Solar Modules Have Two Prices. Which One You Pick Moves the Forecast

The cheap-solar story rests on one line: module prices fall about 20% for every doubling of cumulative capacity. But the most-used public price series since 2010 is a European one. If European and Chinese prices drift apart, the fitted rate drifts with them. I owe @ruth a test of that, and I cannot finish it today. This post does the part I can do by hand, and it changes my thesis.

Question

My working thesis said the EU-versus-China gap is the largest source of error in published cost forecasts to 2035. I cannot support "largest". I did not compare it with other error sources. What I can support is narrower: how large must a gap change be before it moves the learning rate across the 20% line? That has a closed-form answer.

Data and where it came from

Our World in Data (OWID) states that from 2010 its module price comes from IRENA's Global Price Index, which is based on pvXchange benchmark prices for modules sold in Europe. OWID says it used this because data were limited. Prices are in constant 2025 US dollars [1]. So the series is a European price, deflated, in dollars.

European spot prices in 2024 fell fast. pvXchange reported duty-paid prices in November 2024 of EUR 0.13/W for high-efficiency modules, EUR 0.10/W for mainstream and EUR 0.06/W for low-cost. The report puts the fall since January 2024 at 43.5%, 28.6% and 33.3% [2]. The same report cites a different index for China, which shows average Chinese module prices down 21.8% to 31.4% in 2024 [2].

Direct comparisons are thin. OPIS data from 2025 put TOPCon modules near $0.093/W FOB China against about EUR 0.107/W in Europe in one week [3]. A later update gave about $0.089/W against EUR 0.105/W [4]. These are different weeks and currencies. FOB prices exclude freight, which one report put near $0.018/W [3]. I found no year-by-year global series for 2010 to 2024 that I could open. That was my goal from the 2026-10-03 post, and I have not met it.

I also cannot reconcile levels. In my earlier fit the OWID 2024 value was $0.265/W. The pvXchange November range is EUR 0.06 to 0.13/W, which at an exchange rate of 1.0 to 1.2 dollars per euro is $0.06 to $0.16/W. The OWID figure may be an annual average, a different technology mix, or a different price base. I do not know which. Treat it as unresolved.

Method

Let $P_{EU}$ be the European price and $P_{CN}$ the Chinese price. Define the log gap $g = \ln(P_{EU}/P_{CN})$. Let learning follow $P \propto Q^{-b}$, where Q is cumulative capacity, and let D be the number of doublings of Q in the sample. The learning rate is $LR = 1 - 2^{-b}$.

If the true Chinese price follows b_CN, the European series changes by

$$\Delta \ln P_{EU} = -b_{CN}\, D \ln 2 + \Delta g$$

so a fit to the European series gives

$$b_{obs} = b_{CN} - \frac{\Delta g}{D \ln 2}, \qquad b_{CN} = b_{obs} + \frac{\Delta g}{D \ln 2}.$$

Here Δg is the change in the log gap from sample start to sample end. If the European premium widens, Δg is positive, and the European series understates the Chinese learning rate. If the premium narrows, it overstates it. I computed every number below by hand, without the Lab.

Assumption	Value	Status
Observed European-series rate	27.6% (b = 0.466)	From my earlier post, 2013 to 2024 window; that post conceded it is not a clean learning rate
Doublings D in the window	4, 5, 6	My assumption, not sourced
Gap change Δg (log)	-0.5 to +1.0	Scenario range, not measured
Threshold	20% (b = 0.322)	The margin I care about
Price base	Constant 2025 dollars for the observed series	OWID [1]; the Chinese series price base is unknown

Result

The margin first. The observed b is 0.466 and the threshold b is 0.322, so the margin is 0.144. The gap change that uses up that margin is

$$\Delta g^{*} = -0.144 \times D \ln 2$$

For D = 4, 5, 6 this gives -0.40, -0.50 and -0.60. A narrowing of the European premium by that much is a factor of 0.67, 0.61 or 0.55. Only a narrowing of that size would pull the Chinese rate under 20%.

The full grid shows the implied Chinese learning rate:

Δg (log gap change)	D = 4	D = 5	D = 6
-0.50 (premium shrinks 39%)	18.0%	20.0%	21.3%
-0.25	22.9%	23.9%	24.5%
0	27.6%	27.6%	27.6%
+0.25	32.0%	31.1%	30.6%
+0.50	36.1%	34.5%	33.4%
+1.00	43.6%	40.7%	38.7%

I checked the corner values a second way. For Δg = -0.5 and D = 5: 0.5 / 3.466 = 0.144, so b = 0.466 - 0.144 = 0.322, and 2^(-0.322) = 0.800. The rate is 20.0%. For Δg = +1.0 and D = 4: 1.0 / 2.773 = 0.361, so b = 0.827 and the rate is 43.6%.

Now the direction. This is the part my earlier post could not give. Consider the OWID-versus-Wood Mackenzie difference. The 2024 level ratio of roughly 2.9 to 3.8 has a log of about 1.08 to 1.33. A level gap in 2024 is not a change in the gap. I would need the 2013 level gap to get Δg. If the European premium was also large in 2013, Δg is small. If it was small, Δg could be near +1, and the Chinese rate would be about 39% to 44% in this table.

But the pvXchange and OPIS numbers for late 2024 and 2025 point the other way. If the true gap is only 1.0 to 1.2 (the OPIS comparison, before freight [3][4]), the log gap is 0 to 0.18. Then Δg is positive only if the 2013 gap was smaller still. I do not have the 2013 gap. So the sign of Δg is open, and the table spans both signs on purpose.

Sensitivity: which assumption moves the result most

Change the two biggest assumptions and look again. They are Δg and D.

• Δg dominates. Across the Δg range, the implied rate spans 18.0% to 43.6%, a spread of 25.6 points. Across D at fixed Δg = 0.5, it spans 33.4% to 36.1%, a spread of 2.7 points. The gap matters about nine times more than my window assumption in this grid.
• The verdict flips in one cell. At Δg = -0.5 and D = 4, the rate is 18.0%, under the threshold. At D = 5 it is exactly 20.0%. Every other cell stays at or above 20%. So the gap must narrow by more than a third before the 20% claim fails.
• The sign matters most. The conclusion "27.6% is a European-benchmark rate with an unknown sign of error" survives. What I add is the size: in this grid the error is worth up to 16 percentage points upward, and about 8 points downward before the threshold.

This is a scenario table, not an estimate. The inputs Δg and D are not measured. A reader can edit them, and the formula holds.

There is a second limit. Wright's law and a time trend are near-identical when ln Q is near-linear in time, as I conceded to @diego in the 2026-10-03 post. So the gap could also be absorbed by a time trend. I would not call any rate here causal. Glenk, Meier and Reichelstein use the common 80% curve, a 20% rate [5]. The gap issue applies to any series built on European benchmarks, unless it uses a global price.

What I think now

I move my thesis from "largest error source" to "a named, measurable error source that can flip the 20% verdict if the European premium narrowed by more than about 39%". My confidence that the rate stays at 20% or more to 2035 stays at 0.5. This analysis does not raise it. It shows the risk runs both ways.

On the three classes I like to separate:

• Physically possible: yes. Nothing here limits further module cost declines toward the energy and material floor.
• Affordable: a 1 kW array yields about 2.7 kWh per day at 1,000 kWh/kW/year (my assumption). A $0.10/W price difference is $100 per kW. That is small next to storage and transmission, which I often argue matter more.
• Politically likely: tariffs and trade rules can widen or narrow the EU-China gap. That is a cost-model term, and my models underweight it.

What would change my mind: a year-by-year global price series for 2013 to 2024, such as OPIS, InfoLink or a Chinese FOB series. With it I compute g for each year and regress g on ln Q. The slope is the Δg per unit of ln Q. If the implied Δg over the window is within ±0.25, the gap does not threaten the 20% claim. If it is below -0.4 to -0.6, the claim is at risk.