Vehicle database and sales time series
We obtain or infer EV and ICEV prices and characteristics (power, emissions, fuel use) directly from manufacturer websites and sales numbers from Marklines for each of 32 countries and 329 EV, 222 PHEV and 1901 ICEV models (combined across markets) between 2016 and 2022, of which 4 regions are shown here. We define three engine/motor power classes following loosely the Eurostat classification: Economic (<1400cc for ICEVs, <=30 kwh for EVs), Mid-range (>= 1400cc and < 2200cc for ICEVs, >30Kwh and <=70kwh for EVs) and Luxury (>2200cc for ICEVs, >70kwh for EVs). Vehicle prices are roughly lognormally distributed with different means, standard deviations and medians in each country.29 We accumulate vehicle data in this way since 2013 and have a yearly time series of price (and other characteristics) distributions since 2016 (Ext. Tab. 1).
Vehicle sales: The data for vehicle sales by model and technology (i.e. ICEVs, EVs, PHEVs etc.) were obtained from MarkLines, which provides sales data by brand-model since 2004 for more than 32 countries. To verify that MarkLines comprehensively covers all EV and ICEV models available and sold in each country in a way that is consistent with official national statistics, we carefully check that total MarkLines vehicle sales across models match the annual sales data from official sources (IEA, national yearbooks, Eurostat, US Bureau for Transport Statistics).
Vehicle prices: Since vehicle prices can vary significantly between countries for identical brand-models, we visited each manufacturer’s website for each brand-model in China, the US, Germany, France, the UK, India, and collected the suggested retail prices (MSRP), engine sizes (for ICEVs, HEVs and PHEVs) or battery sizes (EVs and PHEVs), energy consumption, weight and driving ranges (EVs only). We find that price variations across EU countries are generally within a range of 20%, and therefore we use prices in Germany as a proxy for several EU countries where data is not otherwise available in languages we can interpret. In cases where a brand-model is sold in some EU countries according to Marklines but is not available in Germany, we use price and specifications data from the manufacturers’ websites in France or the UK. We further verify the brand-model coverage of MarkLines by cross-comparison with the manufacturer website data thus obtained.
Measuring resilience loss prior to a tipping point
As a system loses resilience as it approaches a tipping point, its restorative feedbacks weaken and therefore its ability to return to some equilibrium state declines.54 This change, known as Critical Slowing Down, can be detected with Early Warning Signals (EWS)34. These EWS measure the return rate of a system to equilibrium; a slow return rate indicates a less resilient system which is closer to a tipping point, than one that has a faster return rate. Figure 1b shows how the potential well of the state the system occupies shallows as the system approaches tipping, representing these restoring feedbacks weakening. Visually, one can imagine the ball representing the state of the system taking longer to return to the bottom of the well after being perturbed as the system moves towards tipping.
One way to measure changes in resilience is by measuring the lag-1 autocorrelation (AR(1)) of a time series on the system on a moving window, which measures the correlation between the system at time t and t+1. As the system takes longer to recover, AR(1) will increase as each time step is more correlated to the last, with AR(1) tending towards 1 as a tipping point occurs55.
Variance is also expected to increase over time34, as the restoring feedbacks of the incumbent system degrade, the system can sample more of the surrounding state space under the same level of perturbations. For true early warning of the approach towards tipping, one expects both AR(1) and the variance to increase together,56 in cases where the system is forced with identically distributed white noise. In all cases, these indicators are known to pick up general losses in resilience over time in the system, although they may be influenced by intervention35.
We apply these EWS to measure the loss in resilience of the incumbent regime. For the EV transition the incumbent regime is an ICEV-dominated market, represented here with ICEV market share. This is considered across the national markets in the UK, Germany, France, China and the US. The ICEV share time series are cut at the start of 2020, prior to the tipping points, and have the trend removed using a kernel regression smoother. On a moving window equal to half the length of the time series, we measure the AR(1) and variance on the resulting residual time series, moving this window one month at a time to get a time series of the indicator series. We then use a Mann-Kendall57 test to measure the tendency of the indicators, with Kendall’s tau = 1 suggesting the indicator is always increasing, -1 always decreasing, and 0 no overall trend, this is approach is common in the literature for assessing these resilience trends.
The FTT:transport model
The Future Technology Transformations for transport (FTT:Transport) model is a submodule of the integrated assessment model named E3ME-FTT-GENIE, a simulation framework that covers the economy, technology and the climate.38,58 The FTT family of models consists of FTT:Power36, FTT:Transport,37,38 FTT:Heat,59 and FTT:Steel60 and is a bottom-up representation of the technological change that reproduces past and projects future diffusion patterns for individual technologies calibrated on observed trends. The FTT framework models technological diffusion and investment choices using a set of coupled dynamical diffusion and learning equations that generate path-dependent S-curve technological change profiles.22,23 Under the FTT framework, consumers are proportionally more likely to choose a technology that has a higher market share as a result of its availability, visibility, social influence and network effects, reproducing observed early adopter and early majority impacts on the diffusion lifecycle.
The model is parameterised along many dimensions by extracting averages for each vehicle class (Economic, Mid-Range, Luxury) in each technology category (Petrol, Diesel, Hybrid, EV, compressed natural gas, biofuels, hydrogen) from the database. Levelised costs of transportation values, expressing an expected discounted cost per kilometre driven, are estimated at each time step for each vehicle category, including the effects of learning curves and pecuniary policies (subsidies and taxes of various kinds). An evolutionary discrete choice model is used to determine the choice of agents between all vehicle options that are available to them.38,61 Consumer discount rates of 20% are used, equal between all vehicle categories.
The intangibles component of the levelised costs are then estimated as the value that ensures that the modelled diffusion trajectories match the observed diffusion trajectories.62 The model’s baseline is thus by construction designed to extend into the future the current trajectory of diffusion of the various transport technologies in each of the markets modelled. In some regions, ICEVs are already in absolute decline, in others they are not. Global learning curves are used for each technology, and notably for batteries and EVs themselves without the battery, assuming that vehicle manufacturers operate and carry knowledge across borders. Learning rates are obtained from various sources (see sources in Figure 2).
Modelling policies in FTT:Transport
Fuel economy standards require automakers to design more efficient vehicles, to shift sales toward more efficient models and discontinue inefficient models. Fuel economy regulations are complex as they apply to averages over sales. Given that we don’t model sales specifically per manufacturer, we must make simplifications. Here, efficiency standards are modelled by adjusting the market shares of the technologies exogenously to meet targets. In the presence of a fuel economy regulation, we assume that there are no new market shares gained in the categories being phased out, leaving the existing vehicles to naturally come to end of their statistical lifetime, leading to exponential declines with half-lives or around 12 years.
EV mandates require automakers to sell EVs in numbers exceeding a predefined percentage of their total sales. Here, we set exogenous market share additions at specific points in time to be consistent with the mandate, allowing EV market shares to exceed the mandated value.
Taxes and subsidies are modelled through their effect on the generalised Levelised Cost of Transportation (LCOT), calculated as a probability distributed quantity every time step, driving purchasing decisions in the model.37,38 Pecuniary incentives include fuel taxes, road taxes, ownership taxes and EV/PHEV subsidies. Costs occurring in the future relative to when decisions are taken are assumed discounted by consumers using a consumer discount rate of 20% or less according to scenario choices. Given that consumer discount rates are challenging to measure, we deliberately use a conservatively high value but vary it as a sensitivity to test our model, shown in Ext. Fig. 3 (lower discount rates bring forward cost and price parity).
Modelling parity in EV prices and total cost of ownership
We construct a relationship between battery prices and EV/PHEV prices according to manufacturing costs. Electric vehicle manufacturing costs are estimated based on manufacturers’ suggested retail prices and the cost-to-price markup factor. Vehicle prices are distinguished from vehicle manufacturing costs due to automaker profits and dealer markups. The markup factors (obtained from3) vary across vehicle segments, where markup factors for small vehicles tend to be lower than larger ones.
Manufacturing costs for EVs depend on four key factors: battery price, platform choice, driving range and the vehicle’s energy use efficiency (where more efficient use of electricity allows for smaller and cheaper batteries for a given driving range). Among these key factors, the fall in the cost of batteries typically accounts for 75% of current falls in EV manufacturing costs.3 Battery costs are estimated dynamically quarterly with a learning curve as a function of cumulated sales (see below). Changes in the cost of batteries are represented in vehicle manufacturing costs to determine prices on the basis of markups.
The learning curve (Wright’s law23,45) is based on the empirically observed phenomenon that the unit cost of a technology declines by a constant percentage for each doubling of cumulative production volume (e.g. cumulative installed capacity), as described by the following equation:
Where Pt is the total global cumulative production of the technology at time t (i.e. the total kWh capacity of cells produced), C is the cost per unit (USD/kWh), C0 and P0 are reference cost and cumulative production values at the start time of scenarios, while b is the learning exponent. The latter is related to the cost reduction proportion that results from every doubling of production, what is known as the learning rate (LR):
The learning rate for EVs is measured to be around 20±5%4,17,45. In the initial stages of diffusion, such as is the case for EVs, costs typically decline rapidly as investment and production scales up exponentially. In later stages of diffusion, such as is the case for ICEVs, doubling the capacity becomes less feasible and costs do not decline despite that a learning rate exists for ICEVs as well.
Critical minerals analysis
Engineering models exist that compile information relating physical properties of different batteries to allow the estimation of manufacturing costs based on factors including cell chemistry and critical metal prices. Here we use the Battery Performance and Cost (BatPaC) model63 and the Cell Energy and Cost model (CellEst)64. One missing component in BatPaC are fluctuations over time in the prices of raw metals on the costs of cathode materials. As a supplement to BatPaC, we use CellEst to calculate battery cathode costs based on detailed future metal costs following a scenario approach.
Suppl. Fig. 1 presents ICEV and EV cost analysis, considering various mineral price scenarios and their effect on EV cost parity across six battery chemistries, including variations of Nickel-Manganese-Cobalt (NMC), Lithium-Nickel-Cobalt-Aluminium Oxides (NCA) and Lithium-Iron-Phosphate (LFP), under three price scenarios (see Suppl. Fig.2). Mineral prices were projected based on the methodology found in Zhang et. al.65 Despite high mineral prices, our findings indicate that EVs in all segments are projected to attain price parity between 2025-2030 even under high mineral price scenarios (see Suppl. Fig. 2). Except for the Luxury segment in China, EV and ICEV are expected to reach price parity between 2030 and 2035 in other regions.
Charging infrastructure analysis
In all countries where EV deployment has been observed, both the numbers of EVs and of charge points increase approximately exponentially. However, the absolute number of chargers increases with the adoption of EVs (Ext. Fig. 7) , but the number of chargers per EV decreases with the share of EVs in the fleet (Ext. Fig. 8). Furthermore, the ratio of charging points per EV decreases as the stock share and the proportion of fast chargers increases. Ext. Fig. 8 shows the correlation between the amount of EV charging infrastructure per unit of EV stock and the EV stock shares in the fleet in the US, China, EU and rest of the World (RoW) at the region-year level. When the EV share is low, the number of public chargers per EV tends to be high since initial infrastructure development generally exceeds EV sales to support vehicle electrification1,66. There is significant variation in residential charging accessibility across cities and countries. For example, fewer households in China have access to private chargers than in the US and Norway where there are a lot more single-unit houses. Charging infrastructure per EV is higher in countries where demand for public chargers is high. On the other hand, in countries such as Norway and the US, fewer public chargers are sufficient for a higher number of EVs, although reliance on public charging solutions increases as EVs are increasingly used for long/interurban journeys.66 Eventually, as occurred in Norway, once the demand for and the utilization of charging points is high enough, the deployment of public charging points becomes increasingly based on commercial decisions where government financial support is no longer needed.67
Ext. Figs. 8-9 suggest that initially deploying a high amount of charging infrastructure per EV in the initial stages of diffusion enables to start the diffusion process, in some cases up to two charge points per EV. In most countries, this was done by the public sector. But this initial infrastructure buildup typically makes a small fraction of the total amount of charging infrastructure deployed in the later stages of diffusion, where typically the number of charging points per EV required to sustain a large population of EVs converges to values around 0.2. Therefore, charging infrastructure can form a barrier for the diffusion of EVs in the very early stages of diffusion, but the evidence does not suggest so for the later stages.