Digital Analogs: Computing,Internet,and Spectrum Lessons for New Space Policy

The military years

The computer was invented with military funding. As Richard Rhodes comments: “The first problem assigned to the first working electronic digital computer in the world was the hydrogen bomb.” ¹⁷ Rhodes further states: “The ENIAC ran a first rough version of the thermonuclear calculations for six weeks in December 1945 and January 1946. Los Alamos prepared a half million punched cards of data, enough to keep a hundred people busy for a year at mechanical desktop machines.” For the first decade of computing, the design of atomic weapons provided the major source of computing demand. The U.S. military was the biggest funder of computing, as weapon projects were desperate for more powerful computing tools to analyze the complex design tradeoffs in both fission and fusion devices. ¹⁸

As the Soviets acquired a fission bomb (1949), followed by a fusion bomb (1953), early warning systems capable of detecting Soviet bombers became a critical national priority. Such a system demanded far greater computing capabilities than that existed, with sensors stretched over thousands of miles, and led to tens of billions of dollars invested in the SAGE radar and computing complex. The SAGE system drove many computing breakthroughs, fundamental to later timesharing and interactive commercial systems.

Memory was one of the biggest bottlenecks in computer performance and reliability. Early computers used cumbersome and unreliable display tubes as memory storage. Multiple failures per day, triggered by tube failures or even changes in the weather, were common. A real-time, high stakes system such as SAGE needed vastly better performance and much longer mean time between failure. MIT's Jay Forrester made a major breakthrough with the invention of magnetic core memory. SAGE also pioneered interactive displays, advanced methods of time-sharing, fault-tolerant banks of computers, and other speed, usability, and reliability boosting approaches. ¹⁹

It was during the computer's military years that programming went from wiring diagrams to software compilers. The early breakthrough general purpose machines—ENIAC, Von Neumann's IAS machine, Turing's Mark I, the IBM 701, and the SAGE Whirlwind, all used machine language. These multimillion dollar mainframes were difficult and expensive to program. It was not until April 1957 that FORTRAN finally was finished, and then only available on the IBM 704. Other machines, and other programming languages, followed soon after. ²⁰ By 1960, most computers used recognizable programming languages, with high-level instructions compiled by the computer into its own required machine language.

The early wave of computers was almost exclusively government funded. Between 1949 and 1959, the government was the most important funder of research and development. By the end of the period, the balance was shifting toward commercial self-financing, but all the major commercial systems could trace their R&D closely to military systems. Flamm's data, shown in Figure 1, highlight the heavy reliance on government investing in contractor's R&D during the 1950s.

OPEN IN VIEWER

Fig. 1.

The U.S. Government spent heavily to develop the computer. Data: Flamm. ¹⁸

The military was also paramount in funding the development of the transistor. Although computers made great strides between 1945 and 1955, they still relied on bulky, unreliable, and power-hungry vacuum tubes. Were this to continue, computers would remain high-cost specialty devices operated by only the largest of organizations with the highest value needs. What changed this was the introduction of transistors and integrated circuits. With semiconductor logic, computers became vastly more powerful. At the same time, they became much cheaper to build, own, and operate.

Each successive category of transistors, such as the point contact transistor (1948), the junction transistor (1951), and the planar transistor (1958), was developed on government contract: As Riordan and Hoddeson put it: ²¹

Transistors were absolutely vital for the compact advanced avionics used in missiles, satellites, and manned spacecraft. By the end of the 1950s, transistors had replaced tubes in all new computer designs. Integrated circuits would soon follow, where multiple transistors and other circuit elements were combined on a single silicon crystal.

IBM goes commercial and dominates the mainframe market

IBM was not the earliest commercial computing firm. Eckart and Mauchly, inventors of the ENIAC, always aspired to serve the commercial market. As early as 1947, Prudential Insurance was investing in its company and developing a computer for actuarial uses. ²² IBM was the most successful in translating lessons learned on government contracts to the consumer marketplace. IBM's 1950s government contracts implementing SAGE for the Air Force, the Stretch computer for NSA, and the 7090 for the Ballistic Missile Early Warning System all created a technological base and manufacturing capacity that it could turn to creating a consumer computing market. “Well more than half” of IBM's research and development spending in the 1950s was paid for by U.S. government contract. ²³ The government attached no royalties or residual payments for this backing, and it helped IBM become the world leader in computing. This is a pattern repeated throughout the decades in computing, where public R&D has been central to new computing capabilities.

Beginning in the 1960s, IBM invested heavily in consumer-oriented computing, leading to the breakthrough System/360. At the time, the System/360 was the largest private investment ever in a new product. IBM invested more than $5 billion in a 4-year period, with $500 million on research and development and $4.5 billion for new plant and equipment. ²⁴

Compatibility was a breakthrough feature of the System/360. The same software could run on the largest or the smallest member of the family. This was a radical improvement in usability, and customer demand was extremely high. The System/360 moved IBM into a leadership position it held for several decades.

Even though integrated circuits were available in the early 1960s, the System/360 was transistor based. IBM moved to integrated circuits in the 1970s, with the successor System/370. Although this was several years after its competitors Control Data and NCR, IBM's combination of compatible software and sales force made it safe for IBM to delay its adoption of integrated circuits.

By the end of the 1970s, the computer had existed for nearly 40 years. IBM was the dominant computing firm, facing a modest challenge from the plug-compatible firms. Minicomputer makers, such as Digital and Data General, had also carved out successful offerings with smaller systems tailored to mid-sized businesses and university departments. Computing appeared to be a stable oligopoly. Industry observers were worried about a lack of innovation and a potential threat from Japanese computer companies. All this would change rapidly with another semiconductor breakthrough.

Personal computing and ubiquity

Computing historian Paul Ceruzzi states that “Second only to the airplane, the microprocessor was the greatest invention of the twentieth century.” ²⁵ This was not Intel's perception when it released the microprocessor. Intel viewed the microprocessor as primarily an industrial control device, and did not envision why anyone would want such an underpowered computer. ²⁶ Intel originally developed its first microprocessor, the Intel 4004, under contract to calculator company Busicom. ²⁷ In the next couple of years it released 2 more powerful versions, the 8008 and the 8080. Intel's 8080 and MOS Technology's 6502 would be the chips to launch the personal computer market.

The personal computer's development fell to hobbyists and startups, rather than established computing firms. The personal computer is a classic example of Clayton Christensen's “entry from the bottom.” ²⁸ Hobbyists and startups create a new, inexpensive product that barely qualifies as a working system. Being cheap, it captures the price-sensitive market willing to live with limitations and partial functionality. Over time, as capabilities grow while retaining a low price, the challenger wins more and more share from the incumbent (and expensive) providers.

The acknowledged first personal computer was the Altair 8800, announced in January 1975, and sold by a small company MITS in Albuquerque. It used the Intel 8080. This is the machine that convinced Paul Allen to join MITS, and Bill Gates to leave Harvard and form Microsoft. A version of BASIC for the Altair was Microsoft's first product.

Microprocessors also stimulated the formation of the Homebrew Computer Club in Silicon Valley, which in turn became a hotbed of personal computing hobbyists. Homebrew's most famous alums, Steve Wozniak and Steve Jobs, launched the Apple I in 1975. By 1977, Apple had become a successful venture-backed startup with the Apple II. Apple went public in December 1980.

The largest selling personal computer in the late 1970s was the TRS-80 from Radio Shack. With a national retail presence, and a willingness to advertise the personal computer, Radio Shack was able to sell several hundred thousand TRS-80s per year from 1977 until the mid-1980s.

IBM was much more proactive when confronting the personal computer market than it had been when confronting the minicomputer. Teaming up with Microsoft, the IBM personal computer quickly became the leading business microcomputer. Over time, personal computers grew to dominate the home market as well. By 1990, IBM personal computers and compatibles had nearly a 90% market share.

Personal computers transformed and expanded the computer industry. In 1974, fewer than 115,000 existed in the world. ²⁹ By 1995, ∼300 million units had been sold worldwide. Computing had become an everyday part of business and society.

Digital GPT cost reductions

For hundreds of years, computing technology was a backwater. Decades might transpire between improvements. Yale economist William Nordhaus documents the very slow progress from 1850 to 1940 and the exponential improvements thereafter. A key limitation in this early period was the lack of sufficiently precise manufacturing tools. Charles Babbage provides a striking example. In the early and mid-1800s, Babbage created plans for a “Difference Analyzer” and an “Analytical Engine.” His Difference Analyzer blueprints were sufficiently detailed that the British Museum of History used them in 2002 to produce 2 working devices. Babbage's Analytical Engine was much more ambitious, and would have been a full-fledged computer nearly 100 years earlier than actually occurred. Machining capabilities were too crude to build the machine, and mechanical switches too slow, for the computing device to be practical. While Babbage had developed many of the theories and core design principles, as he himself was forced to concede: “another age will have to be the judge.”

By World War II, electronics had developed to the point wherein computers were now feasible. War time needs in cryptography, radar systems, ballistics, and the Manhattan Project ³³ provided high priority demands justifying the computer development efforts. The previous 20 years of electronics innovations, especially vacuum tubes, created the necessary components for a digital computer.

Figure 2 demonstrates the astonishing rate of progress of the digital GPT, as measured by the declining cost of computation. ³⁴ Modern computers are over a trillion times more powerful than those early models.

OPEN IN VIEWER

Fig. 2.

Cost of computing, 1850–2006, with a trillion-fold cost decline after 1945. Data: Nordhaus ³⁴ , Online appendix, 2006 dollars.

Miniaturization is the driving force behind the cost declines reflected in the Nordhaus graph. The underlying physics were sketched out by Richard Feynman in a 1959 after-dinner speech entitled “There's Plenty of Room at the Bottom.” ³⁵ Feynman pointed out the multiple benefits when circuit elements become smaller. Signals travel shorter distances, current requirements drop, material costs fall, and the entire circuit runs faster. This is true regardless of whether the components are tube or transistor, discrete or integrated. As Feynman put it: “There is nothing that I can see in the physical laws that says the computer elements cannot be made enormously smaller than they are now. In fact, there may be certain advantages.” For the past 70 years, computer technology has benefitted from the virtuous cycle of miniaturization. By the time of Feynman's speech, there had already been 15 years of exponential cost decline. This performance improvement has continued for another 55 years.

Although Feynman provided the direction of the computing industry, a short 1965 essay by Gordon Moore of Intel that appeared in Electronics magazine came to be viewed as setting the pace of this miniaturization and cost reduction progress. ³⁶ Asked to consider the evolution of the industry, Moore sketched his thoughts about cost improvements in the rapidly growing semiconductor industry. What is now known as Moore's Law is the steady doubling of computing power and the rapid fall in computing cost. Engineers and business people struggle to maintain pace, and Moore's Law became the focal point for these design objectives.

Innovation in the components underlying digital computing assisted in this race to smaller, faster, and cheaper. The earliest mainframes filled a room, consumed massive amounts of electricity, and failed often. By the late 1950s, transistors replaced vacuum tubes, dramatically improving reliability, speed, and energy efficiency. The 1960s and 1970s saw the rise of integrated circuits, and their incorporation into computer memory and controls. By the 1980s, integrated circuits were powerful enough to support consumer-oriented personal computers. By the end of the 1980s, there were millions of computers capable of communicating and sharing information. At that point, an Internet was both feasible and highly desirable.

The digital GPT has been a central driver lowering the cost of space by reducing the weight needed to perform a task in space. Both the purchase price of satellites and the required mass for a given satellite function have fallen dramatically due to the digital GPT. The biggest difference between space possibilities now and the Apollo era is not in our rockets, but in our computers.

An extreme example of the interaction between space and the digital GPT is a highly ambitious effort by the Breakthrough Initiative to create deep space nanosatellite probes. Funded by Internet-entrepreneur Yuri Milner, the initiative seeks to combine extraordinarily compact electronics, high-efficiency solar sails, and Earth-based laser arrays into a deep space exploration system capable of near-relativistic speeds. ³⁷

Electronic miniaturization is at the heart of the approach. Digital cameras provide a striking example. In precell phone days, digital cameras were bulky and inefficient. Intense market competition, driven by a global cell phone market supply chain, has resulted in vastly smaller, more powerful, and much more energy efficient detector chips and lenses. The same is true of processors, optoelectronics, and other widely used commercial electronic components. The weight of the Breakthrough Initiative satellite components has already fallen by >99% compared with that in 2000. The 2030 improvements needed are modest in comparison. Much more challenging is the propulsion system, utilizing high-power phased-array ground lasers to “push” the nanosatellite toward its target.

Other new space efforts benefit greatly from ongoing digital miniaturization. Companies such as Planet Labs use clusters of small satellites to provide commercial views of Earth for agriculture, mining, forestry, and other uses. Very large constellations of small but powerful Internet communication satellites are being developed for low-earth orbit. As discussed below, for the foreseeable future the new space market is in essence a digital spinoff.

Digital GPT spinoffs and complementarities

Input substitution is one of the most basic and powerful insights of economics. That is, when an input becomes more plentiful, it is efficient to use more of it. This is true for both business and consumer uses. When a technology improves at the rate of the digital GPT, it leads to widespread substitution in an ever-growing number of industries. This is especially the case when new uses and new applications are possible, rather than just more of the same.

Delong and Summers point out that the digital GPT has been able to transform much of the economy by reinventing and expanding its uses. The first 2 decades of computing were dominated by “complicated and lengthy sets of arithmetic operations.” ³⁸ Scientific computing was central to the earliest usage of computing in weapons research and a national radar defense system. Both computing and the space program owe their early successes to the Cold War nuclear rivalry between the United States and the Soviet Union.

In the 1960s, computing moved from primarily governmental to those funded by commercial demand. Corporate databases provided the second broad area of computing applications, with finance and insurance firms using mainframes to track customer accounts and stock market trades. By the mid-1960s, these mainframes were powerful enough to handle real-time databases, such as airline reservation systems and manufacturing inventory control.

Complementary assets are the third defining characteristic of GPTs. Technological complementarities are those “whose full benefit cannot be reaped until many of the other technologies that are linked to it are re-engineered, and the makeup of the capital goods that embody them are altered.” ³⁹ For computing, the most important of these is software.

Although software has always been essential to computing, the software industry was relatively slow to emerge. For the first few decades, computing companies bundled their main packages with the hardware. Software expertise was valuable, and computer services companies emerged as one of the most important adjuncts to hardware. Firms such as CSC and EDS augmented large organizations' computing efforts. ⁴⁰ IBM announced plans to unbundle its computer systems in late 1968, partially in response to a Justice Department antitrust investigation. Minicomputers and microprocessors always supported third party software.

The personal computer has been especially influential for software innovation. For example, spreadsheets and word processors dramatically altered white collar work. Although it took until 1980 for the U.S. software market to reach $5 billion in sales, by 1990 it exceeded $44 billion and reached $138 billion in 2000. ⁴¹

Some of the most widespread and important software is free. There is a long tradition of open source software, partly driven by AT&T's consent decree with the government restricting it from entering computer services. ⁴² The Unix operating system and the C language emerged from Bell Labs, and became mainstays on university campuses. They would eventually become central to the Internet. The open source movement influenced World Wide Web development, with Tim Berners-Lee and CERN choosing an open source approach to these enabling technologies. ⁴³

Proprietary networks

Almost as soon as computers appeared in large organizations, researchers began experiments to remotely access these machines. Bourne and Hahn document a wide variety of efforts in the 1960s and early 1970s, highlighting the many “firsts.” ⁵⁰ Database search was a key theme, tapping into library catalogs and retrieving scientific journal abstracts. By 1975, there were commercial firms such as DIALOG, LEXIS, and Dow Jones News/Retrieval providing online data access. These companies targeted large businesses with support staff, as the data services were both expensive and difficult to use.

Early U.S. consumer online services faced numerous challenges. When CompuServe launched in 1976, the typical modem speed was a glacial 300 baud. Services had to maintain a very “light”page load. Services distributed CD-ROMs with graphics to augment the online textual content. Usage fees were $12.00 per hour during the day and $5.00 an hour off-peak. ⁵¹ Despite their limitations, online providers The Source, CompuServe, DELPHI, MCImail, GEnie, The WELL, and AOL pioneered consumer online access during the 1980s. Features included online forums, messaging, online encyclopedias, financial services, e-commerce, e-mail, special interest groups, and games.

There were parallel efforts in Europe, especially England and France. In 1981, travel agents could order online through Thomson Holidays. In 1984, a small e-commerce experiment let residents of Gateshead, England, order groceries, prescriptions, and bakery items. ⁵² The French Minitel service was much more extensive, and demonstrated the power of rapidly available information, relative ubiquity, and the tendency to create unexpected services. ⁵³ Minitel utilized text-based information and relied on specialized Minitel terminals. Eventually it would become something of a drag on French Internet adoption during the mid-1990s. ⁵⁴

By the late 1980s, dial-up speeds had increased by an order of magnitude. Costs also fell. The Prodigy online service was able to introduce a relatively low monthly flat rate price ($9.99 at service introduction) for its new service combined with online graphical advertising. The Prodigy Services Corporation was a household shopping and information service that connected home computers with an online network of servers configured to provide a multitude of education, information, and entertainment services. Prodigy was originally developed as part of a joint venture between CBS, IBM, and Sears, Roebuck and Company and launched in certain test markets in 1988 and nationally in 1990. Whereas CompuServe originally started to serve the business community and then branched out to serve home users, Prodigy followed the reverse course. The primary focus of Prodigy was the home user, only later in its history did Prodigy attempt to launch a business services division. ⁵⁵ These early consumer systems were a marketing revelation. Marketers realized that with the rise of personal computers, fax machines, and value-added online services, they had a new and potentially much cheaper and more powerful method of directly interacting with customers. Computers allowed direct marketers to store individual transaction data cost effectively. The new online services suggested they could utilize this personal information to more effectively target and communicate with individual households. Peppers and Rogers pointed out that this had profound effects on the economics of direct marketing: ⁵⁶

The proprietary online services demonstrated latent demand for online capabilities. These limited systems needed a more flexible, cheaper, and interconnected technological approach.

Defense Department's Advanced Research Projects Agency creates the Internet

The commercial Internet and World Wide Web grew out of decades of research, heavily sponsored by Defense Department's ARPA and other research agencies. The Internet was already 25 years old when it was deregulated. It was, and continues to be, an especially powerful and expandable set of technologies for general purpose online communication. ⁵⁷

The governmental origins of the Internet have been well documented, and mirrors early efforts in graphical personal computing. ⁵⁸ The role of the Defense Department's ARPA, especially the Information Processing Techniques Office (IPTO), is again central. ARPA funded the basic packet switching research of Paul Baran, the initial network deployment in 1969, the original IMP servers, and the operation of the early Internet nodes by the consulting company Bolt Beranek and Newman.

Throughout the noncommercial era, the research network continued to develop capabilities we now view as fundamental. Among the notable are e-mail appearing in 1973, message groups in 1975, and TCP/IP in 1983. At each stage of this development, researchers reached consensus on core features, produced working code, and documented their work. An especially valuable innovation was the “Request For Comments” (RFCs) online database. RFCs began with the earliest stage, with RFC 1 published on April 7, 1969. The full catalog of RFCs continues to grow, ⁵⁹ and serves as “the principal means of open expression in the computer networking community.” The success of the Internet approach has strongly influenced open source software generally, with many projects mimicking the governance approach.

The Internet becomes friendly

The World Wide Web originated outside of U.S. funding, but in a very similar setting and with active participation in the RFC process. Tim Berners-Lee was working at CERN in Switzerland in 1990 and was struggling with a method of sharing the huge and varied data emerging from particle accelerators. CERN researchers across the globe needed to share and annotate information, much of it including imagery. From these efforts came the concepts of URLs beginning with http, ⁶⁰ web pages mixing text, images, and links, and many of the specifications so familiar and popular. Other users quickly realized this was a useful tool in many settings, far beyond particle physics. CERN also produced the first web server, with public demonstrations in 1991.

A final critical piece of enabling software was the browser, again emerging from a laboratory setting, this time the National Center for Supercomputing Applications in Urbana-Champaign. Programmers there made a “point and click” interface for the Web, releasing it in 1993, and making Internet use accessible to novices. Many of these NCSA Mosaic programmers relocated to California and formed Netscape. Although there were some disputes regarding naming rights, again a core feature of the Internet began with public funds and was then transferred to the private sector.

By the end of 1993, the main software pieces for a friendly point and click interface to an efficient TCP/IP Internet were in place. What was also needed was access and transmission, which occurred with a modest regulatory change of great consequence.

Freeing the Internet

For much of the 20th century, the telecommunications system on the United States was a regulated monopoly. Had this remained the case, the Internet as we know it would not have happened. “Ma Bell” was very strict with allowable communications. It took a Federal Court decision in 1956 to permit users to attach a snap-on cup to a phone mouthpiece to make a call more private. ⁶¹ More devices were possible after the 1968 Carterfone decision, permitting an acoustic client-side link between the phone system and a radio call. AT&T actively worked against competition of any form in long distance, even if transmission occurred over private microwave lines.

These actions and others led the U.S. government to prosecute AT&T on antitrust grounds. ⁶² The conclusion of the case was a consent decree breaking AT&T into a national carrier and 7 regional Bell operating companies. Although many of these companies have subsequently recombined, with AT&T and Verizon the main surviving companies, ⁶³ during the 1980s and 1990s the U.S. telecommunication network was freer than it had been in a century. In particular, there was no legal presumption against entry or a dominant firm enforcing its choices on the network backbone. This competitive opportunity supported the next critical decision of the government, turning over the Internet backbone to private hands.

For the first 16 years of its existence, ARPA managed the Internet backbone. The first connection occurred in September 1969, linking UCLA and the Stanford Research Institute. By 1971, the network connected a number of the key computer science departments and IT contractors. By 1980, the ARPANET incorporated most of the leading research universities and national laboratories.

In 1985, ARPA transferred control and operation of the nonmilitary nodes of the ARPANET to the National Science Foundation. At first the NSF focused on improving operations of the (now) NSFNET. ⁶⁴ The backbone speed was increased, although minuscule by modern standards. ⁶⁵ It entered into a 5-year managing contract with MCI and IBM/MERIT to run operations.

The NSF insisted that the Internet operate for noncommercial research purposes. The guiding document was the NSF Acceptable Use Policy. In particular, users of the NSFNET were not supposed to advertise or place commercial orders. ⁶⁶

“The purpose of NSFNET is to support research and education in and among academic institutions in the United States by providing access to unique resources and the opportunity for collaborative work.

This statement represents a guide to the acceptable use of the NSFNET backbone. It is only intended to address the issue of use of the backbone. It is expected that the various middle level networks will formulate their own use policies for traffic that will not traverse the backbone.

Under such a policy, most of the modern Internet would be prohibited. There could be no ad-supported Google or Facebook, no ordering through Amazon, no ticket purchases at Expedia, or countless other activities.

Even with these restrictions, the utility and growth of the Internet continued. By 1990, there were ∼1 million Internet hosts. Capacity growth was needed, and the NSF increasingly felt the demands of network operation would conflict with its support of basic science. At the same time, political leaders were pushing to open up the Internet to additional uses.

Greenstein (2015) details the NSF handoff to private industry. ¹¹ After some initial jockeying by lead contractors for a preferential position, NSF reached a final plan for the handover in May 1993. MCI and IBM continued to operate the backbone, but without a requirement for enforcing the Acceptable Use Policy. Various networks were strongly encouraged to interconnect, and firms developed a method of traffic interchange. Several of the core assets of the Internet, such as the DNS registry, were transferred to private hands in a manner eliciting complaints of cronyism. This aside, the transfers happened quickly and smoothly. By 1994, the Internet was open for business, able to carry both research and commercial traffic. One of the largest investment bursts in innovation was about to occur.

Venture-funded innovation

After privatization, entrepreneurs needed backing for their new ventures. To an unprecedented degree, this was provided by VC. As discussed in Hanson (2000), there was a positive feedback loop involving investors, users, and entrepreneurs. Or, as Mowery and Simcoe put it ⁶⁷ :

Large pools of capital, such as pension funds, insurance companies, and other large investors, manage portfolios worth hundreds of billions of dollars. Finance theory shows that efficient and diversified portfolios should devote a (small) portion of their assets to high-risk, high-return investments. Over the past few decades, one preferred method for this risky investment is early stage VC investments in high-technology startups. Some of the largest global corporations, such as Apple, Cisco, Google, and Facebook, trace their origins to this form of risky investing.

Most pension and life insurance companies lack the experience to invest in startups. Instead, they outsource this fraction of their portfolio to VC firms. VC firms gain skill in identifying startup teams with both good ideas and the ability to transform these new approaches into successful companies. The various VC firms organize funds stressing particular areas of emerging technology, such as the Internet, clean tech, or life sciences.

VC firms do not lend money. Rather, they buy an ownership share in a startup. The size of the equity share is in proportion to the funds invested, as well as the current startup valuation. The earliest investments in a new company can be the most lucrative, but they are also the most likely to fail. For example, the earliest investors in Google were able to purchase shares at ∼4 cents a share. These were worth $85 at IPO, or more than $800 per share in 2017. One of these early investors was Jeff Bezos, who invested $250,000 in 1998. If Bezos held on to his Google shares, they would be worth approximately $2.7 billion.

As a startup does not repay a venture investment, which it would with a loan, a VC firm must eventually be able to sell its shares to recoup its original investment and make a profit. Some form of “exit event” is required. For major successes, such as Netscape or Google, this is an initial public offering. More commonly, an acquisition by a more established firm provides the mechanism for selling shares.

Venture investing is a game of averages. In the majority of cases, the venture-backed startup fails to go public and receives only lukewarm acquisition interest. For a venture fund to make money, the few big wins must offset these more numerous failures or modest successes.

The venture capital system is a key reason for the importance and continuing success of Silicon Valley. VC firms and startups have long coexisted in the Bay Area, with veterans of successful companies occasionally starting their own venture operations or joining established firms. Proximity increases the chances of discovering the next big thing, meeting the best entrepreneurs, and bringing in additional investing partners to lower the exposure to any one investment.

Silicon Valley has had the good fortune of operating with the power of the digital GPT at its back. For more than 6 decades, IT has been central to the Valley's economy. Moore's Law created an ongoing stream of new computing opportunities. Engineers substitute digital technology into more industrial devices, production processes, and final goods. This sequence includes such broad areas as mainframe computing, avionics, minicomputing, personal computing, the rise of the Internet, social media, and mobile computing. Perhaps no other technological area has been so rich in the potential for continual improvements and spinoffs, and the association of IT with VC has been very close. Successful venture capitalists became quite good at extrapolating from current possibilities as new price points emerged. Figure 4 highlights just how widespread the importance of the Internet still is among venture-backed startups.

OPEN IN VIEWER

Fig. 4.

Internet venture investments occur in many industry sectors. Data: National Venture Capital Association 2014 Report.

Space investments do not appear separately in Figure 4. If they did, the Internet's influence would be large as well. Startup Space reports $1.8 billion in space-related venture investments in 2015. ⁶⁸ Of this, $500 million went to OneWeb and $900 million from Google to SpaceX. OneWeb is entirely Internet related. Although neither Google nor SpaceX confirms the exact arrangements for their agreement, their agreement appears to be targeting the Internet-related constellation being planned by the firms. This would total $1.4 billion of the $1.8 billion for just those 2 firms alone.

E-commerce pioneered the commercial Internet, but advertising pays the bills

E-commerce was at the core of the first wave of Internet startups. The “dot com” revolution sought to transform both business to business and consumer sales, and attracted a burst of VC funding. Several factors accounted for this emphasis. Facing challenges for Wal-Mart and other “big box” stores and a lack of expertise, existing retailers were slow to adopt online sales. ⁶⁹ This lack of response created a perceived opportunity for startups, especially in consumer categories such as books, music, consumer electronics, and computers. Business marketplaces hoped to challenge existing supply chains in a wide array of industries. A wave of IPOs seemed to validate these entrants, with investors willing to fund retail startups with very little track record and negative cash flow. Many of these startups came to ruin in the 2000–2001 shakeout, never able to effectively reach profitability. Existing businesses, especially in the B-to-B market, woke up to the power of the Internet and used their existing supply chains and marketing expertise to thwart the startups.

Despite the failure of many speculative ventures, there were notable successes. Amazon.com expanded its product lines from books to almost everything, and has managed to expand its sales substantially every year. It is now a retailing powerhouse. EBay serves as the order-taking interface to tens of thousands of small retailers. Dell and other computer makers pioneered direct sales to consumer and build-to-order systems. Music retail pioneer CDNow expanded the available range of titles dramatically over physical stores, and early digital download sites such as Napster showed the power of purely digital delivery to transform the music business. Thousands of other online retailers, both hybrid and online only, expand their reach to a national (and occasionally international) market using the Internet. Consumer e-commerce is still <10% of retail sales, ∼$500 billion in 2016. The e-commerce retail share has followed a (seasonally adjusted) steady upward trend since the Census Bureau began tracking it in 1999. ⁷⁰

Although e-commerce has been a notable commercialization success, the Internet's impact on the advertising market has been profound. As seen earlier, the NSF Acceptable Use policy prohibited advertising. The commonly acknowledged first advertisement using the Web, a banner ad paid for by AT&T, appeared on the hotwired.com web site in October 1994. ⁷¹ It ran on the top of the page and when clicked led the user to the advertiser's web site. This simple format would remain the dominant method for the first decade of online advertising.

Advertising revenues have long been central to the commercial Internet. During the early years, it funded the rise of online portals Yahoo!, InfoSeek, Excite, and America Online. Much of the money spent by the dot com retail startups went to marketing expenses, especially online advertising, driving the development of many of the online portals, search engines, and access providers. Thus, the early growth of online advertising was closely tied to the dot com wave.

The first wave of Web advertising had few of the current capabilities, lacking even the ability to reliably measure impressions. Early online approaches emphasized sponsorship of particular web site sections, especially on leading Internet portal sites Yahoo! and America Online. In addition to exclusivity, a key feature of sponsorship is duration-based payment. Advertising sites specify pricing by time interval, such as cost per month, rather than by viewership levels or user actions. Sponsorship became less common, and performance-based methods more important, as Internet companies developed appropriate metrics.

Search engines, most notably Google, provided a second wave of advertising growth. Search engine advertising reached beyond online retailers and tapped into thousands of small and medium sized traditional businesses. Advertising format simplicity, combined with strong measurement tools, made it possible for these advertisers to shift their advertising dollars from newspaper and Yellow Pages to search spending. For example, Google AdWords generate billions of dollars of revenue using a few lines of text with links to advertisers' landing pages. More recently, social media and video content have grown to be important advertising venues.

Advertising dollars tend to follow consumer time usage, and the many online applications increasingly attracted viewership. Relevancy, especially for search, has made online advertising very effective. Even primitive ads can be highly effective if they reach the right consumer at the right time.

From its humble beginnings in 1994, online advertising revenue is now larger than newspapers, magazines, radio, and even television. In 2016, U.S. online advertising revenues grew >21% and exceeded $72 billion. Some of the most sophisticated technologies in Silicon Valley, such as massive clusters of servers, big data analytics, machine learning algorithms, and artificial intelligence, are devoted to achieving slight improvements in online advertising effectiveness.

Figure 5 shows the rapid growth in U.S. online advertising revenue, with no obvious plateau. ⁷² This is striking for at least 2 reasons. First, early participants did not view the Internet as a major advertising venue. Deregulators were familiar with some of the most popular services on the Internet, such as e-mail, chat groups, and bulletin boards, and expected them to be popular outside of the research community. Web pages also looked promising for sharing commercial material. There was little discussion during the early 1990s that the Internet would be a major advertising venue and, except for modest success by online services such as Prodigy, little historical basis for optimism. Indeed, a best-selling business and marketing book published in 1993 stressed the interactive potential of faxes rather than the Internet. There was also an antipathy to advertising by many early Internet entrepreneurs, including Brin and Page, who did not initially plan to utilize advertising as part of Google's approach.

OPEN IN VIEWER

Fig. 5.

U.S. online quarterly advertising revenue. Data: Internet Advertising Bureau, various years, nominal dollars.

The economics of advertising is a second reason why Figure 5 is so surprising. Aggregate advertising spending is primarily a zero-sum game between advertising venues. ⁷³ For decades, U.S. advertising spending has been ∼2% of gross domestic product (GDP). Growth rate of Internet advertising far exceeds the growth in the economy, coming at the expense of other advertising approaches.

As with the computer, there is a strong element of Schumpeterian surprise in the business applications of the Internet. It is highly unlikely that a committee sitting in 1992 could have foreseen the dot com boom or the dramatic rise in Internet advertising. Yet, by permitting these activities, policy makers allowed a creative wave that has altered many aspects of the retail and advertising industries.

Allocating spectrum

A wide variety of wireless services use spectrum, including broadcast radio, television, point-to-point microwave, satellite, and radar. The U.S. Commerce Department governed the early days of radio broadcast, generally under a “first come, first served” regime and court adjudication of disputes rather than in a regulatory proceeding. With the passage of the 1927 Radio Act, the Federal government asserted more control in assigning licenses to specific operators, designating specific bands for specific uses, and mandating the use of specific technologies.

The 1934 Communications Act assigned to the FCC responsibility for managing non-Federal spectrum “in the public interest.” The National Telecommunications Information Agency is responsible for Federal use of the spectrum. The FCC has historically determined what services and technologies can make use of specific frequencies of the electromagnetic spectrum through an administrative rule making process.

Despite whatever success the FCC might have at determining an optimal combination of service and technology at any point in time, continuing changes in consumer preferences and technology eventually cause that combination to become suboptimal. As the divergence between the value of the current service and potential new uses increases, so do the gains from reallocation. The FCC's increasing reliance on market-based spectrum policies has facilitated such reallocation.

Leo Herzel ⁷⁷ in 1951 and Ronald Coase ⁷⁸ in 1959 set forth the intellectual underpinnings that such licenses could (and should) be awarded by markets rather than by a governmental process. The key insight of Herzel and Coase was that the FCC added no value to the allocation/assignment process as licenses would end up in the hands of those who valued them most highly on the secondary market and the FCC process simply hindered such ultimate allocations.

Using auctions to award licenses for specific services subject to specific FCC technical rules would have been a beneficial but small step forward. In the 1990s the FCC went further, setting up a framework that would allow a competitive market for wireless services and to minimize the role of the FCC in determining the technical and service requirements for licensees. By allowing more flexibility for licensees to change technology and service, operators could attack markets where prices were too high and respond to changes in consumer demand without having to get Commission approval.

Cellular service

One of the many complaints about the old bureaucratically controlled system of license allocation is that it took a long time to respond to market changes and opportunities for new services. The original concept of low-power cellular systems emerged in the 1940s, and rudimentary cellular systems could have been available in the 1970s. ⁷⁹ The FCC did not award the first experimental cellular license until 1982, and more pervasive licensing was not complete until the later 1980s. The FCC awarded 2 licenses per market, first using the comparative hearing approach and then using lotteries.

In 1982, it was not clear that the licenses were extremely valuable. McKinsey Consulting initially thought that there would be 1 million subscribers across the country by 2000, and in the discussions about the division of assets under the AT& T break-up, the negotiators did not care about the placement of the cellular assets. By 1984, increased demand came from the advances in technology due to the microprocessor revolution. Moore's Law created capabilities that led to hand-held phones. The increased value of wireless service manifested itself in prices for the resale of the licenses and the increased number of applications for the remaining available licenses. ⁸⁰

The high value of the cellular licenses demonstrated the inefficiency and inequity of the comparative hearing process. For example, the FCC awarded a Los Angeles cellular license on the basis of 1 additional cell site in the plans on Catalina Island despite the fact that the rest of the systems were identical and ultimately bore no relation to the system that was initially constructed nor to the long-run performance of the system. The FCC moved to award licenses by lottery. Lotteries led to hundreds of thousands of applications. According to McMillan, “In one not atypical case, some dentists won the right to run cellular-telephone service on Cape Cod; they immediately sold their license to a real telephone company Southwestern Bell, for $41 million.” ⁸¹

Initially the FCC planned for a single license per market, awarded to the incumbent wire line telephone provider (expected to be best technically able to develop and operate a wireless telephone system). The Antitrust Division of the Department of Justice pushed the FCC to allocate licenses to at least 2 providers, so there would be competition to provide service. Ultimately the FCC agreed, assigning one of the 20 MHz licenses to the incumbent wire line provider and the other to an entrant. The FCC mandated the use of the Advanced Mobile Phone Service (AMPS) analog technology and banned dispatch service on the cellular frequencies.

Initial prices for cellular service were high. Despite this expense, demand for wireless service was also high and there was a push to allocate more spectrum for wireless service. The FCC's first action was to grant an additional 5 MHz to each of the cellular licensees, compounding any original windfall. Even after this increase, it was clear that both carriers and consumers highly valued mobile wireless spectrum. In 1989, the FCC began a proceeding for a new allocation of spectrum for “Personal Communication Service” (PCS). Around the same time, the FCC began to investigate the introduction of digital technologies to provide more capacity and higher quality service over the same spectrum.

The move to auction

Proposals to auction spectrum were included in budget proposals from Presidents Reagan and Bush in the 1980s and early 1990s, but never adopted by Congress. Congressman and senators with significant influence on the Commerce committees opposed the use of spectrum auctions. For example, in 1987, Senator Daniel Inouye, Chairman, Senate Subcommittee on Communications, wrote to the Chairman of the Senate Budget Committee that a proposal to collect $600 million from a spectrum auction “undercuts the fundamental tenet in communications policy that the airwaves are a limited public resource,” and “it is inappropriate to sell such a resource to the highest bidder.”

Budget pressures helped overcome political opposition. The PAYGO provision of the Budget Enforcement Act of 1990 required any new spending to be “paid for” by revenue increases or spending cuts elsewhere. The projected billions of dollars from selling spectrum became very attractive to the new Clinton Administration, who wanted to increase spending elsewhere. Economists also strongly supported auctions, primarily for efficiency reasons. The only reason that spectrum auctions would fetch high prices is that it was scarce and useful. A high auction price was a strong signal of spectrum constraints and scarcity.

Congress, swayed by the potential revenue and the obvious inefficiencies of the existing system, included a mandate for using spectrum auctions in the Omnibus Budget Reconciliation Act of 1993. Not only did the Act mandate the use of auctions, but also required that auctions commence within 11 months of the passage of the Act.

Auction implementation

Although congressional pressure against auctions had been overcome, there was still political influence on the Commission on the conduct of the highly valuable PCS license auctions. Section 309(j) of the Act required that the Commission consider how to promote the interests of small businesses, businesses owned by women and minorities, and rural telephone companies (“designated entities”) in the award of spectrum licenses. The Act told the FCC to consider using set-asides, bidding credits, and installment payments to promote the interests of the designated entities. Congress saw a need for such measures because the designated entities were not likely to be successful in acquiring licenses in an auction with purely monetary goals. The FCC utilized all 3 mechanisms in the second broadband PCS auction in December 1995 (the “C Block” auction).

Had the FCC simply auctioned 2 or 3 nationwide licenses, a straightforward traditional auction run by a cattle auctioneer would likely have been sufficient. The FCC chose not to allocate nationwide licenses for 2 reasons: smaller bidders would have difficulty acquiring such large licenses at the auction even if they could band together for the bidding and then sell off the undesired portions in their own private auction. The risk would be very high for them about the unknown price of the parts they would want to sell after the auction. The second issue was that the existing cellular providers did not have nationwide coverage with their licenses and systems. Either they would be excluded from bidding across the country or there would be a reduction in the additional competition in areas where they won one of the new PCS licenses. Neither of these outcomes was politically acceptable.

The FCC issued a Notice of Proposed Rulemaking soon after the passage of OBRA’93, discussing several possible auction mechanisms that would be more efficient than a simple sequential auction for the licenses. The responses to the FCC-proposed rule, and an academic conference convened to discuss the auctions, led economists inside the FCC to push for a novel design for spectrum auctions. This was a simultaneous multiround ascending auction (SMR).

There was significant risk in adopting the novel design. It had never been tried before, and the broadband PCS auction was expected to have companies bidding billions of dollars. Failure of the auction design would put spectrum auctions overall at risk, not just risk jettisoning the novel design. The risk of failure was partially mitigated by running 2 narrowband PCS auctions in the summer and fall of 1994, testing the approach and the software, before the first major broadband PCS auction.

As a result of the 2 successful tests, the FCC went forward with the broadband PCS auction using the complex SMR design. Evaluations of the auction were that it allowed bidders to express preferences between substitute licenses and also to aggregate complementary licenses. ⁸² The success of the SMR format led governments around the world adopting it for their spectrum auctions.

Auction concerns

It is important for auction design to be in line with the potential uses of the spectrum. If there is no economically viable use for 1 MHz blocks of spectrum compared with 10 MHz blocks, it would unnecessarily complicate the auction to slice the spectrum into 1 MHz blocks. In a similar way, dividing geographic areas into inefficiently small areas would make aggregating the necessary parts more difficult. The auction design should reflect the most likely use of the spectrum. However, it should also allow companies to subdivide or aggregate the spectrum in terms of both frequency and geography.

One issue with the standard SMR format is the risk of “exposure.” When a bidder values the combination of 2 licenses much more highly than either 1 alone, it is possible that the bidder could pay too high a price if it were to win only 1 of the 2, or to overpay for the combination. This problem occurs predominantly when bidders have different preferences. One solution to the problem of exposure is to allow combinatorial or package bidding. ⁸³

Auctions allocate licenses to those who value them most highly. However, private value for an asset is not necessarily the same as social value. Social value is the sum of producer surplus and consumer surplus, whereas bidders will only consider producer surplus. A license for a very socially useful service may not be the allocation in the auction because the potential provider cannot profit enough to outbid other uses. A bidder may also try to entry. It is quite likely that the private benefit to a monopoly provider of a service is greater than the rents that a second provider of the same service could generate. As a result, the incumbent can often justify a higher bid for spectrum licenses than new competitive entrants. It is important for auction designers to think about competition policy when setting up auctions.