She wasn’t accusing ProSense of fraud. They were probably reinvesting in R&D, absorbing some cost reductions into margin, and genuinely improving the product. But she now had a data-driven starting point for the conversation \u2014 not a vague "we’d like a better price," but a specific, defensible cost trajectory based on decades of empirical research.<\/p>\n
That starting point is the experience curve. And if you’re not using it, you’re negotiating blind.<\/p>\n
From Paper Airplanes to Strategic Weapons<\/h2>\n
The idea that costs decline with experience is older than most people think. In 1936, Theodore Paul Wright, an aeronautical engineer at Curtiss-Wright Corporation, published a study of aircraft manufacturing costs. He found that every time cumulative aircraft production doubled, the direct labor hours per unit dropped by a consistent percentage \u2014 typically around 20%. The pattern held across multiple aircraft types and factories.<\/p>\n
Wright’s discovery was primarily about learning<\/em> \u2014 workers getting faster at repetitive assembly tasks. It was useful for military procurement during World War II (the U.S. government used it to negotiate bomber contracts), but it stayed confined to direct labor in manufacturing.<\/p>\n Three decades later, Bruce Henderson and the Boston Consulting Group blew the concept wide open. In the mid-1960s, BCG analyzed cost data across dozens of industries and found that the decline wasn’t limited to direct labor. Total<\/em> unit costs \u2014 including materials, overhead, distribution, administration \u2014 declined by 20-30% every time cumulative industry output doubled. They called this the "experience curve" to distinguish it from Wright’s narrower "learning curve."<\/p>\n Henderson turned it into a strategic weapon. BCG argued that market share was destiny: the company with the most cumulative production would have the lowest costs, the highest margins, and an insurmountable competitive advantage. This thinking drove the growth-at-all-costs strategies of the 1970s and 1980s. Companies acquired competitors, cut prices below cost to gain volume, and bet everything on riding the experience curve down faster than rivals.<\/p>\n Then the concept fell out of fashion. Critics pointed out that BCG’s advice had led companies into price wars, overexpansion, and neglect of innovation. The experience curve couldn’t explain why IBM lost to smaller, nimbler PC manufacturers, or why high-market-share companies sometimes had lower<\/em> margins than focused niche players.<\/p>\n But here’s the thing: the underlying empirical relationship never went away. Costs do<\/em> decline with cumulative output. The pattern has been documented in semiconductors (Moore’s Law is an experience curve), solar panels (a breathtaking 99% cost reduction since 1976), lithium-ion batteries, wind turbines, genome sequencing, and hundreds of manufactured products. What was flawed was the strategic advice<\/em> built on top of it, not the cost model itself.<\/p>\n With modern data science tools, the experience curve is coming back \u2014 not as a grand corporate strategy, but as a precise analytical tool for procurement, cost forecasting, and supplier management. That’s how Anna Berger used it, and that’s how we’ll use it here.<\/p>\n These two terms get used interchangeably, and that’s a problem because they measure different things. If you confuse them, you’ll underestimate cost reduction potential.<\/p>\n The learning curve says: "Workers get faster at doing the same thing." That’s true, and it matters \u2014 but it’s only one piece of the puzzle.<\/p>\n The experience curve says: "Everything gets cheaper \u2014 labor, materials procurement, process engineering, quality control, logistics, overhead allocation \u2014 as cumulative volume grows." The mechanisms include:<\/p>\n When a procurement team uses a learning curve (75-85% slope) to estimate a supplier’s cost trajectory but the real driver is broader experience effects, they’ll predict too little cost reduction. The experience curve \u2014 encompassing all cost elements \u2014 is the right tool for strategic sourcing.<\/p>\n The experience curve follows a power law:<\/p>\n C(x) = C\u2081 \u00b7 x^b<\/strong><\/p>\n Where:<\/p>\n The learning exponent b<\/em> relates to the learning rate<\/strong> (also called the slope) by:<\/p>\n b = log(learning_rate) \/ log(2)<\/strong><\/p>\n An 80% learning rate means costs fall to 80% of their previous level every time cumulative volume doubles. The exponent b<\/em> for an 80% curve is:<\/p>\n b = log(0.80) \/ log(2) = -0.322<\/strong><\/p>\n The beauty of this model is the log-log transformation. Take the natural log of both sides:<\/p>\n ln(C) = ln(C\u2081) + b \u00b7 ln(x)<\/strong><\/p>\n This is a straight line in log-log space: the intercept is ln(C\u2081) and the slope is b<\/em>. That means fitting an experience curve is just linear regression on log-transformed data. Any spreadsheet or statistical tool can do it.<\/p>\n Let’s make this concrete with ProSense’s sensor production. Here are the key facts:<\/p>\n If ProSense follows an 80% experience curve, we can predict unit costs at any cumulative volume. The first doubling \u2014 from 500 to 1,000 cumulative units \u2014 should bring costs from \u20ac145 to \u20ac116. The next \u2014 1,000 to 2,000 \u2014 to \u20ac93. Then \u20ac74, \u20ac59, and so on. By 64,000 cumulative units (approximately 7 doublings from 500), the predicted cost is:<\/p>\n C(64,000) = 145 \u00d7 (64,000 \/ 500)^(-0.322) = 145 \u00d7 0.21 \u2248 \u20ac30<\/strong><\/p>\n Or more intuitively: 7 doublings at 80% per doubling = 0.80^7 = 0.21. So predicted cost = 145 \u00d7 0.21 = ~\u20ac30<\/strong>.<\/p>\n Wait \u2014 that seems too low. And it would be, if ProSense were a pure assembly operation. But sensors require expensive electronic components (MEMS accelerometers, signal processing ICs) whose costs decline on a shallower curve. This is why component decomposition matters, which we’ll get to in a moment.<\/p>\n In practice, fitting the actual data with regression tells us the real<\/em> learning rate rather than assuming one. That’s what the R code does.<\/p>\n Let’s walk through the process step by step, using ProSense GmbH’s simulated production data.<\/p>\n Step 1: Prepare the data.<\/strong> We need quarterly cumulative production and unit cost. The R script generates realistic data following an 80% experience curve with noise \u2014 because real-world data is never perfectly smooth.<\/p>\n Step 2: Log-log scatter plot.<\/strong> Plot ln(cumulative_production) on the x-axis and ln(unit_cost) on the y-axis. If the experience curve holds, you’ll see a clear linear relationship.<\/p>\n Step 3: Fit the regression.<\/strong> Run Step 4: Interpret the results.<\/strong> Check R-squared (should be above 0.90 for a good fit), confidence intervals on the slope, and residual patterns. A learning rate of 0.80 with tight confidence intervals means you can confidently predict future costs.<\/p>\n The linear-scale chart shows the classic concave curve \u2014 steep cost declines early, flattening as volume grows. But the shape can be misleading because it’s hard to tell whether the relationship is truly a power law or some other functional form.<\/p>\n The log-log plot resolves this immediately. A straight line in log-log space confirms the power law. The slope gives us the learning exponent directly, and the regression line lets us extrapolate costs to future volumes.<\/p>\n For ProSense, our regression yields:<\/p>\n The high R-squared and tight confidence intervals tell us the experience curve is a strong fit for ProSense’s cost data. The estimated learning rate of 80.2% is right in the typical range for electronic components manufacturing.<\/p>\n Here’s where most experience curve analyses go wrong: they treat the total unit cost as a monolith. In reality, different cost components decline at different rates, and understanding this decomposition is crucial for realistic forecasting.<\/p>\n Consider ProSense’s sensor unit cost breakdown at the start of production:<\/p>\n Materials have the shallowest curve because they’re largely purchased components \u2014 your learning rate is bounded by your supplier’s learning rate (and their willingness to share those savings). Direct labor has the steepest curve because human learning effects are powerful and well-documented.<\/p>\n This explains why labor-intensive products (apparel, furniture, manual assembly) typically show steeper experience curves (75-80%) than material-intensive products (electronics, chemicals) where curves are shallower (85-90%).<\/p>\n The component plot reveals what a single aggregate curve hides: labor costs have plummeted by over 50%, but materials costs have only declined by 25%. The aggregate curve sits somewhere in the weighted middle.<\/p>\n The waterfall chart makes the strategic implications clear. At low cumulative volume, ProSense’s cost structure was roughly 55% materials, 30% labor, 15% overhead. At high cumulative volume, the mix has shifted even further toward materials (now over 60% of total) because materials declined less than labor and overhead. Future cost reduction efforts should focus on materials \u2014 negotiating with component suppliers, redesigning for cheaper parts, or vertically integrating.<\/p>\n The experience curve isn’t just an academic exercise. It’s a practical tool for at least four critical supply chain decisions.<\/p>\n This is Anna Berger’s use case, and it’s the most immediately actionable. If you know (or can estimate) your supplier’s cumulative production volume and learning rate, you can calculate what their costs should be \u2014 independent of what they’re charging you.<\/p>\n You’re not accusing the supplier of gouging. You’re showing them that you understand cost dynamics and expect pricing to reflect productivity gains. The conversation shifts from "give us a discount" to "let’s talk about how we share the benefits of your growing efficiency."<\/p>\n Important caveat:<\/strong> suppliers have legitimate reasons for prices not tracking the experience curve exactly \u2014 R&D investment, quality improvements, regulatory compliance, raw material inflation. The curve gives you a starting point for negotiation, not a final answer.<\/p>\n If you’re ramping up production of a new product, the experience curve helps you forecast when you’ll hit cost targets, break even, or achieve target margins. Plot your actual cost data against the fitted curve quarterly. If costs are above the curve, investigate \u2014 you may have process issues preventing normal learning. If below, you’re outperforming expectations.<\/p>\n A company with 3x your cumulative production is 1.58 doublings ahead on the experience curve (log\u2082(3) = 1.58). At an 80% learning rate, their unit costs are approximately 0.80^1.58 = 70% of yours. They can undercut your price by 20% and still maintain higher margins. This is why experience-intensive industries tend toward consolidation \u2014 the leader’s cost advantage is self-reinforcing.<\/p>\n If a contract manufacturer has produced 500,000 units and you’ve produced 5,000, they’re roughly 6.6 doublings ahead. At an 80% curve, their costs are 0.80^6.6 = 23% of yours. You will never<\/em> catch up unless you can find a way to leapfrog their volume. This is a strong signal to buy rather than make.<\/p>\n The industry comparison chart shows why one-size-fits-all assumptions are dangerous. Semiconductors have famously steep curves (around 70%) driven by Moore’s Law dynamics. Commodity chemicals are near 90% \u2014 costs decline, but slowly. Knowing the typical slope for your industry gives you a reasonable prior before you have enough data to estimate your own.<\/p>\n The experience curve is powerful, but it’s not a law of nature. It’s an empirical pattern that holds under certain conditions and breaks under others. Honest analysis requires knowing when not<\/em> to rely on it.<\/p>\n Depleting natural resources.<\/strong> If your key input is a finite resource whose extraction gets harder over time (deep-water oil, rare earth elements, high-grade ores), input costs may rise<\/em> with cumulative volume industry-wide. The experience curve for extraction efficiency exists but can be overwhelmed by geological depletion.<\/p>\n Regulatory barriers and patents.<\/strong> A pharmaceutical company with patent protection has no competitive pressure to reduce prices along the experience curve. Regulatory compliance costs (GMP manufacturing, clinical trials, documentation) don’t always decline with volume because they’re driven by regulatory requirements, not production efficiency.<\/p>\n Mature commodities.<\/strong> When an entire industry is far down the experience curve \u2014 think steel, commodity plastics, basic chemicals \u2014 further doublings of cumulative output yield trivially small cost reductions. The curve has largely been exhausted. Competition shifts to other dimensions (service, delivery, quality, sustainability).<\/p>\n Technological disruptions.<\/strong> A disruptive technology can reset the curve entirely. The experience curve for incandescent light bulbs became irrelevant when LEDs arrived with a new, steeper curve starting from a higher cost point but declining faster. If your supplier is riding a technology that’s about to be displaced, their experience curve is worthless for long-term forecasting.<\/p>\n Products with high intangible value.<\/strong> Luxury goods, branded products, and IP-heavy offerings don’t follow cost-based pricing. Hermes isn’t going to cut handbag prices because their artisans got more experienced. The experience curve applies to costs, not to prices in markets driven by brand, exclusivity, or intellectual property.<\/p>\n The management effort trap.<\/strong> Perhaps the most dangerous misconception: cost decline along the experience curve is not automatic<\/em>. It requires deliberate management action \u2014 continuous improvement programs, capital investment in better equipment, workforce training, supplier development. Companies that assume costs will drop simply because volume grew often find that they don’t. The curve describes what can<\/em> happen with good management, not what will<\/em> happen with passive management.<\/p>\n Explore the experience curve with your own numbers. Adjust learning rates, volume growth, and cost components to see how costs decline \u2014 and set data-driven negotiation targets.<\/p>\nLearning Curve vs. Experience Curve: They’re Not the Same Thing<\/h2>\n
\n\n
\n \n<\/th>\n Learning Curve<\/th>\n Experience Curve<\/th>\n<\/tr>\n<\/thead>\n \n Scope<\/strong><\/td>\n Direct labor hours on a single task<\/td>\n All costs: labor, materials, overhead, distribution<\/td>\n<\/tr>\n \n Driver<\/strong><\/td>\n Individual practice and repetition<\/td>\n Cumulative organizational\/industry output<\/td>\n<\/tr>\n \n Typical slope<\/strong><\/td>\n 75\u201385% (labor-intensive assembly)<\/td>\n 70\u201385% (varies by industry and cost structure)<\/td>\n<\/tr>\n \n Unit of analysis<\/strong><\/td>\n A single worker or workstation<\/td>\n An entire product, factory, or industry<\/td>\n<\/tr>\n \n First documented<\/strong><\/td>\n Wright (1936), aircraft assembly<\/td>\n Henderson\/BCG (1968), cross-industry analysis<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \n
The Math That Matters<\/h2>\n
\n
Worked Example: ProSense GmbH<\/h3>\n
\n
Fitting Experience Curves with R<\/h2>\n
lm(log(unit_cost) ~ log(cum_production))<\/code>. The slope coefficient is your learning exponent b<\/em>. Convert it to a learning rate with 2^b<\/code>.<\/p>\n![\"Cost](\"https:\/\/inphronesys.com\/wp-content\/uploads\/2026\/03\/exp_curve_linear-1.png\")
<\/p>\n![\"Log-log](\"https:\/\/inphronesys.com\/wp-content\/uploads\/2026\/03\/exp_curve_loglog-1.png\")
<\/p>\n\n
Not All Costs Learn Equally: Component Decomposition<\/h2>\n
\n\n
\n \nComponent<\/th>\n Share<\/th>\n Learning Rate<\/th>\n Why<\/th>\n<\/tr>\n<\/thead>\n \n Materials<\/strong> (MEMS, ICs, PCBs)<\/td>\n 55%<\/td>\n 85%<\/td>\n Commodity-ish; limited by supplier’s own curve<\/td>\n<\/tr>\n \n Direct Labor<\/strong> (assembly, testing)<\/td>\n 30%<\/td>\n 75%<\/td>\n Classic Wright learning; high improvement potential<\/td>\n<\/tr>\n \n Overhead<\/strong> (equipment, facilities, quality)<\/td>\n 15%<\/td>\n 80%<\/td>\n Utilization improves; fixed costs spread over more units<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n ![\"Component](\"https:\/\/inphronesys.com\/wp-content\/uploads\/2026\/03\/exp_cost_components-1.png\")
<\/p>\n![\"Cost](\"https:\/\/inphronesys.com\/wp-content\/uploads\/2026\/03\/exp_component_waterfall-1.png\")
<\/p>\nStrategic Applications for Supply Chain<\/h2>\n
Procurement: Know What Your Supplier’s Costs Should<\/em> Be<\/h3>\n
\n\n
\n \nYear<\/th>\n Supplier’s Price<\/th>\n Curve Prediction<\/th>\n Gap<\/th>\n Your Opportunity<\/th>\n<\/tr>\n<\/thead>\n \n 2018<\/td>\n \u20ac138<\/td>\n \u20ac138<\/td>\n \u20ac0<\/td>\n Baseline \u2014 fair price at launch<\/td>\n<\/tr>\n \n 2020<\/td>\n \u20ac131<\/td>\n \u20ac121<\/td>\n \u20ac10<\/td>\n Price lagging cost reductions<\/td>\n<\/tr>\n \n 2022<\/td>\n \u20ac125<\/td>\n \u20ac108<\/td>\n \u20ac17<\/td>\n Supplier capturing 60% of savings<\/td>\n<\/tr>\n \n 2024<\/td>\n \u20ac120<\/td>\n \u20ac98<\/td>\n \u20ac22<\/td>\n Gap widening \u2014 time for a conversation<\/td>\n<\/tr>\n \n 2025<\/td>\n \u20ac118<\/td>\n \u20ac92<\/td>\n \u20ac26<\/td>\n \u20ac26\/unit opportunity = \u20ac1.66M\/year<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Operations: Forecast Your Own Production Costs<\/h3>\n
Competitive Intelligence: Why Market Leaders Win on Price<\/h3>\n
Make vs. Buy: When to Surrender to the Specialist<\/h3>\n
![\"Typical](\"https:\/\/inphronesys.com\/wp-content\/uploads\/2026\/03\/exp_industry_slopes-1.png\")
<\/p>\nWhere the Curve Breaks<\/h2>\n
Interactive Dashboard<\/h2>\n