The same plants produced about $19 billion of EBITDA in 2021 and far less two years later. Management guides capex and shipments, not earnings. On our numbers one year's EBITDA covers capex plus shareholder returns three years in four. In a down-cycle year it is close to a coin flip.
// figures from public reports, rounded · estimates and ranges are ours, not the company's · not investment advice
Some companies do not guide earnings because they cannot. Others do not because the honest range is too wide for one number. Steel at scale is the second kind. Revenue about $60 billion. EBITDA between about $5 billion and $19 billion in recent years, on almost the same assets. Same furnaces, same people, a $14 billion range in outcomes.
The P&L is a spread: steel price out, iron ore and energy in, times roughly 55 million tonnes. The spread is set in Shanghai and Rotterdam. So management guides what it controls: capex, shipments, cost programmes.
Silence on earnings is not missing information. It is a wide range, stated by leaving it out. We wrote it down.
Rounded: revenue about $60bn. Recent EBITDA around $7bn. Capex guided. Shipments guided. A dividend plus buyback programme that management treats as a firm commitment. The open question for a CFO is simple: does one year's cash cover all of it without leaning on the balance sheet? That is a coverage question, and a distribution answers it directly.
EBITDA = steel margin (shipments × spread pool) + mining − fixed costs and SG&A. Plan: $7.0bn, close to recent years. Our ranges:
| Driver | Plan | Range | Shape | Why |
|---|---|---|---|---|
| Steel spread | $9.5bn pool | −30% … +25% | triangular | steel price minus the raw-materials basket; the widest honest range in heavy industry, skewed down |
| Shipments | index 1.00 | −5% … +5% | uniform | volume follows the same demand cycle; ρ +0.4 with the spread |
| Iron ore (mining) | $1.5bn pool | −20% … +20% | triangular | a partial hedge: dearer ore squeezes steel but pays the mines |
One modelling choice matters more than the rest: the +0.4 correlation between spread and shipments. In a down-cycle, both fall together. A spreadsheet sensitivity moves them one at a time. Real years do not.
Hedging is in the model, where it exists. Nobody hedges the steel spread at this scale. There is no liquid contract for hot-rolled coil minus the raw-material basket across fifty million tonnes. The hedge ArcelorMittal does have is its own mines: when ore rises it squeezes the steel margin but pays the mining division. That is in the model as its own line with its own range, and it is why iron ore appears as a driver with upside rather than as a cost.
The bar that matters is capex plus shareholder returns, about $6 billion a year.
Simulated EBITDA versus the $6bn coverage bar
EBITDA, $bn. Blue bar: P10 to P90. Dot: P50. Dashed line: capex plus shareholder returns.
About a quarter of the distribution sits below the bar.
Three years in four, the business funds itself with room to spare. The fourth year comes up short by most of a billion, and the buyback and the credit rating compete for the same cash.
What moves the year
Share of the total EBITDA swing, one driver at a time between its P10 and P90.
Cost programmes are real work. They operate inside the 27% that is not the spread.
Who sets the driver
The same swing as above, grouped by who sets the driver. Bars: share of the total swing.
81% of the swing is set in Shanghai and Rotterdam. The lever is the balance sheet.
When four fifths of the year is set in other people's markets, the lever is not in the P&L. It is in the balance sheet: how much net debt you carry into a down-cycle year, how the buyback flexes with cash, and how much of the spread you lock in with customers on longer contracts. That is how the company behaves, and it is why it does not guide earnings.
Bear (P10, about $5.2bn): the spread compresses and takes shipments with it. Coverage fails by about $800m and net debt absorbs it. This is not a crisis. It is one year in ten, by construction.
Base (P50, about $6.9bn): coverage with about $900m of headroom. The dividend is safe and nobody writes about it. That is the normal outcome for a spread business.
Bull (P90, about $8.5bn): a tight steel market. These years fund the next decade's balance sheet. They arrive about as often as the bear years do.
Confine the spread to the lower half of its range, −30% … 0%. Everything else unchanged, same seed.
Chance of covering capex plus returns
Same model, same seed. Only the spread range changes.
In the down-cycle case the P10 falls to $4.3bn, nearly $2bn short of the bar.
In a down-cycle year the dividend is a coin flip. That explains the company's own behaviour: net debt kept low, returns that flex with cash flow, and no earnings number for the market.
What coverage probability do you manage to? Is the balance sheet sized for the one-in-ten or the one-in-twenty spread year? At what spread does the buyback pause before the dividend does? How much of the 27% non-spread variance do the cost programmes actually reach? None of these need guidance. All of them need a distribution, which the choice not to guide suggests already exists.
"No guidance" is still information. When management declines to name a number, the range is the message. The balance sheet shows how wide they think it is. A distribution turns that silence into something you can use.
// method: simplified driver model, plan values tuned to public accounts; 10,000 Monte Carlo scenarios per run through the ScenarioCast engine, correlations and hedges as stated in the model section; sensitivities one-driver-at-a-time at marginal P10/P90; what-ifs re-run on the same seed. Figures rounded. Our estimates, not the company's. Not investment advice.