In 1968, a researcher named Bob Taylor at the Pentagon’s Advanced Research Projects Agency was tired of switching between three different computer terminals in his office, each connected to a different machine at a different university. His solution — a network that would let one terminal talk to all of them — received a million dollars in funding. A year later, a UCLA student attempted to transmit “login” to a computer at Stanford; the system crashed, and the first message ever sent over what would become the internet was simply “lo.”
This is not how grand national infrastructure projects are supposed to begin. And that, we will argue, is precisely why it worked.
A Graveyard of Confident Bets
The internet emerged during a period when the U.S. government was placing enormous, deliberate bets on the technologies it believed would define the future. Consider the record. The Synthetic Fuels Corporation, created in 1980 and hailed by President Carter as “the cornerstone of U.S. energy policy,” received an $88 billion authorization to produce liquid fuel from coal. Over its six-year existence it spent barely five percent of that budget before closing its doors — none of its four funded projects survive today. Oil prices declined and the entire premise evaporated. The Superconducting Super Collider, approved in 1987 at an estimated cost of $4.4 billion, saw its projected cost balloon to $11 billion before Congress killed it in 1993, leaving behind 14 miles of empty tunnel in the Texas earth and $2 billion spent. DARPA’s own 1980s-era expert-system AI programs — the so-called Fifth Generation initiative that was supposed to deliver reasoning machines through top-down logical rules — consumed years of funding and produced systems so brittle that AI research entered the notorious “AI winter” of the early 1990s.
And then there was the networking protocol that the entire Western policy establishment had backed. By the mid-1980s, the Open Systems Interconnection (OSI) standard had the support of everyone who mattered: computer companies, telephone companies, national governments, international standards bodies, and even the U.S. Department of Defense — the original sponsor of TCP/IP. The Department of Commerce mandated OSI for all government computer purchases after 1990. IBM alone spent hundreds of millions of dollars on it. OSI was the officially sanctioned future of networking. It had everything except working code. TCP/IP, by contrast, followed the motto of the Internet Engineering Task Force: “rough consensus and running code.” While OSI committees debated seven-layer reference models, scrappy engineers were already shipping implementations. Universities wrote in their networking plans that they would use TCP/IP “until OSI becomes available.” Eventually they dropped the final clause.
The Minitel Illusion
France offers the most instructive counterexample. In 1983, on orders from the president, France Télécom began rolling out Minitel — a government-designed telematics system that put the country online a decade before the internet went mainstream in the United States. By the mid-1990s, some 20 million users were connecting to 25,000 services through 6.5 million terminals. It was a genuine technological achievement and, by any near-term measure, a policy success. You could book train tickets, check the news, and — this being France — access the famous messageries roses.
But Minitel’s very success became its trap. The system’s popularity delayed France’s transition to the open internet by roughly a decade. The French government’s allegiance to the domestically developed Minitel impeded the full, widespread adoption of the internet in France. A planned, centralized, government-administered network worked beautifully within its original parameters — and precisely because it worked, it foreclosed the disorderly innovation that produced something vastly more powerful. France had the discipline to plan; it lacked the chaos to discover.
The Paradox at the Core
The standard telling of internet history treats ARPANET as a triumph of visionary government planning — public investment yielding civilization-defining returns. That narrative is not entirely wrong, but it obscures something more interesting. The government agencies that funded early networking research also maintained a telecommunications monopoly that actively impeded its deployment. AT&T, a regulated monopoly over telecommunications, represented a gigantic roadblock to the emergence of data communications — its management believed data networking was irrelevant and focused on protecting its voice telephone network. Until the 1968 Carterfone decision, you couldn’t legally attach an unapproved device to a phone line. One arm of the state funded the research; another arm protected the monopoly that delayed its adoption.
Meanwhile, outside the research labs, Bulletin Board Systems appeared around 1978 as a grassroots phenomenon — individual computers run by hobbyists in their homes, connected via standard telephone lines. They had no government funding, no grand vision, and no connection to ARPANET. They had users who wanted to communicate. The internet ultimately spread not because anyone mandated a national network, but because TCP/IP offered the simplest way to bridge incompatible systems — military networks, academic networks, hobbyist BBSes, and commercial walled gardens all running on different architectures. The protocol won by solving a coordination problem, not by executing a plan.
The Investment Parallel
For readers who evaluate technologies and allocate capital for a living, this pattern — government bets big, market emerges sideways — should be deeply familiar. NSF’s decision to prohibit commercial traffic on the NSFNET backbone had the predictable result of stimulating private, competitive long-haul networks like PSI and UUNET. The privatization of the internet backbone happened not because the government planned it, but because commercial demand outgrew the government’s willingness to carry it.
The uncomfortable lesson is that the most consequential technology platform of the late twentieth century emerged from the productive friction between competing interests — researchers who wanted to share data, bureaucrats who protected phone monopolies, hobbyists who wanted to talk to each other, and entrepreneurs who sensed commercial potential in connecting them all. The Synfuels Corporation, the Superconducting Super Collider, the OSI protocol stack, the Fifth Generation AI program, and France’s Minitel all received more deliberate government support than TCP/IP. All of them either failed outright or became dead ends. The technology that actually changed the world was the one the government funded almost by accident and then alternately nurtured and obstructed.
This should give pause to anyone — regulator, investor, or commentator — who believes that the trajectory of AI can be steered by committee. As we have argued previously, asking today’s AI leaders to design regulations for tomorrow’s competitors is roughly equivalent to asking IBM, GE, and AOL in 1993 to write the rules for the internet. The most important AI applications of 2035 are probably being built right now by people nobody has heard of, solving problems nobody has articulated, using methods that today’s regulatory frameworks cannot anticipate. The question is not whether government should fund research — it should — or whether safeguards matter — they do. The question is whether we have the institutional humility to recognize that the most consequential outcomes will emerge from forces no single actor controls, and to design policies that leave room for that emergence rather than foreclose it.
Bob Taylor just wanted to stop switching chairs. The rest followed.