The Case Against Building New Roads

Over the next few decades we’re going to spend hundreds of billions widening highways and laying new transit rail — most of it for a world that won’t exist by the time it’s paid off. Full autonomy isn’t the trip you take today with an empty driver’s seat. It’s a different transportation system: trips that take a fraction of the time, cities that don’t need so much parking, highways that move several times the cars they carry now. And it’s coming whether or not we pour the concrete.

I think about this every time I’m stuck on Highway 401 in Toronto — one of the widest and busiest highways on the continent, 16+ lanes in places, and at rush hour it still crawls below 30 km/h. The fix on offer is always more concrete: more lanes, more transit. That’s mostly wasted money, because the real fix isn’t concrete; it’s better software and autonomous vehicles. The self-driving technology is nearly perfected. What isn’t finished, and what actually stands between us and that faster world, is the politics and standardization of getting every car onto one coordinated system.

And that has a price tag today. If the destination is a coordinated, software-run road network twenty-five years out, then much of what we’re about to build is for a world that won’t exist by the time it’s paid off. What we build today should serve the world that’s coming, not the one we’re in now.

1. What a world of autonomous vehicles actually looks like

Picture an intersection with no traffic lights. Not "the lights are off" — no lights at all, because the cars don’t need them. The only lights left are for pedestrians. You walk up, cars whizzing by with hardly any gaps between them, and you press the button for the direction you want. It immediately tells you how long it will be before you can cross. Then a crosswalk lights up — maybe one direction, maybe all four at once. The cars don’t screech to a halt for you. There just aren’t any cars there. They knew the crossing was coming, so they either routed around the intersection entirely or hung back a block and never accelerated toward it. The moment the crossing clears, they flow through — fast and tight, one direction streaming through, then the perpendicular direction, in coordinated bursts. Nobody ever comes to a stop.

Inside the car, three things are different. It’s much faster: highway speeds where you used to crawl, with no time lost at lights. The clearances are tight enough to be unnerving at first: cars packed closer together and moving faster than any human driver would tolerate. And you’re not driving — you’re working, reading, on a call, talking to whoever’s riding with you, asleep. That last one is worth a real dollar figure: plenty of people spend two hours or more a day behind the wheel, and handing that time back is genuine economic value. It also changes where people can afford to live: when a long commute is time you get back instead of time you lose, and the commute is shorter in the first place, there’s likely a renewed trend toward cheaper land in far suburbs and smaller towns.

And you probably don’t own the car. In a city where every car is part of the fleet, the nearest empty one is never far away; there’s a robotaxi maybe fifty feet from you, and it comes in seconds. You get out and it drives off to pick up someone else instead of sitting in a parking spot for nine hours. When you want to go home, a different car comes.

That’s the future.

2. The solved part and the unsolved part

Two very different things inside the phrase "autonomous vehicles":

Individual self-driving is one car perceiving the world and driving like an excellent, never-distracted human. It sees the other cars and reacts to them. This is the part the phrase usually points to, and it’s the one that’s basically solved — Waymo has it working, and it’s getting better. But on its own it delivers only modest gains: fewer accidents, better attention, no one sitting through a green light because they’re looking at their phone. Genuine, but incremental. A better driver in every car is still a car stuck in the same traffic; the jams don’t truly clear until the cars start coordinating.

Networked coordination is the other thing entirely, and it’s where the real multiples live. Now every car isn’t just seeing the others — it’s sending its position, its desired destination, and a processed version of what its cameras see, up to a central system. The system does the rest. It knows where every car is and where each one wants to go, and dynamically routes the flow, spreading demand across roads to keep any one of them from jamming. When a car needs to change lanes, it doesn’t slow down and wait for a gap — the lane beside it simply opens one, the car slides over, and everyone keeps their speed. There’s no merge shockwave: the stop-and-go ripple never even starts. Every lane runs at whatever speed it needs to. No one ever has to stop. This is what empties the intersection from the opening paragraph: the cars hang back or route away because the system already knows the crossing is coming and has planned around it.

This is not a central computer steering your car over the internet — that fails the instant your connection hiccups, and no one should accept brakes that depend on a data center. The design is a split. Each car drives itself, in real time, on its own sensors and onboard computer: collision avoidance, lane-keeping, the millisecond loop that keeps you alive. On its own it’s already a fully capable driver that could run tight gaps at high speed with no network at all, and the entire safety-critical job lives right here, in the car.

The networked layer sits on top, and its real job is routing — but routing far more precise than anything you’d recognize as Google Maps. It doesn’t just pick roads; it hands each car a position as a function of time: be here, at this instant, going this speed, so the whole flow interlocks like gears. The car treats that schedule as a strong guideline and follows it whenever it safely can, but its own driving always has priority. If a gap tightens up, the car backs off the plan, and the plan re-solves around it within seconds. It’s re-solving constantly anyway as new riders are entering destinations and fleet cars are peeling off to go fetch them, so the whole position-versus-time schedule is recomputed across the network continuously. That takes serious compute: a permanent, centralized data-center load. The same layer quietly adds safety on top, too. Not the life-or-death loop, which stays in the car, but a real backstop: other cars vouching for where you are when your GPS drops or a camera fogs, plus warnings about hazards too far ahead for your own sensors to see. None of this has to happen in a millisecond, so a round trip up to the network and back is fine.

Now the numbers that turn the vision into a multiple.

Speed. A highway today moves at around 100 km/h, give or take depending on where you are. A coordinated highway runs at around 170 km/h (about 105 mph), and I picked that number deliberately. It’s roughly what the fast lane on a free-flowing German Autobahn already does every day, driven by humans, and coordinated cars do it with no reaction lag at all. I don’t push much past it because the returns diminish fast: each extra bit of speed shaves less and less time off the trip, while both costs that matter keep climbing. A crash gets more violent, since its energy rises with the square of speed, and the car’s energy efficiency drops as speeds increase — already roughly double the energy per kilometer at 170 km/h as at 100 km/h. Past a point you’re paying a lot more, in fuel and in risk, to save a little more time on a trip that’s already fast. So 170 km/h is the sweet spot, and you could probably push it to 180 km/h.

Throughput. Lane capacity depends on two things: how tightly the cars are packed, and how fast they’re moving. Sitting in rush hour, you’re already in tight gaps, bumper to bumper, but you’ve lost the speed. A jammed highway packs tight and crawls below 30 km/h at rush hour; the reason so few cars get through per minute is simply that they’re barely moving. Coordinated cars hold gaps nearly as tight but restore the speed, packed almost as close, at 170 km/h instead of 30 km/h. The raw speed ratio alone is nearly 6:1; give back a little for the slightly larger gap you need at speed, and the throughput multiple probably lands around 4×. I’ll quote 4× as the conservative number and bank the rest as headroom. That’s the unlock, and it’s both levers at once, not speed alone: humans can’t safely hold a tight gap at high speed because they need room to react; coordinated cars don’t, because they aren’t reacting — the car behind already knows what the car ahead will do before it does it.

The 401 in Toronto runs 16 lanes wide in places. Run those 16 lanes coordinated, at 4× the throughput each, and they move what would take 64 of today’s gridlocked lanes. We don’t have a shortage of road. We have a shortage of coordination.

today16 lanes · gridlocked, under 30 km/h

coordinatedthroughput of 64 of today's lanes · 170 km/h

That "4× per lane" is a gridlock number. Throughput — cars per lane per hour — only matters where a road is actually full. On a road that’s already free-flowing it’s irrelevant: the road isn’t the bottleneck, so whatever extra it could carry changes nothing anyone feels. Throughput is the supply argument, and supply only matters where it’s scarce. On the open road, then, forget throughput; the only thing autonomy changes there is speed, and only modestly: 100 km/h to 170 km/h. The dramatic speed story is gridlock — 30 km/h to 170 km/h, where a trip mostly spent crawling turns an hour into fifteen minutes. Speed is the half you actually feel.

3. The gains come gradually, then all at once

The payoff builds gradually as more cars go autonomous, then takes one final jump when the busiest roads are finally made autonomous-only — and most of it comes from coordination, not from the cars simply driving themselves.

The mixed era — now until the switch. This is the long stretch we’re in, and stay in for decades: autonomous cars sharing the road with human-driven ones. It starts where we already are: supervised driver-assistance (Tesla’s FSD, GM’s Super Cruise, Ford’s BlueCruise) legal in all fifty states, though the human is still the legal driver, alongside geofenced robotaxis you can’t buy (Waymo already runs paid service in around ten cities). From there it just deepens — cars you own but never touch the wheel of, more people skipping ownership for robotaxis. The gains rise the whole way, but gently: fewer crashes and better attention from the cars themselves, and, quietly more important, better routing for everyone, since the more cars that drive themselves and feed a shared map, the more of the fleet the system can see and steer (Google Maps does a thin version today). Fewer bad drivers, fewer accidents, smoother flow, traffic eased a little, all well before any tight-gap coordination.

There’s a hard ceiling, though, and it’s the whole reason this stays gradual: while humans share the road, the gaps can’t close. An autonomous car mixed in with human drivers still has to leave human-sized gaps, because the cars around it are unpredictable and packing tight behind one is unsafe, robot driver or not. The tight gaps, the high speed, the 4× throughput — none of it is reachable until every car in the flow is autonomous: same rules, mutually predictable, each sharing its next move with the system before it makes it. That’s the real reason coordination beats raw self-driving. It isn’t that the cars drive better; it’s that they become legible to each other.

The jump — autonomous-only roads. Once a stretch of highway is autonomous-only, you get the whole thing at once: the tight gaps, the 170 km/h, far fewer accidents, the lanes that suddenly carry 4× the cars. This is also where the routing changes character: up to now the system has been suggesting which roads to take, but once every car on a stretch is autonomous it stops suggesting and starts assigning each car an exact position-versus-time, be in this space at this instant, because now every car can actually obey it. This is the contentious step, but it’s the path to the big gains, and you can’t unlock them while a single unpredictable human is in the flow. The jump from the mixed era to autonomous-only is the jump from "a better car" to "a different system."

A little, for a long time — then, the day the busiest roads switch to autonomous-only, a step change all at once: tight gaps, 170 km/h, 4× throughput.

4. It’s not a technology problem — it’s an agreement problem

The technology will be reliable well before it gets implemented and unlocks all the benefits. I’d put genuinely reliable full autonomy before 2040, and the switch — the day large regions go autonomous-only — at 2050 or later. The gap between those two dates isn’t engineering. It’s politics, standards, and coordination. Teaching a car to drive is the hard engineering problem; getting an entire civilization to agree on how they’ll all drive together is a different kind of hard — slower, more political, and the part that actually sets the timeline.

Concretely, three things have to be standardized:

Sensing — what counts as "good enough" to be certified. I think it makes sense to mandate something deliberately conservative: LiDAR plus a defined camera set, even though Tesla bets it can do the whole thing with cameras alone and might be right. The logic is simple: it’s cheaper to over-spec hardware you might not fully use than to under-spec and have to retrofit. Once a car has that sensor set, knowing exactly where it is comes essentially for free — not from GPS, which is only accurate to a few meters, but from matching what its cameras see against a high-definition map, down to the centimeter. And that map builds itself as a byproduct of ordinary cars driving ordinary trips (roughly what Mobileye and Tesla already do), continuously updating, no survey fleet and no expensive per-car GPS. Once most cars carry standard cameras, well into the mixed era, every road maps itself for free.

Control — standardized drive-by-wire, with performance floors. The computer has to actuate steering, throttle, and brakes to a defined precision. Throttle is already electronic in most modern cars; steering and braking are the real barrier, since "steer-by-wire" and "brake-by-wire" are still rare. We’re not retrofitting anyone’s current car for this; that’s the whole reason the switch date sits far out, so the equipped cars are simply the new ones. And it’s not just can the car hit a target. Every car has to comfortably do what the system asks of it even when it’s old and worn, not just the day it left the lot: reach the set speed, brake inside a defined distance, and hold a turn at a given speed without exceeding what its tires and load can take. Performance floors, acceleration and turn-radius minimums, cornering limits at full load — the works. Note that gas vs. electric is irrelevant to all this.

A standardized, open computer interface. Mandate the same physical connector and board on every car, so a better computer can be dropped in later, and reserve the space and wiring for it now. Think of it as an open-source computer slot for cars: you wire the car for autonomy today, with whatever sensors the standard requires, and leave a port the brain plugs into. The brain comes later — a new module, or just a software update. This is the hinge the rest of the plan hangs on.

5. Wire it now, drop the software in later

This is the move that makes the whole transition politically possible. You split the mandate in two: require the hardware now, and let the self-driving software arrive years later, on its own schedule.

The proposal: set a far-future date for the switch — the day the busy roads go autonomous-only — and then, well ahead of it, mandate that every new car sold starting maybe five to seven years from now carries the full sensor suite and the standard interface, even if the self-driving computer isn’t ready yet. A "dummy" computer is fine for now; the point is that the eyes and the wiring are in place. When the software arrives, you drop in a real brain or push an update, and the car you already own becomes autonomous. No one has to buy a new car for the transition.

The reason the date has to be far out is a fairness argument that runs in two directions, and it’s what sets the whole timeline:

You can’t force people to junk working cars. So the switch date has to be far enough out that the cars on the road today have aged out naturally before manual driving ends. If you buy a car today and the switch is twenty-five years out, you don’t care that it isn’t autonomy-equipped, because you were never keeping it twenty-five years anyway.

You can’t disadvantage the carmakers who don’t have the hardware yet. Require every new car to be equipped with drive-by-wire a year from now and you’d hand an advantage to whoever already designed it in and punish everyone else. Give them a real runway instead: a few years to get the sensors, a standard port, and drive-by-wire (the real work) into their new models. By the time buyers care about autonomy, every maker has complied. And you’re only asking for hardware, not software — not "solve self-driving," just "build it to the standard."

So the sequence is: define the protocol now → all new cars carry the hardware within five to seven years → and finally, once the fleet is ready and the software is proven, the switch flips and manual driving ends on the roads where coordination pays off. Somewhere along that line, in parallel, full autonomy for individual vehicles gets legalized (largely here already).

That same principle, build today for the world that’s actually coming, is exactly what should change how we spend money right now.

6. Think twice about building new roads

If you believe the coordinated end-state is coming in twenty-five, maybe thirty years, then you have to underwrite today’s infrastructure spending against that future — and once you do, a lot of it stops making sense.

Roads — think hard before building new ones. Throughput per lane depends on the gaps between cars and their speed. Coordinated cars keep the gaps gridlock-tight but restore full speed, so each lane carries on the order of 4× the traffic it does now, and the roads we already have can cover the demand we’ll get. That’s 4× the supply per lane, with every car moving at full speed instead of crawling, not the usual tradeoff where more cars means slower going. And the usual objection, that new capacity just induces new demand, doesn’t undo it, because it won’t quadruple traffic. Now run the clock on a new road: years to approve, more to acquire the land, more to build. And the building itself makes traffic worse for years, so the road’s first stretch of life is just paying back the congestion it caused before it nets out positive at all. Then add the opportunity cost of the land — a network-coordinated future might want it back, and people may well want to tear lanes out of today’s mega-highways by 2050. Stack all that up and a new road approved today is capturing a small slice of value. None of this means never touch a road: maintaining what we have is necessary, that’s the whole reason the title says new roads, and there will be real exceptions. It means every new-capacity decision should be underwritten against the world that’s coming, and most won’t survive it. Fix congestion with software, not concrete.

Light rail and streetcars — don’t build them. A streetcar is a bus on rails. What a dedicated rail line actually buys you over a bus is a lane free of traffic — and you get most of that by painting a dedicated bus lane, at a tiny fraction of the cost. Under coordinated autonomy even that goes away: a stopped autonomous bus just tells the other cars it’s stopping and they route around it, so a bus in a shared lane causes no slowdown worth paying billions to avoid. And it really is billions — a mile of new US light rail runs anywhere from tens of millions to over $200 million depending on how much of it is tunneled. Cities are pouring those billions into rail that a coordinated road network will make redundant. Same logic as the roads, just a much bigger check. The exception is the dense downtown core, where a subway still makes sense. The streets above just can’t move the masses of people that subways can. Even at higher speeds, they physically aren’t wide enough, and so many crossings have to be held open for pedestrians that cars have fewer paths available.

Robotaxis — far fewer cars, far less parking. Your car sits parked about 95% of the time; the average car is driven roughly an hour a day. The only reason you own it is that you can’t lend it out while you’re not using it — if you don’t own it, something has to bring it to you, and paying a human just to ferry an empty car to your door makes no sense. Remove the driver and that logic collapses: a fleet car comes to you, drops you off, and goes to the next person instead of sitting in a lot. So you need far fewer total cars (still enough to cover rush hour, but far fewer than one per person) and far less parking. On cost, the driver is something like 50–80% of a rideshare fare, so robotaxis should eventually be meaningfully cheaper than Uber.

Freight — the biggest near-term prize. Not because the technology is further along, since driverless taxis already carry passengers, but because highways are the easy case (no pedestrians, no intersections, structured and mapped), and a truck that drives itself attacks one of the largest cost lines in logistics, the driver, while running around the clock without rest breaks. What this means for logistics companies: Once the driving is a commodity anyone can buy, the moat is the physical network: the warehouses, terminals, and distribution footprint that’s expensive and slow to replicate.

Power — make better use of what we already have. Where does the electricity for all these EVs (and all the AI) come from? Part of the answer is to stop wasting what we already generate. Two loads largely disappear in a coordinated network. The first is traffic signal lights: the US runs more than four million of them, drawing about three terawatt-hours a year, and a coordinated network has none. The second power load is lighting. Autonomous vehicles don’t need the road lit the way human eyes do. Their LiDAR sees in the dark on its own, and their cameras have the car’s headlights, which can run on high beam if needed since the only other eyes on the road are cameras and high beams don’t dazzle them. So highway and freeway lighting can switch off (residential streets stay lit for people on foot). All US street and highway lighting runs to one to two percent of national electricity, tens of terawatt-hours. But most of that stays lit for residential streets, so what actually turns off is the highway slice plus the traffic signals, totaling in the ballpark ten terawatt-hours. An EV uses roughly 3,500 to 4,000 kWh a year, so that’s enough to power a few million of them — a real slice of today’s EV fleet, but a rounding error against the quarter-billion-plus cars on US roads. Part of the "we need more power" problem is really a "use what we have better" problem, solved with software rather than new generation. New power generation capacity is the other part of the solution, and I’m not saying to stop power plants.

And the cars themselves get more efficient, especially in the city. A coordinated car almost never brakes; it lifts off the accelerator instead, because it already knows what every car around it is about to do, and brakes only for something genuinely unexpected, like a pedestrian stepping out. Far less energy thrown away as heat at every stop. (An EV’s regenerative braking claws back part of that, but only part; a gas car recovers none.) That shifts the conversation around gas-versus-electric: a gas car that rarely brakes and rarely idles burns far less than one crawling through stop-and-go traffic. (The highway runs the other way — cruising at 170 km/h burns more than at 100 km/h — but the brutal waste in driving is urban stop-and-go, and that’s exactly the part coordination erases.) I’m not saying we should forever rely on gas, but it genuinely shifts the payoff math of switching.

Recycling — the one-time teardown. Push the clock out thirty years and there’s a one-time industrial event waiting: tearing down all the now-useless poles. Street lights, highway lights, traffic signals — the steel poles holding up infrastructure a coordinated network has retired. That’s on the order of 5 to 10 million tons of metal, a rough guess: figure tens of millions of poles at a few hundred kilograms of steel each. It’s overwhelmingly steel with a meaningful slice of aluminum (the concrete and fiberglass poles are low-value, so set them aside). A real recycling industry of its own, spread over years.

Every one of these is the same move. Underwrite the spend against the world that’s coming. The road, the passenger rail line, the parking garage all look like much less favorable investments.

7. Politics, not arithmetic

A lot of infrastructure investment doesn’t pencil out even without autonomy. Plenty of light rail and new highway capacity already fails an honest cost-benefit test — the billions never come close to the time they save, especially once you count the years of construction that make traffic worse before they make it better. Autonomy makes the payoff math even worse, but it’s not the only thing at play.

So why do these projects get built? Because the decision is political, not arithmetic. Big builds mean construction jobs, construction is a large and visible voting bloc, and "we’re not building the road" is a hard sentence for a politician to say. Markets are the right default. But today’s markets misprice a lot of genuine value, and what’s best for society as a whole isn’t what gets built. Paying people to build things we won’t need is barely better than paying them for nothing. The better move is to point the same money and the same workers at things there’s real demand for — housing, schools, the buildings people actually want and will pay off further into the future. There’s no shortage of useful work to do. Roads we’re about to stop needing just aren’t on the list.

Autonomous vehicle infrastructure also unlocks a market structure that better prices roads. Once a central system knows exactly where every car is and which stretch of road it used, you can bill road upkeep to the people who actually wear the roads out, instead of through a blunt gas tax or out of general government revenue/debt. Every road becomes a toll road, but the tolls are tiny and exact, priced to the foot of asphalt you drove and how busy it was when you drove it. That isn’t a money grab; it’s just pricing a shared resource accurately, which is the thing a well-designed market is supposed to do and the thing today’s road funding mostly fails at. And it ripples outward: when the true cost of running goods over a given road shows up in the toll, it shows up in the final price too — freight, delivery, everything priced a little closer to what it actually takes. It’s a small example of a bigger shift: as tracking gets cheap, markets can price things more exactly, and more fairly, than they ever could before.

8. The hard questions

A few possible objections:

"This hands the government a kill switch over where I can go." Every car keeps a manual mode, exactly like any car today: it works offline, needs no destination, and nothing in the system can physically stop you from using it. The car itself can never trap you — you can always just drive off. Drive somewhere you’re not allowed and the police deal with you, the same way they would today for running a red light or heading the wrong way down a one-way street. At that level, autonomy leaves you no more constrained than you already are.

The second layer of the worry is the network. A system that knows where every vehicle is, continuously, is a movement-tracking capability no government has ever had, and you can’t fully opt out: pull the tracker from your own car and the cars around you still register you. That’s a real change in what a state could do. The question it raises is political, not technical: who holds that capability, and what keeps them from misusing it.

In a democracy, two things hold it in check. One is intent: I don’t believe that most governments are hunting for ways to control where their citizens drive. The other is political cost — cutting off millions of people’s movement, even briefly, is the kind of act that ends governments.

Authoritarian states are the hard case. A regime set on restricting someone’s movement can mostly do it already; what the network adds is scale — the means to do it to everyone at once, cheaply. The only real check left is the one that limits those regimes everywhere else: push people far enough and the backlash topples you.

"But people like driving." Some do, and they’re easy to accommodate. You can’t drive yourself on a packed coordinated highway — that’s exactly where forbidding manual control buys all the efficiency. But on open, scenic, low-traffic roads, manual driving stays perfectly legal. And because the system tracks every car, it knows when a road is clear and fast, so it can safely hand control back to you there. You lose the gridlock and keep the open road. On those roads, it’s a mixed-autonomous-human scenario: for people who don’t want to drive, their car can take over like a normal human driver, with larger safety gaps, like Waymo today.

And yes — this displaces a lot of work. Truckers, taxi and rideshare drivers, delivery drivers. That’s a real and serious consequence — but it’s a different essay.

9. Build for the world that’s coming

So picture it again. The intersection with no traffic lights, where the cars simply aren’t there until you’ve crossed. The corridor at 170 km/h where you used to crawl at 30 km/h. The hour of your day handed back. The car that comes in seconds — yours to own if you want, but something most people will rent because it’s cheaper.

There are two kinds of hard here. One is technical: teaching a car to see, decide, and weave through a coordinated swarm at 170 km/h. That’s the genuinely deep problem, the harder of the two by a wide margin, and it’s being solved anyway. The other is just agreement — a unified sensor spec, a driving protocol, and a switch date. The questions there are real and the process is slow, but in pure difficulty it’s nothing next to the tech. We’re cracking the hard problem and dragging our feet on the easy one. So start now, because the expensive mistake is pouring today’s concrete for a world that won’t exist by the time it sets.

From the sensors that should already be standard in the next car you buy, to the highway lane that shouldn’t get built, to the transit rail line that shouldn’t get funded — the move is the same. Wire it now. Drop the software in later. And underwrite infrastructure investment for the world that’s coming.