Average is Over could use more focus on the Zero Marginal Cost Economics of software

Tyler Cowen’s excellent and widely discussed Average is Over argues computers will accelerate our existing trend toward a more stratified society. People adept at teaming up with computers will get richer, while those who aren’t will get left behind. Hence the title. Average is Over. What I found most striking was Cowen’s causal mechanism. He argues stratification comes from the skyrocketing productivity (and relative scarcity) of humans adept at teaming up with computers. Reasonable enough. But what Cowen didn’t emphasize is the odd economics of digital goods, such as digital music, eBooks and most importantly software itself. They have near zero marginal cost. Making copies costs nothing. As computer based digital goods start to dominate the economy, their odd zero marginal cost economics will loom larger and larger. I’ve written about Digital Economics before, so let’s see what it might mean for Cowen’s thesis.

Average is Over is really an extension of Cowen’s previous book, The Great Stagnation, where Cowen argues innovation has already reached the point of diminishing returns. The centerpiece of Stagnation is the Jonathan Huebner chart below showing innovations per capita peaked in 1873, declining ever since.

In his follow up Average is Over, Cowen uses freestyle chess as the canonical example of human/computer pairing. Freestyle chess is played by opposing teams of humans paired with computers. The humans know the strengths and weaknesses of the various chess playing computer programs, and help select the best move in conjunction with computers. These human/computer hybrid teams play at a higher level than either working alone. It’s far and away the most advanced form of chess being played today. He uses this example as a model for other changes coming in the economy, such as computers working with nurses being far more effective at diagnosing illness than human doctors working alone. Note that Cowen avoids words like artificial intelligence on purpose, as he believes computers and humans are good at different things. It’s pairing them up that takes productivity to the next level.

So how do zero marginal cost economics fit in the picture? Well, from economics we know in perfect competition prices tend to drop to marginal costs. Hence the prices of digital goods should tend toward zero as we get closer to perfect competition market structure. In particular this happens when internet competition and distribution remove friction from the supply chain. See the chart below to contrast traditional versus digital goods. Traditional manufactured goods like cars and smartphones are in green. As you ramp up output past the pain point, constraints on factory infrastructure, overtime pay and the supply chain eventually make widgets more expensive per unit to produce. Contrast this to digital goods like eBooks and smartphone apps in red. They just get cheaper and cheaper as you scale. More here.

Let’s use the music industry as an example. In the chart below overall sales of music peaked in 1999, the year Napster kicked-started file sharing. As music moved to digital format and distribution, we got rid of physical stores, physical media and middlemen. Closer to perfect competition. So of course prices fell closer towards marginal cost zero. Net music sales tanked. And the conversion to pure digital distribution is far from complete.

The chart above is an absolute nightmare for music creators. And yet it’s simultaneously heaven for music consumers. Subscription services such as Spotify, Rdio and Pandora are free if you tolerate ads, which 3/4 of people are fine with. For power consumers like me, you can pay a nominal $5 or $10 monthly fee to avoid ads and get extra features. The bottom line is music is now free (albeit with ads) for people who want it. It’s a zero marginal cost success story.

What’s hard to determine from the data is whether we should see this as a pure productivity gain (good thing) or a net drop in GDP output (bad thing). The problem of estimating productivity gains between incomparable goods (CDs versus Pandora streaming) go way beyond music. It’s an unending debate in economics. The problem is especially difficult for industries disrupted by computers. From wikipedia, the productivity paradox is “the apparent contradiction between the remarkable advances in computer power and the relatively slow growth of productivity at the level of the whole economy, individual firms and many specific applications.” I would say the majority of economists agree with Tyler Cowen, that the overall rate of productivity growth is in decline. One way to show this is at a definition level. If decline was not the consensus, economists would agree to tweak how productivity gains between incommensurate goods were calculated. And they haven’t. Yet there remains a respectable minority who believe productivity gains are underestimated. For example see this EconTalk podcast where Russ Roberts interviews Joel Mokyr, and they discuss Cowen’s book. I’m in the Russ Roberts/Joel Mokyr camp obviously. As a similar example of digital disruption, the newspaper industry has simultaneously imploded and entered an information golden age. And in the technology industry itself, software is merely another digital good with zero marginal duplication costs. As smartphone app stores and operating systems remove the cost and friction to shop/install/uninstall/update software (remember dll hell?), phone software is becoming free. And with internet distribution, this trend towards free software is impacting consumer software across the board. Consumer software companies are struggling to find other business models to survive: selling ads on free software, bundling software with hardware devices (since hardware retains a non-zero marginal cost structure), or using freemium pricing models.  Switching costs and friction remain much higher for business software, but even there we’re seeing software-as-a-service reducing switching friction, and pushing pricing downwards.

Let’s move on to a critique of Cowen’s two books from Matthew Yglesias. Yglesias points out that Cowen blames innovation stagnation for income stratification in The Great Stagnation. Then blames innovation acceleration for income stratification in Average is Over. Got you. Changes to the rate of innovation appear to cause stratification either way. I think Cowen could respond by claiming the net level of innovation is indeed going down. But within that lower overall level, computer innovation is special because it has a higher than normal stratification impact. A good response. But this begs the question. Why is computer innovation special?

This is where the zero marginal cost aspect of digital goods comes into play. Recall the case where nurses paired with diagnostic computers can potentially outperform doctors in diagnosing illness. Once written we could push this diagnostic software out to every smartphone on the planet. While this would require a massive up front infrastructure investment, the marginal cost of adding the app to each phone would be zero. Computers and software really are special. Furthermore this provides a response to Yglesias’ critique. We use zero marginal cost economics to explain The Great Stagnation as being a mismeasurement of productivity gains, and then use it again to explain the skyrocketing productivity of human/computer pairing discussed in Average is Over. There’s nothing special about computer technology replacing human labor. Railroads, electricity, plumbing, washing machines, all technology does this. That’s the point. What’s special about modern software is it scales to hundreds of millions of users at almost no added marginal cost. Consider Facebook’s acquisition of Instagram. Much of the focus was on how much money changed hands and who got rich. Wrong lesson. The right lesson is getting your head around the crazy fact that piggybacking on modern internet infrastructure allowed 13 employees to scale to millions of users. Dude. Mind blown.

A useful economic precedent to the economics of software are utility companies, such as electrical power plants or water treatment plans. Deemed “natural monopolies“, these companies have high fixed costs combined with economies of scale which make adding customers very low margin. This comparison is particularly useful in recognizing how the internet giants of today might get regulated like past utilities, which could be a good thing or a bad thing depending on your politics. (Bad thing in my view). It also let’s us understand how nationalist internet champions could be promoted in countries like China and other parts of Asia, which want to retain control of the commanding heights of the economy. Yet there’s huge differences as well. Switching costs for users between utilities such as your water company are massive. It’s often impossible. Switching costs between Microsoft Bing Search and Google Search are negligible. Just a click away. And no physical power plant on Earth has a billion active users like Facebook. The scale of modern software and it’s associated cost structures are way beyond anything utilities could ever hope to achieve.

In the next 10-20 years, as Cowen points out, we should expect computers to impact sectors previously untouched, such as health care (recall the nurse/diagnostic example), transportation (self driving cars), and possibly even government (automation of services with computer voice interaction). Much farther into the future, we can imagine an extreme case where robots build factories which in turn replicate and build more factories. This would push the zero marginal cost structure of software out into the real world of physical goods. My point here is the impact of software’s zero marginal cost structure is profound and yet to be fully exploited.

So where do I stand in regard to Cowen’s thesis? As mentioned, I disagree with The Great Stagnation, being in the (hopefully respectable) minority view that productivity gains are under counted. This ties back to my main theme that the impact of zero marginal costs goods is greatly under appreciated. For Average is Over, I’m generally in agreement with Cowen. Computers are indeed driving more stratification. Where we differ is Cowen never really explains why computers are different in their stratification effects from earlier technologies. Yglesias has a point. What’s implied by Cowen is working with computers enhances the premium on cognitive skills. But again, why? This takes us back to the odd economics of zero marginal cost goods, which radically enhance the scale, productivity and pay of those who work well with computers. Finally, compared to Cowen I’m far less sure how much stratification to attribute to technology versus social, institutional and political factors. It’s a provocative and wonderful book, but left me feeling cautious. Of course like any economist, Tyler Cowen is well aware of the special nature of zero marginal cost digital goods. I just wish he’d written more about it (arguing the impact either way) in his book.

———————-

Edits from Dec 16: on request shortly after posting I wrote a new final paragraph to explicitly summarize my position on stratification relative to Tyler Cowen. Great suggestion. In re-reading, also wish I’d clarified a bit more on how the hidden productivity argument ties to consumer surplus. But wikipedia covers it here.

More Reading: AI and Zero Marginal Cost Minds

If you want to go dig farther into robotic replication and Artificial Intelligence, a good place to start is economist Robin Hanson. He has an upcoming book I’m looking forward to reading, about computers emulating human minds. I’m not actually a believer in computer emulation of minds by the way, as I think it will prove extremely difficult. Traditional AI will get there first, however long that may take. But Hanson’s reasoning around the economics of zero marginal cost minds is brilliant and very relevant to the discussion above. You can go here for an interesting post from Hanson on pundit reaction to Cowen’s book, plus read about his skepticism as to the importance of computer automation in driving stratification. Go here to get a list of all of Hanson’s posts on Emulated Minds (EMs).