The heat nearly shut down a nuclear plant on the Danube. What a mistake by two Nobel laureates has to do with it
Four technologies in three countries came under strain in one week, for a single reason: the level of one river. Finance learned the same lesson in 1998, and paid $4.6 billion for it.
Hungary's Paks nuclear power plant is going dark this week for the first time in its forty-four-year history. Not because of a failure, not because of a political decision, and not because it ran out of fuel. Because of the water level in a river.
To understand why this is one of the most important developments in Europe's power sector this summer, you need to step outside the power sector for a moment. And go back twenty-eight years.
But first things first.
One variable, four technologies
Paks accounts for close to half of Hungary's electricity generation. Four units, 2,000 MW, cooled by water pumped straight from the Danube. The problem here is not the temperature of the water but its level: at water levels this low, the pumps can no longer draw the required volume. On Monday the plant was delivering 240 MW instead of two thousand, and Prime Minister Péter Magyar announced in a video posted on social media that engineers had managed to keep the last turbine running. "This is very good news," he said.
It is worth pausing on that sentence. Very good news, for the prime minister of a European Union member state, is that one turbine is still spinning at his country's only nuclear plant. At that point a full shutdown looked like a matter of hours. Bringing the plant back online could take weeks, because there is no rain in the forecast.
Had it ended there, this would be a story about one plant in one country. It did not end there.
The same water that ran short at Paks flows on. On Monday the Romanian navy fired on its own river to turn it around: one hundred and eighty kilograms of explosive detonated underwater near the Bala Canal, in order to push more flow toward the intake of the Cernavodă nuclear plant. One unit is already offline, shut down as a precaution; the second is operating at its limit. Danube flow in Romania has fallen below a third of the July average.
A little further upstream, in Petrovaradin in Serbia, boats are sitting on dry sand. Serbia's principal hydro plant is running at a fifth of capacity. Prime Minister Djuro Macut convened an emergency meeting and asked citizens to cut back on air conditioning.
And then the point at which this stops being a story about a river. The Romanian government declared a nationwide energy alert for the whole of August. Dacia and Ford brought forward maintenance and halted car production until the nineteenth. Other large industrial consumers agreed to curtail their draw during peak hours. The level of one river rewrote factory schedules in a country several hundred kilometres from Paks.
One distinction is worth recording immediately, because headlines blur it. A month ago in France the problem was river temperature and a legal limit on discharging warm water: an administrative constraint imposed on a fully functional plant, and one that lifts along with the heatwave. On the Danube the problem is the absence of water. That is a physical constraint, indifferent to the time of day, lasting exactly as long as it takes for rain to fall. The first distorts the shape of the day. The second holds output down, day after day. The only thing they have in common is a photograph of a cooling tower on the internet.
On top of that sits a layer that gets discussed least, because it has neither its own power plant nor its own prime minister: at extreme air temperatures, transmission line capacity falls while demand rises. The grid narrows precisely when more has to be pushed through it.
So let us tally up what actually failed here. Not a unit. Not a plant. Not a country. What failed is something that appears on no outage list, because it is not a piece of equipment.
What failed was independence.
Nuclear, hydro, thermal-plant cooling and transmission capacity are four different technologies, built by different companies, in different decades, in different countries. Throughout the working lifetime of everyone now operating and analysing these systems, they behaved like independent events, because they did in fact fail separately. This week they turned out to be exposures to a single common factor, and that factor showed up in several systems at once.
And this is the point at which things start to feel uncomfortably familiar. I sat the first stage of Poland's investment adviser examination in September 1998, while also completing an engineering degree. Russia had defaulted a few weeks earlier, and in the same month that I sat that exam, the New York Fed was assembling a rescue consortium for a fund run by two Nobel laureates. That August was not a chapter in a textbook for me; it was material I was learning as it happened. The markets were working through precisely the same mistake, expensively and in public. It is worth following that story closely, because in a moment we are going to swap out every noun in it.
August 1998. The fund with two Nobel laureates on board
In October 1997, Myron Scholes and Robert Merton collected the Nobel Prize in economics for the formula that allows an option to be priced. One of those achievements that reshape an entire industry and make it into the textbooks while the authors are still alive. At the same time, both were partners in an investment fund called Long-Term Capital Management, run by John Meriwether, a legend of bond arbitrage.
Eleven months after the Nobel, that fund came close to taking the global financial system down with it.
Not because its founders were reckless. Quite the opposite. This is a story about how the best people in the business, equipped with the best tools available, made a mistake that from where they stood was impossible to see.
The fund did not bet on market direction. It did something considerably cleverer: it hunted for small, statistically reliable pricing discrepancies and used leverage to magnify them, spreading the risk across more than a dozen unrelated markets, from US Treasuries to Danish mortgage bonds, Japanese swaps and shares in companies undergoing mergers. Every position small. Every position carefully calculated.
And here is the number without which the whole story is merely an anecdote: for every dollar of its own capital, the fund had roughly twenty-five dollars borrowed. At that kind of leverage you do not need to be very wrong. You need to be slightly wrong everywhere at once.
In August 1998 Russia defaulted on its domestic debt and devalued the rouble. On the face of it a remote event: nobody at the fund had built a strategy around Russian bonds as a core position.
And yet within a few weeks all of those dozen-plus markets moved in the same direction.
Not because they had suddenly become similar to one another. Because after the Russian shock every market participant in the world wanted the same thing in the same hour: to flee into safe assets and cut leverage. The common factor was not the economy. The common factor was how people behave under pressure.
There was a second mechanism too, subtler and more important. The fund's positions were not a secret. Other participants knew roughly what it held and at what scale, so when things began to slide, some of them started trading against it. Correlation did not descend from the sky as a kind of market weather. Part of it was manufactured by everyone looking at the same position and drawing the same conclusion from it.
The fund lost around $4.6 billion in four months, and in September a consortium of fourteen banks organised by the New York Fed had to take it over so that its liquidation would not drag the rest of the system with it.
And now the most important part, the one that usually gets summarised wrongly. This fund's models were not stupid. They were not merely wrong about the positions. They were wrong above all about how those positions would behave together. The correlations had been calculated carefully, on many years of history, in which there was not a single day when everything moved at once.
Diversification did not vanish in August 1998. It had simply always been conditional. It worked in the normal regime, where markets each mind their own business, and stopped working in the stress regime, where everyone minds the same one. Correlation is not a material constant. It is a function of the state of the system.
That is the entire error, in one sentence: the model was not bad. It was estimated in one regime and used in another.
What was learned then
The financial industry's response to 1998 was concrete and expensive. Stress testing existed before that, of course, but in practice it mostly came down to taking each risk factor in turn and asking: what if this one falls by thirty per cent. After 1998 the question shifted. What began to be tested was not so much levels as the correlation matrix: what if things that are independent in the data stop being independent for a fortnight.
Liquidity risk was also taken far more seriously. The issue was not that the fund's positions were worthless. The issue was that there was no way out of them, because everyone wanted out at the same time. Liquidity is available precisely up to the moment it becomes genuinely necessary.
Ten years later, in 2008, the same lesson returned in a more expensive edition. This time it concerned mortgages spread across the whole of the United States, "geographically diversified," until it turned out that they were one bet on one factor.
Back to the Danube
Now please read those paragraphs again, swapping out the nouns.
A portfolio spread across a dozen independent markets is the European generation mix: nuclear, coal, gas, hydro, wind, solar, across several countries at once. The common factor hidden in the background is not investor panic but an anticyclone, the same mass of air that dries out a river in Hungary, warms a river in France, puts the wind to sleep in Germany and drives demand up in Poland.
And liquidity? Imports.
Because that is exactly the silent assumption sitting inside the way we think about the evening peak: that if something were to run short, we would simply buy it in from abroad.
And here we need to be clear about where the criticism is directed, or it lands at the wrong door. System operators do not hold that assumption. Foreign units in the capacity market carry their own derating and participation limits, and European adequacy assessments are run on consistent weather years applied simultaneously to all countries, precisely in order to capture correlated conditions. The tool exists and it is a good one.
The problem is not the absence of a tool. The problem is what it is fed. The tail of the distribution is estimated on a climatic history that is ceasing to describe the climate in which these assets will operate for the next fifteen years. Put the numbers on one scale. The gauge at Paks had a record low of minus 98 centimetres, set in 2018 and treated as an extreme. Minus 134 is the threshold at which the plant has to shut down entirely. Forecasts for this week pointed to minus 144. The extreme of eight years ago now sits some forty-odd centimetres above the emergency threshold. This is not an accusation of ignorance. It is an accusation of stale calibration, which is a considerably more serious matter, because a model with stale calibration does not look broken. It looks like it is working, right up to the day it stops.
One floor down, the problem is worse. In the revenue models of developers, banks and transaction advisers, "imports will cover the evening peak" is not even a proposition. It is background, written into historical data from years in which buying in was always possible. Imports, however, are not an unconditional buffer. They are a variable correlated with the very factor that creates the deficit. The neighbour we reach out to is reaching for the same thing in the same quarter-hour and for the same reason. To be precise: imports do carry genuine hedging value, because systems differ in mix, demand profile and hydrology. It is just that this value depends on the state of the entire region, so in correlated stress scenarios it has to be reduced rather than carried over at the level observed in quiet years.
There is even an exact counterpart here of that second, subtler mechanism from 1998. Power market participants look at the same weather forecast, run it through very similar models, and draw the same conclusion at the same hour. Correlation does not arise from physics alone. Part of it we manufacture ourselves, by all of us looking in the same direction.
The analogy has a limit, of course, and it is better to draw it oneself than to wait for someone else to do it. The fund of 1998 failed not from correlation alone but from correlation multiplied by leverage and by forced position closure under margin calls. Nobody forces a power system into liquidation, and its common factor is primarily physical, only secondarily behavioural. This is a metaphor for risk, not a causal model. But the part of it that concerns the joint distribution carries across directly.
We saw a preview of this on 30 June, when PSE, the Polish transmission system operator, declared capacity-alert periods for the third time in its history. I traced that evening in detail a month ago, quarter-hour by quarter-hour, and I am coming back to it because only now is it clear what that evening was a sample of. It is worth reading the operator's justification literally, because everything is in it: the activation was required because of high demand associated with the heat, low forecast wind generation, and unavailability of capacity in conventional units.
Those are three factors. Air conditioning running on a hot evening, dead wind, and capacity losses in Polish units. On top of that, imports were effectively unavailable: French nuclear was running under thermal discharge limits at the time, and the evening cross-border balance had fallen to almost zero. Four things that appear in every model as separate line items occurred on the same evening. Not by accident. All of them had the same heat dome above them.
It ended well, because the system's resilience found an ally just in time: the wind came up exactly as the solar was fading. I wrote at the time that "this time they shook hands." The most important words in that sentence were the first two.
And one more thing, which a few days ago could not have been written, because it had not yet happened. On the morning of Tuesday 4 August, the operator declared capacity-alert periods for the hours 17:00 to 19:00. That is the fourth such alert in the history of the Polish capacity market and the second within five weeks. The previous two were more than two years apart. The operator's justification is almost word for word the same as on 30 June: high demand associated with the heat, low forecast wind generation, unavailability of part of the capacity in conventional units.
And the point is not that the tool worked, because it worked correctly and that is precisely why it exists. The point is frequency. An event that occurred twice in the entire history of the capacity market, more than two years apart, has now occurred twice in five weeks. Four events are too few to demonstrate a change in the distribution, and it would be inconsistent to claim otherwise in an article about estimation on too short a history. But they are enough to start testing for that change, rather than assuming in advance that we are looking at the same tail as before. Because a model estimated on history does not see a change in the rate at which the tail arrives until it is too late.
What follows from this
I do not know how this week will end, and I have no intention of guessing. That is not the interesting part.
Though you do not need to guess in order to see what is at stake. For Tuesday 4 August, the first fixing on the Polish power exchange priced the quarter-hour from 19:45 to 20:00 at around PLN 2,079/MWh, against roughly PLN 263/MWh in the afternoon hours. I am not claiming this is the Danube. The Polish evening peak has its own well-understood cause in fading solar output, and the exchange does not publish explanations of causes. I am claiming something more cautious and more important: this is a price set the day before, on a forecast, in a week when the entire region is competing for the same energy. How that evening actually settled is in the operator's data for anyone to check, and I am deliberately not making a punchline of it, because a single reading settles nothing. On 30 June the day-ahead market priced the evening peak at PLN 2,290/MWh, while balancing energy settled in those same hours at an average of PLN 374. The distance between those two numbers is the whole difference between assessing a decision and assessing an outcome.
What is interesting is that reality has just delivered a stress scenario free of charge: named, dated, quantifiable, spanning three countries and four technologies at once. Scenarios like this are usually invented around a table as an abstract "minus thirty per cent" shock. This one arrived with full documentation and no invoice.
There are three ways to waste it: to decide it is a Hungarian matter; to decide that since nothing happened in Poland, nothing happened; or to file the event on the outage list instead of in the correlation matrix.
And where does it bite? Everywhere a sentence of the form "if it comes to it, we will buy it in from outside" sits in somebody's model, spreadsheet or contingency plan. That sentence is not wrong. It is simply unquantified. Few people have assigned it a probability, because for twenty years there was no reason to, and things that have always worked stop being treated as assumptions and become background. The difference between an assumption and a calculation is that a calculation can be defended in front of someone who asks.
Because in substance, the investment layer around the power sector is today roughly where finance was in 1998. It already has correlation risk in full view. The vocabulary exists, but it stops at the operator's floor.
And here the sentence we started this section with comes back. "If it comes to it, we will buy it in from outside" is not solely an assumption about availability. It is a hidden claim about diversification: that the neighbour is a separate source. And security does not come from the number of sources, only from the number of independent sources. A portfolio of a dozen positions with a single common factor underneath them has considerably fewer independent positions than counting them would suggest. Not one, because a common factor is not yet a correlation of one. But considerably fewer than a dozen, and it is precisely that difference which determines the tail of the distribution. Imports, meanwhile, are the item we count with the most confidence of all, because they sit on somebody else's balance sheet and cost us nothing. Which is exactly why, outside the operator's floor, it is too rarely checked whether they are genuinely independent.
A month ago I noted here that the Polish operator has described itself, in its strategy to 2040, in two roles: the guardian, responsible for the safe operation of the system here and now, and the architect, rebuilding it for the future mix. The guardian answers the question of how much capacity is standing in reserve this evening, and on 30 June it answered well. The architect faces the harder question, because it can only be verified a decade from now: whether, in the mix currently being built, interconnections with neighbours will be a separate source, or the same bet on the same weather. Counting technologies and megawatts does not answer that.
So perhaps let us start with three terms. Imports as a correlated asset rather than an unconditional buffer, that is, with their own haircut on availability under Europe-wide weather stress. Joint distributions instead of lists of individual factors. And the weather year as a common risk factor rather than one of the scenario parameters.
Three terms are not yet a vocabulary. But one has to start somewhere, and the alternative is well documented. Twenty-eight years ago that vocabulary took four months to write and cost $4.6 billion, even with two Nobel laureates sitting at the table. It is worth finding out whether it can be done more cheaply this time.
Jacek Janiuk, CIIA, founder and CEO of Envalis. Originally published in Polish at e-magazyny.pl on 5 August 2026.