⌕

$75,000 TO $200. AND THE CHEAP ONE DRAWS 2,160 LINES INSTEAD OF 64.

NOV 20, 2026 · SPLASH · TECH · THE COST CURVE
Velodyne and Hesai lidar units

IN 2007, ONE LIDAR UNIT COST $75,000. Today Hesai sells its ATX for around $200 — about 1/375th of the price.

The usual explanation is that China throws volume at it and sells at a loss. But the cheap one draws more lines than the expensive one did: 64 lasers in 2007, and RoboSense's Phoenix chip now produces 2,160 lines from a single chip.

Cheaper and denser at the same time is not what volume does. Something structural changed: the laser stopped being a part and became a chip.

WHAT THE THING ON THE ROOF ACTUALLY DOES

One laser is simple. It fires light and times how long that light takes to hit something and come back. In one nanosecond light travels 30 cm. A car 100 m ahead, there and back, is about 667 nanoseconds. That time is the distance.

One shot, one dot. Do it more than a million times a second and you get a 3D map — a point cloud.

But one laser spinning traces one circle — a single line. So the original stacked 64 lasers, each angled slightly differently, and spun the whole head 5 to 15 times a second. That's 64 rings on the ground. When the industry says “64 channel” or “64 line,” that's what it's counting: lasers.

THE CAN, WHICH WAS BUILT BY HAND

David Hall, an engineer who ran a speaker company, built it. Inside: four blocks of 16 lasers and two blocks of 32 receiving sensors.

Every laser has its own partner sensor. Every line takes a full set — laser, lens, receiving sensor, and a circuit to boost the signal. They put in 64 sets.

His 2007 patent says those pairs are aligned one by one, a third of a degree apart. Which means every unit comes out slightly different — open-source robot software still ships a calibration file with individual values for all 64 lasers.

By early 2008, only a few dozen had sold. Every one of them at $75,000.

THE GAP A TIRE COULD FIT THROUGH

The lines were about 0.4° apart. At 100 m, that's roughly 70 cm between one line and the next — over two feet. A whole tyre could slip through the blind spot between two laser rings.

And the only way to get more lines was to fit more lasers — each one hand-aligned.

It still won. At Velodyne's 2007 driverless-car race, that can sat on the roof of seven of the 11 finalists.

THEN THE PATENT WAR

2019: Velodyne took that patent and sued two Chinese companies, Hesai and RoboSense.

Hesai settled the next year and agreed to pay royalties until 2030.

Then Hesai answered with a lidar that looked nothing like the can. It turned the parts inside — the ones fitted by hand — into chips. Ouster, the American company that merged with Velodyne, now packs 128 lasers into an area the size of a grain of rice.

HOW A LASER BECOMES A CHIP

It starts with which way the light comes out.

Older lidars use edge-emitting lasers: shine sideways out of a wafer that has been diced into pieces. You can't even switch one on until it's cut — so you cut them one at a time, test them one at a time, fit them one at a time.

A VCSEL (vertical cavity surface emitting laser) shines straight up off the wafer surface. You can switch on thousands at once on the wafer and test them all before cutting. And because the light goes up, you can pack them tight like squares on a checkerboard — dozens or hundreds of lasers built into one chip from the start.

The catch: each VCSEL is weak, so for the longest range the sideways lasers still have an edge.

Hesai went the chip route too. It formed a chip team in 2017 and squeezed the circuits that run hundreds of laser channels into a few chips: first chip in 2020, fourth generation in 2024. With that fourth generation Hesai says it put all seven key components — lasers, receivers, driver circuits, through to signal processing — into its own chips. The ATX, the sub-$200 unit, is built on it.

THE HARDER HALF: CATCHING THE LIGHT

Sending is only half of it. The returns are the hard part.

The light can't be blasted as hard as you want — it has to stay safe shining into an eye, which is the Class 1 rating. And returned light falls off with the square of distance: twice as far is a quarter as bright, ten times as far is a hundredth. Far reflections are incredibly faint.

Older receivers used an avalanche photodiode, which amplifies in step with incoming light — so every line needed its own sensor plus its own amplifier circuit.

The replacement is the SPAD — a single photon avalanche diode. One photon sets off an avalanche of electrons and you get a click. The problem is it's so sensitive that a photon of sunlight makes it click too.

The trick: fire at the same spot over and over and stack up the time of every click, bin by bin. Sunlight shows up whenever. The reflection comes back at the same moment every time. Stack enough and one bin towers over the rest — and that bin's time is the distance. You're reading distance off a bar chart.

And that's why it scales: a click is on or off — digital — so it can be made with the same process as smartphone camera sensors. In 2021 Sony put about 100,000 10-micron SPAD pixels on a top layer with the timing circuits on a layer below and stacked them into one chip 6.25 mm corner to corner — smaller than a fingernail.

WHAT THAT DOES TO A FACTORY

Sending is a chip. Receiving is a chip. Now look at what that does to assembly.

RoboSense showed a line doing fully automated alignment at 8 seconds per unit. Hesai said back in 2023 it builds a lidar in 45 seconds.

The sequence is the whole story: merge parts into chips → part count drops → assembly gets simpler → simple assembly can be handed to machines → once machines stamp them out, the more you make, the cheaper they get.

Hesai shipped 1.62 million lidars in 2025 alone. In 2007 the number was a few dozen.

The moving parts shrank too. Velodyne spun the entire head, all 64 laser pairs with it. The ATX spins a single mirror inside. It's 10 cm square, 3 cm thick, 360 g — under a pound. A can that weighed over 10 kg now fits in your palm, which is why it can go behind a windshield or inside a headlight.

AND THAT'S WHY THE LINES EXPLODED

Pull out the 64 hand-fitted pairs and put two chips in the same spot: a VCSEL chip that sends and a SPAD chip that receives.

Now the number of lines isn't set by how many lasers you can fit. Each row of pixels on the receiving chip covers its own height — so a row of pixels becomes a line of the lidar. Pack the pixels tighter and the lines get denser. The line count starts riding the semiconductor curve.

RoboSense's Phoenix chip, unveiled in April 2026: 2,160 lines from a single chip, no stitching chips together — a 2,160 by 1,900 resolution the company says is denser than a 4-megapixel camera, with a 600 m detection range.

Hesai's Picasso chip the same month: also 2,160 lines from one chip. The ETX lidar built on it goes up to 4,320 lines. Both slated for mass production this year.

Do the arithmetic at 100 m: 64 lines gave you a 70 cm gap. Spread 2,160 lines evenly over 27° and they're just over 2 cm apart. A gap a tyre could fall through becomes one you could barely fit a finger into.

THE PART THE HEADLINES SKIP

Three caveats, because the cheap number and the dense number are not the same product.

1. The sub-$200 ATX tops out at 256 channels. It is not the 2,160-line chip.

2. No price has been announced for the 2,160-line chips. Cheap lidar and dense lidar are currently two different purchases.

3. Hesai's chief scientist, Sun Kai, has said 500 lines with 300 m of range is a reasonable benchmark — which suggests the line race is already past what the cars need, and the fight ahead is about proving safety in real numbers, not resolution.

AND TESLA

The take going around is that lidar costs $200 now, so Tesla clinging to cameras must be pride. Look at what Musk actually said in 2019: he called lidar expensive and unnecessary.

At a 375th of the price, the first half has crumbled. The second half he's still holding: people drive with two eyes and a brain, so a car should manage with cameras and AI. He said it again this year.

But here's the twist: in the first quarter of 2024, Luminar named its biggest lidar customer — and it was Tesla. Not for the cars it sells: reportedly mounted on test cars as a measuring stick to check whether the cameras' distance judgement is right.

And the cost of getting out isn't zero either. Volvo put lidar in, then took it out — dropping it starting with the 2026 EX90. Volvo ended the Luminar contract in November 2025, and Luminar filed for Chapter 11 in December 2025, with liquidation plans confirmed in April 2026.

SPLASH'S READ

The lesson isn't “China is cheap.” It's that a $75,000 assembly of hand-aligned parts can never follow a cost curve, and a chip can.

Look at what actually moved: 64 laser pairs, each fitted a third of a degree at a time, tested one at a time — that's a craft product. Replace it with a VCSEL array you can test thousands at a time on the wafer, and a SPAD sensor built on smartphone-camera process, and the price floor stops being a labour hour and becomes a slice of a wafer.

And the density was the tell. Volume explains a cheaper version of the same thing. It does not explain 2,160 lines where there were 64. That's a different kind of product, and it only exists because the thing got re-architected.

Which is the same shape as the AI infrastructure story and the mall: the price is downstream of the structure. Change what the thing is and the number takes care of itself.

Lidar didn't get cheap. It stopped being built.

AD SPACE · RESERVED
MORE ON THE BOARD ▸
💻 All Tech · CoreWeave and Scale AI
Sources & claims: Sources. Price history: Velodyne's HDL-64E, built by David Hall (who came from a speaker company), priced at $75,000 per unit in 2007-08 with only a few dozen sold by early 2008; its 2007 patent describing 64 laser pairs aligned one at a time about a third of a degree apart; four blocks of 16 lasers and two blocks of 32 receivers inside the head; 0.4° line spacing giving roughly 70 cm between lines at 100 m. The unit appeared on seven of the 11 finalists at the 2007 DARPA Urban Challenge. Hesai's ATX is priced from around $200 and is rated at up to 256 channels; Hesai shipped 1.62 million lidars in 2025 and said in 2023 that it was building one in 45 seconds; its chip programme began in 2017 with a first chip in 2020 and a fourth generation in 2024, consolidating seven key components into in-house chips. RoboSense's Phoenix SPAD-SoC was unveiled in April 2026 with 2,160 native receive channels and a 2,160 × 1,900 resolution, a 600 m detection range, and a stated production line cycle of 8 seconds per unit; Hesai's Picasso chip, announced the same month, also delivers 2,160 lines from one chip, with the ETX lidar going to 4,320 lines. The 2007 laser-to-chip transition, the edge-emitting vs VCSEL distinction, Class 1 eye-safety limits, inverse-square falloff, avalanche photodiodes, SPAD histograms and Sony's 2021 stacked SPAD sensor (about 100,000 10-micron pixels, 6.25 mm diagonal) are as described. Velodyne sued Hesai and RoboSense in 2019; Hesai settled in 2020 with royalties running to 2030; Ouster merged with Velodyne. Tesla: Musk's 2019 description of lidar as expensive and unnecessary, repeated this year; Luminar naming Tesla as its largest lidar customer in Q1 2024 for test vehicles; Volvo ending the Luminar contract in November 2025 and dropping lidar from the 2026 EX90, with Luminar filing Chapter 11 in December 2025 and confirming liquidation plans in April 2026. Hesai chief scientist Sun Kai's “500 lines with 300 m” benchmark is as quoted. Video source: Whatzat?, “Lidar Fell From $75,000 to $200: How 64 Lasers Became a Chip” (2zIiAVuAhrQ) — read in full, with every figure checked against the reporting named here. Related: CoreWeave and Scale AI.