We tend to think of the history of computing as a private-enterprise success story, and most of it is. Fairchild and Texas Instruments built the semiconductor industry, and Intel and Motorola turned it into a mass market. IBM, Burroughs, Remington Rand, and DEC built the mainframes; Apple, HP, and Dell built the machines that replaced them. The software came from Microsoft, Adobe, Oracle, and a few thousand companies mant people have never heard of. None of it required a government agency to ask.

What gets left out is where many of those products and technologies originated. The Department of Defense paid for the network over which your packets travel. It paid for the demo where the mouse was first shown in public. It paid for the desert race that trained the people now running self-driving car programs. The arrangement was almost always the same. A federal agency funded the research, built the first working version, or created the program that pushed the thing out of the laboratory. A company then commercialized it years later.

Before there was an industry

There wasn’t much appetite for a company to build a computer in the 1940s. Development would be a huge expense and the payoff would be uncertain, because there was no assurance that there would be customers. The first machines were built because the government wanted a specific answer.

ENIAC. The Army signed the contract with the University of Pennsylvania in June 1943 for “research and development of an electronic numerical integrator and computer.” The stated purpose was artillery firing tables. The first real program run on the finished machine was something else entirely: a Los Alamos thermonuclear feasibility calculation, at John von Neumann’s suggestion, started in December 1945, two months before the public dedication. It consumed roughly a million punch cards, and the problem statement is still classified.

Magnetic core memory. MIT’s Whirlwind project began in 1944 with Navy funding as a flight simulator and evolved into a real-time computer. It nearly died when the Office of Naval Research lost patience with it, and survived because the Soviet atomic test brought Air Force money in. Whirlwind’s electrostatic storage tubes were the bottleneck. They were unreliable, requiring constant refreshing, and expensive to keep alive. Jay Forrester’s replacement for them was a lattice of tiny magnetic rings, each holding a bit stored as the direction of magnetization. This core memory was deployed in August 1953. The machine roughly doubled in speed, and because a magnetized ring stays magnetized, the memory held its contents with the power off, which is a property RAM still does not have. Core remained the standard high-speed memory technology for the next twenty years.

Real-time computing at scale. SAGE (Semi-Automatic Ground Environment), the continental air defense system built out of Whirlwind, was the largest computing project of its era, and MIT Lincoln Laboratory notes that it exceeded the Manhattan Project in both funding and personnel. Each machine contained about 49,000 vacuum tubes and weighed 250 tons; its power supply drew 3,000 kilowatts, enough for about 1,500 houses. Every center ran two of them, one as a hot spare, and that is where fault-tolerant design got its first serious workout. IBM built 56 machines, and its headcount on the project went from about 300 at the end of 1953 to over 7,000 by 1958. The project was a major catalyst for IBM to scale and become a dominant computing company. The system was also obsolete by the time it was fully deployed in 1963, having been designed against Soviet bombers in an era of intercontinental missiles.

Commercial computing as a business. The Census Bureau contracted for a machine in 1948 at a fixed price that turned out to be a small fraction of what building it actually cost, and the shortfall helped bankrupt the Eckert-Mauchly Computer Corporation (J. Presper Eckert and John Mauchly were the engineers who designed the ENIAC) . Remington Rand bought the company’s remaining assets in 1950 to finish the job, and UNIVAC I serial number 1 was accepted in March 1951. Census had done this once before. Herman Hollerith’s tabulating machines cut the 1890 tabulation from the eight years the 1880 census had taken to about two, and the company he built on that contract eventually became IBM.

Networks

Almost every piece of the Internet, from routers to the @ in an email address, was built under federal contract by people who were not trying to build a consumer product.

ARPANET. ARPA (the Advanced Research Projects Agency, later DARPA, with a D for Defense) issued a request for quotation to 140 companies in 1968. Twelve responded. IBM and Control Data both declined, on the grounds that no suitable minicomputer existed at a workable price, and AT&T was openly hostile to the whole idea of packet switching. Paul Baran at RAND and Donald Davies at Britain’s National Physical Laboratory had already published the concept; ARPA paid to find out whether it worked at scale. Four sites were connected by the end of 1969: UCLA, SRI, UC Santa Barbara, and Utah.

Routers. Wiring incompatible mainframes directly to one another would have required custom software for every pair of machines on the network. ARPA’s answer was to put a separate minicomputer in front of each one, a ruggedized Honeywell DDP-516 with 12K words of memory and no job except to accept packets and forward them onward. BBN, the contractor that built them, called them Interface Message Processors. Everyone else eventually called the idea a router.

TCP/IP. The ARPANET’s original protocol assumed a single network under a single administrator. Joining networks that were separately owned, built on different hardware, and unreliable in different ways needed something else. Bob Kahn started that work at DARPA and brought in Vint Cerf, and their design was published in 1974. The Defense Department declared it the standard for military networking in 1982 and set a hard cutover deadline of January 1983, which is the only reason the transition ever finished.

Network email. Ray Tomlinson worked at BBN, an ARPA contractor, and in 1971 he stitched together a local mail program and an experimental file transfer utility so that a message could be sent between machines. He picked @ to separate the user from the host because it could not legally appear in a username on the operating system he was using. Fifty-five years later, people still type it every day.

Academic networking. Computer science departments without a Defense Department contract could not connect to the ARPANET at all, which left most of the field on the outside. NSF (the National Science Foundation) funded the construction of one in 1980. CSNET went live the following year with three sites, grew to serve most of the country’s computer science departments, was told to become self-supporting by 1986, and actually managed it.

The Internet backbone. NSFNET ran alongside CSNET from late 1986 and eventually made it redundant, carrying the traffic that turned a research network into public infrastructure and increasing its capacity from 56 kbit/s to 45 Mbit/s by late 1991. Its acceptable use policy flatly prohibited “use for for-profit activities,” so for most of its life, the backbone of what became the commercial Internet was legally closed to commerce. Lifting that restriction took an act of Congress in 1992. By the time the backbone was retired in April 1995, it connected 50,766 networks in 93 countries.

The web browser as a consumer product. Tim Berners-Lee built the web at CERN, an intergovernmental laboratory. Marc Andreessen and Eric Bina then built Mosaic, the browser that made the web worth using, at the National Center for Supercomputing Applications, an NSF-funded center at the University of Illinois. NSF’s grant aimed specifically at Mosaic did not arrive until 1994, after the browser existed, so the honest version is that NSF paid for the building and the people rather than the product.

PageRank. The Stanford Integrated Digital Library Project was an NSF award to work out how to search library collections, and what it actually paid for was a crawl of 24 million web pages and the disks to hold them. Larry Page and Sergey Brin used that crawl to test the idea that a page’s importance could be inferred from who linked to it. Brin held an NSF Graduate Research Fellowship, and their 1998 paper credits NSF, DARPA, and NASA in the acknowledgments, which is an unusual set of sponsors for a search engine.

The interactive computer

Nearly everything about how we use a computer, as opposed to what it computes, was demonstrated first on ARPA contracts in the 1960s. Almost none of it reached a product for another twenty years.

Time-sharing. Batch processing meant handing over a card deck and getting results back hours later. Project MAC opened at MIT in 1963 on an ARPA grant to solve a different problem: giving many people the convincing illusion that each of them owned the machine. Sixty years later, the scheduler in practically every computer and phone still uses this concept. Project MAC’s flagship system, Multics, was late, slow, and complex enough that Bell Labs pulled out in 1969, which is why Ken Thompson and Dennis Ritchie went off and wrote Unix.

The mouse. Douglas Engelbart’s Augmentation Research Center at SRI ran on money from ARPA, the Air Force, and NASA. The patent, “X-Y Position Indicator for a Display System,” was granted in 1970 and expired in 1987, before mice were a mass-market item. Engelbart personally received about $10,000 for it, and SRI later licensed it to Apple for a reported $40,000. It is among the worst returns on a patent in the industry’s history.

Hypertext and windowed displays. Engelbart’s system, NLS (the oN-Line System), had linked documents and a screen divided into independently scrolling regions by 1968, thirteen years before the Xerox Star shipped and sixteen years before the Macintosh. Xerox PARC (the Palo Alto Research Center) picked up the ideas along with the people, hiring from the ARPA-funded labs and taking Bob Taylor, who had run ARPA’s computing office and commissioned the ARPANET. Steve Jobs then visited PARC twice in December 1979, trading Xerox the right to buy pre-IPO Apple stock in exchange for the demos. Xerox sold the stock before the Macintosh ever shipped.

Videoconferencing and shared-screen editing. In December 1968, at a conference in San Francisco, Engelbart sat on stage in front of a thousand people and worked a computer 30 miles away in Menlo Park, with microwave links carrying video and a projector borrowed from NASA, throwing all of it onto a 22-foot screen. Partway through, a colleague’s face appeared in a window next to the document they were both editing. The event is now called the Mother of All Demos, and ARPA and the Air Force paid for it.

Computer graphics. Ivan Sutherland wrote Sketchpad on the TX-2 at MIT’s Lincoln Laboratory in 1963. The user drew a shape once, stored it as a master, and placed copies of it, all of which changed when someone edited the master. A user could also state a constraint and let the program solve for it. Both ideas are in every CAD package sold today. Sutherland then took over ARPA’s computing office in 1964 and spent two years funding graphics research everywhere else, and the University of Utah program that ARPA money built produced Ed Catmull, Jim Blinn, John Warnock, and Jim Clark. That is Pixar, Adobe, and Silicon Graphics out of a single department.

Silicon

Kilby and Noyce invented the integrated circuit on corporate money. What the government supplied was a customer willing to buy chips by the hundred thousand at prices no one else would pay, back when nobody else wanted any.

The integrated circuit market. Kilby was at Texas Instruments and Noyce at Fairchild, and very little federal research money reached either of them. The demand for their integrated circuits was another story. The federal share of US integrated circuit shipments tells that story on its own:

NASA’s Apollo program consumed roughly 60% of US integrated circuit production in 1963, buying the same Fairchild logic gate about 200,000 times and watching the price fall from $43.50 apiece to $20 or $30. The Air Force’s Minuteman II guidance computer overtook Apollo as the largest single customer by 1965. Every doubling of cumulative output cut unit cost by 20% to 30%, so federal orders at military prices bought the price down for everyone who came later.

Cheap custom chips for researchers. A university group that wanted a chip fabricated in the late 1970s needed a full wafer run costing tens of thousands of dollars, which put custom silicon out of reach of anyone without an industrial sponsor. DARPA paid USC’s Information Sciences Institute to act as a broker, batching many small designs onto shared wafers so each group paid 5% to 10% of a full run. The first batch went out in 1980 with 65 designs from 8 organizations, and by 1988, a single chip could cost as little as $258. The service was called MOSIS, and it is the reason a graduate student in 1985 could use a chip they had designed themselves.

Microelectromechanical systems. Etching moving mechanical parts out of silicon had the same problem as custom chips: no shared route to fabrication, so every group started from nothing. DARPA began paying in 1992 for shared foundry infrastructure on the MOSIS model. The surface micromachining processes developed under that program are how MEMS (microelectromechanical systems) devices are made today, including the accelerometer that tells a phone which way it is being held and the one in cars that decides whether to fire the airbag. MEMS are also used for tire pressure and fuel delivery sensors, flow meters, and micromirror arrays for digital projectors.

Speech, language, and autonomy

DARPA funded artificial intelligence for decades through programs that were often judged failures on delivery, and then ran a series of public contests that did more for the field than the grants had.

Artificial intelligence as an academic field. ARPA money built the AI labs at MIT, Stanford, Carnegie Mellon, and SRI, and for the field’s first two decades, that was most of the field. Then the Mansfield Amendment of 1969 required Defense-funded research to show a direct military application, and open-ended AI work got harder to justify. The AI winter that followed is usually told as a story about researchers overpromising. It is also a story about what happens when a discipline has only one customer.

Speech recognition. ARPA launched the Speech Understanding Research program in 1971, wanting 10,000 words from any speaker, and an advisory board talked it down to 1,000 words in a quiet room from cooperative speakers. Carnegie Mellon’s HARPY managed 1,011 words, running 80 times slower than real time. DARPA declined to renew the program in 1976, partly because no one had defined the evaluation criteria in advance, so whether it had succeeded was itself a fight. When DARPA restarted speech work in the mid-1980s, it insisted on blind benchmarks administered by NIST (the National Institute of Standards and Technology). That benchmark discipline, more than any single algorithm, is what carried the field to the dictation that runs on phones and computers today.

Digital assistants. DARPA’s Personalized Assistant that Learns program ran from 2003 to 2008, and its centerpiece was CALO, the Cognitive Assistant that Learns and Organizes, run by SRI across 25 institutions. The goal was software that watched what you did all day and got better at it, which is a research problem rather than a product. Three CALO researchers left to build the product anyway, incorporated in 2007, and put an app on the App Store in February 2010. Apple bought them ten weeks later and shipped Siri on the iPhone 4S in 2011.

Self-driving cars. DARPA’s approach here was unusual. Rather than fund a program, it announced a 2004 race across the Mojave Desert with a cash prize and no restrictions on who could enter. Not a single vehicle finished. Nineteen months later, five did. By 2007, the cars were driving in traffic and obeying California law. Sebastian Thrun, who won in 2005, was running Google’s self-driving project by 2009; Chris Urmson, from the Carnegie Mellon team that won in 2007, took that project over and later founded Aurora. The first generation of an entire industry came out of three races in the desert.

Robotics and automated security. DARPA kept running contests after the cars, with mixed results. The Robotics Challenge, set up after Fukushima to build machines that could work where people could not, is better remembered for its blooper reel of humanoid robots toppling over than for who won, and the falls did more to calibrate public expectations than the results did. In 2016, DARPA ran a capture-the-flag tournament at DEF CON in which seven machines played against each other with no humans involved, finding flaws in each other’s software, writing working exploits against them, and patching their own code before the exploits could be used back. The winner then entered the human competition at the same conference and finished last, which is a reasonable summary of where automated security stood that year.

Security and anonymity

The ciphers in today’s browsers came from elsewhere: DES from IBM, AES from two Belgians. What the government built and then gave away is a smaller and stranger list.

Hash functions. NSA designed SHA-1 in 1995 and SHA-2 in 2001, published both through NIST as federal standards, patented SHA-2, and then licensed it royalty-free. SHA-1 was broken in February 2017, when a team from Google and CWI Amsterdam produced two different PDF files with the same hash, at a cost of roughly 6,500 CPU-years. SHA-256 has not budged. It signs your TLS certificates, your operating system updates, and every block in the Bitcoin chain, so an NSA design is quietly holding up most of the trust on the Internet.

Access control in Linux. NSA wrote SELinux, the security-enhanced version of Linux, and released it under the GPL in December 2000. It was merged into the mainline Linux kernel in 2003 and has been enforcing access control policies on every Android phone since 2014. The most widely deployed piece of NSA software in the world runs on over a billion devices, and its source code is publicly available.

Vulnerability identifiers. Before 1999, two security tools describing the same flaw had no way to say so, because every vendor kept its own naming scheme. MITRE, a nonprofit that runs federally funded research centers under contract to US agencies, proposed a shared list and published 321 entries in the first year. CVE (Common Vulnerabilities and Exposures) identifiers now underwrite essentially every vulnerability scanner in commercial use, and the whole thing still runs on a federal contract. When that contract was within a day of expiring in April 2025, much of the security industry took notice at once.

Severity scoring. Knowing a flaw exists says nothing about whether to stay up all night fixing it, and through the early 2000s, every vendor answered that question on its own scale. The National Infrastructure Advisory Council, a presidential advisory body, convened a working group of industry security executives that reported in October 2004, and version 1 shipped the following February. Every scanner in the industry now reports a number between 0.0 and 10.0, and three successive revisions of the standard have each computed that number differently.

Onion routing. This is the entry that often surprises people. Tor (originally The Onion Router) is the tool most associated with dark web markets and with evading law enforcement, and it was invented by the US Navy. Encrypting a message hides its contents but reveals exactly who is talking to whom, which, for intelligence traffic, is often the more sensitive fact. Paul Syverson, David Goldschlag, and Michael Reed, working at the Naval Research Laboratory in 1995, found a way to wrap a message in layers of encryption so that no single relay in a chain knows both the sender and the destination. The patent is assigned to the United States Navy. The Navy then released the code publicly, for a reason that is obvious once stated: an anonymity network used only by the government provides no anonymity at all, because anyone using it is a government agent. It needs ordinary traffic to hide in. US government sources still supplied about 35% of the Tor Project’s revenue as recently as fiscal 2024.

Navigation

GPS is the clearest case in the catalog of a technology the government built outright rather than funded. The Air Force built Navstar for the military, launched the first satellite in 1978, and reached full operational capability in 1995. The civil signal was deliberately degraded to roughly 100 meters of accuracy until Bill Clinton ordered the dithering switched off in May 2000, which took civilian accuracy to about 20 meters overnight and is the reason a phone can identify what lane a car is in. Current satellites are built without degradation hardware, and the signal has been free of user fees by policy since 2004.

The common theme

The instinctive framing is that government funds basic research and industry commercializes it. Procurement did more than either.

The largest effect the federal government had on computing came from buying things nobody else could afford yet, at prices no commercial customer would pay, in volumes large enough to move a manufacturer down the learning curve. Apollo and Minuteman paid to work out the bugs in the integrated circuit, and by 1978, nine chips in ten were going to buyers who had nothing to do with either. SAGE did the same for IBM. Research grants produced the ideas; purchase orders produced the industry.

The other thing public money bought was patience. DARPA funded speech recognition for four decades through programs judged failures on delivery, and no company’s board would have carried that. The federal share of American research and development has fallen from 67% in 1964 to 18% in 2023, which is mostly a story about how much private research grew rather than how much public research shrank. But nobody in the private 82% is running a forty-year program on the chance that it eventually works.

References

Cerf, Vinton G., and Robert E. Kahn. “A Protocol for Packet Network Intercommunication.” IEEE Transactions on Communications 22, no. 5 (May 1974): 637–648.

Hafner, Katie, and Matthew Lyon. Where Wizards Stay Up Late: The Origins of the Internet. New York: Simon & Schuster, 1996.

Mowery, David C. “Federal Policy and the Development of Semiconductors, Computer Hardware, and Computer Software.” In The U.S. Government and Technology Policy. Cambridge, MA: National Bureau of Economic Research.

National Research Council. Funding a Revolution: Government Support for Computing Research. Washington, DC: National Academies Press, 1999.

National Center for Science and Engineering Statistics. U.S. R&D Totaled $937 Billion in 2023. NSF 26-314. Alexandria, VA: National Science Foundation, 2026.

Redmond, Kent C., and Thomas M. Smith. From Whirlwind to MITRE: The R&D Story of the SAGE Air Defense Computer. Cambridge, MA: MIT Press, 2000.

Syverson, Paul F., David M. Goldschlag, and Michael G. Reed. “Anonymous Connections and Onion Routing.” Proceedings of the 1997 IEEE Symposium on Security and Privacy (May 1997): 44–54.

Waldrop, M. Mitchell. The Dream Machine: J.C.R. Licklider and the Revolution That Made Computing Personal. New York: Viking, 2001.