. The market's consensus is wrong about the sign.
The deeper blind spot is institutional. The ECB has outsourced the measurement of the most consequential technology shock in a generation to statistics agencies whose sampling frames were designed for an industrial economy. The ledger remembers what the market forgets, and the ECB keeps no ledger of its own. The marginal investor now reads hyperscaler capex guidance, on-chain compute rates, and agent settlement volumes in real time. Frankfurt reads a revised quarterly index.
And notice the framing choice. "Depends who benefits" converts a political question into a technical one. If AI is disinflationary, voters cheer and demand falls short. If it is inflationary, the asset-owning cohort is protected and the wage-earning cohort absorbs the cost. Presenting that as a forecasting problem, rather than a distributive one, is itself a governance act — and Panetta, of all people, knows the difference.
Watch three things, none of which appear in the ECB's dataset. Negotiated wages, to see whether the labour channel has broken. On-chain GPU-hour rates, to see whether AI costs are actually deflating for the operators building the rails. Euro-stablecoin float, to see whether Europe's AI economy is being built in dollars, on dollar rails, under Eurocrat rules.
The ledger is public. The question is who is reading it.
