Empirically grounded technology forecasts
and the energy transition
Decisions about how and when to decarbonize the global energy system are highly
inﬂuenced by estimates of the likely cost. Here, we generate empirically validated
probabilistic forecasts of energy technology costs and use these to estimate future
energy system costs under three scenarios. Compared to continuing with a fossil
fuel-based system, a rapid green energy transition is likely to result in trillions of net
savings, even without accounting for climate damages or climate policy co-
Rupert Way, Matthew C. Ives,
Penny Mealy, J. Doyne Farmer
Empirically validated probabilistic
forecasts of energy technology
Future energy system costs are
estimated for three different
A rapid green energy transition
will likely result in trillions of net
Energy models should be
updated to reﬂect high
probability of low-cost
Way et al., Joule 6,1–26
September 21, 2022 ª2022 The Authors.
Published by Elsevier Inc.
Empirically grounded technology
forecasts and the energy transition
*Matthew C. Ives,
and J. Doyne Farmer
Rapidly decarbonizing the global energy system is critical for address-
ing climate change, but concerns about costs have been a barrier to im-
plementation. Most energy-economy models have historically underes-
timated deployment rates for renewable energy technologies and
overestimated their costs. These issues have driven calls for alternative
approaches and more reliable technology forecasting methods. Here,
we use an approach based on probabilistic cost forecasting methods
that have been statistically validated by backtesting on more than 50
technologies. We generate probabilistic cost forecasts for solar energy,
wind energy, batteries, and electrolyzers, conditional on deployment.
We use these methods to estimate future energy system costs and
explore how technology cost uncertainty propagates through to sys-
tem costs in three different scenarios. Compared to continuing with a
fossil fuel-basedsystem, a rapid green energy transition will likely result
in overall net savings of many trillions of dollars—even without account-
ing for climate damages or co-beneﬁts of climate policy.
Future energy system costs will be determined by a combination of technologies
that produce, store, and distribute energy. Their costs and deployment will change
with time due to innovation, competition, public policy, concerns about climate
change, and other factors. To provide some perspective on the likely future energy
system, Figure 1 shows how the energy landscape has evolved over the last
140 years. Figure 1A shows the historical costs of the principal energy technologies,
and Figure 1B gives their deployment; both of which are on a logarithmic scale. As
we approach the present in Figure 1A, the diagram becomes more congested,
making it clear that we are in a period of unprecedented energy diversity, with
many technologies with global average costs around $100/MWh competing for
The long-term trends provide a clue as to how this competition may be resolved: The
prices of fossil fuels such as coal, oil, and gas are volatile, but after adjusting for inﬂa-
tion, prices now are very similar to what they were 140 years ago, and there is no
obvious long-range trend. In contrast, for several decades the costs of solar photo-
voltaics (PV), wind, and batteries have dropped (roughly) exponentially at a rate near
10% per year. The cost of solar PV has decreased by more than three orders of
magnitude since its ﬁrst commercial use in 1958.
Figure 1B shows how the use of technologies in the global energy landscape has
evolved since 1880, when coal passed traditional biomass. It documents the slow
exponential rise in the production of oil and natural gas over a century and the rapid
rise and plateauing of nuclear energy. But perhaps the most remarkable feature is
CONTEXT & SCALE
Decisions about how and when to
decarbonize the global energy
system are highly inﬂuenced by
estimates of the likely cost. Most
energy-economy models have
produced energy transition
scenarios that overestimate costs
due to underestimating
renewable energy cost
improvements and deployment
rates. This paper generates
probabilistic cost forecasts of
energy technologies using a
method that has been statistically
validated on data for more than 50
technologies. Using this approach
to estimate future energy system
costs under three scenarios, we
ﬁnd that compared to contuinuing
with a fossil fuel-based system, a
rapid green energy transition is
likely to result in trillions of net
savings. Hence, even without
accounting for climate damages
or climate policy co-beneﬁts,
transitioning to a net-zero energy
system by 2050 is likely to be
economically beneﬁcial. Updating
models and expectations about
transition costs could dramatically
accelerate the decarbonization of
global energy systems.
Joule 6, 1–26, September 21, 2022 ª2022 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Please cite this article in press as: Way et al., Empirically grounded technology forecasts and the energy transition, Joule (2022), https://doi.org/
the dramatic exponential rise in the deployment of solar PV, wind, batteries, and
electrolyzers over the last decades as they transitioned from niche applications to
mass markets. Their rate of increase is similar to that of nuclear energy in the
1970s, but unlike nuclear energy, they have all consistently experienced exponen-
tially decreasing costs. The combination of exponentially decreasing costs and rapid
exponentially increasing deployment is different from anything observed in any
other energy technologies in the past, and positions these key green technologies
to challenge the dominance of fossil fuels within a decade.
How likely is it that clean energy technology costs will continue to drop at similar
rates in the future? Under what conditions will this happen, and what does this imply
for the overall cost of the green energy transition? Is there a path forward that can
get us to net-zero emissions cheaply and quickly? We address these questions
here by applying empirically tested, state-of-the-art cost forecasting methods to en-
Historically, most energy-economy models have underestimated deployment rates
for renewable energy technologies and overestimated their costs
, which has led
to calls for alternative approaches and more reliable technology forecasting
Figure 1. Historical costs and production of key energy supply technologies
(A) Inﬂation-adjusted useful energy costs (or prices for oil, coal, and gas) as a function of time. We show useful energy because it takes conversion
efﬁciency into account (see Document S1 section ‘‘End-use conversion efﬁciencies’’). Electricity generation technology costs are levelized costsof
electricity (LCOEs). Battery series show capital cost per cycle and energy stored per year, assuming daily cycling for 10 years (these are not directly
comparable with other data series here). Modeled costs of power-to-X (P2X) fuels, such as hydrogen or ammonia, assume historical polymer electrolyt e
membrane (PEM) electrolyzer costs and a 50–50 mix of solar and wind electricity.
(B) Global useful energy production. The provision of energy from solar photovoltaics has, on average, increased at 44% per year for the last 30 years,
whereas wind has increased at 23% per year. These are just a few representative time series; all data sources and methods are given in Document S1
section ‘‘Data sources for Figure 1.’’
Institute for New Economic Thinking at the
Oxford Martin School, University of Oxford,
Oxford OX1 3UQ, UK
Smith School of Enterprise and the Environment,
University of Oxford, Oxford OX1 3QY, UK
SoDa Labs, Monash Business School, Monash
University, Melbourne, VIC 3800, Australia
Mathematical Institute, University of Oxford,
Oxford OX2 6GG, UK
Santa Fe Institute, Santa Fe, NM 87501, USA
2Joule 6, 1–26, September 21, 2022
Please cite this article in press as: Way et al., Empirically grounded technology forecasts and the energy transition, Joule (2022), https://doi.org/
. Recent efforts have made progress in this direction
largely deterministic in nature. The methods we use are probabilistic, allowing us to
view energy pathways through the lens of placing bets on technologies. After all,
powering modern economies requires betting on some technologies one way or
another, be they clean technologies or more fossil fuels—the best we can do is
make good bets. Which technologies should we bet on, and how likely are they to
pay off? We focus on solar, wind, batteries, and electrolyzers, which we call here
‘‘key green technologies’’, because they could play crucial roles in decarbonization
and have strong progress trends that are well documented in publiclyavailable data-
sets. We also consider the major incumbent energy technologies and compare our
forecasts with projections made by inﬂuential energy-economy models. We investi-
gate three different energy transition scenarios and discuss the implications for
whole system costs and transition pathways.
Figure 1 provides a glimpse into the diverse nature of technological change as
technologies rise and fall from dominance.
It reﬂects how innovation and tech-
nological learning produce different outcomes for different technologies. The di-
versity of rates of technological improvement for energy technologies seen in Fig-
ure 1 is typical of technologies in general.
Roughly speaking, technologies can
be divided into two groups based on their rates of improvement. For the ﬁrst
group, comprising the vast majority of technologies, inﬂation-adjusted costs
have remained roughly constant through time. Fossil fuels provide a good
example: although there has been enormous progress in technologies for discov-
ery and extraction, as easily accessible resources are depleted, it becomes neces-
sary to extract less accessible resources, creating a ‘‘running-to-stand-still’’ dy-
namic in which prices have remained roughly constant for more than a century
(this is true for all minerals
). Another example of a non-improving technology
is carbon capture and storage (CCS); despite signiﬁcant effort, over its 50-year
commercial history for enhanced oil recovery, costs have not declined at all.
contrast, for a select group of improving technologies, costs have dropped
roughly exponentially while deployment has increased exponentially.
of improvement for technologies such as optical ﬁbers and transistors are as
high as 40%–50% per year. Solar PV, wind, and batteries have behaved similarly
but with improvement rates closer to 10% (see Document S1 section ‘‘The hetero-
geneity and persistence of technological change’’). This makes unit costs for these
technologies predictable, even if the speciﬁc technological innovations that lead
to lower costs are not predictable.
Because the behavior of these two groups of technologies is so different, they
require different cost forecasting models. Fossil fuels such as oil and gas are tradable
commodities, and according to efﬁcient markets theory, their prices should follow a
This provides a useful approximation for roughly a decade, but over
longer spans of time they display mean reversion.
This makes autoregressive
models a natural choice, and we use them to forecast oil, coal, and gas prices
(see Experimental procedures and Document S1 sections ‘‘AR(1) process,’’ ‘‘Oil,’’
‘‘Coal,’’ and ‘‘Gas’’).
For the select group of technologies that are improving, improvement rates are
For these technologies there are two dominant methods
for quantitatively forecasting costs based on historical data. The ﬁrst is a generalized
form of Moore’s law, which says that costs drop exponentially as a function of time
(i.e., at a ﬁxed percentage per year).
The second is Wright’s law,which
Joule 6, 1–26, September 21, 2022 3
Please cite this article in press as: Way et al., Empirically grounded technology forecasts and the energy transition, Joule (2022), https://doi.org/
predicts that costs drop as a power law of cumulative production.
is also called an experience curve or learning curve, and cumulative production is
also called experience. (For a discussion of challenges and caveats concerning
Wright’s law, see Document S1 section ‘‘Wright’s law caveats.’’) Multifactor models
have been proposed using additional input variables, such as patenting activity
and research and development (R&D) expenditures, but data are limited and
they require additional parameters. This can lead to overﬁtting, resulting in poor
(see Document S1 section ‘‘Bias-variance trade-off’’).
Multifactor models have so far not been properly tested, and we do not use them
Successful technologies tend to follow an ‘‘S-curve’’ for deployment, starting with a
long phase of exponential growth in production that eventually tapers off due to
Under Wright’s law, during the exponential growth phase costs
drop exponentially in time, as they do for Moore’s law, but when production growth
eventually slows, cost improvement also slows. Improving technologies often spend
many decades in the exponential growth phase, making it hard to distinguish be-
tween Moore’s law and Wright’s law. Forecasts using the two models have similar
accuracy in backtesting experiments.
This brings up the important question of responsiveness to investment. Under
Moore’s law, costs are assumed to change exogenously over time, independent
of policy and investment. Under Wright’s law, costs depend on experience.
Although experience does not directly cause costs to drop, it is correlated with
other factors that do, such as level of effort and R&D, and has the essential
advantage of being relatively easy to measure.
For comparison, the historical
time series displayed in Figure 1 are plotted as experience curves in Figure S17.
The same heterogeneity of improvement rates seen in Figure 1 is evident for
Wright’s law—the fact that fossil fuel prices have not dropped historically means
that experience had no net effect—in stark contrast to key green technologies.
In this paper, we focus on Wright’s law because it satisﬁes the basic intuition that
exerting greater effort induces greater effects. (We repeated all our modeling
experiments using Moore’s law and found that the qualitative conclusions are
similar; see Document S1 section ‘‘Moore’s law results.’’ For a more thorough
discussion of causality, see Document S1 section ‘‘Discussion on questions of
Wright’s law has usually been used to generate point forecasts, meaning that the
forecast is a deterministic function of experience, with no estimate of the error of
the forecast. Early attempts at introducing error bars did not provide apriorifunc-
tional forms, which made the data requirements for out-of-sample testing prohibi-
More recently, apriorierror estimates were derived that predict forecasting
accuracy as a function of historical improvement rates and volatility, and the number
of data points available for forecasting.
Based on comprehensive backtesting,
this method was shown to generate reliable probabilistic estimates of future costs.
This was done by selecting reference dates in the past and then, using only the
data available at the time, making forecasts over all time horizons up to 20 years
different technologies, based on roughly 6,000 forecasts, the empirically observed
forecast accuracy closely matched the aprioriderived estimates on all time horizons
up to 20 years ahead.
Our main contribution in this paper is to systematically
apply this method—which we call the stochastic experience curve or stochastic
Wright’s law—to the energy transition.
4Joule 6, 1–26, September 21, 2022
To test the accuracy of the stochastic experience curve method for forecasting costs
of energy technologies, we applied it to historical data for solar, wind, batteries, and
polymer electrolyte membrane (PEM) electrolyzers; the results are shown in Figure 2.
Data prior to each forecast year were used to estimate parameters, then observed
deployment data in subsequent years were used to generate forecasts conditioned
on experience. The forecasts for solar, wind, and batteries are reasonable: most of
the future values lie within the 95% conﬁdence interval (CI), consistent with the apri-
ori error estimates. As expected, forecast uncertainty decreases for later forecasts
with more historical data. Due to the short dataset and high historicalvolatility, fore-
casts for electrolyzers are not as accurate, but the conﬁdence intervals capture this.
Figure 2. Historical performance of the stochastic experie nce curve forecasting method
(A–D) The four panels show stochastic Wright’s law applied to observed data for (A) solar, (B) wind, (C) batteries, and (D) P2X electrolyzers. Forecasts are
made at regular intervals, using prior cost and deployment data to calibrate the model and ‘‘future’’ deployment data to generate the forecasts.
Forecast medians and 95% conﬁdence intervals (CIs) are shown, and colors denote forecast year, from earliest (dark blue) to most recent (red). Costs are
LCOEs for solar and wind, and capacity costs for batteries and electrolyzers. P2X electrolyzers are assumed to be PEM electrolyzers here.
See Document S1 section ‘‘Data, calibration and technology forecasts’’ for further details and data sources.
Joule 6, 1–26, September 21, 2022 5
It is instructive to compare the accuracy of the forecasts in Figure 2 to the outputs
of inﬂuential global energy-economy models that are used to inform the Intergov-
ernmental Panel on Climate Change (IPCC) and guide global climate policy.
grated assessment models (IAMs) are used to evaluate policies and generate sce-
narios for deployment and cost that are consistent with given climate targets
under the assumption of optimal decision-making by economic agents. Their out-
puts are typically called ‘‘projections’’ to indicate that they are not intended to be
used as forecasts. Figure 3 emphasizes this point. Figure 3A shows past projections
of solar PV energy costs by the International Energy Agency’s (IEA) World Energy
Model and several IAMs and compares them with the observed data.
The projections shown correspond to scenarios with the most aggressive climate
policies and highest rates of technological innovation, i.e., those that produce the
highest rates of key green technology deployment and the most optimistic cost de-
clines. Nonetheless, their projected costs have been consistently much higher than
historical trends. The inset of Figure 3A gives a histogram of all 2,905 projections of
the annual rate at which solar PV system investment costs would fall between 2010
and 2020, as reported by nine separate IAM teams in the AMPERE modeling com-
The mean value of these projected cost reductions was 2.6%,
Figure 3. Historical PV cost projections and ﬂoor costs
(A) The black dots show the observed global average levelized cost of electricity (LCOE) over time. Red lines are LCOE projections reported by the
International Energy Agency (IEA);
dark blue lines are integrated assessment model (IAM) LCOE projections reported in 2014;
and light blue lines
are IAM projections reported in 2018.
IAM projections are rooted in 2010 despite being produced in later years. The projections shown are
exclusively ‘‘high technological progress’’ cost trajectories drawn from the most aggressive mitigation scenarios, corresponding to the largest
projected cost reductions used in these models. Other projections made were even more pessimistic about future PV costs. The inset compares a
histogram of projected compound annual reduction rates of PV system investment costs from 2010 to 2020 with what actually occurred (based on all
2,905 scenarios for which the data are available
(B) PV system ﬂoor costs implemented in a wide range of IAMs. The colors denote the year the ﬂoor cost was reported, ranging from 1997 (dark green) to
2020 (light green). Observed PV system costs are also shown. The cost of PV modules scaled by a constant factor of 2.5 is provided as a reference.
For further details and data sources, see Figures 8 and 9A and Document S1 section ‘‘Solar PV electricity.’’
6Joule 6, 1–26, September 21, 2022
and all were less than 6%. In stark contrast, during this period, solar PV costs actually
fell by 15% per year.
This makes it clear that it would have been a bad idea to treat these projections as
conditional forecasts. By contrast, the stochastic experience curve method produces
reliable conditional forecasts of known accuracy (and a published forecast of 2020
far more accurate than any of the IAM or IEA projections made at the time
). One of
our goals in this paper is to illustrate how such forecasts are useful for planning the
energy transition. (Note that IAM and IEA projections are better for mature incum-
bent technologies such as fossil fuels, but their projections for solar PV, wind, batte-
ries, and electrolyzers have systematically underestimated deployment and overes-
Wright’s law is widely used to generate technology cost projections in IAMs.
However, it is typically used in conjunction with ad hoc constraints such as deploy-
ment rate limits and ﬂoor costs, i.e., ﬁxed levels that costs are assumed to never
fall below. Because IAMs use costs to determine deployment (and vice versa), and
many allow perfect foresight, constraints are necessary to prevent sharp cost de-
clines due to Wright’s law from leading to solutions in which key green technologies
are deployed faster than is physically or socially plausible. It is difﬁcult to know what
constraints are realistic, which leads to ad hoc choices that strongly inﬂuence the
The historical record indicates that the constraints on key green technology deploy-
ment and costs used in IAMs have so far been much too stringent. For example, as
shown in Figure 3B, past ﬂoor costs used in IAMs have already repeatedly been
violated. We know of no empirical evidence supporting ﬂoor costs and do not
impose them (see Document S1 section ‘‘The use of ﬂoor costs in endogenous tech-
nological learning models’’). Similarly, while there are likely limits to how quickly we
can deploy key green technologies, it is difﬁcult to know what they are. The outputs
of IAMs depend critically on these constraints, which always alter the projections in
the same direction, making them more pessimistic about the costs and deployment
of key green technologies. As demonstrated here, the exponential growth of key
green technologies and the relative accuracy of the unconstrained version of
Wright’s law suggest that thus far these constraints have not been binding. The
imposition of excessively strong constraints is likely an important reason why the
projections of these models have not corresponded to the historical record.
Probabilistic cost forecasts for individual technologies
We applied the methods discussed so far to make forecasts of future energy costs
and prices. Given the reasons discussed in the introduction, for solar, wind, batte-
ries, and electrolyzers, we used the stochastic form of Wright’s law; and for oil,
coal, and gas, we used an autoregressive model of order 1 (AR(1)). To generate
experience curve forecasts, parameters for each technology were estimated from
historical data. We then speciﬁed scenarios for the future deployment of each tech-
nology as a function of time and predicted a distribution of future costs.
We deﬁned three representative deployment scenarios that we will explain in more
detail later. The ﬁrst scenario is consistent with the energy system transitioning away
from fossil fuels by around 2050, and so we label this deployment scenario the ‘‘Fast
Transition’’. The second scenario is consistent with eliminating fossil fuels by around
2070, so we label it the ‘‘Slow Transition’’. The ﬁnal scenario is consistent with fossil
Joule 6, 1–26, September 21, 2022 7
fuels continuing to dominate the energy system, so we label it the ‘‘No Transition’’.
Figure 4 shows these three deployment scenarios for each of the four key green tech-
nologies in the context of their historical long-term growth trends. The deployment
trajectories for these scenarios are consistent with the S-shaped curve widely
observed in technology diffusion—the differences between them reﬂect differences
in the timing and abruptness with which the growth of each technology tapers off.
Figure 5 shows probabilistic forecasts for seven particularly important energy tech-
nologies. The main panels of Figures 5A–5D show forecasts for key green technolo-
gies in the Fast Transition scenario, which are made using the stochastic version of
Wright’s law. The insets show costs versus experience and emphasize that median
costs develop identically as a function of experience in all scenarios. The side panels
of Figures 5A–5D illustrate that under Wright’s law, forecast distributions depend on
the scenario; as a result, in a faster transition, we are likely to reach lower costs
sooner. Each Wright’s law technology initially follows its current trend of exponen-
tially decreasing costs, but then progress slows as its rate of deployment drops.
To generate fossil fuel cost forecasts, the AR(1) model was calibrated to observed
data. For fossil fuels, model parameters depend on past data, but forecasts are in-
dependent of deployment, so each technology has a single forecast in all scenarios.
Figure 5 also shows a selection of future cost projections reported by IAM and IEA
studies. As before, we show only their most optimistic projections, i.e., low cost pro-
jections that correspond to high technological progress scenarios. Consistent with
the historical behavior of these models illustrated in Figure 3, the cost projections
are high relative to historical trends. They are also all substantially higher than our
Of course, the deployment corresponding to these cost projections is not the same as
that used to make our forecasts, so they are not perfectly comparable. However, as the
boxplot panels show, the disparities persist across all our scenarios, including the No
Transition scenario. This makes it clear that our cost forecasts are, all things equal, signif-
icantly lower than those used in these highly inﬂuential energy-economy models.
The stochastic version of Wright’s law we use here captures the historical volatility of
technological progress and the associated parameter estimation error, and it
Figure 4. Technology deployment of key green technologies
Observed data up to 2020 are shown, plus three hypothetical growth scenarios up to 2050, corresponding to Fast Transition (blue), Slow Transition
(yellow), and No Transition (gray). The trend line shown is the line through the ﬁrst and last data points plotted and is indicative of the long-run trendso
8Joule 6, 1–26, September 21, 2022
Figure 5. Technology cost forecasts
(A–D) The main panels show cost forecast distributions under the Fast Transition scenario for solar PV, wind, batteries, and PEM electrolyzers; the 50%
conﬁdence interval (CI) is dark blue, and the 95% CI is light blue. Also shown are several representative recent and past projections corresponding to
Joule 6, 1–26, September 21, 2022 9
projects this uncertainty forward in future cost distributions. It thus provides cost
ranges that are supported by empirical evidence, as opposed to the ad hoc ranges
that are often used to design energy scenarios and pathways.
Similarly, the AR(1)
model captures the historical volatility of fossil fuel prices and projects this forward in
forecast distributions. Although it might appear that the range of possible outcomes
in Figure 5 is larger for solar, wind, and batteries than for fossil fuels, they are actually
Fast Transition scenario, for example, ranges from roughly $2 to $40 per MWh, which
is a factor of 20 and an absolute range of $38. By contrast, the price of oil in 2050
ranges from $20 to $110 per barrel, which is a factor of 5.5 but an absolute range
of $90. The uncertainty ranges we forecast for fossil fuels are in line with IEA
Note that the uncertainties for electrolyzers are much higher than for
the other three key green technologies because the historical series is short and
From single technologies to a full system model
To forecast the likely costs of the green energy transition and explore how uncer-
tainty in individual technology costs propagates through to uncertainty in system
costs, we constructed a simple, transparent model of the global energy system
based on well-deﬁned technology deployment scenarios. We added seven more
technologies to the seven technologies already presented: coal-ﬁred and gas-ﬁred
electricity, nuclear power, hydroelectric power, biopower, redox ﬂow batteries, and
electricity networks. While the real-world energy system includes many other tech-
nologies, we used this limited ensemble because (1) it covers most of the current
ﬁnal and useful energy of the system (around 90%), (2) it includes sufﬁciently many
diverse technologies for representing a wide range of energy transition pathways,
and (3) it maintains a level of simplicity suitable for conveying our main results on
future technology costs and their uncertainties.
Since our study is not intended to be comprehensive, but rather to focus on cost de-
clines for key green technologies, we do not consider liquid biofuels, geothermal
power, marine energy, traditional biomass, co-generation of heat, solar thermal en-
ergy, or CCS (our results are nevertheless robust to these modeling choices; see
Our approach to scenario construction differs from that currently used in most stan-
dard energy-economy models, where deployment in one period is used to project
costs in the next, and vice versa. By iterating between costs and deployment in
this way, small errors can quickly get ampliﬁed, leading to scenarios that are incon-
sistent with empirically observed trends. Instead, we followed earlier energy system
and constructed scenarios exogenously by specifying how much energy or
storage will be provided by each technology as a function of time, just as we did
for single technology deployment trajectories earlier (Document S1 section ‘‘Sce-
nario construction’’). We classify energy services into four categories—transport,
industry, buildings, and energy sector self-use (Document S1 section ‘‘Model
Figure 5. Continued
‘‘optimistic’’ mitigation scenarios made by IAMs and the IEA (red lines) (see Figure 9). For batteries, both lithium-ion (Li-ion) consumer cells and Li-ion
electric vehicle (EV) battery packs are shown, although their costs have now converged; our forecasts are based on consumer cells, whereas the IEA
projections shown are based on EV batteries. The boxplots in the right-hand panels compare cost forecasts in 2050 under the No Transition, Slow
Transition, and Fast Transition scenarios. The insets show historical experience curves and forecasts, with learning rates that are independent of the
scenario, and vertical lines that indicate how far each technology moves down the probabilistic experience curve in each scenario.
(E–G) These panels show probabilistic cost forecasts for oil, coal, and gas based on the AR(1) time-series model (see Document S1section ‘‘Data,
calibration and technology forecasts’’ for details of data sources and model calibration).
10 Joule 6, 1–26, September 21, 2022
components’’)—and assume that end-use sector demand grows at an overall rate of
2% per year (we vary this assumption in Document S1 section ‘‘Sensitivity to system
growth rate’’). We impose the constraint that all scenarios must reliably provide iden-
tical levels of energy services througho ut the economy. This straightforward scenario
construction method allowed us to match long-standing technology growth trends.
This is in contrast to scenarios generated by IAMs, which typically do not match his-
torical deployment trends for fast-progressing technologies and often fail to match
wider transition dynamics too.
The three scenarios that we introduced earlier—Fast Transition, Slow Transition, and
NoTransition—areshowninmoredetailinFigure 6. They run from 2021 to 2070 and
were chosen to represent three distinctly different energy system pathways. In the
Fast Transition scenario (Figures 6A, 6D, and 6G), solar, wind, and batteries continue
to grow at rates that are somewhat slower than their long-term growth rates, as indi-
cated in Figure 4, for approximately a decade. Electrolyzers are at an earlier stage of
their S-curve; they behave similarly but stick to their current exponential growth rate
for two decades. Following trajectories similar to standard S-curves, once these
technologies become dominant, deployment slows to grow at 2% per year. Short-
term storage and electriﬁcation of most transport are achieved with batteries,
whereas long-duration energy storage (LDES) and all hard-to-electrify applications
are served by power-to-X (P2X) fuels, i.e., by using electricity for hydrogen electrol-
ysis and either directly using hydrogen or using it to make other fuels such as
ammonia and methane as needed.
This corresponds to an ‘‘electrify almost every-
thing’’ scenario, with full sector coupling.
Under this scenario, as shown in Fig-
ure 6D, emissions quickly get close to zero. If non-energy sources of carbon emis-
sions such as agriculture and land-use change are brought under control, it would
likely meet the 1.5Paris Agreement target (Document S1 section ‘‘Emission reduc-
tions and reduced climate risks’’).
In the Slow Transition scenario (Figures 6B, 6E, and 6H), by contrast, current rapid
deployment trends for key green technologies slow down immediately, so that fossil
fuels are phased out more slowly and continue to dominate until mid-century.
Finally, in the No Transition scenario (Figures 6C, 6F, and 6I), the energy system re-
mains similar to its current form for several decades, as low-carbon energy sources
grow only very slowly. This is similar to the reference or ‘‘no policy’’ scenario used
by many IAMs. To provide some context for these three scenarios, they are
compared with scenarios from the IPCC’s Sixth Assessment Report (AR6) in Docu-
mentS1section‘‘ComparisonwithAR6scenarios.’’ Full scenario details are shown
in Document S1 section ‘‘Scenarios.’’
To understand these scenarios, it is important to distinguish ﬁnal energy,whichisthe
energy delivered for use in sectors of the economy, from useful energy,whichisthe
portion of ﬁnal energy used to provide energy services, such as heat, light, and ki-
netic energy (Document S1 section ‘‘Energy system description’’). Fossil fuels tend
to have large end-use conversion losses in comparison to electricity, which means
that signiﬁcantly more ﬁnal energy is required to obtain a given amount of useful en-
ergy. Switching to energy carriers with higher conversion efﬁciencies (e.g., moving
to electric vehicles [EVs]) signiﬁcantly reduces ﬁnal energy consumption.
the Fast Transition scenario, eventually almost all energy services originate with
electricity generated by solar PV and wind, which is used either directly, via batteries,
or by making P2X fuels for later consumption. As shown by comparing Figures 6G
and 6I, the Fast Transition qualitatively increases the role of electricity in the energy
Joule 6, 1–26, September 21, 2022 11
Our model captures approximately 90% of current ﬁnal energy (excluding energy
carriers that are already renewable, such as bioenergy and biofuels, plus petrochem-
ical feedstock, which is not an energy carrier; see Table S2). Of course though, useful
energy is what matters. The model also covers around 90% of current useful energy,
but this is more difﬁcult to estimate. Under the Fast Transition and Slow Transition
scenarios, this fraction increases with time due to increased electriﬁcation. (See
Document S1 section ‘‘Model description’’ for further details).
To estimate full system costs, we need cost forecasts for all technologies. Although
coal-ﬁred electricity and gas-ﬁred electricity showed signiﬁcant cost declines for
Figure 6. Scenarios
(A–I) The three columns represent the three energy system scenarios. The three rows are: (A–C) annual useful energy provided by each technology as a
function of time; (D–F) annual ﬁnal energy provided by each technology as a function of time; and (G–I) annual electricity generation and storage in grid-
scale batteries and EV batteries. Total electricity generation is divided between ﬁnal electricity delivered to the economy and electricity used to
produce P2X fuels for hard-to-electrify applications and for power grid backup.
12 Joule 6, 1–26, September 21, 2022
some of the 20th century (as power generation components underwent technolog-
ical learning; see Figure 1), in the long run their costs are increasingly dominated by
so we used the AR(1) model for these (Document S1 sections ‘‘Coal elec-
tricity’’ and ‘‘Gas electricity’’). We used stochastic Wright’s law for nuclear power, hy-
dropower, and biopower, although these all have ﬂat or rising costs, so the choice of
cost model is immaterial here, and they have only limited signiﬁcance for energy
transition in this analysis. We also used stochastic Wright’s law for ﬂow batteries,
but for electricity networks, since we did not have historical cost data characterizing
them in enough detail, we pessimistically assumed that unit costs will remain the
same as they are now, bearing in mind that costs are heterogeneous (Document
S1 section ‘‘Electricity networks’’).
How much will each scenario cost?
There are many different approaches to modeling energy system pathway costs.
We used the ‘‘direct engineering costs’’ approach, in which the overall cost of a
scenario is computed by adding up the costs of the component technologies (Docu-
ment S1 section ‘‘Estimating total system costs’’). We summed the costs of direct-use
oil, coal, and gas; electricity generated by seven different technologies; utility-scale
grid batteries and electrolyzers; and additional infrastructure for expansion of the
electricity grid. For electricity generation costs, we used the levelized cost of elec-
tricity (LCOE) metric. This is particularly advantageous here because then the expe-
rience curve formulation inherently captures historical progress trends in all LCOE
components, including capital costs, capacity factors, and interest rates, which
would otherwise be difﬁcult to forecast separately (Document S1 section ‘‘Units
and justiﬁcation for the use of LCOE’’). We estimated infrastructure costs that are
not directly covered by technologies included in the model, for example, for fuel
storage and distribution (Document S1 section ‘‘Fuels infrastructure’’), and for
fueling or charging light duty vehicles (Document S1 section ‘‘Electriﬁcation of trans-
port’’), and argue that they are roughly the same across scenarios.
To apply our probabilistic technology cost forecasting methods in a given scenario,
we employed a Monte Carlo approach, simulating many different future cost trajec-
tories, then exponentially discounting future costs to calculate the expected net pre-
sent cost (NPC) of the scenario up to 2070 (Document S1 sections ‘‘Net present cost
of transition’’ and ‘‘Main case results’’). Figure 7A shows annual system costs through
time for each scenario. The black boxplots represent the full cost forecast distribu-
tions, whereas the colored bars show median expenditures by technology group.
This shows how, in the Fast Transition scenario, expenditures transfer rapidly from
fossil fuels to key green technologies.
Figure 7B shows the annual system cost forecast distributions in 2050. Rapid replace-
ment of fossil fuel technologiesby low-cost key green technologies—in power and trans-
port in particular—causes the expected annual energy system cost in 2050 for the Fast
Transition scenario to be $514 billion cheaper than that for the No Transition scenario,
although the distribution of possible costs for the Fast Transition is wider. After 2050,
as shown in Figure 7A, while the median and interquartile range (IQR) remain relatively
low, the uncertainty of the Fast Transition in relation to No Transition increases. If costs
are in the upper end of theuncertainty range, cheaper alternativeswould be used; we are
not taking this into account, which is a drawback of our method.
Figure 7C shows the forecast distribution of the NPC of each scenario at a ﬁxed dis-
count rate of 2%. Although there is considerable uncertainty, the NPC of the Fast
Transition is likely to be quite a bit lower than that of the No Transition. By contrast,
Joule 6, 1–26, September 21, 2022 13
the Slow Transition is not as cheap as the Fast Transition. This is because the current
high spending on fossil fuels continues for decades, and the savings from key green
technologies are only realized much later. Nonetheless, it also generates savings
relative to the No Transition scenario. Similarly to Figure 7B, the NPC distribution
of the Fast Transition is wider than that of the No Transition. Although this is caused
by higher technology uncertainty, it is important to note that this increased uncer-
tainty is compensated for by the leftward shift in the distribution, due to expected
cost declines associated with scaling up key green technologies.
Figure 7. Scenario costs
(A) Colored bars show median annual expenditures on fossil fuel and non-fossil fuel technologies in each scenario in trillions of dollars (tn USD).
Boxplots show the median and interquartile range (IQR) of total annual expenditures, and whiskers extend from the box by 1.5 times the IQR.
(B) Forecast distributions of the annual system cost in 2050 for each scenario.
(C) Forecast distributions of the net present cost (NPC) of each scenario, for a ﬁxed discount rate of 2%.
(D) Expected net present cost of each scenario relative to the No Transition scenario, as a function of the discount rate. The inset shows the probability
that the NPCs of the Fast Transition and Slow Transition will be lower than that of the No Transition, as a function of the discount rate.
14 Joule 6, 1–26, September 21, 2022
Figure 7D shows how the expected NPC of each scenario varies with the discount
rate relative to the No Transition scenario. The inset shows that there is roughly an
80% chance that the NPC of the Fast Transition is lower than that of the No Transi-
tion, regardless of discount rate. Previous analyses have suggested that whether or
not it makes good economic sense to quickly transition to clean energy technologies
depends on the discount rate.
But here we show a striking result: the Fast Tran-
sition is likely to be much cheaper at all reasonable discount rates.Usingthe1.4%
social discount rate recommended in the Stern Review,
for example, the expected
net present saving is roughly $12 trillion. At the higher discount rate of 5%, the ex-
pected saving is around $5 trillion. Note that there is some evidence that technolog-
ical progress does not slow when technologies reach their saturation phase.
is true, then costs continue to drop at their current pace, according to Moore’s law,
and the Fast Transition saves even more relative to the other scenarios (see Docu-
ment S1 section ‘‘Moore’s law results’’).
We constructed an additional scenario in which nuclear plays a dominant role in re-
placing fossil fuels, but this is much more expensive than the other scenarios. For
example, using a 1.4% discount rate, the expected NPC is about $25 trillion more
than for the No Transition (Document S1 sections ‘‘Slow nuclear transition’’ and
‘‘Main case results’’).
To enhance the credibility of our estimates, we have used consistently conservative as-
sumptions regarding the costs, performance, and operational requirements of clean en-
ergy technologies, and we have done the opposite for fossil fuels. Our requirement that
we ground forecasts on historical data means that in many cases we were forced to
neglect promising solutions, such as demand-side management of power grids, heat
pumps, and end-use efﬁciency, where there are insufﬁcient data.
As a result, the esti-
mated savings presented here should be viewed as lower bounds on the savings likely
to be achieved in reality, as many other innovative technologies and solutions are likely
to be developed (Document S1 section ‘‘Estimating total system costs’’).
Our analysis is based on (weighted) global average costs, but there is wide
geographic variation in energy costs. Within countries, solar and wind tend to be de-
ployed ﬁrst in regions where their costs are favorable, but that is not the case globally
(Document S1 section ‘‘Regional differences in competing technologies’’). In any
case, under the Fast Transition, regional cost differences are quickly overcome
through time. In the historical record of solar PV, for example, it takes less than a
distribution at any ﬁxed point in time. Because costs are summed here, global aver-
ages are sufﬁcient to estimate total system costs, and we expect that future efforts
will take advantage of geographic variation to achieve even cheaper solutions.
Although the Fast Transition happens quickly, it is still possible to replace the energy
system without excessive stranding of capital. Lifetimes of large energy infrastruc-
ture projects typically range from 25 to 50 years, meaning that on average about
2%–4% of capacity needs replacing in any given year. In addition, useful energy de-
mand grows at 2% per year in all our scenarios. These two factors makeit possible for
key green technologies to replace most of the existing energy system in 20 years,
and replace the remaining 5% within a few decades more, without necessarily
stranding assets before their economic lifetimes. Past estimates that suggest the
emissions from existing, planned, and proposed electricity generation infrastructure
will exceed the Paris carbon budget assumed that current utilisation rates of such as-
sets will remain constant in future, despite an increasingly competitive market, and
Joule 6, 1–26, September 21, 2022 15
that all planned and proposed deployments will go ahead, which has become
increasingly unlikely over the last decade (Document S1 section ‘‘Stranded assets’’).
To avoid confusion, we want to be clear about what we have done and what we have
not done. In contrast to most IAMs, which attempt to project both deployment and
costs conditioned on policies, we are less ambitious: we only forecast costs condi-
tioned on deployment. Although we have tried to choose deployment scenarios
that we think are reasonable, we do not attempt to forecast deployment. Our moti-
vation for taking a less ambitious modeling approach is that this allows us to stay
close to the empirical data. For improving technologies,webaseallourforecasting
on methods that have been carefully tested by making out-of-sample forecasts. We
do not assume future technology costs, we forecast them using a well-tested meth-
odology. If the historical data were different, the results would be different. Humility
is always required in making forecasts, butwehavegonetogreatlengthstoground
our forecasts with empirical data, using statistical methods to assess their reliability.
Although our Fast Transition scenario is subjective, we believe it is plausible (Docu-
ment S1 section ‘‘Is the speed of the Fast Transition achievable?’’). The deployment
trajectories are in line with past trends. There appear to be no major obstacles to
bringing the necessary technologies to scale in terms of land use, sea, climate,
raw materials, manufacturing capacity, energy return on energy invested, or system
Nonetheless, there are signiﬁcant institutional challenges, and to stay
on the current growth paths for the next decade, policies that enforce portfolio stan-
dards and/or stimulate demand will likely be needed. Our key contribution here is to
show that if we can stay on these growth paths for the next decade, we will likely
realize substantial savings. The cornerstone of the Fast Transition scenario is the
timely expansion of key green technologies, because only as these are scaled up
can fossil fuels be phased out and the savings be realized. The primary policy impli-
cation of our results is that there are enormous advantages to rapid deployment of
key green technologies. Achieving this is likely to require strong international pol-
icies for building infrastructure, skills training, and making the investments required
to realize future gains.
Our approach is complementary to IAMs. It builds on historical trends directly and
thus provides a counterweight to projections by IAMs. We have demonstrated
that the constraints that are commonly used in IAMs are likely an important cause
of the mismatch of their projections with historical data. Future work could explore
how softening these constraints within IAMs changes their projections.
We want to stress that, unlike IAMs, we are not attempting to ﬁnd optimal solutions.
There are very likely other scenarios that are cheaper than the Fast Transition sce-
nario, which was constructed to explore whether (with sufﬁciently rapid deployment)
a rapid transition can achieve net cost savings, and if so, with what probability. Given
the likely future cost of gas, it could be possible to achieve cheaper scenarios by us-
inggasinplaceofP2Xfuelsinsomelocations and applications, but these of course
would not be zero-emissions systems. Similarly, while fossil fuel prices have not his-
torically trended down, competition from key green technologies may force them
down, although this is feasible only at substantially reduced production levels where
only the cheapest fossil fuel producers are competitive.
This emphasizes the point
that, while most of the Fast Transition is aligned with market forces, policies that
discourage the use of fossil fuels will still be needed to fully decarbonize energy.
16 Joule 6, 1–26, September 21, 2022
Our forecasts are based on time-series methods. It would be preferable to be able to
make forecasts based on ﬁrst principles, but unfortunately this is not possible (even
though Wright’s law can be derived from a simple model
). Certain characteristics
of technologies, such as modularity, are predicted to be associated with more rapid
but despite some evidence that this is indeed true for
modularity, it only explains a small fraction of the variance.
A fundamental limita-
tion of all model-based forecasting methods is data, as this is required for model
calibration. When historical data are too sparse or do not exist, these methods
can not be used. In such cases, expert elicitation methods may be applied,
although this requires care, as retrospective analysis indicates that time-series
methods are more reliable.
Future research to address the issue of sparse data
could consider hybrid approaches that combine model-based forecasts with expert
judgments, or it could investigate the extent to which technology analogs can reli-
ably be used in forecasting. Perhaps most useful though would be a better under-
standing of the role of technology aggregation in model-based forecasting; for
example, how do forecasts of global solar PV costs relate to speciﬁc subtypes of
PV and to regional costs? In any case, the time-series methods that we used here
are currently the most reliable way to make conditional forecasts of technology costs
that agree with historical data.
Although the infrastructure costs for a rapid green energy transition are substantial,
we forecast that they are likely to be more than offset by lower energy costs. The
largest infrastructure cost is for enhanced grid capacity. In 2050, for example, our
estimated electricity network annual expenditure for the Fast Transition is about
$670 billion per year, compared with $530 billion per year for the No Transition.
However, the expected total system cost in 2050 is about $5.9 trillion per year for
the Fast Transition and $6.3 trillion per year for the No Transition. Thus, although
the additional $140 billion of grid costs might seem expensive, it is signiﬁcantly
less than the savings due to cheaper energy. The essential reason that the Fast Tran-
sition is cheaper than the Slow Transition is because it realizes the cost savings due to
cheaper energy sooner—faster deployment increases the probability of rapid prog-
ress in key green technologies, so that savings accrue for longer.
The likelihood of cheaper energy raises the possibility of a rebound effec t. Cheaper en-
ergy may increase global energy demand so that it grows faster than the historical 2%
per year rate assumed here. We view this as a ‘‘good problem’’: while this would raise
overall costs in the Fast Transition, renewables produce clean energy, and cheaper en-
ergy is likely to improve global living standards. To address the potential cost in-
creases, we performed a sensitivity analysis around the system growth rate assumption,
by also considering long-run growth rates of 1% and 3%. The resulting energy systems
vary widely in size and therefore represent a wide range of plausible population, eco-
nomic, and technological pathways in future. We found that our results are robust to
these variations (Document S1 section ‘‘Sensitivity to system growth rate’’).
In response to our opening question, ‘‘Is there a path forward that can get us to net-
zero emissions cheaply and quickly?,’’ our answer is: ‘‘Very likely, and the savings are
probably quite large.’’ Our quantitative analysis supports other recent efforts using
up-to-date data and technology assumptions that conclude that the green energy
transition may be cheap.
The 2022 IPCC AR6 estimates that the additional
cost of decarbonizing the energy system in order to have a greater than 67% chance
of keeping warming below 2C corresponds to a GDP loss in 2050 of 1.3%–2.7%.
Our results suggest that there is likely no cost at all—the transition is expected to be
a net economic beneﬁt, raising future GDP.
Joule 6, 1–26, September 21, 2022 17
We have demonstrated that the models used by the IPCC have in the past consis-
tently overestimated key green technology costs. It is important that this bias is
addressed (although, as shown in Figure 9A, the PV cost projections in the AR6 data-
the probabilistic forecasts presented in this paper, the IPCC database models are
only considering costs that are extremely unlikely—in the pessimistic direction—
and are not fully exploring the space of plausible scenarios. IPCC conclusions thus
appear to be based on an over-sampling of near worst-case scenarios regarding
key green technology costs. While even technologies with strong progress trends
sometimes experience cost increases, as has occurred for solar in the mid-2000s
and the current period due to supply shortages of key production inputs, our results
account for these ﬂuctuations. By carefully characterizing historical progress trends
and volatility, the stochastic methods used here capture both the downside risk that
progress in some key green technologies may stall, and the upside risk—the prob-
ability that via routine invention and innovation some technology costs will fall faster
than historical trends some of the time.
Our analysis indicates that even the downside outcomes of a rapid green energy
transition are not that bad, due to the dramatic cost declines seen already. When en-
ergy system pathways are viewed in terms of bets placed on portfolios of technolo-
we ﬁnd that the Fast Transition scenario has an expected payoff of around
$5–$15 trillion. Moreover, it is also a safe bet, with around an 80% probability that
it will be cheaper than continuing with a fossil fuel-based system (and 82% when
compared with a slower transition). Since the future is uncertain, all public policy
and decision-making is ultimately a question of making the smartest bets we can,
given the often precarious circumstances we face. Our results suggest that deploy-
ing technologies according to the Fast Transition scenario is a very good bet, both in
terms of lowest costs and lowest emissions.
We want to emphasize that our results indicate that a rapid greenenergy transition is
likely to be beneﬁcial, even if climate change were not a problem. When climate
change is taken into account, the beneﬁts of the Fast Transition become over-
whelming. A common simpliﬁed method for estimating economic damages due
The range of proposed values is vast, but just as an example, at a discount rate of
5%, assuming SCC values in the range $30–300/tCO
(rising at 3% per year
total expected Fast Transition savings, up to 2070, of $31–$255 trillion. At a lower
discount rate of 1.4%, the range of expected savings is $88–$775 trillion. Thus,
the beneﬁts of the Fast Transition are likely much larger than the energy system
cost savings evaluated in this study.
The belief that the green energy transition will be expensive has been a major
driver of the ineffective response to climate change for the past 40 years. This
pessimism is at odds with past technological cost improvement trends and risks
locking humanity into an expensive and dangerous energy future. While argu-
ments for a rapid green transition cite beneﬁts such as the avoidance of climate
damages, reduced air pollution, and lower energy price volatility (Document S1
section ‘‘Additional beneﬁts from the Fast Transition’’), these beneﬁts are often
contrasted against discussions about the associated costs of the transition. Our
analysis suggests that such trade-offs are unlikely to exist: a greener, healthier,
and safer global energy system is also likely to be cheaper. Updating expectations
to better align with historical evidence could fundamentally change the debate
18 Joule 6, 1–26, September 21, 2022
about climate policy and dramatically accelerate progress to decarbonize energy
systems around the world.
Correspondence and requests for resources should be addressed to rupert.way@
This study did not generate any new materials.
We employ two time-series models for forecasting technology costs. The ﬁrst is a
ﬁrst-difference stochastic form of Wright’s law, developed and tested by Lafond
which models costs dropping as a power law of cumulative production.
Let ctbe the cost and ztbe the experience of a given technology at time t,andlet
uÞbe an independent and identically distributed (IID) draw from a normal
distribution. Then future costs are predicted using the iterative relationship
log ctlog ct1=uðlog ztlog zt1Þ+ut+rut1:(Equation 1)
This relationship has three parameters. For a given technology, the experience
exponent ucharacterizes the average rate at which costs drop as a function of
experience, and the noise variance s2
ucharacterizes the variability of this relation-
ship. The autocorrelation parameter rcharacterizes the persistence of ﬂuctuations
in cost improvements. To avoid overﬁtting, and to ensure that our forecasts
adhere strictly to the same statistical properties as those tested by Lafond et al.
we use r=0:19 for all technologies, which was found to be a good overall
choice for 50 different technologies. This was necessary because ﬁtting three param-
eters to short data series such as those we have here degrades out-of-sample fore-
casting accuracy. (We also did a comparison of all our results replacing Wright’s law
by a generalized form of Moore’s law [see Document S1section ‘‘Moore’s law
When applying the model to technologies with falling costs, as shown in Figure 5,
two features of the model must be stressed. First, the Wright’s law model does
not simply ‘‘assume’’ that if costs fell in the past then they will fall in future—indeed,
costs are predicted to rise with a non-zero probability that depends directly on
observed data in the past. Second, despite the downward trends, all cost forecast
distributions are always strictly positive, since costs develop in log space.
For fossil fuels we use an AR(1) model:
log ct=log ct1+bðmlog ct1Þ+et;with IID etN0;s2
where m=E½log ctis the unconditional mean of the logarithm of cost, seis the vola-
tility of the noise process et,andbistherateofmeanreversion.Forcomprehensive
details on forecasting methods see Document S1 section ‘‘Technology cost
Additional technology cost projections
Figures 8 and 9show the solar and wind cost projection ensembles that underlie the
speciﬁc cost projections displayed in Figures 3 and 5.
Joule 6, 1–26, September 21, 2022 19
We model the supply of energy services increasing at a ﬁxed rate per year (2% in our
main speciﬁcation). Energy services include heating/cooling, light, mobility, suste-
nance, materials, hygiene, and communications, but since these are hard to measure
and data are sparse, we take useful energy as a proxy for energy services. Energy
transition scenarios are constructed by assuming that growth rates of energy carriers
and technologies follow logistic (‘‘S’’) curves with speciﬁed start and end points
consistent with the growth of total useful energy. We model the relationship
Figure 8. IEA PV LCOE projections
All PV LCOE projections found in the IEA’s World Energy Outlook (WEO) reports are shown in
colors varying from purple through light green (note that ‘‘projection’’ here means conditional
forecast—this is a forecast that is conditional upon a whole array of modeling assumptions
regarding the scenario within which the forecast is embedded). The ﬁrst such projection was found
in the WEO 2001. The four projections we selected to plot in Figure 3 are shown in red and were
chosen as examples of ‘‘high progress’’ projections. The ﬁrst two, published in the WEOs from 2001
and 2008, may be considered high progress projections, because in those reports, cost ranges were
provided, and we simply picked the lowest points of those ranges. The upper ends of the ranges
were signiﬁcantly higher. The second two (beginning in 2015 and 2019) may be interpreted as ‘‘high
progress’’ projections, because they correspond to the highest mitigation scenarios available in
the WEOs from which they are sourced (WEO 2016 and 2020). Note, however, that in those reports,
only region-speciﬁc cost projections were provided, so we have plotted the simple global average
of those values in the high mitigation scenarios. Observed values are from the Performance Curve
Database (described in Nagy et al.
) up to 2010 and from Bloomberg New Energy Finance (BNEF)
See Document S1 section ‘‘Data, calibration, and technology forecasts’’ for more details on data
20 Joule 6, 1–26, September 21, 2022
between ﬁnal and useful energy based on average conversion efﬁciency factors
given by DeStercke.
We take these conversion factors to be static and apply
them on a per-energy-carrier, per-sector basis. The model therefore does not
Figure 9. PV and wind capital cost projections reported by IAMs
Capital cost projections reported by various modeling comparison projects are shown as blue, red, yellow, and green lines for (A) PV and (B) onshore
wind. Each line corresponds to a single scenario. To construct and plot the LCOE projections in Figure 3, we selected two capacity cost projections
reported in 2014 (cyan lines) and two reported in 2018 (magenta lines). For Figure 5, we also added one PV projection reported in 2022 (dark green).
These may all be interpreted as ‘‘high progress’’ projections because they are among the lowest in their cohorts (i.e., the cyan lines are on the low end of
the suite of blue lines, the magenta lines are low relative to the red and yellow lines, and the dark green line is low relative to the green lines). Note that
these are (in eight out of nine cases) global average values, whereas some other projections are region speciﬁc. For PV, the projections plotted are as
follows: (1) model: MESSAGE, scenario: ‘‘AMPERE3-450,’’ region: World (from AMPERE
); (2) model: DNE21, scenario: ‘‘AMPERE3-450,’’ region: World
); (3) model: IMAGE 3.0, scenario: baseline, region: China (from Krey et al.
); (4) model: REMIND-MAgPIE 1.7–3.0, scenario:
SMP_1p5C_early, region: World (from SR15
); and (5) model: WITCH 5.0, scenario: EN_INDCi2030_1000, region: World (from AR6
). For wind, the
projections plotted are as follows: (1) model: MESSAGE, scenario: AMPERE3-450, region: World (from AMPERE
AMPERE3-450, region: World (from AMPERE
); (3) model: AIM/CGE 2.1, scenario: TERL_15D_LowCarbonTransportPolicy, region: World (from SR15
and (4) model: REMIND-MAgPIE 1.7–3.0, scenario: SMP_1p5C_early, region: World (from SR15
). To calculate LCOEs, we used technology lifetimes
and discount rates reported in Krey et al.,
and operations and maintenance (O&M) values from the original studies where possible (and if not, then
Krey et al.
values again). We used global average capacity factors of 0.18 for PV and 0.3 for wind, based on recent data reported by IRENA
Observed data sources for PV are given in Table S21. Wind data are from IRENA.
Joule 6, 1–26, September 21, 2022 21
include improvements in the underlying efﬁciency of each energy carrier at
providing energy services, but it does allow efﬁciency improvements on a per-sector
basis, via energy carrier substitution. For example, the conversion efﬁciencies of oil
and electricity in the transport sector are assumed constant, yet the efﬁciency of the
sector as a whole may be improved by switching from oil to electricity. The end point
of each scenario in 2070 is deﬁned by the shares of technologies providing electricity
generation and the shares of energy carriers providing ﬁnal energy. The start points
for all scenarios are identical and match the current shares in the base year, 2020.
Details of all growth rates, timings, and energy carrier mixes for each scenario are
given in Document S1 sections ‘‘Scenario construction’’ and ‘‘Scenarios.’’
Our modeling approach is based on two key design principles: (1) include only the
minimal set of variables necessary to represent most of the global energy system
and the most important cost and production dynamics, and (2) ensure all assump-
tions and dynamics are technically realistic and closely tied to empirical evidence
(Document S1 section ‘‘General approach’’). This means that we focus on energy
technologies that have been in commercial use for sufﬁcient time to develop a reli-
able historical record for forecasting purposes. This is also an essential constraint
because deployment of technologies typically takes a long time, and only technol-
ogies that have track records are positionedtoplayanimmediateroleinconfronting
We choose a level of model granularity well suited to the probabilistic forecasting
methods used, i.e., one that allows accurate model calibration and ensures overall
cost-reduction trends associated with cumulative production are captured for
each technology. Our model design can be run on a laptop, is easy to understand
and interpret, and allows us to calibrate all components against historical data so
that the model is ﬁrmly empirically grounded. The historical data do not exist to
do this on a more granular level.
We omitted several minor energy technologies. Co-generation of heat, traditional
biomass, marine energy, solar thermal energy, and geothermal energy were omitted
either due to insufﬁcient historical data or because they have not exhibited signiﬁ-
cant historical cost improvements, or both. Liquid biofuels were also excluded
because any signiﬁcant expansion would have high environmental costs (Document
S1 section ‘‘Bioenergy, solar thermal energy, marine energy and geothermal en-
ergy’’). Finally, CCS in conjunction with fossil fuels was omitted because (1) it is
currently a very small, low growth sector, (2) it has exhibited no promising cost im-
provements so far in its 50-year history, and (3) the cost of fossil fuels provides a
hard lower bound on the cost of providing energy via fossil fuels with CCS (Docu-
ment S1 section ‘‘CCS’’). This means that within a few decades electricity produced
with CCS will likely not be competitive even if CCS is free. There may of course be
some role for CCS in non-energy, direct-emission applications, but this is outside
the scope of this paper.
Since renewable energy production is variable, storage is essential. In the Fast Tran-
sition scenario we have allocated so much storage capacity using batteries and P2X
fuels that the entire global energy system could run for a month without any sun or
wind (Document S1 section ‘‘Energy storage and ﬂexibility requirements’’). This is a
sensible choice because both batteries and electrolyzers have highly favorable
trends for cost and production (Document S1 sections ‘‘Batteries’’ and ‘‘Hydrogen
and electrolyzers’’). From 1995 to 2018 the production of lithium-ion (Li-ion) batte-
ries increased at 30% per year, while costs dropped at 12% per year, giving an
22 Joule 6, 1–26, September 21, 2022
experience curve comparable to that of solar PV.
Currently, about 60% of the cost
of electrolytic hydrogen is electricity, and hydrogen is around 80% of the cost of
so these automatically take advantage of the high progress rates for so-
lar PV and wind.
We ensure system reliability constraints are met—including robustness to seasonal
demand variations—by providing sufﬁcient levels of energy storage, ﬁrm capacity
resources, over-generation of variable renewable energy (VRE) sources, and network
(Document S1 section ‘‘Energy storage and ﬂexibility requirements’’).
To be speciﬁc, when VRE penetration is high, we ensure enough utility-scale battery
storage is available to store 20% of average daily electricity generation (though note
that daily generation is much higher than daily end-use consumption, because
excess generation is used to produce P2X fuels). Flow batteries are able to store a
further 10% of average daily generation. In addition, when VRE penetration is
high, transport is electriﬁed, which as well as being a ﬂexible demand source, could
also act as another storage source (though system reliabilityconstraints are met here
without relying on it). Excess VRE is used to produce P2X fuels in sufﬁcient quantities
to supply all end-use sector requirements and also to provide global power grid
backup for 1 month each year.
Data and code availability
We used data from a wide range of sources. Many of these were free and openly
available on the internet, but some were accessed via standard university-wide sub-
scription licenses held by the University of Oxford. Sources include: the IEA,
the International Renewable Energy Agency,
the U.S. Energy Information
Bloomberg New Energy Finance, Bloomberg L.P. (via Bloomberg
Terminal), and several academic papers. For more details, see Document S1 section
‘‘Data, calibration and technology forecasts.’’ All data will be made available upon
request (unless legal restrictions exist).
The code used in this analysis will be made available upon request.
Supplemental information can be found online at https://doi.org/10.1016/j.joule.
This work was supported by funding from Partners for a New Economy (R.W. and
J.D.F.); Baillie Gifford (J.D.F.); the European Union’s Horizon 2020 research and
innovation programme under grant agreement no. 730427 (COP21 RIPPLES)
(R.W.); and the Oxford Martin School, through the Institute of New Economic
Thinking and the Post-Carbon Transition programme (LDR00530) (M.C.I. and
P.M.). This research was also funded by the Economics of Energy Innovation and Sys-
tem Transition project (EEIST), which is jointly funded through UK Aid by the UK Gov-
ernment Department for Business, Energy, and Industrial Strategy (BEIS) and the
Children’s Investment Fund Foundation (CIFF) (R.W. and M.C.I.). The contents of
this manuscript represent the views of the authors and should not be taken to repre-
sent the views of the UK government or CIFF. The authors gratefully acknowledge all
these sources of ﬁnancial support and, additionally, the Institute for New Economic
Thinking at the Oxford Martin School for its continuing support. The authors also
thank Lucas Kruitwagen and Jing Meng for valuable contributions to the work.
Joule 6, 1–26, September 21, 2022 23
Conceptualization, J.D.F.; methodology, R.W. and J.D.F.; software, R.W. and M.C.I.;
investigation, R.W.; data curation, R.W. and M.C.I.; formal analysis, R.W. and M.C.I.;
writing—original draft, J.D.F., R.W., M.C.I., and P.M.; writing—review & editing,
J.D.F., R.W., M.C.I., and P.M.; visualization, R.W.; supervision, J.D.F.; funding acqui-
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: December 23, 2021
Revised: May 16, 2022
Accepted: August 19, 2022
Published: September 13, 2022
1. Perlin, J. (1999). From Space to Earth: the Story
of Solar Electricity (Aatech Publications).
2. Creutzig, F., Agoston, P., Goldschmidt, J.C.,
Luderer, G., Nemet, G., and Pietzcker, R.C.
(2017). The underestimated potential of solar
energy to mitigate climate change. Nat.
Energy 2, 17140. https://doi.org/10.1038/
3. Xiao, M., Junne, T., Haas, J., and Klein, M.
(2021). Plummeting costs of renewables - are
energy scenarios lagging? Energy Strategy Rev
35, 100636. https://doi.org/10.1016/j.esr.2021.
4. Jaxa-Rozen, M., and Trutnevyte, E. (2021).
Sources of uncertainty in long-term global
scenarios of solar photovoltaic technology.
Nat. Clim. Change 11, 266–273. https://doi.
5. Hoekstra, Auke, Steinbuch, M., and Verbong,
G. (2017). Creating agent-based energy
transition management models that can
uncover proﬁtable pathways to climate change
mitigation. Complexity 2017, 1–23. https://doi.
6. Shiraki, H., and Sugiyama, M. (2020). Back to
the basic: toward improvement of
technoeconomic representation in integrated
assessment models. Clim. Change 162, 13–24.
7. Victoria, M., Haegel, N., Peters, I.M., Sinton, R.,
¨ger-Waldau, A., del Can
˜izo, C., Breyer, C.,
Stocks, M., Blakers, A., Kaizuka, I., et al. (2021).
Solar photovoltaics is ready to power a
sustainable future. Joule 5, 1041–1056. https://
8. Stern, N. (2016). Economics: current climate
models are grossly misleading. Nature 530,
9. Pindyck, R.S. (2013). Climate change policy:
what do the models tell us? J. Econ. Lit. 51,
10. Farmer, J.D., Hepburn, C., Mealy, P., and
Teytelboym, A. (2015). A third wave in the
economics of climate change. Environ. Resour.
Econ. 62, 329–357.
11. Gambhir, A., Butnar, I., Li, P.-H., Smith, P., and
Strachan, N. (2019). A review of criticisms of
integrated assessment models and proposed
approaches to address these, through the lens
of beccs. Energies 12, 1747. https://doi.org/10.
12. McCollum, D.L., Gambhir, A., Rogelj, J., and
Wilson, C. (2020). Energy modellers should
explore extremes more systematically in
scenarios. Nat. Energy 5, 104–107. https://doi.
13. Lovins, A.B., U
¨rge-Vorsatz, D., Mundaca, L.,
Kammen, D.M., and Glassman, J.W. (2019).
Recalibrating climate prospects. Environ. Res.
Lett. 14, 120201.
14. Pye, S., Broad, O., Bataille, C., Brockway, P.,
Daly, H.E., Freeman, R., Gambhir, A., Geden,
O., Rogan, F., Sanghvi, S., et al. (2021).
Modelling net-zero emissions energy systems
requires a change in approach. Clim. Policy 21,
15. Stern, N., Stiglitz, J.E., and Taylor, C. (2021).
The economics of immense risk, urgent action
and radical change: towards new approaches
to the economics of climate change. J. Econ.
Methodol. 29, 1–36.
16. Grubler, A., Wilson, C., Bent o, N., Boza-Kiss, B.,
Krey, V., McCollum, D.L., Rao, N.D., Riahi, K.,
Rogelj, J., De Stercke, S., et al. (2018). A low
energy demand scenario for meeting the 1.5C
target and sustainable development goals
without negative emission technologies. Nat.
Energy 3, 515–527. https://doi.org/10.1038/
17. Victoria, M., Zhu, K., Brown, T., Andresen, G.B.,
and Greiner, M. (2020). Early decarbonisation
of the European energy system pays off. Nat.
Commun. 11, 6223. https://doi.org/10.1038/
18. He, G., Lin, J., Sifuentes, F., Liu, X., Abhyankar,
N., and Phadke, A. (2020). Rapid cost decrease
of renewables and storage accelerates the
decarbonization of china’s power system. Nat.
Commun. 11, 2486. https://doi.org/10.1038/
19. Bogdanov, D., Ram, M., Aghahosseini, A.,
Gulagi, A., Oyewo, A.S., Child, M., Caldera, U.,
Sadovskaia, K., Farfan, J., De Souza Noel Simas
Barbosa, L., et al. (2021). Low-cost renewable
electricity as the key driver of the global energy
transition towards sustainability. Energy 227,
20. Arthur, W.B. (2011). The Nature of Technology.
What It Is and How It Evolves (Free Press).
21. Fouquet, R. (2008). Heat, Power and Light:
Revolutions in Energy Services (Edward Elgar
¨bler, A., Naki
c, N., and Victor, D.G.
(1999). Dynamics of energy technologies
and global change. Energy Policy 27, 247–280.
23. Koh, Heebyung, and Magee, C.L. (2006). A
functional approach for studying technological
progress: application to information
technology. Technol. Forecasting Soc. Change
73, 1061–1083. https://doi.org/10.1016/j.
24. Koh, H., and Magee, C.L. (2008). A functional
approach for studying technological progress:
extension to energy technology. Technol.
Forecasting Soc. Change 75, 735–758. https://
25. Nagy, B., Farmer, J.D., Bui, Q.M., and Trancik,
J.E. (2013). Statistical basis for predicting
technological progress. PLoS One 8, e52669.
26. U.S. Geological Survey (2022). Mineral
Commodity Summaries 2022. Technical report.
27. Trancik, J.E., Jean, J., Kavlak, G., Klemun, M.M.,
Edwards, M.R., McNerney, J., Miotti, M.,
Brown, P.R., Mueller, J.M., and Needell, Z.A.
(2015). Technology Improvement and
Emissions Reductions as Mutually Reinforcing
Efforts: Observations from the Global
Development of Solar and Wind Energy.
Working paper (MIT Press).
28. Reiner, D.M. (2016). Learning through a
portfolio of carbon capture and storage
demonstration projects. Nat. Energy 1, 1–7.
24 Joule 6, 1–26, September 21, 2022
29. Rubin, E.S., Davison, J.E., and Herzog, H.J.
(2015). The cost of CO
capture and storage.
Int. J. Greenhouse Gas Control 40, 378–400.
Special Issue commemorating the 10th year
anniversary of the publication of the
Intergovernmental Panel on Climate Change
Special Report on CO2 Capture and Storage.
30. Malkiel, B.G. (1973). A Random Walk down Wall
Street (Norton & Co).
31. Pindyck, R.S. (1999). The long-run evolution of
energy prices. Energy J 20, 19449089. http://
32. Shaﬁee, S., and Topal, E. (2010). A long-term
view of worldwide fossil fuel prices. Appl.
Energy 87, 988–1000. https://doi.org/10.1016/j.
33. Farmer, J.D., and Lafond, F. (2016). How
predictable is technological progress? Res.
Policy 45, 647–665.
34. Moore, G.E. (1998). Cramming more
components onto integrated circuits. Proc.
IEEE 86, 82–85.
35. Wright, T.P. (1936). Factors affecting the cost of
airplanes. J. Aeronaut. Sci. 3, 122–128.
36. Thompson, P. (2012). The relationship between
unit cost and cumulative quantity and the
evidence for organizational learning-by-doing.
J. Econ. Perspect. 26, 203–224.
37. Witajewski-Baltvilks, J., Verdolini, E., and
Tavoni, M. (2015). Bending the learning curve.
Energy Econ 52, S86–S99. https://doi.org/10.
38. Alberth, S. (2008). Forecasting technology
costs via the experience curve—myth or
magic? Technol. Forecasting Soc. Change 75,
39. Lafond, F., Bailey, A.G., Bakker, J.D., Rebois,
D., Zadourian, R., McSharry, P., and Farmer,
J.D. (2018). How well do experience curves
predict technological progress? a method for
making distributional forecasts. Technol.
Forecasting Soc. Change 128, 104–117.
40. IPCC (2022). Mitigation of climate change.
Contribution of working Group III to the sixth
assessment report of the Intergovernmental
Panel on Climate Change. In Climate Change
(Cambridge University Press).
41. Riahi, K., Kriegler, E., Johnson, N., Bertram, C.,
den Elzen, M., Eom, J., Schaeffer, M.,
Edmonds, J., Isaac, M., Krey, V., et al. (2015).
Locked into Copenhagen pledges —
implications of short-term emission targets for
the cost and feasibility of long-term climate
goals. Technol. Forecasting Soc. Change 90,
42. Krey, V., Guo, F., Kolp, P., Zhou, W., Schaeffer,
R., Awasthy, A., Bertram, C., de Boer, H.-S.,
Fragkos, P., Fujimori, S., et al. (2019). Looking
under the hood: a comparison of techno-
economic assumptions across national and
global integrated assessment models. Energy
172, 1254–1267. https://doi.org/10.1016/j.
43. Huppmann, D., Kriegler, E., Krey, V., Riahi, K.,
Rogelj, J., Calvin, K., Humpenoeder, F., Popp,
A., Rose, S.K., Weyant, J., et al. (2019). C
Scenario Explorer and Data Hosted by IIASA.
Technical report (IIASA).
44. Ferguson, C.D., Marburger, L.E., Doyne
Farmer, J.D., and Makhijani, A. (2010). A US
nuclear future? Nature 467, 391–393. https://
45. Wilson, C., Grubler, A., Bauer, N., Krey, V., and
Riahi, K. (2013). Future capacity growth of
energy technologies: are scenarios consistent
with historical evidence? Clim. Change 118,
46. Anandarajah, G., McDowall, W., and Ekins, P.
(2013). Decarbonising road transport with
hydrogen and electricity: long term global
technology learning scenarios. Int. J. Hydr.
Energy 38, 3419–3432.
47. Heuberger, C.F., Rubin, E.S., Staffell, I., Shah,
N., and Mac Dowell, N. (2017). Power capacity
expansion planning considering endogenous
technology cost learning. Appl. Energy 204,
48. DeCarolis, J., Daly, H., Dodds, P., Keppo, I., Li,
F., McDowall, W., Pye, S., Strachan, N.,
Trutnevyte, E., Usher, W., et al. (2017).
Formalizing best practice for energy system
optimization modelling. Appl. Energy 194,
49. Fricko, O., Havlik, P., Rogelj, J., Klimont, Z.,
Gusti, M., Johnson, N., Kolp, P., Strubegger,
M., Valin, H., Amann, M., et al. (2017). The
marker quantiﬁcation of the shared
socioeconomic pathway 2: a middle-of-the-
road scenario for the 21st century. Global
Environ. Change 42, 251–267. https://doi.org/
50. Gritsevskyi, A., and Naki
cenovi, N. (2000).
Modeling uncertainty of induced technological
change. Energy Policy 28, 907–921.
51. Trutnevyte, E. (2016). Does cost optimization
approximate the real-world energy transition?
Energy 106, 182–193.
52. Davis, S.J., Lewis, N.S., Shaner, M., Aggarwal,
S., Arent, D., Azevedo, I.L., Benson, S.M.,
Bradley, T., Brouwer, J., Chiang, Y.-m., et al.
(2018). Net-zero emissions energy systems.
Science 360, eaas9793. https://doi.org/10.
53. Brown, T., Schlachtberger, D., Kies, A.,
Schramm, S., and Greiner, M. (2018). Synergies
of sector coupling and transmission
reinforcement in a cost-optimised, highly
renewable European energy system. Energy
160, 720–739. https://doi.org/10.1016/j.
54. Eyre, N. (2019). Energy efﬁciency in the energy
transition (European Council for an Energy
Efﬁcient Economy Summer Study
Proceedings), pp. 247–254.
55. McNerney, J., Doyne Farmer, J., and Trancik,
J.E. (2011). Historical costs of coal-ﬁred
electricity and implications for the future.
Energy Policy 39, 3042–3054.
56. Krey, V. (2014). Global energy-climate
scenarios and models: a review. WIREs Energy
Environ 3, 363–383. https://doi.org/10.1002/
57. Edenhofer, O., Lessmann, K., Kemfert, C.,
Grubb, M., and Ko
¨hler, J. (2006). Induced
technological change: exploring its
implications for the economics of atmospheric
stabilization: synthesis report from the
innovation modeling comparison project.
Energy J. SI2006, 19449089. http://www.jstor.
58. Nordhaus, W.D. (2017). Revisiting the social
cost of carbon. Proc. Natl. Acad. Sci. USA 114,
59. Broome, J. (2008). The ethics of climate
change. Sci. Am. 298, 96–102. http://www.jstor.
60. Stern, N. (2007). The Economics of Climate
Change: the Stern Review (Cambridge
61. Woo, J.R., and Magee, C.L. (2021). Relationship
between technological improvement and
innovation diffusion: an empirical test. Technol.
Anal. Strateg. Manag. 1–16. https://doi.org/10.
62. Lowe, R.J., and Drummond, P. (2022). Solar,
wind and logistic substitution in global
energy supply to 2050 – barriers and
implications. Renew. Sustain. Energy Rev. 153,
63. Mercure, J.-F., Pollitt, H., Vin
Edwards, N.R., Holden, P.B., Chewpreecha, U.,
Salas, P., Sognnaes, I., Lam, A., and Knobloch,
F. (2018). Macroeconomic impact of stranded
fossil fuel assets. Nat. Clim. Change 8, 588–593.
64. McNerney, J., Farmer, J.D., Redner, S., and
Trancik, J.E. (2011). Role of design complexity
in technology improvement. Proc. Natl. Acad.
Sci. USA 108, 9008–9013.
65. Wilson, C., Grubler, A., Bento, N., Healey, S.,
De Stercke, S., and Zimm, C. (2020). Granular
technologies to accelerate decarbonization.
Science 368, 36–39. https://doi.org/10.1126/
66. Verdolini, E.