The Cost of More · Why AI ROI fails before it scales

Open the AI Cost Model
The state of enterprise AI today

Everyone is spending more on AI. Almost nobody can say what it buys.

Consumption is climbing an exponential nobody chose. Value arrives later than cost. Who survives the gap is decided by one number — what a finished unit of work costs — and it appears on no invoice.

Scroll
The ROI illusion

Spending more is not the same as knowing what it buys.

The budget is approved, itemised, correct — and blind. Cost arrives with every unit now; value arrives later. Whether you are still there when it does is decided by a number the budget does not contain.

🧾A budget you can defend

approved, and still blind
  • Tells you what you committed.
  • Benchmarked in full, and still unattributable.
  • Justifies the spend — says nothing about what one finished thing costs.

📐A margin you can state

one denominator
  • Tells you what one finished thing costs you.
  • One denominator — so scale compounds instead of bleeding.
  • Tells you what scaling will do, before you commit to it.
Demand has never saved anyone. Demand multiplied by the wrong margin is just a faster way to die.
— the lesson of the largest industrial bankruptcy in Europe's recent history
One factory, one promise

The flagship had everything. Except one number.

Northvolt was Europe's answer: whoever makes the cells owns the transition. $14 billion raised. A $55 billion order book. The cheapest green power in Europe. And a continent's purpose behind it: never again to depend on another region for the thing everything else is built on.

Demand was never the problem — the consumers were multiplying on a curve nobody controlled, and the value sat years out. That is what makes the margin the whole game: exponential volume, deferred payoff — the cost of one valid cell decides your runway. The flagship could not state it.

Every fact public. Every line real. The one figure the whole bet rested on is not among them.

THE FLAGSHIP · PUBLISHED FACTS EVERYTHING BELOW WAS REPORTED · 2016–2025 ORDER BOOK $55B CONTRACTED DEMAND DESCRIPTION FIGURE STATUS Capital raised equity, debt and a $5B green loan $14B+ public Contracted orders carmakers, multi-year, per the 2022 annual report $55B public Input power green northern hydropower — the quoted price: excellent cheap public Capacity announced the first plant alone 60 GWh public Production reached as of late 2023, against plan ~0.5% reported Everything above: knowable from outside ? What one finished, sellable cell cost no report could state it — week to week, line to line CASH · 2023 MAR 2025
Where the money entered

Seven places cost enters a cell, and one number that never came out.

All seven run at once — that is a ramp. Outside: what was raised, what was announced. Inside: one figure nobody could split — what this week's valid cells actually cost.

Tap a station. Then set each layer on the line itself — and watch what the seven make together.

The cost that owned the ramp is one of theseand it was never the one being argued about
Two published readings
All seven, end to end 70.0% .95·.95·.95·.95·.95·.95·.95 nobody reports this one
Cost per finished cell 1.43× = 1 ÷ 0.700 what one valid cell has to carry
Written off, every week 30,000 cells' worth of complete seven-layer cost — on a line where nothing looked broken. Scale is the multiplier — started per week100,000 The same factory, the same demand, at 96% end to end: 4,000.

Seven green numbers. Every owner is telling the truth. The layers do not add — they multiply, and the number they produce is the one that decides survival. Perfect any single layer and watch how little it moves. Four points per layer is the whole distance between the two companies.

Seven layers in the price of one valid cell. None of them is the answer. Their product is.

⚠ Where it broke

The order book was never the problem. The number nobody could state was.

They argued about the visible — hiring, timelines, the next site, the next raise. The cost that owned the ramp never announces itself.

So effort landed on what could be seen, and the number underneath did not move. They were not choosing badly. They were choosing blind.

Then the largest customer walked — €2 billion, cancelled over cells that did not arrive as valid cells. The bet broke quietly, per cell, long before it broke publicly.

The site sat on the cheapest power in Europe. Being careful about the price you were quoted does nothing about the cost per finished thing you were never shown.

March 2025: bankruptcy — $5.8 billion of debt, $30 million of cash. Nine years, fourteen billion dollars, and the number was still unknown. You can measure forward, never backwards. The unmeasured months are not delayed — gone.

Four months of truth

The winner never argued with totals. It measured one line, one week, one cell at a time.

CATL runs on a discipline, not a secret: the cost of a valid cell, to the decimal, across every layer — materials, energy, throughput, idle hours, repetition — week by week. As reported: “We can raise a factory's yield to 96% in four months.” The flagship took four years to reach 70.

1How much was raised.everyone has this one
2What capacity was announced.most programmes get here
3What a cell should cost at scale.some get here
4What a finished cell cost this week, on this line.answered weekly — by the winner
Stopping at three was never a failure of ambition. It was a missing instrument. The winner had one.
THE FLAGSHIP · AS REPORTED 4 YEARS → 70% NO WEEKLY NUMBER MEASUREMENT THE WINNER · AS REPORTED 4 MONTHS → 96% W1 W2 W3 W4 ONE NUMBER · EVERY LINE, EVERY WEEK · A READING ORDER

The measurement saved nothing. Not one cent. It bought the end of guessing. The difference was never the machines — it was knowing, every week, where the money left the line: which station, which layer.

Judgement, only after transparency

Same cell. Same demand. A tenth of the cost.

Then — and continuously — the winner changed things. No order was turned away, and no line stood still. Each move works a different layer of one valid cell's cost, because at exponential volume the margin of a single unit is the whole business — and the margin is more than yield.

1Production moved into the cheap hours and the learning curve.re-time
2Every step placed where it runs cheapest — the supply chain pulled inside the walls.re-place
3Chemistry and process chosen again, generation after generation — whichever does the same cell for less.re-choose
4Every lesson paid for once — a flaw fixed on one line is fixed on every line.re-use
PACK COST · 2008 PACK COST · 2024 1 2 3 4 ≈ −90% THE SAME CELL · A DECADE OF JUDGEMENT · NOTHING SCALED DOWN

The next gigafactory arrived priced before it happened. A decade of judgement cut pack cost by roughly ninety percent. The cost stopped being weather and became a decision — and an earned margin became permission to scale.

The map has a name

This was never really about batteries.

Your estate is on the same curve: exponential consumption, deferred value, competitors engineering their margin. Pick one ordinary week and state what one valid unit of AI-assisted work cost. If the honest answer is the budget divided by guesswork — this story is about your factory too.

A cell's price has seven layers. A valid unit of your work has the same anatomy — inference, context, movement, idle capacity, contention, retries — and nobody owns the question.

The number is finite and separable. It takes a denominator, an order of work, and the honesty to price what you cannot yet see. That is the companion model.

A $55B order book, and no cost per finished cell
One denominator
Cost per unit of finished work — the figure no budget line carries, and the one the gap is crossed on.
A ramp that multiplied every cost it touched
The phase ladder
Why the climb is exponential — actors × turns, context and retries — on four named, editable assumptions.
A scrap channel with no weekly number
Hidden spend, priced
Every line is Hidden, Transparent or Optimized — and what you cannot see is reported in money, never scored as zero.
The cheapest power in Europe — and it didn't matter
Your own rates, your own split
The baseline is your invoice. The model carries no prices of its own — only illustrative estates, which your first real figure replaces.
Effort landing on what was visible
The dominant axis, computed for you
So the fix lands on what is expensive, not on what is argued about.
Four months, measured — then permission to scale
What changing it would return
Recoverable spend, per line, from your own figures — the published basis and the vendor uplift shown apart, each expanding to its arithmetic.

Stop defending the budget. Start pricing the work.

Get one denominator, find the line that actually owns the cost, and see what scaling does to it — computed from your own numbers, not ours. The margin is the runway; the model states it.

Open the AI Cost Model