Energy Systems That Compound

Volume IV — Production and Regeneration

THE PLATE

A watercolour of rolling hills in late summer, a farmhouse and a line of trees on the far ridge. — Plate IV.05The Ten Thousandth.The ten thousandth panel costs less than the first. Nobody economised. Ten thousand people each learned something small and none of them could tell you what.

THE LETTER

The energy transition is usually argued as a moral question, and it is an arithmetic one. That is not a complaint about moral argument. It is an observation that the arithmetic has been settled for some time, in public, against measurements anybody can download, and that the people who read it early made a great deal of money while the people who read it late are still calling it optimism.

This chapter is about one equation and what follows from it. The equation says that the cost of making a thing falls by a roughly constant proportion each time the total number ever made doubles. Theodore Wright published it in 1936 about airframes. It has since been measured across dozens of technologies, and in solar photovoltaics it has held for forty-seven years through four industry collapses, three countries of manufacture and two changes of dominant cell architecture.

What follows from it is a distinction that belongs in every capital committee in the world and is currently in almost none of them. An energy source whose cost falls with cumulative deployment is not a cheaper version of an energy source whose cost rises with cumulative extraction. It is a different asset class. They have opposite exponents on the same variable. Discounting them at the same rate, forecasting them with the same model, or holding them in the same line of the same schedule is a category error, and it is the category error that caused twenty years of institutional forecasts to be wrong in the same direction.

Chapter I.01 computed how much energy arrives and what it costs to fetch it. Chapter III.02 asked whether energy can serve as a unit of account. Neither asks what this chapter asks: what kind of asset is a technology whose price is a function of how much of it you have already built?

You should want three things by the end. The rate, measured, with its interval. The projection, and an honest statement of what it is actually sensitive to. And the instrument — a contract that pays for a learning curve instead of hoping for one.

— The Editors

DISCOVERY

What is already working

Begin where the evidence is oldest, which is not where the argument usually starts.

Theodore Wright, 1936. An aeronautical engineer at Curtiss-Wright published a paper in the Journal of the Aeronautical Sciences observing that the labour required per airframe fell by a stable proportion each time the cumulative number of that airframe doubled — about a fifth. He was not proposing a law of nature. He was reporting a production statistic, in order to price a contract. Nearly a century later it is still the best single predictor of unit cost that industrial economics has, and it is still usually called by the name of the man who noticed it while doing something else.

Kenneth Arrow, 1962. Arrow gave the effect its economics: if experience produces improvement and improvement is not fully appropriable, then production creates a public good, and a market will systematically under-produce it. That sentence, written before a single commercial solar cell was sold, is the complete theoretical case for deployment policy. It is worth reading it again, because almost every subsequent argument about subsidy is a rediscovery of it.

Béla Nagy, Doyne Farmer, Quan Bui and Jessika Trancik, 2013. They tested the competing functional forms against sixty-two technologies with real cost and production histories and found Wright's — cost against cumulative production — performed best, though narrowly, and that the improvement was measurable and forecastable rather than anecdotal. Farmer and Lafond extended it in 2016 to a distribution of forecast errors, which is the move that turns a curve into an instrument: not a point estimate, but an interval you can underwrite.

Gregory Nemet's history of the module, 2019. Nemet traced where the solar cost decline actually came from and found no single cause and no single country. American research, Japanese appliance-scale manufacturing, German demand policy, Australian cell science and Chinese factory scale each did a distinct thing, and the decline required all of them in sequence. Goksin Kavlak, James McNerney and Trancik decomposed it numerically in 2018 and found the weighting moved over time — research dominant before about 2001, scale and plant size dominant after. The curve is not magic. It is a named list of mechanisms, and each one is purchasable.

The German feed-in tariff. Germany's Erneuerbare-Energien-Gesetz, from 2000, guaranteed a price per kilowatt-hour to anybody who built a generator and connected it. It was expensive, it was domestically unpopular in its later years, and it did something no research programme had managed: it created a predictable, bankable, multi-year order book. A factory will not be built for a market that might exist. It will be built for one that has already signed.

And then the auctions, which are the part people miss. Once the curve was running, competitive procurement in India, Chile, Saudi Arabia, Portugal and the Gulf discovered prices that no forecaster had put in a scenario. The mechanism that found them was not a subsidy. It was a sealed-bid auction with a creditworthy offtaker and a short grid connection, which is to say: an institution. The hardware price had already fallen worldwide; the auctions revealed how much of the remaining cost was local friction.

Australia's rooftops. The clearest natural experiment in the field. A residential system in the United States costs on the order of three dollars a watt fully installed; in Australia it is around one dollar for the same panels from the same factories — a ratio of about 3.0×. The hardware is a world price. Everything else is a local institution: permitting, inspection, interconnection, customer acquisition, sales tax treatment, electrician licensing. That gap is not a scandal. It is a worked list of the things a jurisdiction can change without inventing anything.

Six findings, one pattern. In every case the cost fell because somebody guaranteed volume, and the volume taught the factory. Nobody optimised their way down this curve. They bought their way down it, and what they bought stayed bought.

THE ARITHMETIC

What works, what does not, and where the line sits

The equation, first, with its terms.

        C(Q) = C0 · (Q / Q0)^(−b)

  Q   cumulative units ever produced
  C   cost per unit
  b   the learning exponent
  LR  the learning rate, the fraction taken off per doubling
      b = −log2(1 − LR)

A learning rate of 20 percent means b = 0.322; three doublings leave you at 0.8³ = 0.512 of where you started. Everything else in this chapter is that line, measured or contradicted.

The measurement. Take the published crystalline-silicon module price series in real 2019 dollars per watt against cumulative installed capacity in megawatts — eleven points from 106 dollars a watt at 0.3 MW in 1976 to 0.15 dollars a watt at 1,419,000 MW in 2023 — and regress log price on log cumulative capacity. The module in lib/verify/IV_05.py does exactly this and prints the inputs first.

  n                                  11
  OLS slope (log2 price on log2 Q)  −0.3679
  standard error of the slope        0.0271
  t(0.975, df = 9)                   2.262
  R²                                 0.9534
  LEARNING RATE                      22.5 %      95% interval 19.1 % to 25.7 %
  learning exponent b                0.3679

Twenty-two and a half percent per doubling, with an interval you can put in a credit paper. Over the same period cumulative production ran through 22.17 doublings. A second route — endpoints only, no regression, no intermediate points, sharing none of the regression's assumptions — gives 25.6 percent, a disagreement of 3.1 points. The endpoint figure is higher because the middle of the series fell more slowly against deployment than either end. Neither number is the truth alone. Quote the interval.

The spread across the literature is wider than any single estimate, and that is the honest picture. Rubin, Azevedo, Jaramillo and Yeh reviewed the published learning rates for electricity supply technologies in 2015:

  technology                 low     high    typical
  solar PV modules           10 %     47 %     23 %
  onshore wind (capital)    −11 %     32 %     12 %
  offshore wind               5 %     19 %     12 %
  pulverised coal             6 %     12 %      8 %
  lithium-ion pack           15 %     24 %     19 %

Read the wind row twice. The band contains zero and goes negative. A learning rate is a measurement of a particular technology in a particular production system over a particular window. It is not a property of being renewable, and a paper that assumes one because the technology is virtuous has assumed the conclusion.

The measured cost series, which is what actually decides projects. IRENA's global weighted averages of commissioned projects, 2010 against 2023 in 2023 dollars:

  utility solar PV, LCOE      $0.460/kWh  ->  $0.044/kWh     −90.4 %
  onshore wind, LCOE          $0.111/kWh  ->  $0.033/kWh     −70.3 %
  offshore wind, LCOE         $0.203/kWh  ->  $0.075/kWh     −63.1 %
  utility solar, installed    $5,124/kW   ->  $758/kW        −85.2 %
  onshore wind, installed     $2,179/kW   ->  $1,160/kW      −46.8 %

And here is a discipline worth more than any of those figures. Onshore wind ran through 2.42 doublings between those dates. Compute its learning rate three ways and you get three different numbers for one technology:

  naive LCOE learning rate        39.4 %
  capital-cost learning rate      23.0 %
  published turbine learning rate 12 %       (Rubin et al.)

None is wrong; they measure different things. The decomposition shows why: capital cost fell to a ratio of 0.5324 while the global weighted-average capacity factor rose from 27 to 36 percent, a ratio of 0.7500. Multiply and you predict an LCOE ratio of 0.3993; the observed ratio is 0.2973, leaving a residual of 0.7446 for operating cost, cost of capital, project life and the shift in the country mix. An LCOE learning rate and a capital-cost learning rate are different quantities and are quoted as one every week. When somebody hands you a learning rate, the first question is what was in the numerator.

Batteries, the same shape. BloombergNEF's volume-weighted pack price ran from $1,160/kWh in 2010 to $139/kWh in 2023 in real 2023 dollars — a fall of 88.0 percent, a compound 15.1 percent a year — and to $115/kWh in 2024. Ziegler and Trancik, by a wholly independent route over 1992 to 2016, find about 13 percent a year in real terms. Two methods a decade apart, agreeing to about two points.

Now the forecasts, which are the most persuasive numbers in this volume.

Annual global solar additions were 17 GW in 2010, 346 GW in 2023 and 452 GW in 2024 — compound growth of 26.1 percent a year. Successive institutional outlooks from the late 2000s onward projected annual additions roughly constant at the then-current level; Way, Ives, Mealy and Farmer reproduce the pattern edition by edition. A flat-line forecast assumes a growth rate of 0.0 percent a year for a quantity growing at 26.1.

  cumulative 2023 if additions had stayed flat from 2010      261 GW
  cumulative 2023, measured                                 1,419 GW
  ratio                                                         5.4 x

The 2014 roadmap for solar photovoltaics projected about 1,700 GW installed by 2030. The world passed that level in 2024, six years early, with 1,865 GW.

And the shape of the error is the finding, not the size of it. It was not one bad forecast. It was the same forecast, in the same direction, for roughly fifteen consecutive editions by careful people with good data. A run of errors that never changes sign is not a run of mistakes. It is a property of the model class — an additive model fitted to a multiplicative process — and a model-class error can be corrected once, for every technology that comes after.

The projection, and what it is actually sensitive to. From 1,865 GW and $0.044/kWh, carried forward at the three rates of the measured interval:

  cumulative   doublings   LR 19.1%   LR 22.5%   LR 25.7%     $/kWh
   4,000 GW      1.10       0.791      0.755      0.721     0.0348 / 0.0332 / 0.0317
   8,000 GW      2.10       0.640      0.585      0.535     0.0282 / 0.0258 / 0.0236
  16,000 GW      3.10       0.517      0.454      0.398     0.0228 / 0.0200 / 0.0175

State the sensitivity honestly and it is the useful part. The whole measured interval on the learning rate moves the 16 TW answer by 1.30×. Moving the deployment assumption between 4 TW and 16 TW moves it by 1.67×. The projection is more sensitive to how much gets built than to how fast the technology learns — which means an argument about the learning rate is usually a displaced argument about deployment policy, and should be conducted as one.

The ceiling nobody puts on the slide. A module at $0.15/W sits inside a total installed cost of $0.758/W — 19.8 percent of the system. If modules alone learned, the system learning rate would be 4.5 percent, and free modules would leave $0.0353/kWh standing. Balance of system does learn, more slowly; soft cost barely learns at all. That is where the 3.0× residential gap lives, and it is why the remaining work is institutional rather than technological.

Here is the cut. Hold the annual arithmetic beside the deployment support that produced it. The world generated 1,630 TWh of solar in 2023. That same quantity costs $71.7 bn at 2023 costs and $749.8 bn at 2010 costs — an annual difference of $678.1 bn. Cumulative German differential cost attributable to solar is carried here as a band, €150–250 bn, because the accounting boundary moves between sources; at 1.08 dollars to the euro that is $162 bn to $270 bn, repaid by the annual difference in 0.24 to 0.40 years. Take a deliberately wider and unverified band for all global deployment support, $500–1,000 bn, and the payback is 0.74 to 1.47 years.

So it was never a subsidy. It was a purchase. A subsidy is a transfer that buys a quantity of something now. This bought a permanently lower price for everybody afterwards, including everybody who contributed nothing, and it continues to pay every year without further payment. Germany did not buy German electricity with the feed-in tariff. It bought the world's solar price, at a price that now looks like a rounding error, and the asset it acquired cannot be repossessed. Whatever else is true about that policy, the accounting treatment was wrong in every contemporary account of it: it was capital expenditure misfiled as consumption.

And now what does not work, because the positive claims are only load-bearing if this part is said plainly.

LCOE omits system integration cost, and the omission grows non-linearly with penetration. Levelised cost is a true statement about a generator and an incomplete one about a grid. It prices energy, not energy-at-a-time-and-place, and the difference is small at low penetration and dominant at high.

The measured anchors. CAISO curtailed 0.19 TWh of wind and solar in 2015 and about 3.40 TWh in 2024 — 17.9× in nine years, roughly 6.8 percent of utility-scale solar output. Germany in 2023, at about 52 percent renewables, curtailed about 10.5 TWh of some 272 TWh of renewable generation — 3.9 percent — with congestion-management costs of roughly €3.1 bn against 465 TWh of consumption, or €6.67/MWh system-wide. That German figure is congestion management in total, not renewables alone; the boundary is stated rather than hidden.

Read those two together and notice what they refuse to say. Germany curtails a smaller share at a higher penetration than California. Network depth and interconnection, not penetration alone, set the number. Anyone offering you a single monotone curve of curtailment against renewable share is selling something.

The peer-reviewed form of the non-linearity is in the market value. Hirth finds the market value of wind falling from about 1.1× the average power price at zero penetration to 0.5–0.8× at a 30 percent share, and solar falling faster — from about 1.3× to roughly 0.5× by about a 15 percent share. Ueckerdt and colleagues put integration cost at €25–35/MWh at a 40 percent wind share, which is the same order as the generation cost itself.

And the cost of a shifted kilowatt-hour, computed rather than asserted. Take a four-hour utility battery at $250/kWh installed, 300 full cycles a year, 85 percent round-trip, fifteen-year life, 7 percent cost of capital:

  capital recovery factor, 7% / 15 yr        0.1098
  annual capital charge per kWh capacity    $27.45
  kWh delivered per kWh capacity per year    255.0
  COST OF SHIFTING one kWh                  $0.1076
  delivered shifted kWh, all in             $0.1516
  against prompt solar at                   $0.0440      -> 3.4 x

The kilowatt-hour that has to wait costs three and a half times the one that does not. At low penetration almost none of them waits. At high penetration most of them do. On an illustrative shape — and it is labelled illustrative everywhere it appears, because three real systems are an ordering and not a regression — a blended cost of $0.0440/kWh at 10 percent penetration becomes $0.1263/kWh at 80 percent, a factor of 2.9×.

That is the honest negative and it does not go away. It is also not a refutation, because the firming technology is itself on a steep curve: the battery pack fell 88.0 percent while the solar module fell further, and the shifting charge above is a 2023 number falling at roughly the rate of section four. The correct statement is the uncomfortable one that holds both halves. Generation cost is solved and integration cost is not, and the second is now the larger half of the remaining problem.

One more negative, and it is the one that keeps this from being a slogan. The French pressurised-water programme built 58 reactors, largely to a small number of designs, by one utility — 5.86 doublings — and real overnight construction cost rose by a factor of about 3.5. That is a learning rate of −23.8 percent. Grubler calls it negative learning by doing; Lovering, Yip and Nordhaus find greater heterogeneity across countries, and both belong in the record. Malhotra and Schmidt give the mechanism: small, modular, mass-produced units of low design complexity learn quickly, and large, bespoke, site-assembled units do not. Cumulative production is not a cause. It is an opportunity that a particular manufacturing architecture is able to take. A solar module is made a hundred million times a year indoors. A reactor is made a few dozen times in a career, outdoors, in the rain.

DREAM

What becomes ordinary

In the economy that has absorbed this, the capital committee's papers carry two exponents.

Every long-lived supply commitment arrives with a measured learning rate beside its price, in the same way it now arrives with a credit rating, and with the interval attached rather than a point estimate. Nobody finds this unusual. The question what is the exponent on cumulative volume, and what is the standard error is asked in the room the way what is the covenant is asked, and a proposal that cannot answer it is not rejected — it is sent back for measurement.

Procurement understands that it is buying two things. It is buying kilowatt- hours, and it is buying a share of the world's next cost reduction, and the second is priced. Volume commitments are written where the learning rate justifies them and withheld where it does not, and the difference is visible in a column. When a supplier's implied learning path sits below the lower bound of the measured interval, somebody says so out loud in a negotiation, with the regression on the table, and the price moves.

The treasury function holds the distinction that this chapter is about. Assets whose cost falls with cumulative deployment are financed long, because time is on their side. Assets whose cost rises with cumulative extraction are financed short and amortised fast, because time is not. Nobody discounts them at the same rate, because they are not the same kind of thing, and the schedule shows two columns where it used to show one.

Integration cost is priced separately and openly. A grid operator publishes what it costs to move a kilowatt-hour four hours, and the number is in the newspaper the way a mortgage rate is. Generators bid firm or prompt and are paid differently, and the difference is not a penalty — it is information that lets a developer choose where to put a battery. Curtailment is a measured input to siting decisions rather than an embarrassment to be argued about after the fact.

Soft cost is a published per-jurisdiction figure, quarterly, in dollars a watt, and cities compete on it. A mayor who has taken a dollar a watt out of the local permitting process is understood to have done something equivalent to attracting a factory, because the arithmetic says so and the arithmetic is on the front page of the report. The three-times gap between one country's rooftop and another's has closed to something like one and a half, not because of any new technology, but because the gap was made of forms.

And the firms that build physical things run the same regression on their own lines. A manufacturer knows its own learning exponent the way it knows its gross margin — measured from its own production records, updated annually, argued about seriously. Volume decisions are made against it. The curve stops being a fact about solar panels and becomes a standard management instrument, which is what it was in 1936 before anybody thought of it as an energy argument.

DESIGN

The structure that gets there

Four moves, in order. They compound, which is the point, and each is available without anybody's permission.

One — measure your own exponent before you borrow anybody else's.

Take your own production records: cumulative units on one axis, real unit cost on the other, both in logs. Run the regression. Print the slope, the standard error and the R². If the fit is poor, that is a finding and not a failure — it usually means the product changed underneath the series, and the cure is to cut the series at the change rather than to average across it. Do it for three things: a component you make, a service you deliver repeatedly, and an input you buy. You will find rates between zero and about a quarter, and you will find at least one negative, and the negative will be the most informative thing on the page.

Two — buy volume, not price.

This is the whole demand-side mechanism in four words. A purchaser who negotiates hard on unit price against an uncertain volume gets a supplier who cannot invest. A purchaser who commits cumulative volume gets a supplier who builds the line that makes the price fall — and who can be held to a price path that reflects it. Volume is the input to the exponent; price is the output. Negotiating the output is negotiating the wrong variable.

The governance is straightforward and should be written down. A volume commitment is a liability; treat it as one, size it against the balance sheet rather than the enthusiasm, and set it where the implied learning is defensible against the measured interval.

Three — price integration separately, and build for it deliberately.

Do not let a generation LCOE do a system's work. Any procurement above about a fifth of your consumption should carry a separate line for firming, computed the way section seven computes it — capital charge, cycles, round-trip, life — and should state the penetration at which the blend was struck. This is not pessimism about renewables. It is the condition under which a renewable procurement survives contact with the operations team, and the reason most of them that fail, fail.

The design consequence is that storage, demand flexibility, interconnection and oversizing are substitutes for one another and should be tendered against one another. The cheapest firm kilowatt-hour is frequently the one that was never stored, because a process moved four hours.

Four — attack the part that does not learn.

Hardware is a world price and you will not beat it by negotiating. Soft cost is a local institution and it is entirely yours. Permitting time, inspection regime, interconnection queue, standard designs, pre-approved equipment lists, one-visit sign-off: each of these is a line in a document that a person can change. The three-to-one gap between comparable countries is the measure of how much is available, and none of it requires an invention.

The sequence matters. Measure first, because the volume commitment in move two has to be sized against a number. Commit volume second, because it takes a year for a line to respond. Price integration third, because it is the thing that will be used to attack the programme in month eight and it should already be in the paper. And do the institutional work throughout, because it is slow, free, and the only part nobody else can do for you.

DESTINY

How it holds when nobody is pushing

It sustains because the exponent is self-funding. Each doubling pays for the next one — that is the whole of the mechanism, and it is why the transition stopped needing advocates somewhere around the point where solar became the cheapest new generation in most of the world. A cost decline that is a function of volume does not require conviction. It requires order books.

Three failure modes, named honestly.

The curve flattens at a material floor. Nothing declines forever. Modules are already a fifth of a utility system's cost; silver, glass, aluminium, copper, land and labour do not follow the same exponent. Watch for the point where the fitted slope shallows for three consecutive periods, and re-fit rather than extrapolate. A curve that has been extrapolated through a regime change is how a careful firm gets an incautious answer.

The world market fragments and each fragment learns slower. This is the serious one and it is underway. The learning rate measured above is a global rate on a global cumulative volume. Tariffs, local-content rules and industrial policy split one world curve into several national ones, and a national curve runs on a smaller Q and therefore doubles less often. A country may rationally choose that trade — for supply security, for employment, for sovereignty — but it should do so knowing the price, and the price is a slower exponent for everybody including itself. Say the number when you make that choice.

Learning is confused with virtue. The French reactor programme is in this chapter for exactly this reason. The moment a learning rate is assumed because a technology is on the right side of an argument, the instrument has been inverted and it will produce confident nonsense. The discipline that keeps it honest is the one this chapter uses throughout: measure, print the interval, and name what the number did not look at.

The thing that makes it durable, in the end, is that the arithmetic is public. The series are downloadable, the method is a regression, and the answer does not depend on anybody's good faith — which is why it kept being true for twenty years while the forecasts said otherwise.

DELIGHT

What it feels like

There is a particular pleasure in a straight line on a log-log plot, and it is not the pleasure of being right. It is the pleasure of a mess resolving. Forty- seven years of price collapses, trade disputes, factory bankruptcies, silicon shortages, three countries of manufacture and two cell architectures — and when you put cumulative volume on one axis and cost on the other, it is a line. The noise was real and it was all vertical.

Then the second pleasure, which arrives later and stays longer: watching somebody's face when the exponent lands. It is usually a finance person, and it is usually quiet. They have spent a career with assets that get more expensive as you use them up, and the idea that the act of buying a thing makes the next one cheaper is not, at first, a fact about solar panels. It is a fact about what kind of world they are operating in, and it takes a moment.

And the small daily one. A factory floor where the ten thousandth is better than the first and nobody can tell you exactly why is a pleasant place to stand. Something is being learned that nobody wrote down, by people who would not call it learning. The compounding is not in the equipment. It is in the hands, and the equation is only how we noticed.

OPERATIONALIZE THIS

At the level of finance

The instrument: a deployment-indexed offtake agreement.

A conventional power purchase agreement fixes a price per megawatt-hour for fifteen years, and both parties then argue about the level. The argument is unresolvable because they are forecasting different learning rates and neither says so. This instrument makes the disagreement explicit and prices it.

The structure. A long-term offtake whose price steps down against a published index of cumulative world deployment for the technology, rather than against a calendar. The buyer commits volume; the seller commits a price path with a stated implied learning rate; the index does the arbitration.

The mechanics.

Volume. Firm annual quantity, ten years. Make the commitment real — a take-or-pay floor at 80 percent of nominal is what makes a factory bankable, and it is the consideration the buyer is actually offering.
The index. A named public series of cumulative installed capacity, identified by publisher and edition, with a stated successor if publication ceases. Never an index either party maintains.
The price path. P(t) = P0 · (Q(t)/Q0)^(−β), with β negotiated. β is the whole contract: it is the seller's implied learning rate, and it is comparable to the measured interval, in public, by anybody.
Collar. A floor protecting the seller against a deployment collapse and a cap protecting the buyer against an index revision. Both stated in dollars, not in adjectives.
Term. Ten years, with a re-fit review at five that re-estimates the learning rate on the intervening data and re-strikes β inside the collar.
Firming, priced separately. A second schedule for shaped delivery at the computed shifting cost, escalated on its own index. Never blended into the headline price, because a blended price cannot be re-negotiated in parts.

The worked case. Five hundred gigawatt-hours a year for ten years. A flat offer at $52/MWh is $26.00 m a year, $260.00 m over the term. World deployment has been doubling about every 3.3 years. Index the same contract and the ten-year total becomes $226.15 m at an implied 10 percent learning rate, $198.07 m at 19.1 percent, $188.22 m at 22.5 percent — a saving of $33.85 m to $71.78 m against the flat path, for no change in volume and no change in counterparty.

The balance-sheet treatment. The volume commitment is a purchase obligation and is disclosed as one. The indexed price path is not a derivative if it settles in physical delivery for own use — take advice early, because the own-use determination is where this transaction either stays simple or becomes a fair-value exercise every quarter. Where the buyer also funds capacity, the contribution is capitalised against the asset it creates, not expensed: the whole argument of this chapter is that deployment support is capital expenditure, and the accounting should say so.

The counterparty. Start with an existing supplier on an existing product where you already have five years of cost history — because move one of the Design requires a measured exponent, and you cannot measure a counterparty you have never bought from. Energy is the obvious application and it is not the first one. The first one is a component you already buy ten thousand times a year.

The first ninety days.

Day	Action	Artifact
1–15	Pull cumulative volume and real unit cost for three bought-in items	Three series, cleaned
16–30	Run the regression on each. Print slope, standard error, R²	The measured exponents
31–45	Choose the one with the tightest interval and the largest spend	One-page case
46–60	Draft the indexed path and the collar; name the public index	Term sheet
61–75	Negotiate `β` against the measured interval, in the room, on paper	Signed volume floor
76–90	Add the separate firming schedule; agree the five-year re-fit	The executed agreement

The number that decides it. One inequality, on the front page:

     β implied by the seller's price path   <   lower bound of the measured
                                                learning rate interval

If the seller's implied learning rate sits below the measured lower bound — below 19.1 percent on solar modules, on the estimate above — you are being asked to pay for a cost reduction that the rest of the world is about to hand you for nothing. If it sits above the upper bound, the seller has promised something the evidence does not support and the collar is the only thing protecting your delivery. Between the bounds is a real negotiation, and it is the only region in which this contract should ever be signed.

APPRECIATIVE QUESTIONS

Twelve, for a room

Discovery — what is already working

What do we make or buy most often, and what has happened to its real unit cost as the cumulative number has grown? Who here has watched that happen and could describe it?
Where have we already committed volume rather than negotiated price — and what did that supplier do with the certainty we gave them?
Which of our costs has fallen without anyone running a programme to make it fall? What was actually learned, and by whom?

Dream — what becomes possible

If every long-term commitment in this organisation arrived with a measured learning rate beside its price, what would be the first decision to change?
Imagine our own production data fitted as cleanly as the solar series. What would we be willing to promise a customer that we cannot promise today?
If the cost of firming energy were published as openly as an interest rate, what would we build differently on this site?

Design — what we build

Which three series could we assemble this month from records we already hold — and who has them?
Where could we convert a price negotiation into a volume commitment without increasing our total exposure by a single unit?
What is the part of our cost base that does not learn, and what would it take to change the form rather than the price?

Destiny — how it holds

What would we watch to know that a curve we rely on has flattened — and how many periods of evidence would we want before re-fitting?
If we chose a local supplier over a world price, what is the number we would want on the table beside that choice, so it is a decision and not a drift?
Who here would be willing to say out loud that a learning rate we assumed was not measured? What would make that an easy thing to say?

WORKS CITED

Wright, T. P. (1936). "Factors Affecting the Cost of Airplanes." Journal of the Aeronautical Sciences, 3(4), 122–128.

Arrow, K. J. (1962). "The Economic Implications of Learning by Doing." The Review of Economic Studies, 29(3), 155–173.

Boston Consulting Group (1972). Perspectives on Experience, 3rd edn. BCG.

Nagy, B., Farmer, J. D., Bui, Q. M. and Trancik, J. E. (2013). "Statistical Basis for Predicting Technological Progress." PLoS ONE, 8(2), e52669.

Farmer, J. D. and Lafond, F. (2016). "How Predictable Is Technological Progress?" Research Policy, 45(3), 647–665.

Way, R., Ives, M. C., Mealy, P. and Farmer, J. D. (2022). "Empirically Grounded Technology Forecasts and the Energy Transition." Joule, 6(9), 2057–2082.

Rubin, E. S., Azevedo, I. M. L., Jaramillo, P. and Yeh, S. (2015). "A Review of Learning Rates for Electricity Supply Technologies." Energy Policy, 86, 198–218.

Kavlak, G., McNerney, J. and Trancik, J. E. (2018). "Evaluating the Causes of Cost Reduction in Photovoltaic Modules." Energy Policy, 123, 700–710.

Nemet, G. F. (2019). How Solar Energy Became Cheap: A Model for Low-Carbon Innovation. Routledge.

Ziegler, M. S. and Trancik, J. E. (2021). "Re-examining Rates of Lithium-ion Battery Technology Improvement and Cost Decline." Energy & Environmental Science, 14(4), 1635–1651.

International Renewable Energy Agency (2024). Renewable Power Generation Costs in 2023. IRENA, Abu Dhabi.

International Renewable Energy Agency (2025). Renewable Capacity Statistics 2025. IRENA, Abu Dhabi.

International Energy Agency (2014). Technology Roadmap: Solar Photovoltaic Energy — 2014 Edition. OECD/IEA.

BloombergNEF. Annual Lithium-Ion Battery Price Survey. Successive editions.

Lazard (2024). Levelized Cost of Energy+ (LCOE+), Version 17.0.

Hirth, L. (2013). "The Market Value of Variable Renewables: The Effect of Solar Wind Power Variability on their Relative Price." Energy Economics, 38, 218–236.

Ueckerdt, F., Hirth, L., Luderer, G. and Edenhofer, O. (2013). "System LCOE: What Are the Costs of Variable Renewables?" Energy, 63, 61–75.

Joskow, P. L. (2011). "Comparing the Costs of Intermittent and Dispatchable Electricity Generating Technologies." American Economic Review: Papers & Proceedings, 101(3), 238–241.

Mills, A. and Wiser, R. (2012). Changes in the Economic Value of Variable Generation at High Penetration Levels: A Pilot Case Study of California. Lawrence Berkeley National Laboratory, LBNL-5445E.

Denholm, P. and Margolis, R. M. (2007). "Evaluating the Limits of Solar Photovoltaics (PV) in Traditional Electric Power Systems." Energy Policy, 35(5), 2852–2861.

Grubler, A. (2010). "The Costs of the French Nuclear Scale-up: A Case of Negative Learning by Doing." Energy Policy, 38(9), 5174–5188.

Lovering, J. R., Yip, A. and Nordhaus, T. (2016). "Historical Construction Costs of Global Nuclear Power Reactors." Energy Policy, 91, 371–382.

Malhotra, A. and Schmidt, T. S. (2020). "Accelerating Low-Carbon Innovation." Joule, 4(11), 2259–2267.

Arthur, W. B. (1989). "Competing Technologies, Increasing Returns, and Lock-in by Historical Events." The Economic Journal, 99(394), 116–131.

Northey, S., Mohr, S., Mudd, G. M., Weng, Z. and Giurco, D. (2014). "Modelling Future Copper Ore Grade Decline Based on a Detailed Assessment of Copper Resources and Mining." Resources, Conservation and Recycling, 87, 190–201.

National Renewable Energy Laboratory. Annual Technology Baseline, successive editions; and U.S. Solar Photovoltaic System Cost Benchmark, successive editions.

California Independent System Operator. Managing Oversupply: Monthly Curtailment Reports. Successive editions.

Bundesnetzagentur. Monitoringbericht and Netzengpassmanagement, successive editions.

Ember (2024). Global Electricity Review 2024.

Aurecon (2019). Hornsdale Power Reserve: Year 2 Technical and Market Impact Case Study.

Note on figures. Every figure above is computed in lib/verify/IV_05.py, which prints its inputs with their units and sources before any result. The learning rate is an ordinary least-squares fit on eleven published points with its standard error and a 95 percent interval, checked against a second route that shares none of the regression's assumptions. Pre-2000 module prices are read off the published curve and rounded to two significant figures, and the robustness check is printed. The blended-cost-against-penetration table is an illustrative model and is labelled as one wherever it appears. One figure — cumulative global deployment support for solar — could not be verified to a primary source; it is carried as a band, the conclusion is shown to hold across the whole band, and the gap is named in the module's output rather than rounded away.