Saturday 10 October 2026
The Track and the Train
A train capable of 260 km/h1 runs on a railway where its average speed is closer to 110. The train is new. The problem is not.
Yesterday, the computer scientist Arvind Narayanan told journalist Ezra Klein, on his New York Times podcast, that artificial intelligence is a »normal« technology:2 powerful, but something that will change society slowly, because the world around it changes slowly. His example was a railway. Amtrak’s new Acela trains, built by the French manufacturer Alstom, can run at about 260 km/h. But the Northeast Corridor, which runs some 735 km from Washington via New York to Boston, was laid out when a fast train meant 80 to 100 km/h, and its curves are built for that reality. The average speed is still about 110. It took fourteen attempts at computer modelling before the new trains were allowed to begin testing on the line. The new had to be shown to fit the old.3 The train got faster; the journey didn’t. AI, Narayanan said, is mostly the train rather than the track.
»the gap between maximum and actual performance exposes a familiar confusion: mistaking the capability of a single component for the capability of the entire system«
That doesn’t mean the trains were a bad investment. They carry about a quarter more passengers,3 which matters even if speed doesn’t transform the journey. But the gap between maximum and actual performance exposes a familiar confusion: mistaking the capability of a single component for the capability of the entire system.
The historian of technology Thomas Hughes called a part that has fallen behind the rest of a system a »reverse salient«.4 Components advance at different rates, and the one that falls behind draws the inventors’ attention. The railway seems a textbook case. But if the bottleneck is so visible, why don’t we put all our effort into removing it? A train can be bought, tested, delivered, and photographed at a launch. A railway corridor is a tangle of land, bridges, curves, signals, operating rules, and budgets. Improving it means coordinating institutions for decades. The train is a product. The track is a political settlement.
The sums show the difference. The new trains cost about $1.6 billion; bringing the corridor up to what they can do was estimated in 2024 at more than $100 billion.3 The track is not being ignored, only worked on to a different clock. Under the Hudson, a $16 billion tunnel, the first new rail tunnel beneath the river in more than a century, launched its boring at the end of September. It is due to open in 2035.5 The existing tunnels opened in 1910 and were flooded by Hurricane Sandy in 2012. In February, when frozen federal funding ran out, work stopped for weeks.6 The trains took nine years from order to passengers. The track takes decades.
Stewart Brand, a maverick environmentalist, entrepreneur, and writer, offers another way to see this in How Buildings Learn. Buildings, he argues, are made of layers that change at different speeds: the site endures, the structure lasts for decades, while services and fittings are replaced again and again. When the layers are tied tightly together, the slow ones block the quick.7 A railway is a similar stack of time. Trains are replaced every few decades (the first Acela fleet has run for more than twenty-five years), while some of the bridges and tunnels they pass through are more than a century old.
The economic historian Paul David made a related argument about electrification. Electric motors didn’t immediately produce the productivity gains their technical advantages seemed to promise. Factories had to be reorganised around them. The gains came from changing the system around the component.8 Nor does the work end there: bottlenecks move. Clear one and another becomes decisive. This is the expensive part of progress. It’s easier to buy faster trains than to straighten a curve through a densely settled corridor. It’s easier to install new software than to change a planning process, a supply chain, or a division of responsibility. The new part works; the world around it doesn’t change.
The people who sell the train and the people who look after the track want different things. A manufacturer delivers a finished object against a contract. A railway authority has to pay costs now, with disruption, for benefits that come years later. The first can point to what it has delivered. The second has to explain what must be moved or torn up before anything gets better.
Acela trains lean into curves, so they can take bends faster without every bend being straightened. It’s ingenious, and a sensible adaptation to what the past left behind. But the tilting mechanism leaves the larger question intact. What, exactly, has been improved? The train can do more than the railway permits. Both the achievement and the limit are real. Progress is not always held back by a lack of ingenuity. Sometimes ingenuity is concentrated on the part of the system that can change, while the part that sets the limit remains untouched.
The train leans, politely, into a century-old curve.
1 I’m using the metric system here as I live in Europe. If you prefer imperial, 735 km is approximately 457 miles, 80 to 100 km/h is 50 to 60 mph, and 260 km/h translates to around 160 mph; 110 km/h is roughly 70 mph.
2 Ezra Klein & Arvind Narayanan (2026) »Intelligence Isn’t Power«. The Ezra Klein Show, The New York Times, 9 October 2026. https://www.nytimes.com/2026/10/09/opinion/ezra-klein-podcast-arvind-narayanan.html
3 Mark Walker (2024) »After years of delays, Amtrak moves toward faster trains in the Northeast«. 13 January 2024 (republished by The Boston Globe). https://www.bostonglobe.com/2024/01/13/nation/after-years-delays-amtrak-moves-toward-faster-trains-northeast/
4 Thomas P. Hughes (1983) Networks of Power: Electrification in Western Society, 1880–1930. Baltimore: Johns Hopkins University Press.
5 »Hudson Tunnel Project begins major boring milestone«. NJBIZ, September 2026. https://njbiz.com/hudson-tunnel-project-tunnel-boring/
6 Bryan Gottlieb (2026) »Gateway Hudson Tunnel Work Resumes Even as Funding Uncertainty Mounts«. Engineering News-Record, 10 March 2026. https://www.enr.com/articles/62642-gateway-hudson-tunnel-work-resumes-even-as-funding-uncertainty-mounts
7 Stewart Brand (1994) How Buildings Learn: What Happens After They’re Built. New York: Viking. Chapter »Shearing Layers«.
8 Paul A. David (1990) »The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox«. American Economic Review 80(2), pp. 355–361.