Misunderstandings and misstatements perennially pervade any debate about public funding of research and development. That must be so for any topic involving public money, almost by definition, but arguments about funding for scientific research and development contain a unique and special irony.
Well-working government funding is, by definition, difficult to assess, because of two criteria common to subsidies for R&D at virtually all western governments: specifically, governments seek to fund activities yielding large benefits, and these activities should be actions not otherwise undertaken by the private sector.
The first action leads government funders to avoid funding scientific research with low rates of return. That sounds good because it avoids wasting money. However, combining it with the second criteria does some funny things. If private firms only fund scientific R&D, where the rate of return can be measured precisely, government funding tends to fund activities where returns are imprecisely measured.
That is the irony of government funding of science. Governments tend to fund scientific research in precisely the areas where the returns are believed to be high, but where there is little data to confirm or refute the belief.
This month’s column will illustrate, with a little example, the server software Apache. As explained in a prior column (“How Much Apache?”), Apache was borne and invented with government funding. Today, it is rather large and taken for granted. But how valuable is it? What was the rate of return on this publically funded invention? It has been difficult to measure.
How much did it cost?
Apache started life as the server software that helped the browser operate. Like the Mosaic browser, it was invented at the National Center for Supercomputing Applications (NCSA) at the University of Illinois.
During the National Science Foundation’s (NSF) management of the research-oriented Internet (approximately 1985–1995), the agency invested $200 million dollars in creating Internet-related technology. That budget covered the costs of the Internet backbone, operational expenses, and the supercomputer centers, of which NCSA was one. Apache originated out of NCSA around 1995, so the rate of return calculations will start counting benefits after 1995.
Before the NSF programs, the defense department funded most of the early invention related to the Internet in the 1970s and early 1980s. Although the financial commitment was considerable, no historian has made a precise estimate of the size of the commitment to networking, which was one of several programs it funded. Government secrecy about budgets prevents any historian from uncovering further details.
Nonetheless, we do know something useful. The entire expenditure for the Information Processing Techniques Office (IPTO), the agency within DARPA that funded most of the Internet, did not exceed approximately $500 million over its entire existence (1963–1986), and the funding for what became the Internet was but one of many IPTO projects. To be conservative, add another $200 million to the costs of creating the Internet, which is much more than the likely cost at that time.
This calculation includes the most direct costs for creating the Internet, and excludes numerous other costs for research. For example, this estimate of costs does not include a range of other experiments in computer science that NSF paid for (out of different budgets) and from which the general community of researchers learned.
Cost and benefit
That sets up the first cost-benefit calculation. The prior column estimated that Apache is worth between $2 and $12.2 billion dollars in 2012, 17 years later. As noted then, $2 billion and $12.2 billion are implausibly low and high. So for this cost-benefit calculation we consider two more plausible levels, $7.1 billion and $10 billion, where the former is the midpoint between the two numbers and the latter is as high as plausible.
What constant rate of growth would generate such levels of benefit after $400 million investment 17 years earlier? Using the current dollar numbers, the former generates a rate of growth of 17 percent, and the latter 19 percent. To be sure, inflation (which averaged 2 to 3 percent a year in this period) would diminish some of that gain, but the rate of return is quite high for only one output from NCSA, and there were many more benefits than just this one.
A second cost-benefit calculation examines the scale of economic growth attributable to Apache, using productivity gains from investment in software, much as discussed earlier. This is too much to show in a column; an interested reader can see the full research elsewhere, written with Frank Nagle (here or here).
Consider the question: if additional software were added to the capital stock of software, how much economic growth would it generate? In each year, it generates more income, and that continues to accumulate after 1995. Consider the total through 2012. This estimate must make assumptions about the rate of growth of Apache, which we do not observe directly.
To illustrate how this calculation works we begin with one year, the estimate for 2012. In 2012, software contributed 0.16 percent of labor productivity growth. Lest that sound small, consider that there was only 1.56 percent total labor productivity growth per year.
Software is even more impressive than it sounds. It accounts for only 3.75 percent of the income share in that year, but software makes a big contribution to productivity growth in comparison to the size of software in use.
The low and high estimates for the value of Apache would place it at 1.3 and 1.8 percent of the stock of software in 2012. Although the increase in labor productivity due to Apache has to be comparatively small, the US economy is so large that even a small improvement can yield a substantial economic gain.
In this case, we assume that labor does not increase, but only software does, and we simulate how much GDP would grow with more software. This approach estimates that additional software stock would generate a low/high estimate of $1.0 billion and $1.8 billion in additional income in 2012.
Gains over time
Apache also generated income in all the years between 1995 and 2012. For those estimates, we must make an assumption about the rate of growth in Apache servers over this period, which we did not observe directly. We make an estimate from public data on the growth of webpages supported by Apache over this time period.
Netcraft’s public data (http://www.netcraft.com) shows webpages grow at one rate over the latter part of the 1990s, and then grow at a slower rate in the last decade, with Apache generally supporting 60 percent of all webpages. This history suggests that a constant rate of growth over all 17 years is probably slower than the true rate of growth, so we make the assumption of linear growth in order to be conservative. Next we estimate the incremental contribution of Apache for each year, just as we did for 2012.
Now we can ask: “How much economic activity did the investment in the Internet generate by 2012?” Such a calculation requires aggregating all future benefits into the same dollar units. We assume a 10 percent discount rate on the future from 1995, so we can add up the contemporary dollars. This assumption will weight short-term gains against those that come many years later. Again, that is conservative.
By that estimate, Apache generated economic activity equivalent to between $2.6 and $4.5 billion by the end of 2012. On an investment of $400 million, that is a rate of return between 10.5 and 14 percent.
There are important qualifications to this second set of estimates. First, this simulation stresses the effect of gains from the use of IT, often called capital deepening. If Apache generated productivity gains to the production of IT, the above estimates would not capture those gains.
Second, no reader should quote the rate return on Apache as precisely 10 to 14 percent. The estimate depends on assumptions, such as of linear growth and 10 percent discounting. A small increase in the rate of discount would mildly lower the rate of return, as would later growth in Apache, and, similarly, vice versa.
So why make such an estimate? These estimates provide a good sense of the scale of the gains, and small changes in those assumptions do not alter the quality of the answer. Anything in this range has to yield a large rate of return, which is the key point.
Apache is but one effect of public funding of the Internet. It is hard to measure, and therefore easy to overlook. Some detective work and calculation indicate that the return on investment for Apache was quite high. Since this is only one program, this leads us to conclude that the returns from federal R&D invested in the Internet must be underestimated.
The bigger point should not get lost in the detail. By design, government funded the research for creating the Internet because private firms would not do it, but it was regarded as beneficial for society.
After the fact, there is no way to reach any other conclusion but a positive one. One offshoot from the program, such as Apache, yielded a return that justified the entire expense.