To understand why the internet exists the way it does (why certain cities became hubs, why data centers cluster along rivers, why your traffic probably touches a building in northern Virginia), you have to go back to the mid-1800s. The internet didn't invent its own geography. It inherited it.

The Wire Came First

When Congress authorized the transcontinental telegraph in 1860 via the Pacific Telegraph Act, the deal was simple: build it fast, build it along the railroad right-of-way, and the government would pay. Western Union completed the line on October 24, 1861: 27,500 poles, 2,000 miles of iron wire, connecting Omaha to Carson City via Salt Lake City. The moment it closed, the Pony Express shut down. It had been operating for 18 months. The telegraph made it obsolete overnight.

The first Pacific Coast telegraphs were even more instructive about how demand drives infrastructure. Before any transcontinental ambitions, the earliest telegraph wires on the West Coast connected gold mining camps in the Sierra Nevada to the market cities of Sacramento and San Francisco. Miners needed to know the price of gold. Buyers needed to know what was coming down the mountain. That economic pressure, a commercial need to communicate faster than a horse could travel, is the engine behind every communications infrastructure build in American history, right up through the fiber and data center boom of the 2020s.

The railroads and telegraph were inseparable. Congress wrote the requirement directly into the 1862 and 1864 legislation authorizing the transcontinental railroad: railroad companies had to operate telegraph facilities alongside their tracks. Western Union handled the actual operation and paid the railroads for access. The result was a physical spine of communication that followed the cheapest possible path across the continent: the same grades and passes the trains used.

That pattern never broke. It just changed materials.

SPRINT Is Literally a Railroad Company

Here is one of the great underappreciated facts in telecommunications history: Sprint, one of the major carriers that built America's fiber backbone, started as a railroad company.

SPRINT stands for Southern Pacific Railroad Internal Networking Telecommunications. Southern Pacific had been operating microwave relay towers along its rail lines since at least the 1960s, again following the right-of-way logic. In 1972, they started selling surplus capacity on those towers to corporations as private long-distance lines. In 1978, they formalized this as Southern Pacific Communications, which became the original Sprint. When GTE bought it in 1983 for $1.2 billion, the brand went with it. By 1988, SP Telecom had a dedicated construction train laying fiber optic cable along railroad rights-of-way across the country.

The same story played out across the industry in the 1980s and 1990s. AT&T, MCI, Qwest: all of them paid railroads for easement rights and then buried fiber in the same dirt that had held telegraph poles a century earlier. The infrastructure logic was identical: railroads had already solved the hard problem of getting a continuous linear right-of-way across private land, mountain passes, and rivers. Telecom companies just paid for a lane in that same corridor.

This is why, if you look at a map of major fiber routes in the United States, they trace the old railroad lines almost exactly. The Donner Pass route in California, the Union Pacific corridor across Wyoming, the Santa Fe route through the Southwest: all of them carry fiber. Some of those conduits have been in the ground since the 1980s and are still in use today, lit with far more advanced lasers than anyone imagined when they buried them.

ARPANET and the Birth of Packet Switching

On October 29, 1969, at 10:30 PM Pacific time, a UCLA student named Charley Kline sat down at an SDS Sigma 7 computer and tried to log into a machine at Stanford Research Institute 350 miles away. He typed "l" and "o," and the SRI system crashed. An hour later, after Bill Duvall at SRI adjusted some parameters, Kline tried again. This time it worked.

The first message transmitted on ARPANET was "lo." Not by design. By accident.

ARPANET was funded by the Defense Advanced Research Projects Agency, which wanted a communications network that could survive a nuclear strike. The actual technical innovation was packet switching, the idea developed by Paul Baran at RAND and independently by Donald Davies in the UK. Instead of dedicating a circuit between two points (which is how telephone calls worked), you break data into packets, send each one independently across whatever path is available, and reassemble them at the destination. No central point of failure. Lose a node, the packets route around it.

By December 1969, there were four nodes: UCLA, SRI, UC Santa Barbara, and the University of Utah. By 1971, there were 15 nodes at universities and defense contractors. The rest of us weren't invited yet.

NSFNet: The Public Spine

The transition from ARPANET to something approaching a public internet happened in stages throughout the 1980s. The National Science Foundation started NSFNet in 1985 to connect university supercomputing centers, initially at 56 kilobits per second. By 1988 it was running at T1 speeds (1.5 Mbps). By 1991, NSF removed commercial restrictions, allowing the first commercial internet service providers to connect.

On April 30, 1995, NSF shut the backbone down. By that point, commercial carriers (MCI, Sprint, UUNET, ANS) had built parallel infrastructure that had already absorbed the traffic. Four Network Access Points were established in New York/New Jersey, Washington D.C., Chicago, and San Jose as the handoff points where these private networks exchanged traffic with each other. The government stepped back. The internet became a commercial product.

That transition, more than any other single moment, is what made the modern internet possible. And it's what made the real estate market in northern Virginia very, very valuable.

Why Ashburn, Virginia Became the Center of the Internet

The NAP in Washington D.C. was called MAE-East: Metropolitan Area Exchange East. It was established in 1992, initially in Vienna, Virginia, and later anchored in Ashburn. In the mid-1990s, MAE-East was reported to carry a majority of the world's internet traffic, because so much of the early commercial internet peered there.

AOL was in Loudoun County, Virginia, the same county as Ashburn, because Steve Case had started the company there and didn't move it. AOL at its peak was the dominant consumer internet company in America, so its network peered with everyone. That made Ashburn a mandatory stop for any carrier that wanted AOL traffic. Equinix built one of its first large data center campuses in Ashburn in 1998. More networks connected there. More data centers followed. It compounded.

A figure you will see repeated is that "70% of the world's internet traffic" passes through Loudoun County. That number is not supported by the data; TeleGeography, which actually measures interregional traffic, puts the share routed through the region closer to 20 percent, still enormous, but not the headline figure. Either way, Loudoun's "Data Center Alley" is the densest internet hub on the planet. It is not the cheapest land in America. It has traffic congestion and power constraints. Companies build there anyway because that's where the fiber terminations are, that's where the exchange points are, and leaving means paying for cross-connect fees to reach the networks you need. The economics of co-location create a gravity well. Once a place is a hub, it stays a hub.

Los Angeles and the Pacific Rim

Los Angeles became a major internet hub for a geographic reason that has nothing to do with the entertainment industry or the tech scene: it is the nearest major American city to Asia.

Submarine cables land where geography tells them to land, and then build-out radiates from there. Hermosa Beach, Malibu, and the LA area have become landing stations for systems connecting Southeast Asia to the U.S. mainland. This is why Los Angeles hosts major data centers and internet exchange points despite having some of the highest real estate costs in the country. The alternative is to pay for long-haul transport from the cable landing to somewhere cheaper, and at the latency requirements and traffic volumes involved, that cost adds up fast. You put your data center where the cable comes ashore, or close to it.

The same logic explains Singapore, Hong Kong, and Mumbai as Asian internet hubs. They are cable convergence points where geographic reality overrides economic preference.

The Dark Fiber Decade

Between roughly 1996 and 2001, the telecommunications industry went genuinely insane.

The Telecommunications Act of 1996 opened the local loop to competition and triggered a wave of investment that had no precedent in the industry's history. Companies poured more than $500 billion, most of it debt-financed, into laying fiber across the continent. Qwest's construction trains ran day and night along Burlington Northern Santa Fe tracks. Global Crossing laid transoceanic cables. WorldCom, before it collapsed in the largest accounting fraud in American history, built one of the largest private networks ever assembled.

The justification was a number: internet traffic was doubling every 90 days. That number was wrong. Network researchers found that backbone traffic was doubling roughly once a year. Still extraordinary growth, but about 3-4x slower than the figure driving investment decisions. The entire $500 billion buildout was predicated on a statistic that turned out to be a myth.

When the bubble burst in 2000-2001, somewhere between 85% and 95% of the fiber that had been laid sat unused: "dark fiber," meaning it was physically in the ground but never lit with laser traffic. Bandwidth prices collapsed by 90%. Dozens of carriers went bankrupt.

The irony is that the overbuilding created the foundation for everything that came after. When bandwidth is essentially free, you build applications that use a lot of it. YouTube, Netflix, video calling, cloud computing: none of those businesses work economically in a world where bandwidth is expensive. The ruinous overinvestment of the late 1990s turned out to be a gift to every internet company that came after. The fiber was in the ground. Someone was going to use it.

Google on the Columbia River: Cheap Power and Cold Air

In 2006, Google built its first data center outside Silicon Valley. They chose The Dalles, Oregon, a small city on the Columbia River at the eastern edge of the Columbia River Gorge, about 80 miles east of Portland.

The reasons have nothing to do with proximity to users or network topology. They are physical and economic.

In 1938, the federal government began building Bonneville Dam on the Columbia River. The Dalles Dam followed in 1957. The hydroelectric infrastructure that came from these projects produced cheap, abundant power for the Pacific Northwest, so cheap that it attracted heavy industrial users. Harvey Aluminum built a massive smelter at The Dalles in 1958 because smelting aluminum is extraordinarily electricity-intensive and the power there was cheap enough to make the economics work.

When Google arrived, they built on the former site of a Martin Marietta aluminum smelter. The same cheap hydroelectric power that had made aluminum smelting viable there made data center operation viable. Data centers are, fundamentally, industrial consumers of electricity. Google has invested over $1.8 billion in the campus through multiple expansions.

This pattern repeats across the industry. Iowa has cheap wind power and land. The Carolinas have tax incentives. Iceland runs data centers on geothermal power and uses ambient air for cooling. Phoenix has attracted data centers despite the heat because land and power are cheap and the desert climate's low humidity makes certain cooling approaches more efficient. Economics overrides climate if the incentives are large enough.

The GPU Accident That Changed Everything

Neural networks had been theorized since the 1940s and productively researched since the 1980s. They required enormous amounts of computation: essentially, large matrices of numbers being multiplied together repeatedly. The problem was that CPUs, optimized for sequential instruction execution, were not particularly good at this. Training a deep neural network on a CPU could take weeks or months.

GPUs had been designed for an entirely different purpose: rendering 3D graphics in video games. A game needs to compute the color of millions of pixels simultaneously, 60 or more times per second. That requires massive parallel arithmetic, specifically the kind of large matrix operations that neural network training also requires. The hardware was built for one application but happened to be exactly what another application needed.

Nvidia recognized this crossover potential and in 2007 released CUDA, a programming model that let developers write general-purpose software to run on Nvidia GPUs.

The breakthrough moment came in 2012 at the ImageNet competition. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto submitted AlexNet, trained on two consumer Nvidia GTX 580 GPUs. AlexNet won with an error rate of 15.3%, beating the runner-up by over 10 percentage points. The second-place system used traditional computer vision techniques that researchers had spent decades refining. AlexNet destroyed it with a few weeks of GPU training.

Nvidia's market capitalization, roughly $5 billion before AlexNet, grew to over $1 trillion by the early 2020s. Jensen Huang had cofounded the company in 1993 to make graphics cards for video games. He ended up building the essential hardware for artificial intelligence, not because he predicted it, but because the mathematical structure of the two problems happened to be the same.

The data center industry transformed again. The hyperscalers had built for CPU compute: standard racks, standard power densities, air cooling. GPU clusters for AI training require dramatically more power per rack, liquid cooling in many cases, and different networking architectures. The Columbia River data centers, the Ashburn campuses, the facilities everywhere, all of them are being retrofitted or replaced to handle the power densities that AI training demands.

The pattern that started with mining camps needing telegraph lines and railroads needing right-of-way has repeated in every generation: the application creates the demand, the demand creates the infrastructure, and the infrastructure reshapes the geography of where things happen and why.

Sources