What a Dead LiDAR Bet on a Volvo Roof Teaches Us About Building Things That Don’t Break, and Why VAST Bet the Way It Did
A quick disclosure before I get into this: I’m a big fan of both companies in this story. I rent a Volvo whenever I’m traveling abroad. Back home, I drive a Tesla Model Y. So none of what follows is written from the outside looking in with an axe to grind against either brand. It’s written by someone who trusts both enough to hand them the wheel.
There’s a small plastic dome that used to sit on top of the Volvo EX90’s windshield, looking a little like a rally car’s roof scoop. It housed a Luminar LiDAR sensor: a laser rangefinder that was supposed to be Volvo’s answer to the hardest problem in the car industry, making a machine see well enough to be trusted with a human life.
That dome is empty now. Volvo has ended its relationship with Luminar entirely, pulled LiDAR from the EX90 and ES90 for the 2026 model year, and reportedly started compensating customers in markets like Norway (around €1,800 each, according to affected owners) for a feature that will simply never arrive. There’s no retrofit. And Luminar, the company that spent a decade convincing the industry that LiDAR was non-negotiable for safe autonomy, has filed for Chapter 11 bankruptcy protection.
That’s not a supply-chain footnote. It’s the end of a five-year architectural bet on how to close the gap between “the car can see pixels” and “the car can be trusted with your kids in the back seat.” And the way it ended has something to teach anyone building systems that need to be trusted, not just admired.
A decade of believing the eyes weren’t enough
Cameras give you a two-dimensional projection of a three-dimensional world; turning that into an exact distance, an exact velocity, an exact probability that something is about to step off the curb is an inference problem, and inference can be wrong. LiDAR sidesteps the inference: it fires pulses of laser light, times their return, and hands you a literal, measured point cloud of the world. No guessing, no neural network to trust, just physics telling you exactly how far away that shape is. For a long time, that felt like the only responsible way to build a life-safety system. Waymo built on LiDAR. Cruise built on LiDAR. Volvo, a brand whose identity is built on the word “safety,” partnered with Luminar in 2020, took an equity stake to make the relationship stick, and by 2022 was marketing the EX90 as “the safest Volvo ever built,” with LiDAR standard, not optional.
The bet wasn’t crazy. It was, for years, the responsible-sounding choice: more raw truth has to be safer than trusting an algorithm’s guess.
The unraveling
In October 2025, Luminar disclosed in an SEC filing that Volvo would make LiDAR optional rather than standard, and had pushed its decision on Luminar’s next-generation Halo sensor from 2027 to “at least 2029.” Luminar filed a damages claim, suspended further commitments, and, in a sign of how thin the margin for error had become, stopped paying its own sensor manufacturer, triggering a breach notice in return.
On November 14, 2025, Volvo terminated the relationship outright, citing, in its own words, Luminar’s “failure to meet its contractual obligations.” The fallout was fast: a 25% workforce cut, the CFO’s departure, loan defaults, an SEC investigation, an eviction lawsuit. On December 15, 2025, Luminar filed for Chapter 11, with CEO Paul Ricci citing a court-supervised sale as “the best path forward.” The bankruptcy court confirmed Luminar’s liquidation plan in April 2026.
Meanwhile, actual customers were left holding actual promises that would never be kept. In Norway, owners of LiDAR-equipped EX90s and ES90s report being compensated around €1,800 apiece for features that will simply never arrive. There’s no retrofit path. If your car shipped with the dome, you keep the dome: a permanently inert, non-functional roof bump on a car that no longer has a company behind the sensor inside it.
Two ways to answer the same hard question
Strip away the automotive specifics and every self-driving strategy is trying to answer one question: cameras give a flat projection of the world, and a car needs to reason in three dimensions with essentially zero tolerance for error. There were two ways to close that gap. Option one: bolt on more truth, a sensor that directly measures depth. This is the bet Volvo made, and for years it was treated as the responsible, adult-in-the-room choice. Option two: make the core smarter, push the whole problem into the intelligence layer until the system can derive depth and intent from cameras roughly the way a human does. This is Tesla’s bet, and for years it sounded like hubris.
Tesla removed radar from its cars in 2021 and ultrasonic sensors in 2022, consolidating its perception stack around cameras and neural networks, what it calls Tesla Vision. Critics, and there are still credible ones, point out that human vision is backed by hundreds of millions of years of evolutionary tuning that no current neural network fully replicates, and Tesla’s Full Self-Driving system has had a genuinely bumpy, heavily scrutinized public record, including regulatory investigations that remain unresolved.
Volvo’s retreat isn’t proof Tesla was right about everything. It’s proof that the patch has a failure mode the core-intelligence bet doesn’t.
Why bolt-ons fail in a way cores don’t
A bolt-on introduces a supplier relationship exactly at the point where failure is least tolerable, without giving you control over that supplier’s roadmap, cost curve, or survival. You’ve taken your hardest, highest-stakes problem, the exact place where you can least afford surprises, and outsourced part of the solution to an external party whose incentives, financial health, and execution are not fully yours to manage. Contrast that with the cost curve of “make the core smarter”: compute gets cheaper on a decades-long trend, training data compounds, model architectures improve as software updates rather than hardware swaps. A strategy anchored to the intelligence layer gets better every year almost for free. A strategy anchored to a specialized external sensor is only as good as one company’s ability to keep manufacturing it, and if that company stumbles, your roadmap stumbles with it, on a timeline you don’t control.
Luminar didn’t fail because LiDAR is bad technology; by most accounts its sensors did roughly what they were built to do. It failed because being a single-purpose hardware bolt-on for one dominant customer is an inherently fragile business position. That’s a pattern with a name in venture circles: the point solution that gets absorbed or outcompeted once someone solves the underlying problem at the foundation instead of at the seam.
Two responses to the same gap in capability: one you can find all over the technology industry, not just in cars.
The pattern is bigger than cars
At VAST, the founding architectural bet is structurally identical to Tesla’s, aimed at a different problem. For decades, the storage industry’s answer to “handle more data, faster, at scale” was to bolt something on: a caching appliance in front of a slow array, a separate block gateway, a separate file gateway, another system again for the vector database that AI workloads now demand, all stitched together with brittle data-movement jobs. Every one of those bolt-ons is, structurally, a Luminar: a vendor you don’t fully control, a firmware version to track, a supply chain dependency sitting in your most critical path.
The same logic applies to any storage product built to do just one protocol well, whether that’s a system purpose-built only for block, only for file, or only for object. A niche product tuned for exactly one access pattern is, in this sense, no different from a niche sensor tuned for exactly one job: it can be genuinely excellent at that one thing and still not survive the transition that’s underway. AI workloads don’t respect protocol boundaries the way older applications did. A single training or retrieval pipeline can need fast block access for checkpoints, file access for datasets, object access for cold archives, and vector access for search, sometimes all within the same hour. A product that only speaks one of those languages forces the same bolt-on pattern onto its customers that LiDAR forced onto Volvo: bridge it to the others with gateways and synchronization jobs, and hope the seams hold. Single-protocol specialists also inherit the same cost-curve problem a single-purpose sensor has. Their whole value proposition is being better at one narrow thing, not at everything, so they can’t ride the broader trend the way a unified platform can. As workloads get more multi-protocol by the year, the market for a product that only does one protocol well keeps shrinking, the same way the market for a car that needs an external laser rangefinder to see kept shrinking once the alternative got good enough. Being excellent at a narrow slice of the problem turns out not to be the same as being necessary.
Building the core instead of the patch
VAST’s DASE architecture (Disaggregated, Shared-Everything) is the storage industry’s version of the “make the core smarter” bet. Instead of stacking specialized hardware and software around the historical limits of spinning disk and rigid protocols, VAST rebuilt the foundation: one platform, one data path, that treats QLC flash as the center of the architecture rather than a cache bolted in front of something older. Compute and storage scale independently instead of being welded together in fixed appliance ratios, and every node sees the same global namespace instead of a private, siloed slice of it. No tiering logic to babysit. No separate system to keep in sync. No Luminar sitting in your data center’s most critical path, one bankruptcy filing away from becoming your outage.
The legacy stack accumulates seams. A unified data platform removes them at the foundation instead of patching around them.
This matters more right now than ever, because AI workloads are exceptionally good at finding every seam in an architecture and applying pressure until it cracks. Training jobs, retrieval-augmented pipelines, and inference at scale all need block, file, object, and vector data to move together coherently, without a human babysitting synchronization jobs at 2 a.m. A legacy, bolted-together stack can be made to work for AI, the way a car without LiDAR can be made to drive itself with enough workarounds. But “can be made to work” and “is the architecture you want underneath a mission-critical AI platform for the next decade” are different claims.
Why this is the bet worth making
When you build the shortcut, the bolt-on sensor, the caching appliance, you’re making a quiet admission: we don’t fully trust the core to be enough, so we’ll patch around it and hope the patch holds. That can work for a surprisingly long time, right up until it accumulates fragility exactly where you can least afford it, at the seams, under real load. When it finally breaks, it breaks the way Luminar broke: fast, publicly, with real customers left holding a non-functional bump on their roof and a refund check for a promise that will never be honored.
When you build the harder thing instead, the smarter model, the unified architecture, you’re betting that the fundamental problem is solvable at the foundation, and that solving it there is worth it precisely because everything downstream gets simpler and more resilient. That bet takes longer to pay off. But it doesn’t leave a bankrupt supplier and a stack of refund checks in its wake five years later.
Better to solve the hard problem once, in the foundation, than to be the company writing refund checks for a feature that was never coming.
The bond that actually matters
I write this because I’ve watched the alternative up close for long enough to know how the story ends, and Volvo just gave the entire industry a very public case study of the ending. When your data platform is the foundation underneath your AI initiatives, the thing you genuinely cannot afford is to discover, three years in, that a bolt-on you didn’t fully control has become your Luminar. You want to build on a foundation that got harder to build in the first place, specifically so it never becomes the seam that breaks under you later.
The dome on the roof of a 2025 Volvo EX90 will keep sitting there for years, empty now, doing nothing: a small monument to a promise that ran out of foundation to stand on. The lesson underneath it isn’t really about LiDAR, or cameras, or even cars. It’s about which kind of bet you’re willing to make when the problem in front of you is genuinely hard, and whether you’re willing to do the harder, slower work of solving it at the root.
Sources & Further Reading
Volvo ends relationship with Luminar, drops LiDAR for 2026 models (The Drive): https://www.thedrive.com/news/volvo-has-dropped-luminar-and-lidar-for-2026-models
Volvo abandons LiDAR technology after dropping supplier (The Truth About Cars): https://www.thetruthaboutcars.com/cars/news-blog/volvo-abandons-lidar-technology-after-dropping-supplier-45132346
Volvo drops LiDAR system amid supplier problems (Carbuzz): https://carbuzz.com/volvo-cuts-lidar-ex90/
Luminar says it stopped payments on Iris LiDAR products for Volvo (TipRanks / The Fly): https://www.tipranks.com/news/the-fly/luminar-says-stopped-payments-with-respect-to-its-iris-lidar-products-for-volvo-thefly
LiDAR-maker Luminar files for bankruptcy (Yahoo Finance): https://finance.yahoo.com/news/lidar-maker-luminar-files-bankruptcy-173455096.html
Volvo stops using LiDAR, supplier is pissed off (Autojosh): https://autojosh.com/volvo-stops-using-lidar-as-supplier-are-pissed-off/
Norwegian EX90/ES90 owners report ~€1,800 compensation for dropped LiDAR (@RoamingNorway on X, unverified by Volvo directly): https://x.com/RoamingNorway/status/2081954682696577184
Luminar Technologies initiates voluntary Chapter 11 proceedings (investor press release): https://investors.luminartech.com/news-events/press-releases/detail/110/luminar-technologies-inc-initiates-voluntary-chapter-11