Learning Rates: Why Some Technologies Get Cheap and Others Don’t 

Why do some technologies become dramatically cheaper over time, while others remain stubbornly expensive? Understanding this question is crucial for decisions on energy, climate, and innovation policy. The concept of technological learning offers a powerful framework for explaining long-term cost reductions and forecasting the future competitiveness of emerging technologies. In this article, we explore what learning rates are, why they differ across technologies, how they can guide investment decisions, and some of the most common misconceptions surrounding their use.

What is ‘learning’ in technologies? 

A technology’s learning rate indicates how much the unit cost of a technology decreases each time cumulative production doubles. A learning rate of 20% means every doubling of cumulative production is accompanied with lowered costs by 20%. Different technologies exhibit very different learning rates. Solar photovoltaics (PV) has followed a learning rate of roughly 23%, with prices falling by more than 99%, from $100 per Watt to less than $0.25 per Watt, between 1976 and 2026.  Lithium-ion batteries have also exhibited a high learning rate of 20%. Onshore wind sits at around 12%, while nuclear power plants’ costs have largely remained unchanged since first recorded, thus having a learning rate of close to 0%. 

The “learning” in the learning rate stems from the learning-by-doing phenomenon: the more we perform a task or produce something, the more efficient we become, leading to lower costs over time. In turn, the gradual improvements lead to long-term cost declines. This contrasts with cost improvements associated with new inventions or technological breakthroughs.  The “learning” phenomenon was most famously documented for aircraft manufacturing in 1936, when planes’ production costs were first plotted against their cumulative production, thus coining the term ‘Wright’s Law’. Such cost-cumulative production curves are now known as ‘learning curves’ for technologies. It is also worth noting that ‘experience rates’ and ‘experience curves’ are frequently used as synonyms in the field. 

What determines learning rates? 

The most common way to obtain learning rates for cost projections has been to fit historical cost data to a learning curve for individual technologies. Alternatively, experts can be asked to project technology costs. However, this has been proven less reliable than using learning rates. So, how do we obtain learning rates for early-stage technologies when there is no historical cost data to estimate a learning curve? 

Recent innovation literature proposes a way forward: Malhotra & Schmidt developed a technology typology that classifies technologies based on their inherent characteristics: design complexity and need for customisation, showing that more complex and customised technologies have exhibited slower learning (See Fig. 1). Other studies have also shown that technologies with larger unit sizes exhibit lower learning rates. Based on these theories, early-stage technologies, such as direct air capture (DAC) technologies and fusion power, have been assessed for their inherent characteristics to determine a likely learning rate. But why do we care about learning rates? 

Why getting learning rates right matters 

Imagine a policymaker choosing between two novel clean technologies to invest in today. Technology A is cheaper today, but early evidence suggests a low learning rate. Technology B is expensive today, but will likely have a high learning rate. If the policymaker chose not to invest in B because of its higher costs today, B would not be able to travel down its learning curve to become far cheaper than A. Underestimating B’s learning rate would have led to sacrificing long-term inefficiency for short-term efficiency. 

This scenario is not just hypothetical. For decades, solar PV was seen as a niche energy source due to its high costs. Germany nevertheless supported early deployment through solar feed-in tariffs, driven primarily by energy security and climate objectives. China then scaled the industry massively. As a result, solar PV is now one of the world’s cheapest sources of electricity. This highlights the importance of accurately understanding learning rates: they were widely underestimated for solar PV at first. With accurate learning rate estimations, we could identify the next ‘solar PV’ with high learning rates to drive deployment through policy and achieve long-term cost savings. Similarly, we could also identify any technology we are committing to today that will remain only as ‘Technology A’ with low learning rates. 

Misconceptions and unknowns of learning rates 

As researchers on technological learning, we’ve come across some common misconceptions about the learning rate that we would like to clarify in this blog post. 

First, a common misconception is that cost reduction from learning will simply happen over time. However, cost reductions occur over production rather than over time. Without policy or market demand to drive production, costs remain stagnant. For example, when DAC technologies were projected to fall to around $350/CO2t, some readers concluded that this would simply be the cost in due time. However, it is important to consider that these projected costs are contingent on a 2,000-fold increase in cumulative production, and someone must invest in this as well.   

Second, learning rates are a long-term tool and are not well-suited for short-term forecasting. For example, in 2021, lithium-ion battery prices briefly rose on the back of a lithium shortage and a spike in lithium prices. Some claimed that the learning mechanics had been broken, that prices had now bottomed out, and, consequently, that electric vehicles would never compete head-on with combustion cars. However, these pessimists were quickly proven wrong, and in recent years, learning has continued just as before, with no foreseeable halt. To some, it seems unimaginable that cost reductions are still possible in new technologies when the costs have already become so low. However, this pattern is baked into the learning rate concept: consistent cost reductions will continue as long as production grows exponentially, as is expected for solar panels and batteries. Thus, we expect their costs to continue decreasing. 

While we are confident in clarifying these common misconceptions, many key questions about technological learning remain unanswered. For example, are there technologies for which cost reduction happens only in a local context (within a country, region, etc.)? If so, why and how can we determine those specific degrees? How can we best measure the technological characteristics of design complexity and need for customisation? These may (or may not) be covered in future blog posts, so stay tuned!