A couple of days ago, on 1 August, Brad Setser posted that The Economist had picked a good week for a currency cover, and that the Big Mac index agreed with his own work in putting the yuan around 30% undervalued. I replied.

Setser is the Whitney Shepardson senior fellow at the Council on Foreign Relations, and before that he was deputy assistant secretary for international economic analysis at the US Treasury and a senior adviser to the US Trade Representative.

He has spent two decades tracking global capital flows, reserve accumulation, and currency intervention, and his blog Follow the Money is one of the few places where the actual balance of payments data get read properly. I have a great deal of respect for Brad’s work, which is why the disagreement is worth having in public rather than in a footnote.

His number is exactly what the index says. In the July 2026 edition a Chinese Big Mac costs 31% less in dollars than an American one, and in January it was 33% less. The question is what that gap is a fact about.

The proposition is not a throwaway. It says that a persistent gap between a currency’s market rate and its purchasing power parity rate is a fact about the country’s price level rather than a fact about its currency, and if that is right then a good deal of what is written about undervalued currencies is a category error.

Let me be clear about where I start, because the reply reads as though I were reaching for Balassa-Samuelson first. I am not. I start with purchasing power parity, and I start there because it is the decisive factor. Prices and exchange rates move together, and everything else in currency analysis operates on the remainder.

I have forecast exchange rates professionally for a quarter of a century, and PPP has been the starting point of every model I have built. Balassa-Samuelson comes second, as the correction that says the parity a country converges to depends on how productive it is. I have tested that correction many times, in many formats, on many datasets. Never on the Big Mac index.

Which is a strange omission, because the index turned forty this summer.

It first appeared in the print edition of 6 September 1986, in a table of thirteen countries, under the heading of a light-hearted guide to whether currencies were at their correct level. Pam Woodall, then the paper’s economics editor, devised it as a semi-humorous illustration of purchasing power parity, and it gave the language the word burgernomics.

She cannot have expected it to still be running four decades later, in more than fifty countries, twice a year. The anniversary was as good an occasion as I am going to get to run the test again on data I had never used. And with a bit of help from Kaggle, Github and Claude it can now be done relatively fast.

What follows is the result: 2,036 observations, 54 currencies, and thirty years of the index, together with income, productivity, institutions, and governance data for every one of them.

What the index actually measures

Start with the theory, because almost every argument about the Big Mac index is really an argument about purchasing power parity, and most of the people having it have not thought carefully about which version they mean.

Purchasing power parity is one of the oldest propositions in international economics.

Gustav Cassel, who is one of my absolute favourite economists, gave it its name and its modern form during the First World War, when the gold standard was suspended and the practical question was which exchange rates Europe should return to afterwards.

And his answer was that the equilibrium rate is simply the ratio of the two countries’ price levels. So if a basket of goods costs 100 kroner in Denmark and 20 dollars in America, the equilibrium rate is five kroner to the dollar.

The Big Mac index applies exactly this logic to a single good: take the local price of a Big Mac, divide it by the American price, and you have an implied exchange rate that can be compared with the actual one. If the actual rate is weaker than the implied rate, the currency is called undervalued.

Note what this is not. It is not a measure of whether a currency will rise. It is a measure of whether the price level and the exchange rate are consistent with each other. Any adjust might happen through changes in the price level or the exchange rate – or both.

Figure 1 puts every observation on a single chart, with the Big Mac parity rate on the horizontal axis and the actual rate on the vertical, both measured in local currency per dollar and both on a logarithmic scale.

That covers seventy-four currencies over forty years and a span of seven orders of magnitude, from the Swiss franc at one end to the Vietnamese dong at the other.

Figure 1. Parity holds across seven orders of magnitude

The regression slope is 1.0162, which means a forty-year test of Cassel’s proportionality postulate cannot reject a coefficient of one, and the explained variance is 0.982.

Cassel was right!

The bit the chart hides

But now look at the same picture from the other side, because the median absolute deviation from the 45-degree line is 23%.

On a chart spanning seven orders of magnitude that deviation looks like nothing at all, while to anybody actually running money a 23% misalignment is enormous. And that gap is what the rest of this piece is about.

The standard explanation is Balassa-Samuelson, after Béla Balassa and Paul Samuelson, who published it independently in 1964.

The mechanism is straightforward once you separate goods into two kinds. Traded goods such as steel, wheat, and semiconductors have their prices arbitraged internationally, while non-traded goods and services such as haircuts, rent, and waiting a table do not.

Rich countries are rich because they are productive in the traded sector, and that productivity bids up wages across the whole economy, including in the barber shop where no productivity gain has occurred. So the price of non-traded goods rises with income, the overall price level rises with it, and rich countries end up expensive.

A Big Mac is itself mostly non-traded, being beef and bread assembled by local labour in rented premises. Not to mention electricity and taxes.

But Balassa-Samuelson is a statement about productivity, and productivity can be measured in a dozen ways that do not agree with each other.

So I ran the exchange rate equation with each of them in turn, alongside the burger price, a governance term, and a dummy for the four city states (See more on that further below). A negative coefficient means a stronger currency.

Eight of the ten carry the sign the theory predicts with a t above two – meaning it statistical significant at a 5% level.

Whether productivity is measured at purchasing power parity, at constant prices, per worker, per hour worked, as total factor productivity, or as how long people live, a more productive country has a stronger currency than the burger parity rate implies.

Human capital and years of schooling are the two that fail, and both are stocks of education rather than measures of output, which is not the same thing.

Output per hour is the theoretically correct variable, and in the latest Penn World Table it covers 52 of the 55 currencies, which it did not in earlier versions. Total factor productivity gives the largest coefficient of all, at −27.52, though it is also the least precisely measured.

Life expectancy is the one I keep coming back to. It contains no prices, no exchange rate, and no monetary quantity of any kind, and it still gets a t of −4.71.

Note what is not in the table. GDP per head measured at market exchange rates is the variable most people reach for, and it has no business being here. Convert income to dollars and you have multiplied by the exchange rate, which is the thing on the left-hand side. The relationship that produces looks strong and means nothing.

Figure 2 shows what the surviving relation looks like, with productivity measured as GDP per head at constant purchasing power prices, which is the measure that ends up in the model.

Figure 2. Richer countries have dearer burgers

The slope is 0.23. A country twice as productive as another has a burger price around 16% higher, which is the Balassa-Samuelson effect in its plainest form.

Note the spread around the line, though. At any given level of productivity the burger price varies by 50 log points or more.

Institutions are hard to separate from income

Every empirical exchange rate economist reaches for institutions at this point, and I am no exception.

The trouble is visible in one chart. Figure 3 plots the Fraser Institute’s measure of legal system quality and property rights against productivity. The correlation is 0.75, which leaves 43% of the variation in one that the other does not account for.

Figure 3. Institutions and income move together

I want to be careful about what that does and does not mean.

It is not that institutions are income, and nothing here says that good institutions are produced by being rich. Rich countries have good institutions and well governed countries are rich, and forty years of burger prices will not tell you which way that runs. What the correlation does mean is that the two are very hard to hold apart in estimation.

That shows up as instability. Put several institutional measures in together and they take turns having the wrong sign, because they are all proxies for the same underlying thing and each one absorbs whatever the others leave behind. Which measure survives a general-to-specific search depends on the order in which the others are removed.

So instead of picking one, here is every institutional measure I have, entered one at a time into the same equation.

All ten point the same way. Three are individually significant. The first principal component of all ten, which takes 84% of their common variation, enters at −3.61 with a t of −1.98.

That is the honest version of the institutional result, and it is a better one than picking the survivor of a search. The sign consistency across ten measures from two independent sources is the finding. The magnitude is not: the coefficients span a factor of fifty because the measures are on different scales, and which one is individually significant depends on nothing more interesting than measurement noise.

The equation below uses voice and accountability because it survived a general-to-specific search. I would not defend that choice. Rule of law or control of corruption has at least as good a theoretical claim, and both give a similar answer.

The model

Put it together and the equation is written the way a currency forecaster writes one, with the exchange rate versus the US dollar on the left.

100 × ln(local currency per dollar) = 2.58 + 0.989 × relative burger price − 15.68 × productivity − 0.437 × governance + 27.89 × city state

n = 2,036, 54 currencies, 1996 to 2025. R-squared 0.9868. Standard errors clustered by currency.

Note what is not on the right-hand side: no exchange rate, in any form, anywhere. The relative burger price is measured in local currency, the productivity term comes from the IMF in constant purchasing power prices, and the governance term is a percentile rank.

A negative coefficient means fewer local units per dollar, which is a stronger currency. So the reading is that productivity and governance both pull a currency above its parity rate, and that the four city states sit around 32% below either.

Every candidate had a sign attached to it before estimation, and anything that came out with the wrong one was removed. Twenty-two candidates went in and four came out. I want the model to be statistically correct and economic theoretically correct.

That procedure is not innocent, and the reported t values are too high because of it.

Selecting on the estimated sign and then quoting the significance of what survives overstates the evidence, and it does so most for the governance term. Read the burger price and productivity coefficients as estimates and the governance coefficient as an indication.

The burger price coefficient is 0.989 with a t against one of −0.99. Again highly significant.

Cassel’s proportionality postulate cannot be rejected, in an equation with no exchange rate on the right, on 2,036 observations across 54 currencies and thirty years.

One more test before the story is told, and it is the one that matters most.

The equation above is estimated in levels across currencies, which means it can be read as a statement about why some countries are dearer than others.

Add currency fixed effects and the question changes to whether the relation holds within a currency over time. It is a harder test, and two of the four terms do not survive it.

The burger price coefficient barely moves, at 0.97 either way, and productivity survives once vintage effects absorb the common movement in the world price level.

Governance does not survive. It changes sign and becomes significant in the wrong direction. Whatever the governance term is picking up, it is a difference between countries and not something that moves a currency when a country’s own institutions change.

I said above that I would treat it as an indication rather than an estimate. This is why.

That is close to being the whole story. Purchasing power parity, plus The Balassa-Samuelson effect, plus a governance term that is real but hard to size, plus a dummy for four places that are cities rather than countries.

Burgernomics is no evidence of Chinese currency manipulation

Which brings me to the argument that started all of this. Brad Setser as mentioned above that the Big Mac index agrees with his own work in showing the yuan substantially undervalued, and every American administration for twenty-five years has said something similar in considerably stronger terms.

The raw index does appear to show it, since China’s Big Mac has averaged 42% below the American price over the whole period.

However, conditional on the model, I find nothing of the sort.

The yuan has averaged within a few percent of the rate the equation implies, and the sign of the gap flips depending on which productivity measure goes into it: slightly undervalued on GDP per head, slightly overvalued on output per hour, and on the final specification 11% overvalued across the sample with the latest reading essentially on the line. Account for Chinese productivity and there is no unexplained undervaluation left to explain.

Figure 4 puts every currency in the January 2026 edition on one scale, with the United States given a residual of its own rather than assumed to be correctly priced.

Figure 4. Price level and exchange rate misalignment

Meanwhile the two largest unexplained undervaluations in the entire dataset are Taiwan at 21% and Hong Kong at 19%.

The Taiwanese burger is now 61% cheaper than the American one against a model prediction of 21%, and the gap has widened by around 55 percentage points since 1994 while Taiwanese income rose sharply.

The country everybody argues about is the one the model fits best. Just have a look below.

Figure 5 puts the fourteen most traded currencies side by side and shown as they are quoted: units of local currency per dollar.

The dashed line is the rate the model implies, which is the Big Mac parity rate adjusted for productivity, governance, and the city state term. The shaded area on the right-hand axis is the gap between the two, in percent.

Figure 5. The fourteen most traded currencies against the dollar

China is among the tightest fits in the panel. The yuan has averaged 7% away from the rate the equation implies, with the Canadian dollar at 0.5%, the New Zealand dollar at 3.7%, the won at 3.3%, and the yen at 5.5% alongside it.

Taiwan is the failure. The Taiwan dollar goes from roughly fair in the mid-1990s to 47% undervalued today, while the rate the equation implies barely moves. The Hong Kong dollar is the same story without the trend: 55 observations, undervalued in all but one of them, and the city state term takes only part of the gap. It is tempting to say that geopolitics might play a role here – both for Taiwan and Hong Kong.

The Swedish krona has averaged 31% overvalued and the krone 37%, which sounds like a standing feature of both economies. It is not. The krona peaked at 94% in 2011 and reads 9% undervalued in the latest edition.

The Norwegian krone peaked at 116% in 2006 and is now 2% below the line.

The euro has gone from 65% in 2008 to half a percent. All three converged on the model from above, at between one and five percentage points a year, and all three arrived.

Switzerland is the exception that did not arrive. The franc is the standing example of an overvalued currency, and the Swiss Big Mac is the dearest in the world at 9.08 dollars.

It has been converging too, from 106% in 1995 to 22% today, but at 1.7 points a year it is the slowest of the four and it still has two decades to go at that rate. Whatever makes Switzerland expensive is more durable than whatever made Scandinavia expensive. That being said – for all these currencies there are clearly elements of safe haven effects that the model does not fully account for.

The commodity dollars went the other way. Both the Canadian and the Australian dollar were 30% or more above the line during the commodity boom of 2011, and both are now below it, at 9% and 20% respectively. For these two the model reads the cycle rather than a level, and the cycle is the price of iron ore and oil.

How the gap closes

An equilibrium relation is only interesting if deviations from it come back. So the last question is whether they do, and how fast.

The object that has to revert is the real exchange rate: the dollar price of a Big Mac relative to the American one.

I estimated an error correction (EC) equation for it on the annual panel, with the lagged deviation from the model, one lag of the dependent variable, and changes in productivity and governance. 1,154 annual changes across 55 currencies, with years of collapse and hyperinflation removed.

It reverts, and the estimate is stable. Roughly nine percent of a deviation closes each year, which puts the half-life between seven and eight years, and adding short-run dynamics moves it hardly at all.

The country-level picture is consistent with the pooled one.

The adjustment coefficient has the mean-reverting sign in 40 of the 41 currencies with at least twelve annual changes, though only 16 of those are individually significant, and the median country half-life is three years rather than eight. That gap between the pooled and the median country estimate is the usual one in this literature: the pooled coefficient is dragged towards zero by the currencies that barely move.

Push the deviation further back and it fades in the way it should. Measured from two years earlier the coefficient is 0.078, from three years earlier 0.052, both still significant on the same sample.

Seven to eight years is slow, but it is also squarely in the range the purchasing power parity literature has settled on since Rogoff’s survey put the consensus at three to five years for the real exchange rate, and rather slower than that, which is what you would expect from a single traded good with a large non-traded component.

What I cannot do with these data is say which side does the adjusting.

The index is relative prices minus the nominal exchange rate, and estimating a separate equation for each side requires putting one of them on the right-hand side of the other, where it is endogenous by construction. A regime split would help, since a pegged currency cannot move by definition, but the available classifications describe each currency against its own anchor rather than against the dollar, and half the currencies my model calls pegged are pegged to the euro. That is a question for a different dataset – or a later blog post.

What forty years of burgernomics say

Four findings, in the order of how much weight they will carry.

Prices and exchange rates move together, and that is most of the story. The coefficient on relative prices in an equation for the nominal rate is 0.99, and it is 0.97 when currency fixed effects are added, so it is not an artefact of the pooling.

Deviations from the implied rate revert at 8 to 9% a year. Cassel’s proposition, formulated in 1918 to work out what exchange rates Europe should return to after the war, is the single most robust thing in this data.

Productivity moves the parity, as Balassa and Samuelson said it would. The elasticity is around 0.2. Eight of ten measures give it, from output per hour to total factor productivity to how long people live. The coefficient survives currency and vintage effects.

This is the correction to Cassel, and it is a real one, but it operates on the residual and not on the main relation.

Institutions probably matter and I cannot prove it. Ten institutional measures from two independent sources all carry the predicted sign, and three are individually significant. That consistency is worth something. But add currency fixed effects and the coefficient changes sign, as the model section showed. Institutional quality also correlates 0.75 with productivity, and no amount of econometrics on 55 currencies will separate the two.

And the index is not a currency forecast. It measures the consistency between a country’s price level and its exchange rate, and when the two are out of line the historical record says the gap takes seven or eight years to close. The model cannot say which side gives. A currency 40% from its implied rate is not a prediction that the currency will move 40%.

That last point is the one I would press hardest, because it is the one the index is routinely used against.

The Economist has presented it since 1986 as a guide to whether currencies sit at their correct level, and that framing invites the reader to treat a negative number as a trade. It is not. It is a measurement of how expensive one country is relative to another, and it is a rather good one.

So the index deserves a promotion and a demotion at once. As a currency signal it is weak. As an indicator of relative price levels between countries it is one of the longest and most consistent series anyone has.

Which brings me back to where I started – my reply to Brad Setser was right about the mechanism – poorer countries, i.e. countries with relatively lower productivity, have “undervalued” currencies.

A currency that has been cheap for decades is usually a poor country, and China is the clearest case in the data. Condition on Chinese productivity and the yuan sits within a few percent of where this framework puts it, with the sign of the gap flipping depending on which productivity measure goes in. The raw 30% is income, not manipulation.

On its fortieth birthday, then, the index says something rather unfashionable. Cassel was broadly right, Balassa and Samuelson were right about the correction, the yuan is roughly where Chinese productivity puts it.

References

Source data for the index are published by The Economist in its big-mac-data repository and are the basis for everything here.

•  Balassa, B. (1964), “The Purchasing-Power Parity Doctrine: A Reappraisal”, Journal of Political Economy 72(6)

•  Cassel, G. (1918), “Abnormal Deviations in International Exchanges”, Economic Journal 28(112)

•  Feenstra, R. C., Inklaar, R., and Timmer, M. P. (2015), “The Next Generation of the Penn World Table”, American Economic Review 105(10). Version 11.0 is used here, covering 185 countries to 2023, published in October 2025

•  Rogoff, K. (1996), “The Purchasing Power Parity Puzzle”, Journal of Economic Literature 34(2)

•  Samuelson, P. A. (1964), “Theoretical Notes on Trade Problems”, Review of Economics and Statistics 46(2)

Data

The Fraser index and the Penn World Table both end in 2023. For those series the last observation is carried forward. The main model uses only the Big Mac index, IMF income at constant PPP prices, and the governance indicators.

Second-quarter GDP figures arrived this week for both the US and the euro area, and the coverage has been almost entirely about a single number.

US real GDP grew at an annualised rate of 1.5% in the second quarter, down from 2.1% in the first, and it came in weaker than the euro area, which managed 1.8%.

The US economy is slowing, it is now slowing faster than Europe, and the natural conclusion might be that the Federal Reserve has kept policy too tight for too long.

However, that conclusion is totally wrong. The Fed does not control real GDP growth – the Fed controls nominal GDP growth and that actually accelerated from 5.8% annualised growth in Q1 to 7.9% annualised growth in Q2.

So the Fed has actually been EASING rather than tightening monetary policy.

I have seen almost nobody make that point this week, and I find that genuinely puzzling, because the nominal number is not a technical footnote to the real number. It is the only one of the two that the Federal Open Market Committee decides.

What a central bank can and cannot do

Start with the identity that every monetarist starts with. MV = NGDP = PY. The money stock times the velocity of money equals nominal spending in the economy, which in turn equals the price level times real output. It is true by definition, so on its own it explains nothing at all, and in my experience that is precisely why it gets waved away by economists who ought to know better.

But it stops being empty the moment we add a single proposition, and the proposition is that real output over the medium term is determined by the supply side rather than by the central bank. By technology, the labour force, the capital stock, regulation, and the quality of institutions. I would concede that the Fed can push real output around for a few quarters, and that it can wreck an economy outright if the error is large enough, but it cannot raise the growth rate of productive capacity by printing money in the long run. Said in another way in the long run the Phillips curve is vertical.

Call that supply-determined growth rate Y*. If the Fed cannot change Y*, and if it has committed itself to 2% inflation, then the Fed’s job reduces to a piece of arithmetic: it has to deliver nominal spending growth of Y* plus 2%. Deliver more than that and you get more than 2% inflation, deliver less and you get less, and that is very nearly the whole of the theory of inflation a central bank actually requires.

Hence we do not need the Phillips curve here, and we do not need an estimate of r*, of u*, of the output gap, or of any of the other unobservable variables the FOMC spends its meetings arguing about. Take nominal GDP growth, subtract trend real growth, and look at what is left over.

That residual is the inflation rate the current stance of monetary policy is delivering, and I will call it implied inflation.

Choosing the starting point

The whole exercise turns on Y*, so let me be open about how I have arrived at it rather than pull a number out of a hat.

I have measured Y* as the average annualised quarterly growth rate of US real GDP from the first quarter of 2022 onwards. Not a Congressional Budget Office projection, not a Fed staff assumption, and not a number chosen because it produced the answer I wanted, but simply what the US economy has in fact delivered over the past four and a half years.

The reason I have chosen this period is that it is post-Covid distruptions. The choice that actually matters is the starting date, and my criterion is that the real economy should be roughly in balance at that point, because a base period taken during a collapse or a rebound measures the cycle rather than the trend.

Take a look at where that criterion leads. Going back to 2019 would mean averaging in the collapse of 2020 and the rebound of 2021, which tells us about lockdowns rather than about productive capacity. Starting in 2023 would mean measuring the supply side over a window short enough that a single strong quarter moves the answer. By the end of 2021 the labour market had normalised, unemployment had returned to something close to the structural rate, and the output gap had closed.

So the first quarter of 2022 is where the real economy stops telling us about the pandemic and starts telling us about the supply side. That gives Y* of 2.4% and therefore an implied NGDP target of 4.4% (2.4% + 2%) a year.

Note that the choice is conservative rather than convenient, because a lower estimate of trend growth would make the Fed look worse than it does here.

US nominal GDP indexed to Q1 of each year against a 4.4% implied target

Each coloured line is one calendar year of US nominal GDP, indexed to 1.00 in the first quarter and carried through to the first quarter of the following year, while the black dashed line is the 4.4% path. The figure in brackets beside each year label is the implied inflation that year’s nominal growth delivers.

Implied inflation came in at 8.9% in 2021, then 5.4% in 2022, then 3.1% in 2023, then 2.2% in 2024, then 3.7% in 2025, and the first half of 2026 is running at 5.5% annualised.

Not one of the twenty-one quarters sits below the line. Not one. So consider what this measure would have been telling the Fed in real time, because the sequence carries more information than any single reading does.

Through 2020 the Fed did the right thing and did it fast, because nominal spending collapsed, the Fed responded aggressively, and the collapse had been reversed inside a year.

From early 2021 the same measure turns from reassuring to alarming, and implied inflation of 8.9% is neither a rounding error nor a data artefact. Furthermore, the Fed could have seen it coming, because markets were pricing it well before the CPI showed it, and it spent that year explaining instead that the pressures were temporary and supply-related.

Then 2022 to 2024 is the stretch the Fed deserves real credit for. Implied inflation falls from 5.4% to 3.1% to 2.2%, which is a genuine disinflation achieved without a recession, and by 2024 nominal spending was within two tenths of a percentage point of the rate consistent with the target.

And then it reversed, with 2025 coming in at 3.7% and the first half of 2026 running at 5.5% annualised. Consider what the Fed was doing while that reversal was under way. It cut the funds rate three times in the autumn of 2025, in September, in October, and again in December, seventy-five basis points in total, into a year in which nominal spending was already accelerating away from the target, and it has held at every meeting since.

Said in another way, the Fed eased into an overshoot it had spent three years correcting.

The explanation everybody reaches for

There is a war in the Middle East, crude oil rose more than 20% in July alone, and the FOMC statement blamed energy prices explicitly. That argument is nonsense, and it is nonsense in a way that ought to be embarrassing for anybody who lived through 2021.

Hence we heard exactly the same argument five years ago, when inflation was supply chains, semiconductors, shipping containers, and pent-up demand for used cars. Every one of those explanations was a claim about relative prices dressed up as a claim about the general price level, and the general price level rose 20% anyway. Milton Friedman’s line survives contact with the evidence rather better than the alternatives do: inflation is always and everywhere a monetary phenomenon.

That being said, supply shocks obviously do move the inflation data from one month to the next, and I would never argue otherwise. What they cannot do is move nominal spending. Oil changes how a given quantity of money chasing goods divides itself between prices and output, and it does not change the quantity.

So don’t tell me that 7.9% nominal GDP growth is an oil story.

The control group is called the euro area

If the US overshoot were the product of an energy shock, then the shock ought to be visible in the other economies exposed to it, because the war in the Middle East is not a US event.

In fact the euro area imports a far larger share of its energy than the US does, so if the oil price were doing the work here, Europe should look worse rather than better.

Euro area nominal GDP indexed to Q1 of each year against a 3.1% implied target

Trend real growth in the euro area since the start of 2022 comes out at 1.1% on exactly the same measure, which gives an implied NGDP target of 3.1%. Implied inflation there runs 9.0, then 6.9, then 3.3, then 2.7, and then 1.7% in 2025, so the 2025 line finishes below the target path with three of its four quarters below it as well.

Take a look at the two starting points before anything else, because in 2021 implied inflation was 8.9% in the US and 9.0% in the euro area. Hence the initial error was not a US error and it was not a European error. It was the same error, made at the same time, on both sides of the Atlantic, and it was enormous.

Anybody who wants to argue that one of these two central banks was prudent in 2021 has not looked at the data.

Implied inflation, US and euro area

Then look at what came afterwards. Through 2024 the two track each other closely, with the euro area consistently a shade higher, and in 2025 the sign flips to 3.7% against 1.7%, while in the first half of 2026 the gap widens to two and a half percentage points.

That is about as close to a controlled experiment as macroeconomics ever gets, with the same shock, the same starting error, and the same global supply conditions producing opposite outcomes. Whatever explains the divergence differs between Washington and Frankfurt, and there is only one serious candidate.

It is not a supply shock. It is US monetary policy that is far too easy.

Credit where it is due

I have spent a lot of the past two decades criticising the European Central Bank and I have not been polite about it, so let me be equally direct in the other direction: the ECB has handled this well.

In June the Governing Council raised rates by 25 basis points, taking the deposit rate to 2.25%, and in doing so it became the first major central bank to tighten in the face of the energy shock. In July it held, unanimously, while making clear that several members had already asked whether a further increase was warranted and that September remains open.

Note that the question everybody in Frankfurt was arguing about is beside the point. Transitory or persistent, second-round effects or not, look through the oil price or do not: in my view none of it matters. Trend real growth in the euro area is somewhere between 1% and 1.5%, which puts the implied NGDP target between 3% and 3.5%, and the job is to keep nominal spending there whatever crude does.

So the transitory debate is not a hard question that the ECB happened to answer correctly. It is a question a central bank aiming at nominal stability never has to answer at all.

Measured against that target the ECB has delivered, because euro area nominal GDP grew 2.8% in 2025, which is at or below the bottom of the range whichever end of it you prefer.

Furthermore, Frankfurt had the harder task, since a target of 3% to 3.5% leaves very little room and the euro area therefore had to squeeze nominal demand considerably harder than the US did to reach the same inflation outcome. It did, and it worked. That European trend growth of barely 1% is dismal remains a damning verdict on European supply-side policy, but it is not a verdict on European monetary policy.

There is an irony here I cannot resist. In 2011 the ECB hiked twice in response to an energy price rise caused by a supply shock and very nearly killed the euro doing it. In 2026 it has tightened citing energy once more and happens this time to be right, not because the reasoning about oil has improved but because nominal spending was where it should be rather than collapsing.

Unfortunately the first half of 2026 is running at something like 4% nominal growth in the euro area on my estimate, so nobody in Frankfurt should be relaxing either. But the difference is one of degree and the degree is large, because an overshoot of half a percentage point is a monetary policy worth watching, while an overshoot of three and a half percentage points is a monetary policy that has lost its anchor altogether.

What happened on Wednesday

The FOMC left the funds rate unchanged at 3.5-3.75% on 29 July, for the fifth consecutive meeting, and three members dissented in favour of a quarter-point increase, which is the most dissents in one direction in a decade.

What followed was a press conference about forward guidance and about whether the chairman should publish his own projection.

Warsh was asked about a monthly consumer price print that had fallen 0.4%, and he answered that it was not much of a consideration for him while calling inflation elevated in the same breath. It is hard not to find Warsh’s performance internally contradictory.

My optimism has cooled sharply

I was optimistic after Warsh’s first press conference, and I said so at the time. I had been sceptical of him for months, because he spent last year campaigning for the job by telling the White House what it wanted to hear about lower rates and an AI-driven productivity boom.

And then his first meeting produced a much shorter statement, no forward guidance at all, and a single blunt sentence committing the Committee to deliver price stability, which is a large part of what I have argued for since I started writing about monetary policy.

I have not totally changed my mind about Warsh. But that optimism has cooled very sharply, and the reason is not the one people assume. The problem is not that he held rates, because holding is a decision like any other and there are meetings where it is the right one.

The problem is that he held rates while nominal spending grows at 7.9%, and then declined to say which number would make him act. A chairman who places credibility above every other virtue has now had two opportunities to tell us what he is targeting, and he has used both of them to explain why he will not.

Of course he is entitled to dislike forward guidance, and in fact I dislike it too. But refusing to pre-commit to a path for the policy rate is one thing and refusing to name the variable you are trying to control is quite another. The first is discretion about instruments. The second is discretion about objectives, and giving that up is what makes a central banker rule-based rather than merely well-spoken.

The pressure is real, and so is the arithmetic

US President Donald Trump has been explicit and public about what he wants from the Fed, including on the day of the meeting itself, when he backed the chairman and called the institution political in the same breath. The debt stock is around 100% of GDP, and the previous chairman faced a criminal referral for refusing to cut.

Unfortunately that configuration has a name. When the fiscal authority moves first and the monetary authority is left to finance whatever gap remains, controlling inflation stops being a question of will and becomes a question of arithmetic, and the arithmetic is unpleasant. So some part of what we are watching is a Fed that would like to tighten and finds that it cannot, and that part will grow rather than shrink as debt service rises.

But this is where I part company with the fatalists. Unpleasant arithmetic is a constraint rather than an excuse, and it binds through expectations, which are exactly what a central bank chairman is paid to manage. A chairman who names his target, explains his reaction function, and then demonstrates that he will act on it makes the fiscal authority’s problem harder and his own easier, while a chairman who names nothing at all makes precisely the opposite trade.

Hence the refusal to give guidance is not a technical preference about communication. It is the surrender of the one instrument that still works.

What I am asking for, and what it costs

I am not asking the Fed to bring nominal GDP back to the pre-2020 path. That path now sits some 15% below the actual level of nominal GDP, and closing the gap would require a contraction severe enough to produce a deep recession and, in all likelihood, a financial crisis alongside it. The mistake of 2021 cannot be undone, because it has been baked into the price level, and pretending otherwise is not toughness. It is vandalism.

What I am asking for is much smaller. Stabilise the growth rate, bring nominal spending back to roughly 4-4.5% a year and keep it there, say so publicly, and then let the market do most of the work of getting you there.

But I should be honest about the cost. Nominal growth has to fall by three and a half percentage points, and a slowdown of that size cannot be promised painless, since some of it comes out of prices and some out of output depending almost entirely on whether the announcement is believed.

Hence the argument for saying it clearly and saying it soon. The longer nominal spending runs at 8%, the more of the eventual adjustment lands on real activity rather than on inflation. And 2024 showed that the target is achievable, because the Fed came within two tenths of a percentage point of it without a recession.

Volcker, Greenspan, Burns and Miller

Let me be precise about what impressed me in June. The emphasis on rules rather than discretion, the scepticism about how far the Fed’s remit had expanded, the refusal to be sorted into either the hawk box or the dove box, and the insistence that money has something to do with monetary policy. That is the language of the Volcker and Greenspan era, and it had been absent from the Eccles Building for a very long time.

But six weeks and one meeting later, the language is still (partly) there and the policy is not.

Volcker’s distinguishing feature was not that he talked about credibility, because every central banker does that. It was that he named a nominal target, accepted an enormous real cost in order to hit it, and did not flinch when the US president’s party lost seats over it. Burns and Miller talked about credibility too, and what they actually did was let nominal spending run while explaining, meeting after meeting, that the inflation was caused by oil, by unions, by food prices, by anything at all except the money they were printing.

So Warsh has now had two meetings. At the first he asked for a good fight and got one, and at the second he got three votes for tightening from his own committee, which he treated as institutional health rather than as a signal that his colleagues think he is late.

Nominal GDP is growing at 7.9%, trend real growth is 2.4%, and the implied inflation embedded in the current stance of US monetary policy is above 5% and rising in a year when real output is growing below trend.

If nominal GDP continues to growth at close to 8% annually then US inflation will soon be heading forward 5% or more.

Warsh, you are way behind the curve. And the longer you refuse to say what you are targeting, the more the comparison stops being Volcker and Greenspan and starts being Burns and Miller.

If you look at the financial markets, you can “read” what investors are thinking about the economic outlook – are the shocks positive or negative, and are they supply or demand shocks?

Over the past week three things have happened, all of which may have moved the markets:

1. A deal between the Trump administration and the regime in Iran to halt the fighting and gradually reopen the Strait of Hormuz.
2. The new Fed chairman, Kevin Warsh, has held his first FOMC meeting – and the signal was clear: “The Committee will deliver price stability” – the implication being that inflation is above the 2% target, and the Fed is going to do something about it.
3. SpaceX has been floated – the largest IPO in history. It went well for the first few days, but the past few days have not been pretty.

The first event is a positive supply shock – and oil prices therefore fall. That is good news for the world economy.

The second is, by contrast, a negative demand shock – the Fed is signalling clearly that rate rises have moved much closer. A negative demand impulse sends both growth and inflation down. Lower growth is not good for oil prices either, so down they go a little further.

And finally SpaceX – this is the story of an equity market that has become too expensive. It is neither a demand nor a supply shock, but it can develop into a negative demand shock if it leads to financial distress, in which investors increase their demand for “safe assets” – primarily dollars and government bonds.

If that happens, a larger equity-market correction should produce a stronger dollar and lower bond yields (and lower inflation expectations).

1. Lower equities (pointing to overvaluation of the equity market and/or a negative demand shock).
2. Lower inflation expectations and lower yields (a positive supply shock and/or a negative demand shock).
3. A stronger dollar (tighter monetary policy and/or rising risk aversion as a result of the repricing of the US equity market).
4. Lower oil prices (tighter monetary policy, lower growth and/or a positive supply shock).

Taken together, then, it looks as though this is mostly about Kevin Warsh – and, to some extent, Elon Musk’s space project.

Investors are reacting to the fact that Warsh was considerably more hawkish than expected. The Fed has a clear mandate to secure price stability, and it has not lived up to that mandate for more than five years. But Warsh and the rest of the FOMC now appear to be sending a clear signal that something will be done about it. Credibility is to be restored.

That also fits, incidentally, with what we are seeing in the yield curve – long yields, which measure growth and inflation expectations, are falling relative to short yields, which are governed more by expectations about monetary policy.

There is, by contrast, not much effect from Trump’s “peace deal” with the Iranians. If that had dominated, equities should have risen. That has most certainly not happened.

“If he’s alive or dead it doesn’t matter. If he’s dead, just prop him up and put some dark glasses on him like, like ‘Weekend at Bernie’s.’”

That was John McCain on Alan Greenspan in 2007, about a year and a half after Greenspan had stepped down as chairman of the Federal Reserve. The joke worked because it was almost true. For most of two decades, the markets behaved as though Greenspan’s mere presence was the policy.

Greenspan took over the Fed in 1987 and stood at the helm of the world’s most powerful central bank for nearly twenty years. He became the symbol of American economic primacy, and on the trading floors he was treated as a rock star – bigger than any other economist or policymaker in the 1990s and the early 2000s.

I have often been critical of Greenspan’s economic thinking. But the death of a man who shaped the world economy for a generation is a moment for fairness, and in his case the fair assessment is far more interesting than either the hagiography or the easy dismissal.

The Man Who Made Low Inflation the Fed’s Job

Greenspan cemented two things we now take for granted. The first is central bank independence. The second is that securing low and stable inflation is the Fed’s responsibility, and nobody else’s.

That was not obvious when he arrived. Arthur Burns, who ran the Fed in the 1970s, believed inflation was a cost-push problem of union power and a permissive welfare state, to be managed through “incomes policies” – direct government meddling in private price-setting. Greenspan rejected all of it.

Inflation is always and everywhere a monetary phenomenon. Greenspan ran the Fed as if he believed exactly that, even while he refused to say so plainly.

And he refused to say almost anything plainly. “Since I’ve become a central banker, I’ve learned to mumble with great incoherence,” he told a Congressional subcommittee in 1987. “If I seem unduly clear to you, you must have misunderstood what I said.” It became the most quoted thing he ever said, and it was a tell – the cryptic surface concealed a very clear underlying rule.

Leaning Against the Wind, with Credibility

The rule has a name, and the name comes from my old friend Bob Hetzel, the one of the most respected monetary economists the Federal Reserve ever produced, and a man who wrote his Chicago thesis under Milton Friedman.

Hetzel’s history of the Fed divides the postwar period into competing versions of a single procedure that William McChesney Martin invented in the 1950s and called “leaning against the wind” – raising the policy rate when the economy runs hot, cutting it when the economy weakens.

There are two ways to run it. Under Martin and later Burns, the Fed practised what Hetzel calls leaning against the wind with trade-offs. It waited for inflation to appear in the data, treated low unemployment and low inflation as competing objectives, and tried to buy one with the other. It failed. The 1970s were the bill.

Volcker and Greenspan ran the other version – leaning against the wind with credibility. The Fed moved the funds rate preemptively, before inflation showed up, and in doing so it handed the price system back the job of setting employment and output. Hetzel dates this regime from 1979 to 2006, and it produced the Great Moderation.

The defining moment was 1994. Greenspan raised the funds rate sharply with no inflation visible in the data, purely to convince the bond market that he meant it. The “inflation scare” passed, and the Fed’s credibility was secured. That, after all, is what credibility buys you: the freedom to act on a forecast rather than wait for a fact.

Take a look at what actually happened.

The funds rate doubled, from 3% to 6%, across 1994. The 30-year yield jumped first – the inflation scare – peaking above 8% as the bond market challenged the Fed. And then it fell back below 6%, even as the funds rate stayed high. The market tested Greenspan, lost, and capitulated. Long yields fall when the market believes you, not when you stop tightening.

This is, in my view, Greenspan’s real achievement, and it is the one most of his admirers misread. It was not his opacity. It was not his famous “feel” for the data. It was a rule – preemptive, forward-looking, and disciplined by the bond market – and he had the discipline, for a while, to follow it.

Greenspan Understood Nominal GDP Better Than He Let On

Here is the part almost nobody remembers. I dug it out years ago and wrote about it on this blog in 2013.

At an FOMC meeting in November 1992, Greenspan told his colleagues that the old monetarist philosophy would have required 4.5% money growth to drive nominal GDP at 4.5%, but that he was “basically arguing” the Fed was using a nominal GDP goal directly, with the money-supply relationships as mere technical mechanisms. And he added that you would know the markets were convinced about price stability when the 30-year Treasury sat near 5.5%.

Read that again. Greenspan was describing nominal GDP targeting, by name, in 1992 – two decades before a small group of economists – including myself – took it up as a cause.

We – what later became known as the Market Monetarists – never claimed it was new. That was the whole point. Nominal GDP stability was what had produced the Great Moderation in the first place, and our argument was simply that making it the Fed’s explicit, rule-bound mandate was the way to restore that stability after it was lost in 2008. Greenspan had been running, by instinct, the thing we wanted written into law.

Through the Great Moderation nominal GDP grew at about 5.5% a year. Potential real output grew at roughly 3%. The 30-year Treasury yield sat near 5%. Inflation ran close to 2.5%.

That is what price stability looks like when it sits on top of a fast-growing supply side. The numbers moved together, year after year, for the better part of two decades.

The chart above makes the point. Actual nominal GDP tracked a steady 5.5% growth path for the better part of two decades. That 5.5% line was never an announced target – it was a de facto one. And sitting on top of roughly 3% real growth, it delivered exactly the low, stable inflation Greenspan is remembered for.

Five and a Half Then, Four Now

The number that secured price stability in Greenspan’s era will not secure it today. And that is not a change of doctrine. It is the same arithmetic applied to a slower economy.

Nominal GDP growth minus trend real growth is, roughly, inflation. In the Greenspan years the sum was about 5.5% nominal, 3% real, and 2.5% inflation. The Fed could run nominal demand at 5.5% and still deliver price stability, because the supply side was growing at 3%.

That extra point of real growth is gone. The Congressional Budget Office now puts US potential growth a little above 2%, drifting toward 1.8% by the 2030s. Run nominal GDP at 5.5% against that, and you do not get 2.5% inflation. You get something closer to 3.5%.

So the right anchor today is lower – nearer 4% nominal GDP growth than 5.5%. In my view a 4% path against roughly 2% trend real growth delivers the same 2% inflation Greenspan delivered. The target moves because the economy underneath it moved.

This is what the nostalgia for Greenspan misses. He was not more relaxed about inflation than a central banker should be now. He simply had 3% real growth to work with, and his successors have barely 2%. The discipline is worth copying. The number is not.

Where the Market Knew Better

But Greenspan trusted his own judgement more than he trusted the market, and that is where it went wrong.

In a post – There never was a bond market “conundrum” – some years ago (2012) I ran a simple thought experiment. A nominal-GDP-targeting Greenspan in 2005 would have said something like this: we have raised rates 150bp, inflation expectations are well contained, long yields are near 4%, and since we are targeting a 5% path for nominal GDP, there is a real chance we have overdone the tightening.

That is not what he said. The real Greenspan questioned the market’s judgement. He was more optimistic about future nominal GDP growth than the bond market was – and the bond market, it turned out, was quietly forecasting a sharp slowdown. The market was the better forecaster. Greenspan was not.

This is the recurring weakness of even the best discretionary central banker. The rule works precisely because it constrains the judgement of the person running it. Greenspan followed the rule when it agreed with his instincts and overrode it when it did not.

The Crisis Came After He Left

Greenspan is usually blamed for 2008. The standard charge is that easy policy in the early 2000s pumped up house prices and equity prices, and that the resulting imbalances blew up in the crisis. I agree with this – partly. Only partly.

I have never believed that a central bank can reliably spot an asset bubble in real time, still less that it should set the policy rate to deflate one. That is not what monetary policy is for. The Fed has one nominal anchor to manage, not a portfolio of asset prices to police.

And here is the detail the bubble story ignores. Greenspan left the Fed at the start of 2006. The catastrophic error of 2008 – keeping policy far too tight relative to the natural rate of interest while nominal GDP was collapsing – happened on someone else’s watch.

Hetzel’s verdict on the Great Recession is that it was monetary disorder, not market failure. The Fed fell back into its old ways in 2007 and 2008, too focused on headline inflation, too slow to see that the natural rate of interest had gone deeply negative. The Great Moderation did not die of natural causes. The Fed killed it.

So the honest charge against Greenspan is narrower than the popular one. He did not cause 2008. But he let the discipline of the rule slip in his final years, and his successors finished the job.

And this points to the failure that is genuinely his. Greenspan ran a nominal GDP rule, but he never wrote it down. He kept it as instinct, as judgement, as the Maestro’s feel for the data – and instinct does not survive the man who has it. Had he turned his de facto target into a transparent, explicit rule that his successors were bound to follow, 2008 need not have happened. The rule was sound. What was missing was the constitution that would have outlived him.

It Happened Again

We know the cost of leaving the rule unwritten, because the Fed paid it twice.

From 2010 to 2019 the Fed ran another de facto nominal GDP rule – a 4% path. I have argued for years that this is what the data show: from 2009 onwards, US nominal GDP behaved as though the Fed had a 4% level target, tracking it almost perfectly, quarter after quarter, just as Greenspan’s Fed had tracked 5.5%. And just as before, it was never written down. It was the practice of a particular set of people, not a commitment the institution was bound to.

So when COVID came and with that massive fiscal and monetary expansion, there was nothing to anchor to.

Nominal GDP did not return to the 4% path once the shock passed. It exploded – growing at about 10.5% a year across 2021 and 2022, a one-off step up in the price level that the Fed never reversed. That was the inflation of 2021-22. And it was not bad luck. It was the predictable cost of running a good rule informally.

An explicit 4% nominal GDP level target would have caught it. The rule would have told the Fed, in early 2021, that nominal demand was running far above the path and had to be pulled back – and it would have acted before inflation became entrenched, not eighteen months too late. The catastrophe could have been avoided, or at the very least made far smaller.

Two regimes, three decades apart. The same de facto rule, the same success while it held, and the same failure the moment discretion took over. That is the case against the Maestro model in a sentence.

A Quieter Achievement: Ending the Draft

There is one part of Greenspan’s career that has nothing to do with monetary policy, and it may be the part that did the most direct good.

In 1969 President Nixon convened the Gates Commission to study whether the United States could end military conscription. Greenspan served on it alongside Milton Friedman, W. Allen Wallis, William Meckling, and Walter Oi – a remarkable concentration of free-market economic talent. They were the intellectual core that made the case for an all-volunteer force.

The argument was elegant, and it was Greenspan’s kind of argument. The draft, the economists showed, was not free. It was a hidden tax levied on the young – the gap between what a conscript was paid and what he would have had to be offered to serve willingly. A volunteer army, properly paid, was not only more just. It was more efficient, because it let the price system do what conscription did by coercion.

In February 1970 the commission recommended, unanimously, that the draft be abolished. The economists won the argument. Conscription ended in the United States in 1973, and more than fifty years later the all-volunteer force is so settled that few Americans remember it was ever in doubt.

Consider what that meant. Generations of young Americans were spared being compelled into uniform against their will, on the strength of an economic argument that Greenspan helped make. He spent most of his life as the most powerful central banker in the world. But on the Gates Commission he was something rarer – one of a handful of economists who used the discipline to expand human freedom directly. That, too, is the Friedmanite in him: the conviction that the price system allocates better, and more humanely, than the state’s compulsion.

The Rule and the Man

No one is going to prop Greenspan up, fit him with dark glasses, and install him back at the Fed. But the thing worth keeping is not the man. It is the rule he ran.

A central bank does not need a Maestro. It needs a rule that is preemptive, predictable, and credible – and the discipline to follow it. Greenspan had the rule, and for most of two decades he had the discipline too. That is no small thing to leave behind.

A central bank does not need a Maestro. It needs a rule – transparent, explicit, and binding on whoever holds the chair next. We have now watched the same lesson twice: a good rule, run on instinct, that worked until the instinct left the room. In 2008, and again in 2021, the cost of never writing it down came due.

So here is the fair verdict. The lesson of Greenspan’s career is not that we should hope for another like him. It is that we should never have to. Build the rule well enough, and the Maestro becomes unnecessary. He will be remembered as the Maestro. He deserves, just as much, to be remembered for the rule he ran – and to stand as the warning about the rule he never quite dared to write down.

Kevin Warsh’s first press conference as Fed chairman was far better than I expected.

The statement is shorter, forward guidance is gone, and the Committee has promised to deliver price stability. That is most of the prescription I have made for at least fifteen years. But the question that actually matters – what the Fed should be trying to deliver, and the wreck of a regime Powell has left him – is still open.

Kevin Warsh held his first press conference as chairman of the Federal Reserve yesterday, and it was immediately clear that he has put himself firmly in the driver’s seat.

The statement was unlike anything Jerome Powell ever signed off on. It was substantially shorter. It dropped the elaborate “forward guidance” that had been the house style of the Fed for the better part of two decades. And in its place came a single, blunt sentence: “The Committee will deliver price stability.”

I have to admit I did not expect to be impressed. I have been a sceptic of Warsh for months, and I have said so in writing. So before I explain why he changed my mind, let me explain why I doubted him.

Why I Was Sceptical

Warsh did not look, until yesterday, like a man who would face down the White House and put the Fed’s credibility first.

He spent last year campaigning for the job, and he campaigned by telling the White House what it wanted to hear. He was outspoken in favour of lower rates, exactly as Trump has been demanding, and he justified it with the claim that an AI-driven productivity boom would expand supply and pull inflation down on its own.

That argument never persuaded me, and not because I doubt what AI might do to productivity. My objection is more basic. A productivity boom is a positive supply shock, and a central bank should not be reacting to supply shocks at all, positive or negative. If AI does raise productivity, the right response is to hold nominal spending on its path and let the gain come through as lower inflation and faster real growth. Cutting rates because supply is improving is using a supply-side story to justify a demand-side easing – exactly the confusion a rule-based Fed exists to remove.

There was also the question of qualifications. Warsh was a Fed governor from 2006 to 2011, the youngest ever appointed, but his record on inflation forecasting in those years was weak, and prominent NGDP advocates, Scott Sumner among them, were openly sceptical that he was the right man. As was I at the time.

And there was the politics. His IMF speech last April read, frankly, like a job application written for Trump’s ear – a list of grievances about institutional drift, wrapped in language about independence that pointed towards less of it rather than more.

There is, too, the family dimension, which under normal circumstances I would ignore entirely. Warsh is married to Jane Lauder, daughter of Ronald Lauder, a long-standing personal friend of Donald Trump.

These are not normal circumstances. The President has spent a year demanding lower rates and used a renovation investigation as a pressure tactic against Powell. When his preferred candidate to replace the chair turns out to be connected to his inner circle by marriage, that belongs on the analytical record.

There is one side of Warsh’s thinking I have always had sympathy for. He has been a consistent critic of the Fed’s mission creep into financial-stability and macroprudential territory. Those are my positions too.

But in May, on the day Powell chaired his final meeting, my complaint was precise: Warsh was full of diagnosis and empty of framework. He knew what he wanted to remove. He had said nothing about what would replace it.

So that was where I stood. Sceptical, unconvinced by the AI story, worried he would be Trump’s man, and waiting to see whether he had a regime in mind or only a grievance.

He did better than that. And that is why I am writing this.

Letting the Market Do the Work

The first thing you notice about Warsh’s statement is the length. It is substantially shorter than anything Powell put out. That is not cosmetic. The old statement told you what the Committee thought, how it read the data, and where it expected to move. The new one gives the target range, reaffirms ample reserves, notes inflation is still above 2%, and states – flatly – that the Committee will deliver price stability.

Warsh described it himself with understatement: a bit shorter, a bit simpler, dispensing with some older language. Gone, in particular, is forward guidance. And he declined to submit his own dot to the projections, on the grounds that a published rate path locks the central bank into a promise it cannot keep.

Clearly, this is the right instinct. So let me give it a name.

Years ago I defined the Chuck Norris effect like this: you do not have to print more money to ease policy if you are a credible central bank with a credible target.

The point generalises. A credible central bank does not have to do very much, because the market does the work for it. If everyone believes the Fed will deliver, expectations adjust on their own, and the threat to act is usually enough.

Warsh has not announced a rule. But by stripping the statement to a commitment and refusing to spell out a path, he is telling markets to read his reaction function rather than his quarterly mood. He is asking the market to do the lifting.

In its purest form, a fully credible Fed would hardly need to hold a press conference at all. It would be enough to say, once and for all: we will always, at any time, do exactly what is needed to deliver price stability.

The market does the rest. It is the market that implements the policy, and it is the market that makes the real forecasts, through its expectations.

The comparison that comes most naturally to me is not American. It is Danish. Denmark, where I live, runs a fixed-exchange-rate policy against the euro. The Danish central bank – Danmarks Nationalbank – publishes no elaborate forecasts and no guidance. It does not need to.

Everyone knows it will do whatever it takes to hold the krone fixed against the euro. That credibility is the policy, and the market enforces it. Warsh is now saying something structurally identical: we will deliver price stability, you do not need our forecasts, you need only believe us.

Contrast the opposite approach. For years the ECB under Mario Draghi, and Jean-Claude Trichet before him, insisted it would “never pre-commit” to any future action. I always thought that was precisely the problem. A central bank that refuses to say what it is trying to achieve forces the market to guess, and guessing is not a credible target.

Mission Creep and the End of Fine-Tuning

This is also the most refreshing thing about Warsh’s debut, and it goes beyond the statement. The Fed has spent the better part of fifteen years trying to do too much.

Since 2008 it has drifted into questions that have nothing to do with its mandate – financial-stability ambitions, distributional concerns, even climate-related risk. The mandate is price stability and maximum employment. It is not the management of every problem that happens to be fashionable.

So the narrowing Warsh implies is welcome.

There is an old principle, associated with Jan Tinbergen, that each policy target needs its own instrument; a central bank chasing five objectives with one rate will fail at most of them. Goodhart’s Law adds the rest: when a measure becomes a target, it stops being a good measure – which is exactly the trap macroprudential policy walks into.

Warsh, to his credit, seems to grasp both. The Fed has one instrument and should pursue one nominal objective. Everything else is someone else’s job.

This is, fundamentally, an argument about rules versus discretion. The Great Recession and its aftermath were an exercise in central-bank discretion – improvisation dressed up as sophistication. I have always thought we should go the other way, towards a Fed that behaves automatically, like the computer Milton Friedman wanted to run monetary policy. All central banks need to do this.

A Fed that fine-tunes is a Fed that surprises. A Fed that follows a rule is a Fed the market can anticipate. And a Fed the market can anticipate barely has to act at all.

Nominal Stability, Not Price Stability

But a great deal is still missing, and Warsh knows it. Which instruments should the Fed actually use – the money base, or the policy rate? Which measure of inflation is the right one? Should there even be an inflation target at all? My own answer, and my favourite, is a target for nominal GDP. So this is where my renewed optimism meets the hard question. Clarity is worth little if the target is wrong.

Warsh committed the Committee to price stability, which the Fed reads as 2% inflation. Notice what the statement leaves out. The mandate is dual – stable prices and maximum employment – and the new statement mentions only the first. It is a price-stability statement, not a dual-mandate one.

And here I would go further than Warsh did. He committed to price stability. I would commit to nominal stability instead, and the two are not the same thing.

Nominal stability – a stable path for total nominal spending – delivers price stability over time. Price stability pursued directly does not necessarily deliver nominal stability at any given moment. You can hold prices down in the short run by letting nominal spending collapse, or by tightening into a supply shock. That is not stability. It is the opposite, dressed up as discipline.

So it is not really prices a central bank should stabilise. It is nominal spending. Get the nominal path right, and prices look after themselves over the medium term.

My answer to the target question has not changed for more than fifteen years.

The Fed should target the level of nominal GDP – a 4% path, consistent with 2% inflation and around 2% trend real growth. It has two advantages over an inflation target, and both bite right now.

First, it handles supply shocks: a level target lets inflation rise temporarily after an energy shock while holding nominal spending on its path, instead of tightening into it the way the ECB infamously did in 2011.

Second, it makes up for past misses. An inflation target lets bygones be bygones; a level target claws the overshoot back. That is the difference between an anchor that holds and one that drifts.

Now look at what Warsh is staring at. Consumer price inflation hit 4.2% in May, the highest in three years. Producer prices ran above 6%. The Iran war drove oil from around $67 in late February to an intraday peak near $119 in March. The Committee’s median projection now has the policy rate ending the year at 3.8%, a quarter point above today, where three months ago it expected a cut.

So the dot plot has flipped from a cut to a hike, and it is hard to read Warsh’s debut as anything other than a hike now being more likely than a cut.

Here is the part most commentators will miss. The current inflation has two sources, not one. Part of it is the energy supply shock – and a central bank should look through that, not tighten into it.

But part of it is excess nominal demand the Fed never wound back – nominal spending that has run above trend since 2021 and is still drifting higher, as I will show in a moment. A pure inflation-targeter cannot tell the two apart and risks getting both wrong. A nominal GDP target separates them automatically – look through the oil, lean against the excess nominal demand. It is the one framework that gets this moment right.

This, by the way, is why the whole “hawk versus dove” framing is, frankly, useless. To be hawkish or dovish is already to assume policy should be set by someone’s discretionary feel.

What should concern us is not the temperature of the chairman. It is the regime.

The Five Task Forces

In May my complaint was that Warsh offered no framework. Yesterday he began building one, or at least the machinery to build one.

He announced five task forces, each on a core area of policy: communications, the balance sheet, data and data sources, productivity and jobs in an era of transformation including AI, and inflation frameworks examined from first principles.

Each one, Warsh said, would draw on the best minds inside and outside the economics profession, backed by Fed staff, and report to the policymakers.

Two things stood out by their absence. There was no hint – none – of the easiest dovish escape available to him: quietly raising the inflation target, or hiding behind its “flexible average” wording, to make five years of overshoot vanish on paper.

He could have conjured the problem away. He chose not to. And he barely mentioned AI, even though it has a task force of its own. The man who last year leaned on an AI productivity boom to justify the rate cuts Trump wanted did not reach for that argument once.

Which leaves the question I keep coming back to: who. If the inflation-frameworks task force is serious about first principles, the names I would want in the room are obvious – Scott Sumner, David Beckworth, Peter Ireland, and Josh Hendrickson on nominal GDP and policy rules; George Selgin on the balance sheet, where the ample-reserves regime is his territory; and, for any of it, my friend Bob Hetzel, the finest monetary historian the Fed has produced ever. Appoint people like that and “from first principles” might mean something. Appoint the usual committee and it will not.

What I Have Proposed

So let me be concrete about what I would actually have the Fed do, because I am not asking the task force to invent anything. I wrote it down almost exactly ten years ago, in January 2016, and I would not change a word.

First, a 4% nominal GDP level target, defined on the expected level of NGDP eighteen to twenty-four months out. That one target delivers both price stability and maximum employment. No other is needed.

Second, the Fed should stop producing its own forecasts and instead read the market’s: a prediction market for NGDP twelve and twenty-four months ahead, the surveys of professional forecasters, and market-based models of NGDP expectations. Let the market tell the Fed where nominal spending is heading, not the other way round.

Third, give up interest-rate targeting – the dot plot included – and use the monetary base as the instrument. Announce a permanent base-growth rate set to hit the 4% path, justified by one number only: expected NGDP against the target. Leave interest rates entirely to the market.

The consequences follow directly.

The policy is rule-based, not discretionary. It is transparent, with the market doing most of the implementation – the Chuck Norris effect made operational.

There is no zero-lower-bound problem, because control of the base can ease even when interest rates sit at zero. The endless and, frankly, silly talk of bubbles, moral hazard, and macroprudential fine-tuning stops, because the Fed gives up pretending it can do a better job than the market in ‘forecasting’. The Fed stops reacting to supply shocks, positive and negative. And the FOMC could, at last, be handed to the computer Milton Friedman wanted to run it.

I called it, at the time, a forward-looking McCallum rule. The label matters less than the shape of the thing: one target, read from the market, hit with one instrument, justified by one number.

Look at Warsh’s five task forces and you will see the same agenda, broken into committees – inflation frameworks is my first change, communications and data my second, the balance sheet my third. The only real question is whether they follow the framework to where it leads, or spend a year rediscovering why the Fed has always preferred its own discretion.

What He Inherited

All of which raises the obvious question. Why does any of this matter so much? Why is the regime question existential rather than academic?

The answer is the Fed that Warsh has inherited. Because it is not a healthy institution. It is a weakened one, and the weakness is to a large extent self-inflicted.

Trump has spent the past year arguing that Powell ran policy too tight. The data says the opposite.

By my reckoning, Powell ran the easiest monetary policy in modern Fed history, and the cleanest way to see it is not through inflation or the funds rate but through the level of nominal spending – nominal GDP.

From 2010 through 2019, US nominal GDP grew at almost exactly 4% a year. That was the de facto regime under Ben Bernanke and Janet Yellen, broadly consistent with their stated 2% inflation goal given trend productivity and labour-force growth.

I have argued for years that this implicit anchor, and not the dual mandate the Fed talks about, was what actually held the system steady. The Covid shock knocked nominal spending below the path in 2020, and the Fed’s emergency response that spring was, in my view, exactly right.

Then came 2021.

On 29 April 2021, I published a post on this blog titled “Heading for double-digit US inflation.”

I argued that the explosion in US broad money, combined with the largest fiscal expansion since the Second World War, would produce a fast, large, one-off jump in the price level – and that what happened next would depend entirely on whether the Fed moved to anchor expectations or let its credibility erode.

The forecast was right. Headline CPI inflation peaked at 9.1% in June 2022, the highest in four decades. The point is not that I got it right. The point is that the call was available to anyone willing to read the data.

The broad-money numbers were public. The fiscal arithmetic was public. The lags between money and inflation that Milton Friedman identified decades ago are textbook material.

And the Fed, under Powell, chose to do nothing.

Worse than nothing. In 2020 the FOMC had adopted Flexible Average Inflation Targeting, which committed it to running inflation above 2% for a while after running it below. Defensible in theory. A catastrophe in practice.

Through 2021 and into 2022 the Fed held the funds rate at zero and kept buying $120 billion of assets a month while nominal GDP grew 11% in 2021 and close to 10% in 2022 – more than two and a half times the established 4% rate. That is not flexible average inflation targeting. It is the monetary equivalent of price-fixing, and all the while the Fed insisted the inflation was transitory, a word it kept using long after the indicators had made it indefensible. This was not a close call.

Powell then spent 2022 and 2023 cleaning up the mess, and he deserves credit for it. From near-zero in March 2022 the Fed reached 5.25-5.50% by July 2023. Inflation came down, the labour market did not collapse, and that successful disinflation is the one thing from the Powell era I will give him unambiguous credit for.

But the disinflation did not restore the regime. The price-level jump is permanent – undoing it would take a Volcker-scale recession, and with federal debt above 100% of GDP and interest payments above $1 trillion a year, that is no longer arithmetically possible without risking a sovereign default.

The growth rate, though, is a separate matter. Having allowed the one-off jump, the Fed could at least have re-anchored nominal spending at 4% from there. It did not. On my numbers, NGDP growth has averaged about 5.4% a year since the start of 2023, and from the start of 2025 it has run above 6%. The rate is drifting up, not settling.

So the Fed of 2025 and 2026 is not running a tight policy that has restored discipline. It is running a policy that still accommodates above trend. The “tightening” belongs in scare quotes: it is tight relative to the Fed’s own 2021 error, not relative to any credible framework. This is the excess nominal demand I pointed to earlier. The Fed is still too easy.

Why the Attack Lands

So why, if Powell ran the easiest money in modern history, does Trump’s accusation that he was too tight gain any traction at all?

It lands because the Fed proved, in real time, that it cannot be trusted to act on the inflation indicators when its own framework tells it not to.

That is the credibility deficit Warsh inherits, and it is worth more to Trump than any number. A Fed that had read the money data correctly in 2021 and headed off the inflation would now be defending its independence from overwhelming strength. Trump’s attacks would land in empty air. Instead the Fed defends independence from partial credibility, having handed the public a 9% inflation to absorb. The easy money of 2021 and the political crisis of 2025-26 are not two stories. They are one.

The historical rhyme is the early 1970s. Nixon leaned on Arthur Burns – the tapes record instructions, not requests, to expand the money supply before the 1972 election – and Burns complied, explaining at each meeting why the next move could wait. Each decision sounded reasonable on its own. The cumulative result was a decade of inflation and, eventually, a Volcker disinflation that drove unemployment towards 11%.

The parallel to Powell is only partial. He raised rates hard, the disinflation was real, and he resisted Trump’s public pressure throughout. On the political dimension he was not Burns – or rather at least not quite as bad a Burns.

But on the analytical dimension – the willingness to ignore the monetary indicators because the prevailing framework said they did not matter – the parallel is uncomfortably close. Burns ignored the aggregates because his framework dismissed them. Powell ignored them in 2021 because FAIT said the inflation was wanted. The justification differs. The shape of the failure is identical.

And underneath all of it sits the deeper threat: fiscal dominance.

Federal debt held by the public is above 100% of GDP in the US.

Net interest has gone from around $350 billion in 2020 to over $1 trillion in 2025. Every 100 basis points on the policy rate eventually adds about 1% of GDP to the interest bill. The federal government now has a powerful, mechanical reason to want low rates regardless of what the macroeconomy needs – and Trump has been unusually explicit that he wants rates down for the budget.

That is fiscal dominance: the fiscal authority’s needs crowding out the central bank’s independence. It is the condition behind nearly every serious inflation in history, from the German hyperinflation of 1923 to the post-Soviet Russian inflation, chronic Argentine inflation, and the Turkish episode under Erdoğan.

Not Trump’s Fed Chairman

So set Warsh’s debut against all of that, and the most important thing he did yesterday is the thing I least expected. He defied the White House.

Trump appointed him after attacking Powell for not cutting, and Warsh had spent the previous year signalling sympathy for exactly those cuts.

The natural bet was that he would ease, or at least signal easing – the cuts the President wanted. Instead he held rates on a unanimous vote, let the dots tilt towards a hike, and gave a press conference the markets took as anything but the relief Trump was hoping for.

The S&P 500 fell. Bond yields jumped. Asked whether he had spoken to the President since taking the job, he said only: “I don’t have anything for you.” He has a mandate, and he intends to deliver on it. Whatever Trump says.

Set that against everything stacked the other way – the campaign for lower rates, the IMF speech, the Lauder connection, and the fiscal-dominance pressure that gives any modern president a mechanical hunger for cheap money. On his first day, with all of that pushing one way, Warsh pushed back. This is the Arthur Burns fear answered, at least for one meeting.

And there is a deeper point underneath it. The more rule-based and automatic policy becomes, the less it matters who the chairman is, and the less influence any president has over him. A discretionary Fed can be lobbied, pressured, and bullied. A rule-based Fed cannot.

Taking discretion out of monetary policy is not only better economics. It is the strongest protection of central bank independence there is – which, with fiscal dominance bearing down on the Fed, is no small thing.

Warsh also made a point of noting that inflation has now run above the 2% goal for more than five years. He inherits that overshoot, the same five years and the same failure. If he means what he says, it has to end. A central bank that overshoots for five years and shrugs has no credibility, and without credibility none of this machinery works.

The Hard Part Has Started

So I come away more optimistic than I expected, and far more than I was a month ago. The communication is sharper. The mission creep is being reversed. He defied Trump on day one. And the instinct – say what you will deliver, then let the market do the work – is exactly right.

But the easy part is the part Warsh has done. Shortening a statement, dropping a forecast, and holding the line against Trump for one meeting is style and nerve. Choosing the right thing to be clear about is analysis, and it is unresolved.

Friedman taught that monetary policy works with long and variable lags and cannot fine-tune the economy. Scott Sumner updated it: long and variable leads, because the work runs through expectations. Both point to the same conclusion. What matters is the regime, not the meeting-to-meeting decisions. A Fed that spent five years being careful about each meeting lost the regime entirely – and that, not any single rate call, is what Powell leaves behind.

So it was never really a rate decision. It is a regime decision. Warsh has made a good start on the things that are mostly nerve. The thing that is analysis – whether the Committee anchors on nominal spending or on an inflation number that has already failed for five years – is still to come. And in practice, reading all of this together, it is hard not to see a rate hike now as more likely than a cut.

The chairman changed in May. The hard part started yesterday.

Trump says Powell has been too tight. The data says the opposite. Powell has run the easiest monetary policy in modern Fed history, and the credibility cost is what makes the institution so vulnerable today

Later today the Federal Open Market Committee will almost certainly leave its target range for the federal funds rate unchanged at 3.50 to 3.75%, and markets are pricing a 100% probability of the hold. The only suspense in the room is whether Stephen Miran dissents again, as he did at the March meeting, in favour of a quarter-point cut.

That is the part everyone agrees on. The interesting question, in my view, is everything else – and the ritual focus on the rate decision actually obscures what is happening in the institution today.

This is, in all likelihood, Jerome Powell’s final meeting as Federal Reserve chair, with his term ending on May 15. Kevin Warsh has been nominated by the President to replace him, and the Senate Banking Committee is voting on the nomination this morning (US time), four hours before the FOMC’s own decision.

The political timing is, of course, not an accident. Last week the U.S. attorney for the District of Columbia, Jeanine Pirro, transferred her investigation of the renovations at the Fed’s headquarters to the Fed’s own office of inspector general – a procedural move that cleared the obstacle Senator Thom Tillis had placed in front of the Warsh nomination.

So here is what Wednesday actually is. It is not a rate decision. It is the end of an era at the Federal Reserve, and the beginning of a regime under conditions that have not existed in the United States since Arthur Burns sat in the chair.

What the dual-mandate framing misses

The Fed’s dual mandate looks roughly balanced if you are willing to squint at it. Core PCE inflation is running around 3%. The labour market is soft but not collapsing – weak hiring, unemployment at 4.3%, and payroll growth that has slowed without breaking. Brent oil is around $110 a barrel. The energy shock is not yet fully fed through to headline inflation.

That is a reasonable summary of the technical position. And it is also, in my view, almost beside the point.

The Fed has been above its 2% inflation target for five years. Five years. There is no recent precedent for a major central bank running this far above target for this long without either delivering the disinflation or formally rebasing the target. The institution has done neither. It has waited. The technical case for waiting is fine. The institutional consequence of having waited for five years is something else entirely.

A picture of the failure

Trump has spent the past year arguing that Powell has run policy too tight. The data tells the opposite story.

The chart below shows nominal GDP, in level terms, from 2010 through the end of 2025. The dashed grey line is a 4% growth trend anchored at the start of the period.

I have argued for years that the post-2008 Federal Reserve under Ben Bernanke and Janet Yellen ran a de facto 4% NGDP target, and that this target was broadly consistent with their stated 2% PCE inflation objective, given underlying productivity and labour-force growth.

Look at what the picture shows.

From 2010 through the end of 2017, US nominal GDP grew at almost exactly 4% per year. The actual line and the trend line are essentially indistinguishable. This is what a predictable monetary regime looks like – the price level and real output combined growing at a steady, predictable rate that markets can plan around.

Powell took over in February 2018.

NGDP continued at roughly 4% growth through 2019. The Covid shock in Q2 2020 produced a sharp negative deviation, and the Fed’s response in March 2020 was, in my view, exactly right. By the end of 2020, NGDP was already approaching the trend line again.

Then came 2021.

The April 2021 forecast

On this day five years ago, April 29, 2021, I published a blog post on this blog titled “Heading for double-digit US inflation.”

I want to quote what I actually wrote, because the framing matters.

I argued at the time that the explosion in US broad money supply, combined with the largest fiscal expansion since the Second World War, would produce a sharp one-off jump in the US price level – not necessarily a permanently higher inflation rate, but a level shift that would be very fast and very large.

I wrote that the fixed-income market was mispricing the inflation risk. I noted that the Fed seemed to be deliberately behind the curve. And I wrote that what would happen after the price level jump would depend entirely on the Fed’s response.

Either the Fed would move aggressively to anchor expectations and contain the shock as a one-off, or it would allow a serious erosion of credibility and longer-term inflation expectations would jump as well.

That was the forecast. The red dot on the chart marks where it was made.

The forecast was correct on the inflation outcome. Headline CPI inflation peaked at 9.1% in June 2022 – the highest rate of inflation in four decades. The forecast was correct on the timing – the price level moved fast and sharply through late 2021 and 2022.

The point is not that I got the forecast right. The point is that the forecast was available to anyone willing to read the data.

The P-star model I used was developed by Hallman, Porter and Small at the Federal Reserve in 1989. The broad money numbers were public.

The fiscal arithmetic was public. The relationship between monetary expansion and subsequent inflation, with the lags Milton Friedman had identified decades earlier, was textbook material. The Federal Reserve had access to all of this information, employed dozens of monetary economists with sophisticated models, and was looking directly at the same data I was looking at.

And the Federal Reserve, under Powell, chose to do nothing.

Worse than nothing. The Fed had, in 2020, adopted a new framework called Flexible Average Inflation Targeting – FAIT – which committed the FOMC to running inflation moderately above 2% for some period after running it below 2%. The framework was defensible in theory. The application in practice was a catastrophe.

Through 2021 and into early 2022, the Fed held the funds rate at zero and continued QE asset purchases at $120 billion per month even as nominal GDP growth ran at 11% in 2021 and nearly 10% in 2022. Pre-Powell average growth was 4%. Powell’s Fed ran NGDP growth at more than two-and-a-half times the established regime rate.

That is not flexible average inflation targeting. That is, as I have written elsewhere, the monetary policy equivalent of price fixing – holding the policy rate so far below its market-clearing level that the resulting inflation was not an accident but a mechanical consequence.

Powell’s Fed told the public, repeatedly, that the inflation that was already arriving was transitory. The word was, by late 2021, no longer defensible on any reasonable reading of the monetary indicators. It was deployed anyway, and the Fed maintained the framing through most of 2021 and into early 2022, well after the empirical case for it had collapsed.

This was not a close call. This was not a difficult judgement. This was the textbook monetary failure – too much money chasing too few goods, with the central bank insisting that money does not matter and that the inflation will go away on its own.

The reversal that did not anchor anything

Powell then spent 2022 and 2023 trying to undo the damage. From a target range of 0-0.25% in March 2022, the Fed reached 5.25-5.50% by July 2023. Inflation came down, and the labour market did not collapse. That was, by historical standards, a successful disinflation – one of the few things from the Powell era that I will give him unambiguous credit for.

But here is where the chart tells a story that the press releases do not.

Look at the three red dotted lines on the chart. Each one is a 4% trend rebased at a different starting point – Q1 2023, Q1 2024, and Q1 2025. They show what NGDP would have done if Powell’s Fed, having allowed the price-level jump, had at least re-anchored the growth rate at 4% from each of those points forward.

The actual path lies above all three.

I am not, to be clear, arguing that Powell should have tried to bring NGDP back to the original pre-2020 trend. That would require at least a Paul Volcker-scale disinflation, and the current fiscal arithmetic – debt above 100% of GDP, interest payments above $1 trillion – makes that mathematically impossible without causing a US sovereign default (more on that below). The price-level jump is permanent. We have to live with it.

But the growth rate is a separate question. After the price-level jump, the Fed could have re-anchored at 4%. It chose not to. From Q1 2023 to today, NGDP growth has averaged 5.4% per year – 1.4 percentage points above the established regime rate. From Q1 2025 alone, the rate is 6.2%. The growth rate is, if anything, drifting higher rather than settling.

So the Fed in 2025 and 2026 is not running a tight policy that has finally restored discipline. It is running a policy that continues to accommodate above the long-run trend. Powell tightened relative to the catastrophe of 2021-22, yes.

But he has not returned the regime to anything resembling the Bernanke-Yellen anchor. The “tightening” is in scare quotes for a reason. It is tightening relative to the Fed’s own previous error, not tightening relative to a credible monetary framework.

The Burns parallel

The historical episode that maps most directly onto the current situation is the early 1970s. President Nixon, facing reelection in 1972, applied direct and well-documented pressure on Federal Reserve Chair Arthur Burns to ease monetary policy. The Nixon tapes contain explicit instructions – not requests, instructions – that Burns increase the money supply ahead of the election.

Burns complied. He explained, meeting by meeting, why the next move would have to wait, and each individual decision sounded reasonable in isolation. The cumulative result was a decade of inflation, requiring a Volcker disinflation that took unemployment to nearly 11% to break.

The Burns analogy applies to Powell only partially. He raised rates aggressively in 2022 and 2023, the disinflation was real, and he resisted Trump’s public pressure throughout the second Trump term. He has not been Burns on the political-pressure dimension.

But on the analytical dimension – the willingness to ignore the inflation indicators when the prevailing framework said not to act – the parallel is uncomfortably close.

Burns ignored the monetary aggregates because his framework said they did not matter. Powell ignored them in 2021 because FAIT said the inflation was wanted. The institutional failure has the same shape, even if the specific theoretical justifications differ.

The cost of the failure is what we are living through

Here is the structural problem the dual-mandate framing does not capture, and which I think is the single most important thing about the current moment.

Trump has spent the past year publicly arguing that Powell has run policy too tight, that interest rates should be lower, and that the Fed has been hostile to the administration. The chart above shows that the empirical case for this is essentially nil. Powell has run the easiest monetary policy in modern Fed history. NGDP growth has been substantially above the regime rate since 2021. The federal funds rate is below where most reasonable models would put neutral, given the inflation conditions of the past five years.

So why does the political attack land?

It lands because the Fed has, over the same period, demonstrated in real time that it cannot be trusted to act on the inflation indicators when its preferred theoretical framework tells it not to. The credibility deficit Warsh inherits is not just a number. It is the gift that the Powell Fed has handed to anyone, including Donald Trump, who wants to argue that the institution does not deserve the deference that has historically been extended to it.

A Fed that had not produced 9% inflation in 2022 – because it had read the broad money data correctly in 2021 and acted on it – would now be defending its independence from a position of overwhelming strength. Trump’s attacks would land in empty air. Markets would dismiss them. Congressional Republicans would be reluctant to follow.

Instead, the Powell Fed defended independence from a position of partial credibility, after producing a major inflation that the public and the political branch had to absorb. The independence defence has held so far. But the defence is much weaker than it would have been, and the institution is much more politically vulnerable than it would have been, because of the 2021 failure.

This is the link between the easy monetary policy of 2021 and the political crisis of 2025-26 that almost no commentator is drawing. They are not separate stories. They are the same story.

The fiscal-dominance story

Federal debt held by the public now exceeds 100% of GDP. Net interest payments on the federal debt have surged from about $350 billion in 2020 to over $1 trillion in 2025.

Every 100 basis points of higher policy rates adds roughly 1% of GDP to the federal interest bill over time. The federal government has a powerful, mechanical reason to want lower rates – independent of whether lower rates are appropriate for the macroeconomy.

Trump has been remarkably explicit about this. He has stated publicly, more than once, that lower rates would help the federal budget. He simply wants lower rates, and I think he wants them for fiscal reasons, not for macroeconomic ones.

This is the condition economists call fiscal dominance – a regime in which the fiscal authority’s needs constrain the monetary authority’s ability to conduct independent monetary policy.

It is, in fact, the structural condition behind essentially every major historical inflation episode I can think of, from the German hyperinflation of 1923 through the Latin American inflations of the 1970s, the post-Soviet inflation in Russia, the chronic Argentine inflation, and the Turkish episode under Erdoğan. The 1970s American and European inflations were milder versions of the same dynamic.

What changes with Warsh

Some have argued that Kevin Warsh will be a hawk because he dislikes QE, and that reading is, in my view, partly right and substantially wrong at the same time.

The right part first. Warsh was a Fed governor from 2006 to 2011, and he has, as a private citizen, been a critic of the Fed’s enlarged balance sheet and of what he has called mission creep into financial-stability and macroprudential territory. Those are positions I have considerable sympathy for.

What economists call The Tinbergen rule says that each policy target requires its own policy tool, and a central bank that pursues price stability, full employment, and financial stability all at once – while adding new macroprudential objectives on top – is going to do at least one of them badly.

Goodhart’s Law adds the further point that when a measure becomes a target, it ceases to be a good measure. The macroprudential indicators the Fed has been adding to its toolkit are precisely the kind of measures Goodhart’s Law was written about.

So a Warsh chairman who shrinks the institutional ambition of the Fed and pushes back on macroprudential creep is, in my view, doing useful work, and I would actively welcome that part of the transition.

The wrong part, and the more important part, is what Warsh proposes to replace it with. And here, in my view, the answer is that he does not propose anything specific.

Warsh’s diagnosis of where the Fed has gone wrong over the past 15 years is the easiest part of the analysis. What is missing is the framework that would replace the discretion he wants to remove.

He has not said what monetary policy rule he wants the Fed to follow. He has not said how he would handle the next zero-lower-bound episode. He has not said whether he prefers inflation targeting, NGDP targeting, price-level targeting, or something else. He has not said how he would respond if the federal government pushed the debt above 130% of GDP and demanded accommodation.

Warsh’s IMF speech from April 2025 was, in many ways, a job application written for Trump’s ear – a list of grievances about institutional drift, paired with suggestive language about Federal Reserve independence that pointed in the direction of less independence rather than more.

There is also, frankly, the family dimension. Warsh is married to Jane Lauder, daughter of Ronald Lauder, who is a long-standing personal friend of Donald Trump.

Under normal circumstances I would consider this irrelevant. But these are not normal circumstances. The President has spent the past year publicly demanding lower rates.

The investigation into Powell was, by most credible accounts, a political pressure tactic. So when the President’s preferred candidate to replace the chair turns out to have a family connection to the President’s inner circle, that fact deserves to be on the analytical record.

The Friedmanite reading

I take a fairly simple lesson from all of this, and it is the point on which I want to land.

Milton Friedman wrote, in his Nobel lecture and many times before and after, that monetary policy works with long and variable lags and cannot reliably fine-tune the economy. Scott Sumner has updated this with the observation that monetary policy works with long and variable leads, because the transmission runs through expectations rather than through mechanical adjustments to short rates.

Both formulations point to the same institutional conclusion: what matters is the regime, not the meeting-to-meeting decisions. A Fed that has spent five years being technically careful about each individual meeting has lost sight of the regime question altogether.

So watch today’s press conference for what Powell does and does not say about his own future. Watch the June meeting for the first signal of what the new regime actually looks like. And watch, after that, for whether the Warsh Fed acts like an independent central bank that happens to share some of the President’s preferences, or like a central bank whose preferences are increasingly the President’s by another route.

It is not a rate decision. It is a regime decision.

The chairman leaves today. The hard part starts in June.

The chief executives of the world’s most powerful artificial intelligence companies have been saying remarkable things about the future of work.

Dario Amodei of Anthropic has warned, in multiple interviews, that AI could eliminate half of all entry-level white-collar jobs within the next few years.

Sam Altman of OpenAI has spoken of AI agents “joining the workforce” and of intelligence eventually becoming “too cheap to meter”, a utility as abundant as electricity, capable of rewriting the rules of the economy.

Jensen Huang of Nvidia, perhaps the most ebullient of the three, has described a future workforce of “humans and digital humans,” predicted that companies will hire and onboard AI agents just as they do people today, and suggested that the current infrastructure buildout is the largest in human history.

These are not idle remarks. They come from people who run companies at the very centre of the AI industry, who speak at Davos and on 60 Minutes and to US congressional committees, and whose words move markets. They deserve to be taken seriously, and that is precisely why they deserve to be examined seriously.

I want to offer three arguments against the prevailing Silicon Valley consensus on AI and jobs. The first is empirical: the evidence does not support the predictions. The second is technical: the predictions rest on a misunderstanding of what large language models actually are. The third is economic: even setting aside the first two objections, the predictions ignore the most fundamental question of all: at what price?

What the Evidence Shows

The most rigorous attempt to date to measure AI’s actual effects on firms and workers was published in early 2026 by a team including Nicholas Bloom and Steven J. Davis, drawing on surveys of thousands of senior executives across four major economies. The overwhelming majority reported no measurable effect of AI on either employment or productivity over the preceding three years. The effects, where reported at all, were vanishingly small.

This is not an isolated finding. A Yale Budget Lab study, reviewing Bureau of Labor Statistics data through late 2025, found no significant differences in employment outcomes between occupations with high and low AI exposure.

Sam Altman himself acknowledged at a recent conference that companies are blaming AI for layoffs “whether or not it really is about AI”, an admission that ought to give pause to anyone constructing a narrative of AI-driven displacement.

Robert Solow observed in 1987 that you could see the computer age everywhere except in the productivity statistics. The paradox named after him has not gone away. It has simply acquired new occupants.

This should not be surprising to anyone who studies the history of general-purpose technologies. The personal computer arrived in the early 1980s, but the productivity gains it enabled only became measurable in the 1990s. The pattern is consistent across technological revolutions: the gap between a technology’s demonstrated capability and its measurable economic impact is large, and it is measured in decades rather than years.

Altman predicted in early 2025 that AI agents would “join the workforce” and materially change company output within the year. They did not. The prediction has now been quietly extended to 2026, then perhaps 2027.

What Language Models Actually Are

There is a deeper problem with the displacement narrative, which concerns the nature of the technology itself. Large language models are genuinely impressive. But the source of their impressiveness is also the source of their limitation, and that limitation is structural, not a matter of scale.

At their core, these systems are prediction machines. They are built to estimate, given a sequence of words, what word is likely to come next, a process trained on an enormous corpus of human-generated text. The outputs can be fluent, coherent, and occasionally brilliant.

But the mechanism is statistical pattern completion, not reasoning. When a language model produces an analysis of a legal question or a financial situation, it does so not because it understands the question, but because it has encountered vast quantities of text in which similar questions were discussed in similar ways.

The financial industry has been doing something broadly analogous for decades, using quantitative models to find patterns in data and generate predictions. Nobody called those systems intelligent, and nobody suggested they would replace lawyers and analysts wholesale. The AI revolution, in important part, is the democratisation and broadening of such methods, not a qualitative leap into something categorically different.

This matters enormously for the displacement question. The tasks at which these systems genuinely excel are those resembling sophisticated pattern completion: drafting standard documents, summarising lengthy texts, generating code from specifications, producing first drafts from structured inputs.

The tasks at which they remain genuinely poor are those requiring abstract reasoning, causal inference, judgement under genuine uncertainty, and the kind of theoretical model-building that underlies the higher-value components of professional work.

Apple’s research division published a paper in 2025 testing frontier models on logical puzzles requiring genuine reasoning, and found that performance collapsed at high complexity even when the correct method was provided explicitly. The METR research group found, in a randomised controlled trial, that experienced software developers were measurably slower when using AI assistance than without it, and that is this is the domain where AI is supposed to perform best.

Jensen Huang is fond of arguing that AI will enhance rather than replace professionals: the radiologist, he says, will use AI to handle routine work and focus on judgement and care, making hospitals more productive and creating more jobs. This is a reasonable description of what AI does well. But it is precisely not the scenario of mass white-collar displacement. Enhancement and elimination are different economic mechanisms, and they have different implications.

The Price Nobody Mentions

The third objection is the one that I find most decisive as an economist, and the one that receives almost no attention in the public debate.

The current price of AI services does not reflect the true cost of producing them. The leading AI companies are, by their own internal projections, running significant losses and do not expect to reach positive cash flow until the late 2020s at the earliest, and those timelines have already been revised once.

They are sustained by a continuous flow of investor capital at valuations that require extraordinary future growth to justify. The largest technology companies are collectively committing hundreds of billions of dollars annually to AI infrastructure, a scale of capital deployment without precedent in the history of the technology industry.

What this means, economically, is that the price businesses are paying today for AI services is heavily subsidised, not by governments, but by investors who are betting on a future in which these services become vastly more valuable.

The analogy I find useful is this: imagine that every morning a helicopter with a pilot arrived at your door to take you to work, entirely free of charge. Your productivity would rise. You would reorganise your working life around it. You might even let go of some arrangements that no longer seemed necessary. But if you had to pay the actual market cost of a private helicopter and pilot, the calculation would look entirely different. Many of the apparent gains would evaporate.

This is the position businesses are in today with AI. They are restructuring around a technology priced far below its true cost.

When prices normalise, as they must if the companies providing these services are ever to become profitable, many applications that currently appear economically attractive will prove not to be. The entry-level professional who seemed redundant next to a free AI agent may look considerably less redundant next to a properly priced AI agent.

There is also a resource constraint that the displacement narrative tends to ignore. Running AI at the scale Amodei and Altman envision requires enormous quantities of electricity, specialised chips, water for cooling, and capital for infrastructure.

The International Energy Agency projects that global data centre electricity consumption will roughly double by 2030, reaching the equivalent of Japan’s entire annual consumption, and that projection does not assume anywhere near the scale of white-collar displacement being predicted.

If one took seriously the claim that half of all entry-level white-collar work would move to AI within five years, the implied demand on physical infrastructure would be orders of magnitude larger than anything currently being planned. AI does not abolish scarcity. It relocates it.

The Incentive Structure

It would be unfair not to note that Amodei, Altman, and Huang are not disinterested parties in this debate. They are the chief executives of companies whose valuations, fundraising capacity, and competitive positioning all depend on a compelling story about AI’s transformative economic impact.

Sustaining investor confidence in the capital deployed requires a narrative of imminent disruption. Amodei himself has acknowledged that he is deeply uncomfortable with a small group of people making decisions about technology that will affect everyone, yet he continues to make exactly those kinds of claims about economic transformation, from exactly that position.

Altman was at least admirably honest at a recent conference, saying of the current moment: “If there was an easy consensus answer, we’d have done it by now, so I don’t think anyone knows what to do.” That is a notably different register from predicting that AI agents will join the workforce within the year, or that intelligence will become too cheap to meter. The gap between private uncertainty and public prophecy deserves attention.

What Will Actually Happen

None of this is an argument that AI will leave the economy unchanged. It will not. These are genuinely useful tools, and their usefulness will grow as the technology develops and as institutions learn to integrate it into their workflows in ways that are reliable, safe, and cost-effective.

The appropriate historical frame is not the industrial revolution but something more modest and more instructive: the spreadsheet. The spreadsheet did not eliminate finance departments. It changed what finance departments did, making certain kinds of analysis cheaper and faster while freeing human attention for the work that actually required judgement. Demand for financial analysis expanded to fill the additional capacity. Employment in finance did not collapse.

The Jevons paradox, named for the nineteenth-century economist who observed that more efficient steam engines led to more coal consumption rather than less, is worth keeping in mind here.

If AI genuinely makes junior professionals more productive, the likely consequence in many sectors is not that firms need fewer of them, but that demand for their services expands. Lower effective cost stimulates demand. The structure of employment changes; the aggregate volume does not necessarily decline.

I should be transparent about my own position. I use these tools every day, and they have made me more productive in concrete and specific ways. Writing this piece itself involved Claude, which is of course Anthropic’s product.

What I fear most is not mass unemployment. It is a cycle of inflated expectations followed by disillusionment.

After the dot-com crash, many businesses retreated from internet investment at precisely the moment when the genuine long-run benefits were beginning to materialise. The internet did ultimately transform banking, retail, and media but it did so over fifteen to twenty years, not in the two-to-five year windows being promised in 1999.

AI will likely follow the same arc. The worst outcome would be a premature rush driven by subsidised pricing and exaggerated predictions, followed by retrenchment, delaying the genuine benefits by a decade.

The technology is real. The potential is genuine. But Solow’s paradox did not disappear because the predictions got louder. And the entry-level lawyer, the junior consultant, and the graduate analyst may prove rather more resilient than the prophets of Silicon Valley believe, not because AI is unimpressive, but because impressive technology and economically viable technology are not the same thing, and because scarcity, as economics has always insisted, cannot be wished away. It can only be moved.

But Anthropic’s Claude certainly is an amazing product, because it helped me write a lot of this post. Then again, Claude was trained on, among other things, this very blog. Maybe I should ask for a discount on my Claude subscription. I should also confess that writing this article required a fair amount of time correcting Claude’s hallucinated figures and citations that simply did not exist in reality. Perhaps Amodei hallucinates too.

Lars Christensen is an economist, Head of Analysis and co-founder of PAICE, and external lecturer at Copenhagen Business School’s Department of Digitalization. He is the originator of Market Monetarism and writes The Market Monetarist blog.

Contact: LC@paice.io

Regular readers of this blog will know that most of what I do here is explain, analyse, and argue about macroeconomics, monetary policy, and increasingly artificial intelligence. That work is public and free – and I intend to keep it that way.

But over the past year, we have increasingly been asked – through PAICE, the consultancy I co-founded – to bring that kind of thinking directly into organisations. Investment committees, boards, executive teams, and strategy sessions.

So we have formalised that offering under the name Expert Briefings.

What is an Expert Briefing?

It is not a standard presentation. It is a tailored, interactive session where your organisation gets focused access to specialist knowledge on the topics that matter most to you right now. We start from your industry, your risk exposure, and your questions – and work from there.

At PAICE, we cover three interconnected areas:

Macroeconomics, monetary policy, and financial markets – global growth, inflation, central bank policy and communication, interest rates, currencies, commodities, and financial imbalances. The question we always come back to: what does any of this actually mean for your business, your investments, and your risk picture?

Geopolitics and business risk – trade conflicts, energy markets, security policy, and geopolitical shifts. And critically: the concrete implications for companies and investors who have to make decisions in an uncertain world.

Artificial intelligence and technology – the latest developments in AI, what they mean for your specific sector, and how you position yourself strategically when the pace of change is this fast.

These three areas can be addressed individually or in combination, depending on what your organisation needs.

Who is this for?

The organisations we work with range from pension funds and asset managers who need regular macro and market input for investment committees, to CEOs, CFOs, and boards who want an independent external perspective on economics, monetary policy, technology, and geopolitics. Exporters and multinationals dealing with currency risk and trade policy. Banks building internal analytical capacity. Technology companies navigating AI regulation and competitive dynamics.

The common thread is that they want direct access to expertise – not a consultant’s slide deck, but a genuine conversation with someone who has spent decades thinking about these questions.

Formats

We try to be flexible about how this works in practice. Some organisations want a regular monthly or quarterly cadence to stay continuously updated. Others prefer an on-demand retainer – a standing arrangement that allows them to call on a briefing when the need arises, without going through a full procurement process each time. And sometimes the need is simply a one-off session for a strategy day or board meeting.

For topics that genuinely benefit from multiple perspectives, we can also convene an expert panel – two or more specialists combining, for instance, macroeconomics with AI and technology, or geopolitics with energy markets.

Who delivers the briefings?

The briefings are anchored by me. I spent fifteen years as Head of Emerging Markets Research at Danske Bank, including co-authoring the 2006 “Geyser Crisis” report that identified the risks building in the Icelandic banking system ahead of the 2008 collapse. Today I serve as co-founder, co-owner, and Head of Analysis at PAICE, where our work sits at the intersection of macroeconomic analysis, monetary policy, AI, and data.

For broader panels and cross-disciplinary sessions, I draw on PAICE’s network of specialists across macroeconomics, monetary policy, financial markets, artificial intelligence, technology, and geopolitics.

Getting in touch

If any of this sounds relevant for your organisation, we are happy to have an initial conversation about what might make sense.

Contact: hello@paice.io

Speaker platform: http://www.globaletanker.dk (in Danish)

Blue Owl is one of Wall Street’s big names in private credit – a manager of nearly $300 billion in assets. The company’s logo is an owl: the animal that, according to legend, can see everything, even in the dark.

On February 19, 2026, Blue Owl restricted withdrawals from one of its retail-focused funds and quickly sold $1.4 billion in loans to raise liquidity. Investors who wanted out couldn’t get out. The stock has fallen nearly 60% over 13 months.

The blue owl in the coal mine had not seen it coming.

It reminds us of something we have seen before.

In the mid-2000s, many were warning about the American housing market. Lending standards were too loose. Too much capital was chasing too few good loans. And those bearing the risk often didn’t know they were doing so. We know how that story ended.

There is now a new part of the financial system that deserves the same attention. It is called private credit. Most people have never heard of it – and that is itself part of the problem.

Regulatory arbitrage and monetary policy created this market

Private credit is fundamentally a child of two policy choices.

The first was regulation. Banks were subjected to far stricter capital requirements via Basel III after 2008. The intention was understandable enough – but the consequence was predictable: capital and credit demand do not disappear because banks withdraw. They move precisely to where regulation does not follow.

Private credit funds operate without equivalent capital requirements, without the same transparency requirements, and without meaningful macroprudential oversight.

This is the definition of regulatory arbitrage – and it is a foreseeable consequence of asymmetric regulation, not an accidental side effect. The IMF noted in its Global Financial Stability Report in April 2024 that insurance companies were also incentivized to move into private credit precisely because the capital charges are lower and less risk-sensitive than those applicable to commercial banks. Regulation did not reduce risk. It relocated it.

The second was monetary policy – but let us be precise here. This is not the story of a decade of near-zero rates after 2008. That is the wrong diagnosis.

This is the story of what happened from 2020. The COVID response triggered the largest expansion of the American money supply in peacetime history. M2 grew by nearly 27% year-over-year in early 2021 – the highest peacetime rate since the Federal Reserve was founded in 1913. Some of this expansion was justified given the lockdowns. But it continued far too long.

That sent a tsunami of capital into private credit, because institutional investors desperately sought returns in a world where traditional fixed income products yielded nothing.

The market grew from $2 trillion to $3 trillion in precisely that period. When the money supply and rates finally turned from 2022, enormous sums were already locked into illiquid structures held by borrowers priced for a world of extraordinarily cheap money.

Regulation that does not eliminate risk, but merely displaces it – combined with a tsunami of liquidity. That is precisely the cocktail that mixed the subprime crisis.

The Austrian school element: the AI boom as malinvestment

It is worth drawing on an older analytical tradition here – but with an important caveat.

Friedrich Hayek and Ludwig von Mises were the two central figures of the Austrian school – a tradition in economic thinking that flourished in interwar Vienna before spreading to London and Chicago. Hayek received the Nobel Prize in Economics in 1974.

Their theory of the business cycle – known as Austrian Business Cycle Theory (ABCT) provides an explanation of what happens when central banks hold interest rates artificially low for too long.

The argument is simple. When the rate is lower than the market would have set itself, a false signal is sent to investors: capital is cheaper than it really is.

This attracts investment into projects that only look profitable at artificial financing costs – not at the natural rate. Hayek and Mises called this malinvestment – misallocations that look sensible during the boom, but are exposed brutally when monetary policy normalises.

But here is the caveat – and it is important. Austrian Business Cycle Theory is a theory of the unsustainable boom. It explains how the misallocations arise. It does not explain what happens next – and in particular, it does not tell us whether the bust will become a catastrophe. That depends on something else entirely: monetary policy.

The AI boom is a textbook example of the first part of that story.

What we could all ‘The Tesla boom’ of 2020-21 was to a large extent what financed the training of ChatGPT. The COVID liquidity injection sent capital into tech equities and pushed financing costs for AI companies below the natural rate – precisely the Hayekian mechanism, just with a modern transmission. Credit channels kept the expansion going far beyond what the underlying productivity numbers could justify.

Now we are in the middle of the massive buildout of data centres necessary to make the business models profitable – the four largest tech companies spent $360 billion on AI infrastructure in 2025 and are planning $650 billion in 2026. And the financing story is becoming increasingly speculative: AI companies are seeking capital in the Middle East, the Trump administration is talking about using Fannie Mae and Freddie Mac to finance the sector. It resembles the Icelandic banks in 2006-07, rolling around finding new liquidity sources while the warning lights were flashing.

The underlying problem is that the business model that must repay the debt is far weaker than assumed. Microsoft’s AI chief Mustafa Suleiman recently promised that AI would automate most office jobs within 12-18 months. A large study from the National Bureau of Economic Research of 6,000 senior executives shows that nearly 90% report AI has had no measurable effect on employment or productivity. Penn Wharton estimates AI’s contribution to productivity growth at 0.01 percentage points in 2025.

That is not an argument against AI as a technology in the long run – I am after all a huge fan of Large Language Models and a addictive user of LLMs – but it is an argument against pricing $650 billion in annual infrastructure investments on promises that the data consistently contradict.

Private credit and the AI boom are not two separate stories. They are two symptoms of the same misallocation of capital – both created by the same COVID fiscal and monetary expansion, both now under pressure as the bill is presented.

Moral hazard under the Trump boom

On top of regulatory arbitrage and the COVID monetary expansion came a third element: moral hazard on a grand scale.

I have for some time been warning about the dangerous fusion of the Trump administration and parts of American business – particularly the tech sector. When Trump administration AI czar David Sacks says “we can’t afford to go backwards”, I read it as an implicit promise: the government will underwrite the AI boom.

And when large tech companies are increasingly defined as strategically important – as “too big to fail” – there arises precisely the incentive problem that economist Robert Hetzel identified in his analysis of the financial crisis: financial institutions take greater risks when they know others bear the consequences. Gains are privatised. Losses are socialised.

This is not a new phenomenon. It is the same dynamic we saw with Fannie Mae and Freddie Mac before 2008. And it almost always ends the same way.

The cockroaches

The warning lights began flashing in earnest in the autumn of 2025. Let us go through the events in chronological order – because the pattern is more troubling than any individual episode in isolation.

Tricolor Holdings collapsed first. The company operated as both a used car dealer and a subprime lender – packaging high-yield car loans to credit-impaired borrowers into AAA-rated securities and selling them on to investors. When the repayments failed, the whole construction collapsed.

Fifth Third Bank is now accusing Tricolor of fraud, claiming the company pledged the same assets as collateral for multiple loans simultaneously. Investigators are reviewing what may prove to be a manipulated loan database.

First Brands Group followed shortly after. Just weeks before the bankruptcy, the company was marketed by Jefferies Investment Bank as an opportunity for $6 billion in lending – and the company was said to have nearly $1 billion in cash.

It collapsed anyway. It had borrowed massively for acquisitions, then borrowed again against invoices and inventory – and when tariff pressure hit imported components, there was no margin left. Jefferies, Millennium Management, JPMorgan, Barclays, and Fifth Third Bank are all exposed.

Jamie Dimon put it precisely: “When you see one cockroach, there are probably more.”

Then came a series of events that together sketch a pattern. BlackRock wrote down a private loan from full value to zero in three months. This is not as surprising as it sounds: private credit loans are not traded on markets but valued using internal models – what is known as mark-to-model. The IMF documented in its Global Financial Stability Report in 2024 that adjustments to private credit valuations are systematically smaller and slower than in public markets, and that it takes at least four quarters for prices to converge after a shock.

Losses are there before they are visible. A related warning sign is the sharp rise in payment-in-kind interest – where borrowers pay their interest by adding it to the loan principal rather than in cash. The IMF found that the payment-in-kind share in business development company portfolios doubled between 2019 and 2023. Borrowers are not defaulting. They are deferring. Blue Owl restricted withdrawals and sold assets to raise liquidity. Blackstone’s large private credit fund BCRED experienced redemption requests in early 2026 that exceeded the fund’s quarterly cap. And Morgan Stanley and Cliffwater have both been forced to limit withdrawals from their retail-focused funds after investors tried to redeem more than the structures allow.

The official default rate has risen steadily: from 1.76% in Q2 2025 to 1.84% in Q3 and 2.46% in Q4 2025 according to Proskauer’s Private Credit Default Index. That still sounds low – but when debt restructurings and creative loan extensions are included, the real rate approaches 5%.

Fitch puts the actual default rate in private credit at 5.8% through January 2026 – the highest since the index began. In February 2026 alone, 11 default events were recorded, nearly double the monthly average for all of 2025.

Mohamed El-Erian describes it as the “ATM scenario”: investors who cannot exit illiquid positions begin selling what they can sell – regardless of asset class. “If you can’t sell what you want, you sell what you can.”

That is the classic contagion mechanism – not losses, but liquidity pressure that propagates. It resembles a traditional bank run.

And lurking behind the share prices is a mechanism that has not yet fully played out.

The large private credit managers are today rated at the lower end of investment grade. One or two notches down, and they fall below the investment grade threshold – and this is not just a question of prestige. Pension funds, insurance companies, and large bond funds operate under mandates that forbid them from holding securities below investment grade.

A downgrade to junk triggers automatic forced selling from all these institutional investors – not because they want to, but because they are regulatorily obligated to. The IMF estimated in 2024 that pension funds and insurance companies globally had more than $600 billion invested in private credit funds – a figure that has grown rapidly since.

Some of the world’s largest pension funds, with combined assets exceeding $7 trillion, have significantly increased their allocation to private credit while simultaneously raising their financial leverage. The IMF identified this combination – illiquid assets and leveraged balance sheets – as a specific systemic risk. When collateral calls come, these institutions sell what is liquid. That is El-Erian’s ATM scenario in institutional form.

UBS estimates that in a severe AI disruption scenario, the US private credit default rate could hit 13% – twice the stress level for leveraged loans.

That creates precisely the cascade we know from 2008: forced selling pushes prices down, which pushes other funds toward the same threshold, and suddenly it is no longer just private credit under pressure, but all liquid assets that investors must sell to create room on the balance sheet. El-Erian’s ATM scenario, but in a rules-based and automated version.

The blue owl in the coal mine was not the only bird in the mine.

The near-perfect copy of 2008

One must be precise here. $3 trillion sounds like a lot, but in global financial context, the private credit market is relatively modest. The American equity market alone is about 15 times larger. The global bond market is more than 30 times larger.

But that is the wrong way to frame the question.

The subprime market was not the world economy’s largest market in 2006 either.

It still triggered the worst financial crisis since the 1930s. Then-Federal Reserve Chairman Ben Bernanke said himself in 2007 that the problems were “contained”.

They were not – because it was not about subprime’s absolute size, but about its connections to the rest of the financial system and what it revealed about the broader misallocation of capital. It is worth noting that the IMF’s own Global Financial Stability Report from April 2024 concluded that the financial stability risks from private credit “appear contained at present”. That is a precise echo of Bernanke’s formulation. It may prove equally accurate.

That is precisely the same point here. Private credit may not be large enough in itself to trigger a global crisis. But it can be the symptom of a far larger misallocation of capital – driven by the same COVID expansion, the same artificially low rates, and the same moral hazard – that will manifest as losses elsewhere in the financial system as reality catches up with the valuations made in a world that no longer exists.

This is where the data matters. U.S. nominal GDP growth remained relatively stable at around 5-6% through 2006 and into early 2007 – even as the first subprime warning signs appeared and global imbalances were becoming visible. The private credit market was wobbling. But the economy was not yet in crisis. It was only when those financial tensions translated into a de facto monetary contraction – and NGDP growth began to fall sharply through 2007-8 – that a financial correction became a macroeconomic catastrophe.

Critically, central banks made it worse. Spooked by what they perceived as bubble re-ignition – and facing rising headline inflation from an oil price surge – they turned hawkish at precisely the wrong moment (a number of central banks even hiked interest rates during the Summer of 2008). The result was a collapse in nominal spending that turned a necessary market correction into the Great Recession.

Secondary deflation, as Hayek called it, is not a natural consequence of a bust. It is a consequence of monetary policy failure.

The same risk exists today. Private credit may well be mis-priced. AI investment may well include substantial malinvestment. These corrections can be painful. But they do not have to become systemic crises. The key variable is not the size of the bust. It is whether the Fed ensures nominal stability through it.

On February 28, 2026, the US and Israel launched “Operation Epic Fury” – a coordinated strike on Iran that killed Ayatollah Ali Khamenei and threw the country into chaos.

Iran responded with drone and missile attacks on the Gulf states and effectively closed the Strait of Hormuz – the narrow passage through which 20% of the world’s oil consumption passes daily. Brent crude, which was trading below $70 per barrel at the start of February, hit above $100 earlier this week. That is a price increase of over 35% in fewer than four weeks.

Private credit is already cracking, as described above. The AI boom is under pressure from rising financing costs – and higher energy prices is not exactly good news for the energy intensive AI sector either. And now the oil price shock is a reality – the classic stagflation scenario that economists since 1973 have feared repeating.

In that situation, the Fed faces a near-impossible choice. The mandates point in opposite directions. Inflation is too high to ease.

The credit tightening calls for easing. This is the classic monetary policy dilemma that institutional rules and mandate structures make nearly impossible to resolve correctly – and it is precisely the dilemma that in 2008 turned a credit crisis into an economic catastrophe.

There is a way out of this dilemma – but it requires a framework the Fed does not have. A 4% NGDP level target would cut through the confusion between supply-side inflation and demand collapse.

Under such a framework, a central bank does not respond to rising oil prices by tightening – because oil price inflation does not represent excess nominal demand. It responds to falling nominal spending – which is the actual threat.

In 2007, the Fed had no such framework. It does not have one today either. As the German-American economist Rudi Dornbusch put it: “No postwar recovery has died in bed of old age – the Federal Reserve has murdered every one of them.”

And if Kevin Warsh – Trump’s candidate as the new Fed chairman –-is at the helm, it becomes even harder.

Warsh is ideologically sceptical of quantitative easing – the central bank’s purchases of bonds to pump liquidity into the economy when rates hit zero – and has historically viewed the Fed’s balance sheet expansion as a problem rather than a solution.

In the scenario where rates hit the zero lower bound and conventional monetary policy runs out of road, the political and ideological resistance to reaching for precisely this instrument will be maximal – at precisely the moment it is most needed.

Add to this a chaotic White House that is actively undermining the central bank’s credibility and institutional independence, and the picture is complete.

This is not a prediction. Most scenarios probably end with a gradual correction – losses at the weakest actors, tightening of standards, consolidation. That is the normal credit cycle.

But the pieces are placed in a way that is uncomfortably reminiscent of 2008. Only with a markedly worse political situation, a Fed chairman who is not Ben Bernanke, and a supply shock that Bernanke never had to contend with.

Two weeks ago, Microsoft’s AI chief Mustafa Suleyman told the Financial Times that most white-collar work will be fully automated within 12 to 18 months.

Lawyers, accountants, project managers, marketing professionals – everyone who “sits at a computer.”

Anthropic’s CEO Dario Amodei has warned that AI could eliminate half of all entry-level office jobs. Ford’s Jim Farley predicts a halving of white-collar jobs in the US.

In early February, Anthropic launched Claude Cowork – an AI agent capable of performing legal work, among other things.

Thomson Reuters fell 16% in a single day, LegalZoom dropped 20%, and Atlassian lost 35% in a week. According to JP Morgan, it was the largest non-recession-driven decline in software equities in over 30 years.

These are dramatic claims. They deserve a serious economic response.

I should say upfront: I am not an AI sceptic. I use AI extensively in my own work – for research, for drafting, for editing and building a whole lot of macroeconomic, financial and geopolitical simulation models.

This very article was produced with substantial AI assistance. It raises the quality and saves time. But the claim that all office work will disappear within 18 months rests on a fundamental misunderstanding of what large language models are, what they can do, and what they cost.

In short: AI is a tool, not a thinker. It can automate the routine part of office work, but not the thinking part. And even as an automation tool, it is not free: it requires capital and energy on a scale that makes it far from certain that it is cheaper than human beings.

Solow’s paradox, revisited

“You can see the computer age everywhere but in the productivity statistics.”

So wrote Nobel laureate Robert Solow in the New York Times in 1987.

Nearly 40 years later, his observation is once again strikingly relevant. In February 2026, a team of prominent economists – including Nicholas Bloom and Steven J. Davis of Stanford – published the first representative international study of AI’s actual effects at the firm level.

The data come from central bank-sponsored surveys in four countries – the US (Atlanta Fed), the UK (Bank of England), Germany (Bundesbank) and Australia – covering nearly 6,000 senior executives, recruited by telephone rather than through paid online panels (which are riddled with fraud).

Over 90 per cent report no effect on employment over the past three years. 89 per cent report no effect on productivity. The average employment effect is literally zero – 0.00 per cent – and the productivity effect a modest 0.29 per cent.

The Solow paradox is alive and well.

This should not surprise anyone who understands what a large language model actually is.

Language models are econometrics in disguise

A large language model such as ChatGPT or Claude is built on the transformer architecture from Vaswani et al.’s “Attention Is All You Need” (2017).

At its core, the technology is next token prediction: the model predicts the next word in a sequence based on statistical patterns in the training data. When ChatGPT produces an analysis of the causes of inflation, it does not do so because it understands inflation.

It does so because it has seen millions of texts where certain words appear together in certain patterns.

This is – and this is my central point – fundamentally no different from econometrics. A regression model says: given these variables, what is the most likely value of Y?

A large language model does the same: given these words, what is the most likely next word? The mathematics is more complex, the dimensionality higher, but the principle is identical.

The financial sector has been using quantitative models for decades without calling it “intelligence.” VAR models, credit risk algorithms, algorithmic pricing – enormously useful tools that help trained analysts make better decisions. The AI revolution consists of democratisation and broader application, not a qualitative leap.

Gary Marcus, professor emeritus at NYU, has argued for over 25 years that neural networks lack the capacity for abstract reasoning.

His central claim – that language models operate on pattern recognition rather than genuine understanding, and that scaling does not solve this problem – was powerfully confirmed by Apple’s research paper “The Illusion of Thinking” (June 2025), which systematically tested reasoning models on controllable puzzles. At high complexity, all models – including the so-called “thinking” models – collapsed to zero per cent accuracy.

Even when given the correct algorithm as an explicit instruction, performance did not improve. Marcus’s conclusion stands: large language models do not build abstract models of the world. They are built to complete sentences.

The models have not fundamentally improved

The transformer architecture is the same in 2026 as it was in 2017.

The advances since ChatGPT 3.5 in November 2022 are primarily more training data, more parameters and better fine-tuning (RLHF). The underlying mechanism – next word prediction via statistical pattern recognition – is unchanged.

The hallucination problem is the best evidence. Three years and hundreds of billions of dollars later, the models still fabricate facts. They have become better at sounding authoritative, but the underlying tendency to invent things is intact. They are better at lying than a teenager about their homework – they politely but firmly insist that their fabricated answers are correct.

And the fundamental failures are banal. Ask ChatGPT: “I need to wash my car. The car wash is 150 metres away. Should I walk or drive?”

ChatGPT advises you to walk – it’s only 150 metres. But you need to wash the car. You rather need to bring it along. This kind of basic logical understanding is still missing. It is difficult to see this technology replacing every bank clerk within 18 months.

The GPT-5 launch in August 2025 illustrated the point. Reddit threads titled “GPT-5 is horrible” received thousands of upvotes. 3,000 users signed a petition to regain access to the older GPT-4o.

On the prediction market Polymarket, the probability that OpenAI would have the best model dropped from 75% to 14% in a single hour. The event was dubbed “Gary Marcus Day” on social media.

Three mechanisms that undermine scaling

Any econometrics student learns about overfitting in the first semester. When you add more variables to a statistical model, R² always rises – but it is a deception. The model has learnt the noise along with the structure.

When you test it on new data, precision often falls. Precisely the same happens with language models. Scaling – more data, more parameters, more compute – yields better reproduction of what the model has seen, but not necessarily better handling of what it has not.

Then there is context rot. A Microsoft/Salesforce study from May 2025 documented that all tested language models performed 30-40% worse in multi-turn conversations than in single-turn queries.

Chroma Research demonstrated that performance degrades systematically with increasing context length — even when the model can identify all relevant information with 100% accuracy.

This is the opposite of genuine understanding. An economist who receives more information typically delivers a sharper analysis. A language model does the opposite. This is a fundamental limitation, not an engineering problem to be solved with more hardware.

And finally, model collapse – a problem Hans Christian Andersen described better than any AI researcher.

In his story “It’s Quite True!” a hen loses a single feather. The story is retold from perch to perch, and with each retelling it grows: one feather becomes five hens that have plucked each other to death from unrequited love. Andersen captured something fundamental: when information passes through successive stages, it is not just signal that is lost – new patterns are invented.

A 2024 Nature paper (Shumailov et al.) documented that when language models are trained on text generated by previous models, they degenerate through successive iterations. The rare phenomena in the tails of the distribution disappear first – the unexpected formulations, the unique perspectives, the nuances.

The dominant patterns are amplified and distorted. An ICLR 2025 Spotlight paper sharpened the conclusion: even the tiniest fraction of synthetic data – as little as 1 in 1,000 – can trigger collapse. And larger models amplify the effect rather than dampening it.

As of April 2025, 74% of newly created web pages contained AI-generated text. The internet – the primary training source – is increasingly filled with text produced by previous models.

For an economist, this is a classic endogeneity problem: the dependent variable feeds back into the independent variables. The result is biased and inconsistent estimates that diverge from reality over time. Andersen’s henhouse on an industrial scale.

Theory is a modelling strategy

Here is what the entire AI debate is missing: an understanding of the role of theory.

Theory is not something abstract floating above empirics. Theory is the deliberate decision about what to include, what to exclude, and how to structure the relationship between variables.

It is a modelling strategy. Language models have none. They take everything in and search for statistical regularities. This is the opposite of theory – it is inductivism on an industrial scale.

Two of the 20th century’s greatest economists – who often disagreed on methodological matters – agreed on precisely this.

Milton Friedman argued in “The Methodology of Positive Economics” (1953) that a scientific theory must be judged on its predictions, not on the realism of its assumptions – but that one starts with a theory, not with data.

Ludwig von Mises, in “Human Action” (1949), went further: economic theory can be derived logically from fundamental axioms about human action, and empirics alone can never generate theory. They disagreed about method but agreed on the conclusion: pure inductivism – data without theory – is a dead end. That is precisely what language models do.

Try it yourself. Ask ChatGPT to analyse why food prices are rising. The result is not an analysis. It resembles an average article in the Financial Times.

All the right words are there. The structure is tidy. The arguments are recognisable. But there is no original thought, no unexpected angle, no theoretical framework organising the argument in a new way.

It is the average of everything written on the subject – because that is precisely what a language model produces: basically the statistical mean of its training data.

And average can often be good enough. Many people need a competent first draft or a quick summary. But nobody wants to listen to an average economist. Nobody wants to hear an average song – unless it is background music in a supermarket.

What creates value is precisely what deviates from the average: the sharp analysis that sees something others miss. Language models cannot produce that by definition, because they are built to find and reproduce the average.

The exceptional lies furthest from the centre of the training data distribution – and is therefore precisely what the model is worst at.

No, Suleyman – AI will not take all office jobs

Suleyman’s claim deserves a serious economic response. Think of office work as an aggregate of human labour and AI agents.

Economists use CES (constant elasticity of substitution) production functions to model how easily one factor of production can replace another. Even if we assume high substitutability – that AI can perform many office tasks – it does not follow that humans will be replaced.

Because AI agents are not free. Their “wage” is capital and energy.

Every token a language model generates consumes electricity. Every model requires data centres, GPUs, cooling, maintenance. Office workers are “powered” by relatively cheap energy – food, housing, a salary. AI agents are powered by semiconductors, data centres and electricity. It is far from given that the robot is cheaper than the office worker.

The economic logic is clear: even a perfect substitute only replaces human labour if it is cheaper per effective unit of output. And here two independent constraints apply: rising marginal costs (the more we replace human labour with AI, the more the energy market is squeezed) and the inability to think abstractly (even a free AI agent cannot replace the part of office work that requires theory, causal understanding and strategic judgement). Even perfect substitution on the routine dimension combined with zero substitution on the abstract dimension yields a far more modest effect than Suleyman imagines.

I have experienced the productivity paradox personally.

I sat down to write an article about AI and productivity. I chatted with Claude, iterated back and forth, fine-tuned – and suddenly I had spent five hours on something I would previously have written in one. AI felt faster. It was not. The personal productivity paradox mirrors the macroeconomic one.

Moreover, current prices are artificially low. It costs OpenAI more to run many ChatGPT subscriptions than they receive in revenue. They are running at a loss, funded by venture capital (an extremely easy monetary policy in the US in 2020-21). Sooner or later it must become profitable.

When the subscription price goes from $20 to perhaps $200 a month, many applications that currently feel productive will prove not to be economically viable. According to the IEA, global data centre electricity consumption will double to 945 TWh by 2030 – equivalent to Japan’s entire annual electricity consumption.

The evidence: AI makes experienced programmers slower

For those who believe “vibe coding” and AI agents are already replacing programmers, the METR study from July 2025 is a sobering corrective.

In a randomised controlled trial, researchers followed 16 experienced open-source developers across 246 tasks on mature codebases. Developers with AI tools were 19% slower than those without. But – and this is the key – they believed they were 20% faster.

Before the study, they predicted a 24% speed gain. External experts predicted nearly 40%. Everyone was wrong.

69% of developers continued using the AI tools after the study. Not because they were faster, but because it felt easier. The experience of AI productivity and actual AI productivity can be entirely different things. And remember: programming is arguably the domain where AI performs best, because code can be tested externally – you can run it and see whether it works.

If AI makes experienced programmers slower there, what happens in domains without external verification? Law, strategy, analysis, management?

What the data actually show

There is not a single example in economic history of large and sustained technological unemployment. Not one.

The steam engine replaced hand weavers but created factory workers. The car killed the horse and carriage but created mechanics, road builders, suburbs and the service economy.

The computer automated bookkeeping but created an entire IT industry and thousands of new job functions nobody had imagined. The pattern is always the same: the structure changes gradually. Some jobs disappear, new ones emerge, and the transition can be painful for those affected. But mass unemployment as a consequence of technological progress? It has never happened. And there is no reason to believe AI breaks this pattern.

John Maynard Keynes – arguably the most influential economist of the 20th century – predicted in his famous 1930 essay “Economic Possibilities for Our Grandchildren” that his grandchildren’s generation would work just 15 hours a week.

He got most of it right. Living standards in the advanced economies have risen six to eightfold since 1930, precisely as he predicted. But we work far more than 15 hours a week.

Think of it this way: when productivity rises, we receive an extra pound. We can spend that pound on material goods – higher wages, more consumption.

Or we can buy leisure with it – work less for the same pay. You might expect roughly a 50-50 split. But we have consistently chosen consumption over leisure. We would rather have a larger house, a new car than a three-day working week.

Keynes, who literally worked himself to death during the Bretton Woods negotiations in 1946, underestimated this drive. AI will contribute to a gradual reduction in working hours as we become richer and choose more leisure. But it will be a prosperity-driven choice – not involuntary mass unemployment.

Bloom et al.’s study confirms the broader picture. 69% of firms use at least one AI technology, but the average senior executive spends just one and a half hours per week on AI. 28% do not use it at all. Adoption and genuine integration are two very different things.

The software meltdown of February 2026 is itself revealing.

If the market truly believed AI would broadly replace office work, we should have seen a sharp rise in, say, banking equities – banks would stand to save enormously from cheaper software. That did not happen.

The decline in software equities reflects a general nervousness about overvalued technology stocks rather than a genuine expectation of mass automation. AI profit margins remain concentrated in Big Tech , but the broader stock markethas barely moved.

Suleyman is a salesman, not an economist

One ought to ask what Suleyman’s prediction actually is: an economic analysis or a sales pitch?

He is not an economist. He is the head of the AI division at a company investing hundreds of billions of dollars in AI infrastructure, with a clear interest in convincing investors and customers that their product will revolutionise the world.

And one should look at what Microsoft is actually doing – not what Suleyman says.

While he predicts that office work will vanish within 18 months, his company is investing heavily in developing precisely the office tools that AI supposedly renders obsolete.

Throughout 2026, new features and deeper Copilot integration are being rolled out across Word, Excel, PowerPoint and Outlook.

On 1 July 2026, Microsoft will raise prices on Microsoft 365 subscriptions. The company maintains long-term product roadmaps and support programmes that still require human IT staff to handle bugs and compatibility issues.

If Microsoft genuinely believed in full automation by 2027, why raise prices on subscriptions that presuppose human users? Why invest in product updates that AI supposedly makes unnecessary? The answer is obvious: because Microsoft is betting that Office and Azure will continue generating enormous revenues for years to come, with AI as a supplement – not a replacement. The bold predictions are for investors. The business model is for the balance sheet.

The same applies to Anthropic’s Amodei and all the other technology CEOs competing to outdo each other in dramatic predictions.

Their business models depend on investors and markets believing in the AI revolution. That does not make them liars, but it makes them interested parties – and their predictions should be treated accordingly. One would not uncritically accept an oil company’s assessment of climate policy. One should not uncritically accept an AI company’s assessment of AI’s economic consequences.

Implementation is the real bottleneck

I recently spoke with a large international hedge fund. They wanted to use AI internally, but rules around IT security, data protection and client information prevented them from simply opening up the systems. Banks cannot hand their trading systems to something that hallucinates. Hospital records cannot run on a model that invents diagnoses. These are not irrational barriers – they are the reality firms face.

I encounter far too many companies that believe they have “adopted AI” because they have rolled out Copilot in their Microsoft systems.

If that is the strategy, the firm is in trouble. But the opposite fear – of being overtaken – is also exaggerated. Three years ago, I myself believed it would move much faster. I spoke with business leaders who were eager to get started. A year later, I asked the same leaders how it was going. The typical answer: “Well, there were some other things, and it’s a bit difficult, and we actually don’t want to throw away the company culture.” All perfectly rational choices. But it means implementation takes years, not months.

The parallel to the internet is striking. In the mid-1990s, every company needed a website. But nobody knew what to use it for. It became a glorified address book with a phone number and a postal address. Then people began saying you could perhaps sell things through it. Today you can buy everything online. But it took 15 to 20 years. E-commerce only truly broke through in the 2010s – far later than any of us believed in the mid-1990s. AI will likely follow the same pattern.

This is Solow’s paradox once more: the computer’s productivity gains came only decades later, when firms learnt to reorganise their work processes. Implementation, not faster processors, drove the growth. For AI investors and decision-makers, the implication is clear: if the value is in implementation, people with domain knowledge become more valuable, not less.

The last 5% is everything that matters

The first 90–95% of many tasks is routine. Writing a standard letter, producing background music for an advertisement, drafting a market analysis, generating boilerplate code, summarising a document. This is assembly-line work in the knowledge economy. AI can do it more cheaply and quickly. Those gains should be captured.

But the last 5-10% is everything that matters. The original insight in an analysis, the unexpected move in a composition, the theoretical framework that gives data meaning, the strategic judgement that correctly weighs risks. Abstract thought. Precisely what language models cannot do.

And here we hit something civilisational.

The capacity for abstract thought does not arise in a vacuum. It develops through practice – through wrestling with problems, making mistakes, revising one’s understanding, building intuition over time.

Think of the assistant analyst in a bank’s research department who is asked to photocopy and bind presentations. It is routine work – but in the process she reads the material, walks over to the economist who assigned the task, and they discuss what it actually means. That is on-the-job training. That is how one learns.

If AI takes over the routine, that learning process vanishes. A young economist who has never struggled to specify a model, understand residuals and think about what the data are telling her will never develop the deep intuition that enables her to deliver the crucial 10%.

Just as in music: one cannot compose anything original without having practised scales, played other people’s music, understood harmony and structure from the inside. The repetitive exercises are the precondition for the creative leap. If AI takes over the exercises, the foundation for the leap disappears.

The real risk is not that AI replaces thought now – it cannot. It is that it is just intelligent enough to tempt us to stop thinking for ourselves. Slow atrophy, not sudden disruption. Far harder to measure and counteract than job losses.

AI’s greatest threat is not that it is too intelligent. It is that it is just intelligent enough to fool us into believing it is intelligent.

Conclusion

Mustafa Suleyman is wrong. Not about everything – AI is changing office work, and it is doing so already. But the claim that most office work will vanish within two years is economically untenable. And one should remember that Suleyman is not an economist – he is a salesman for the product he predicts will revolutionise the world.

AI cannot think. It is pattern recognition, not abstraction. It lacks theory, causal understanding and the ability to weigh relevance. It delivers average analyses, never exceptional ones.

It is not free. It requires capital and energy on a scale that makes it far from certain that it is cheaper than humans – especially for the complex tasks where value is created.

The empirical evidence shows it. The METR study shows that AI makes experienced programmers slower.

And there is not a single example in economic history of large and sustained technological unemployment. Keynes predicted in 1930 that we would work 15 hours a week. We chose to work more and consume more. AI will not change that dynamic.

AI will deliver real productivity gains – primarily concentrated in domains that are already ripe for automation. Valuable, but not a new industrial revolution. For economists, strategists and decision-makers, the message is clear: AI is a tool, not a replacement for thinking. The most important asset in the knowledge economy remains the human one: the ability to formulate a theory, weigh relevance and understand abstract quantities. No language model can do that.

Data without theory is noise — whether the tool is a simple regression or the world’s largest language model.