German Carmakers Have Surrendered the Future

Where are the German self-driving cars?

This article is co-written with Andrew Miller, who writes the newsletter Changing Lanes (he’s excellent, check him out). Andrew covers innovative mobility, I cover European innovation policy and technology. We divided the work accordingly; he took the driving-automation material, I took the broader German industrial story. We’ve wanted to work together on something for a while, and this felt like the right subject.

In September 2025, Mercedes-Benz CEO Ola Källenius said on a podcast that his company was “at the front“ of automated driving. Listeners would be forgiven for finding the claim persuasive. After all, as he pointed out, Mercedes had 30 years of experience in developing and deploying driver-assistance systems. Not only that, Mercedes was the first manufacturer to certify an advanced driver-assistance system capable of handling all driving tasks under very specific conditions (what the SAE refers to as a Level 3 automated-driving system) in Germany. It was the first non-Chinese OEM to obtain an automated-driving permit in China as well. “Every new Mercedes has the technology and the onboard computer to drive autonomously,” Källenius said.

Four months later, in January 2026, Mercedes killed the system.

Mercedes confirmed that month that Drive Pilot would be dropped from the facelifted S-Class and EQS, the avowed reason being that Luminar, the LiDAR supplier Mercedes had depended on, had failed to meet its contractual requirements (Luminar subsequently filed for bankruptcy in December 2025). A few weeks later, BMW followed by scrapping its own “Personal Pilot” from the 7 Series facelift.

With no German manufacturer publicly selling a car with a Level 3 automated system (as per the SAE), German AVs seem, for all practical purposes, dead.

The choices that Mercedes has made are not unique, but part of a pattern, one that extends well beyond one model line or one company. Germany built the first automated cars. It holds every conceivable advantage, including engineering talent, manufacturing scale, regulatory influence, and established brands; and yet it is poised to license the automotive future from companies in California and Beijing, leaving the 716,000 workers in Germany’s automotive sector at the mercy of decisions taken elsewhere.

The reason for this surprising forfeit is that German OEMs lack the capability to train and scale an AI system that perceives an unpredictable environment and makes real-time driving decisions. This stack improves through more data, more compute, and more iteration cycles. German OEMs are increasingly absent from this stack as original developers.

Understanding why this is happening requires looking not at what Germany failed to do, but at what kind of industrial economy it has, and the different kind that the twenty-first century requires.

A Huge Lead, Squandered

Germany beat everyone else to driving automation. As long ago as 1986, Ernst Dickmanns at the Bundeswehr University Munich began outfitting Mercedes vehicles with cameras and custom processors to teach them to read the road. By 1994, his twin automated sedans were driving themselves on European highways at 130 km/h in normal traffic. A year later, a modified S-Class completed a Munich-to-Denmark run at speeds reaching 175 km/h, with the system in control for 95% of the journey. The Eureka Prometheus program that funded this work included Daimler-Benz and every other major European automaker.

Germany presided over the birth of automated driving, but left rearing the child to others. There was no commercial pull, no institutional equivalent of DARPA’s Grand Challenges to force the technology out of the lab and onto the road, and Europe’s capability dispersed.

Källenius blamed regulation, not capability, for the slow rollout: “If you develop a product you’re not allowed to sell, no market emerges.” But that can’t be the whole story. China and the USA permitted automated fleets years ago. Waymo, Tesla, Baidu Apollo Go, and WeRide already offer Level-4 automated rides in cities like Beijing and San Francisco. According to Källenius himself, Mercedes was the first non-Chinese manufacturer to get a permit operating AVs in China.

The reason Mercedes and other German manufacturers fail in yet another transformative technology is not regulation.

German manufacturers remain world-class at what one might call the ‘carrier layer’ of automated vehicles: platform construction, crash safety, powertrain integration, supplier management, regulatory certification, and manufacturing quality at scale. Mercedes was among the first manufacturers to sell a car that could parallel-park itself without the driver touching the steering wheel, a feature that struck buyers in the mid-2000s as either impossible or magic; its remote-parking feature is equally impressive. Mercedes’s automated valet parking system, approved commercially in a Stuttgart parking garage in 2022, demonstrated genuine Level-4 capability in a structured environment. Drive Pilot itself is a legitimate engineering achievement: 35+ sensors including LiDAR, redundant steering and braking, centimetre-accurate positioning, and the world’s first series-production Level 3 certification. Germany knows how to build safe, well-integrated vehicles and how to navigate the most demanding regulatory regimes on earth.

But as strong as German OEM capability is, their weakness, namely their inability to build an AI sophisticated enough to automate the driving task, more than offsets it.

That weakness may not be obvious at first glance, given the course Mercedes is taking. The firm is pursuing a bifurcated approach. In the US and Europe, its new MB.Drive Assist Pro system runs on Nvidia’s computing architecture and AI software stack—a Level 2++ system that works in city traffic, on highways, and in parking scenarios. In China, Mercedes has partnered with Momenta, a Chinese AI startup, for the same capability. BMW and Audi have made the same choice: BMW has also partnered with Momenta for China-tailored driver assistance on its Neue Klasse vehicles, and with Qualcomm for broader markets. Audi uses Momenta for its new EV line developed with SAIC, and Huawei’s Qiankun system for its legacy gasoline models.

These partnerships are impressive, but point to a common frailty. All three German luxury brands have licensed core driving intelligence from the same Chinese startup for the same market for the same reason: they cannot match what domestic Chinese competitors are shipping.

The trainable driving stack—the system that perceives, predicts, and plans—is the component most likely to generate compounding returns through data accumulation and algorithmic improvement. One caveat worth noting is that robotaxi economics have not yet been proven at scale, and German OEMs still earn substantial margins selling vehicles. But the evidence that we have is suggestive: Waymo, Baidu Apollo Go, and Tesla are each building fleets or datasets that improve their systems with every kilometre driven. The value of the driving stack compounds, but the value of the vehicle platform does not. A manufacturer that licenses the former while owning only the latter is positioning itself as a supplier to whoever controls the learning system. And not only do suppliers earn lower margins than platform owners, they have no leverage: Nvidia or Momenta can put their driving stack into a different manufacturer’s cars more easily than Mercedes can build their own driving stack. The dependency runs one way.

By licensing the autonomy stack rather than building it themselves, German OEMs are surrendering the future.

An Engineering Economy, Not a Platform Economy

The reason German OEMs do not build their own full-stack autonomy systems lies less in technical incompetence than in a structural mismatch between two types of economy. Engineering economies reward precision, pre-validation, exhaustive testing, and incremental improvement. Platform economies reward aggressive data collection, rapid iteration, high risk tolerance, and network effects. The German industrial system is optimized for the first type, and produces excellent component suppliers and integrators.

It does not produce AI-native platform developers.

Volkswagen’s experience makes the structural point most bluntly. Its in-house software unit, CARIAD, spent years attempting to build an integrated software stack and largely failed, prompting what the Financial Times described as a complete reset, a pivot from indigenous development to integrating partner technology. VW had also co-funded Argo AI, a full-stack autonomy venture with Ford; when Argo shut down in 2022, VW’s response was to partner more, not invest more.

The pattern across German OEMs is consistent: when internal development falls behind, the response is to lease capability rather than build it.

The contrast with American technology companies is instructive (let us stipulate that Tesla is one of these; it has much more in common with them than with Ford or GM). General Motors shut down Cruise, and Ford killed Argo AI; one might cite either retreat as evidence that full-stack autonomy is difficult, and no one has an edge. But those were OEM-backed moonshots. Waymo continues to expand its commercial driverless service in multiple US cities, while Tesla’s effort to build a truly Unsupervised Full Self-Driving system accumulates training data from millions of vehicles on the road every day. It’s true that the American OEMs retreated, but American technology companies did not. Germany, not having the latter, needs to rely on the former.

None of this would matter much if automated driving remained a niche feature for early adopters, but the trajectory of automotive technology suggests it won’t. Power steering, anti-lock brakes, and power windows each arrived first as a luxury add-on and became a baseline expectation within a generation; backup cameras and Bluetooth links to a driver’s mobile phone are following that path now. We have every reason to expect that driving automation will do the same. The consumer paying six figures for a Mercedes in 2030 will expect the car to drive itself, at least in some circumstances; the one paying low-five figures for a compact will expect it by 2040. At that point, every car Mercedes sells will carry a software licensing cost to Nvidia or Momenta; a rent extracted from every unit, in perpetuity, by whoever owns the learning stack.

German manufacturers, among all players, should know this the best, and should be fighting the hardest to avoid losing the next battle over a market-defining technology. They have found no answer to Chinese dominance in electric vehicles. Instead, they are lobbying to delay the combustion phase-out, advocating for a slower transition. They are fighting to preserve a capability that their customers will abandon, while ceding to their Chinese competitors the capability everyone will want. Given those circumstances, the fixation on regulation is understandable: German OEMs can’t build competitive EVs, and they can’t make consumers not want to buy EVs, but they can push back against EU regulations that promote EVs. Faced with a difficult problem, they are pulling on the only lever they can reach, and that move is shaping their responses to other problems.

That over-reliance on what they know best is hampering them in other ways beyond technology development. American and Chinese automated-driving leaders are building business models around fleet operations: Waymo runs robotaxis, Baidu’s Apollo Go has reached per-vehicle profitability in Wuhan, and Chinese firms are expanding into the Middle East and Southeast Asia. The logic is to accumulate driving data, improve the system, and monetize through transportation-as-a-service. Meanwhile, German OEMs remain focused on selling vehicles and earning margins per unit.

One might reply that specialization is rational. If Nvidia and Momenta are better at automated-driving AI, why shouldn’t Mercedes license driving ‘software’ from them, while focusing on what they and their workers are best at, namely building the ‘hardware’? Carmakers haven’t been completely vertically integrated since the early days of the industry, and insisting they should become so integrated now seems foolish.

The rejoinder is that automakers have indeed always outsourced certain elements of their vehicles… but never the basis of their brand value. Mercedes commands a price premium because its brand makes a claim about engineering: these cars are built to exacting tolerances by engineers who know how to build them better than anyone else. That claim will remain true of the physical platform, but it will not be true of the software that increasingly defines what the car does. A Mercedes whose driving intelligence comes from Santa Clara and Beijing is not making the same brand claim as the one that justified its price premium in the past. Certainly, the marginal improvements that will make a 2035 Mercedes more desirable than its rivals—meaning the ones consumers will actually pay for—are almost certainly software improvements, not mechanical ones.

When the dominant monetization model shifts from unit sales to fleet-based data services, being an excellent vehicle manufacturer positions you first as a supplier to the platform, not the platform owner. Ultimately, an automaker’s brand value rests on the claim that ‘we know how to build these things and we can iterate the next generation.’

Increasingly, Germany can only assert the first part.

More Than an Automotive Problem

Källenius’s September rhetoric pointed to regulation as the binding constraint: “What we need in Europe is a considered, thoughtful discussion to bring more of this technology to the road.” Germany’s 2021 Autonomous Driving Act does impose strict requirements: Level 4/5 vehicles are permitted only in approved operational areas, require a technical supervisor, and must undergo extensive certification. The US system, by contrast, allows relatively easy testing permits across multiple states. China has issued over 16,000 test licences and opened 32,000 kilometres of roads for automated-vehicle testing.

But the regulatory argument has a fatal flaw: Mercedes itself chose to retreat from Level 3 not because regulation prevented it, but because the commercial case didn’t work. And the system it is replacing Drive Pilot with—MB.Drive Assist Pro—will launch first in China, then in the US, and only later in Europe. If European regulation were the primary bottleneck, one would expect firms to be pushing hardest for regulatory reform. Instead, the entire European public discourse revolves around stopping the EU’s combustion-engine phaseout mandated after 2035.

Germany has no equivalent of Waymo, no full-stack automated-driving developer, and no AI platform champion in any domain. The German Mittelstand produces excellent components, middleware, and specialized tools (sensor companies, simulation providers, remote-driving solutions). It’s telling that Momenta has opened a research centre in Stuttgart. The Chinese firm is recruiting German engineering talent to work in Germany, but on Chinese autonomy systems. The Fraunhofer system produces world-class applied research, and German universities train outstanding engineers, yet the value chain consistently terminates before it reaches the platform layer.

And it isn’t just automotive that’s struggling. For example, Germany played a major role in developing the software foundations of modern robotics, particularly through research institutions such as the German Research Centre for Artificial Intelligence and the Fraunhofer Society. Since the early 2000s, these institutes have pioneered robot perception, machine vision, sensor fusion, and human-robot collaboration. Their work enabled robots to interpret visual data, map environments, and safely interact with humans, so today’s warehouses can be automated, factories inspected, and robots do logistics and manufacturing for us. Only, not really in Germany.

While Germany advanced much of the underlying research, companies like NVIDIA built the core AI and simulation platforms used to train and deploy robots, while firms like Boston Dynamics and Amazon developed large-scale robotics systems and deployed them in real-world settings. Germany remains strong in industrial robotics hardware and automation engineering, but the scalable software stacks, AI ecosystems, and platform economics that now shape the robotics industry have largely been driven by US and Chinese technology companies.

A similar platform, pioneered in Germany and later lost to its industry, connected manufacturing to cloud services. The promise of ‘Industrie 4.0’ was that connecting factory machines to cloud platforms would allow manufacturers to pool operational data across entire production networks, use AI to predict equipment failures before they happen, optimize processes in real time, and continuously improve output. Whoever controlled that data layer would shape how the world’s factories run. German incumbents had every conceivable advantage in smart manufacturing: the world’s best machine tool builders, deep automation expertise, and anchor firms such as Siemens and Bosch. They blew it.

Siemens built MindSphere as a platform play, but refused to commoditize its own hardware. MindSphere was always biased toward Siemens equipment, which is exactly what prevented it from becoming the horizontal standard. AWS, Azure, and Google Cloud had no such conflict; they don’t sell factory machines, so they could be genuinely vendor-agnostic, which is what manufacturers actually wanted.

The broader German response, Gaia-X, was supposed to create a European cloud substrate; instead, it produced what one co-founder called “a crushing failure, a colossal waste of time” and what the CEO of Nextcloud described as “a paper monster” of standards documents with no working product. Today, MindSphere (quietly rebranded “Insights Hub”) runs on AWS and Azure infrastructure; the Mittelstand increasingly pipes its factory data through American cloud services; and only about four percent of global cloud capacity is European-owned.

Germany had the factory floors, saw the opportunity, defined the concept, built the machines, funded the research, created a national strategy with cabinet-level backing… and then watched American cloud companies capture the data platform.

The Carrier Layer Is Not Enough

Automotive is the hardest possible case for this thesis. In no other sector does Germany hold more of the traditional advantages: decades of engineering excellence, deep supplier networks, massive production capacity, global brand equity, and a domestic regulatory framework shaped around its own industry. If these advantages cannot produce an AI-native platform developer in automotive—the sector Germany has dominated for a century—there is little reason to expect one to emerge in domains where Germany has no comparable legacy position.

It would be good to be proven wrong. The argument would be falsified if a German OEM were to announce a full-stack automated-driving subsidiary—genuinely owned, not a majority-minority partnership with a foreign platform provider—backed with Waymo-scale capital and operating driverless miles on public roads. Such an announcement would be evidence of a structural shift. The Mercedes–Uber robotaxi collaboration, tentative as it is, is not that evidence.

The structural mismatch between Germany’s carrier economy and the platform economy goes deeper than capability. One major factor keeping German carmakers from investing in breakthrough R&D such as automated driving systems is the ruinous cost of failure. Restructuring is significantly costlier in Germany than in the US: estimates cited in a recent article for Works in Progress by Pieter Garicano put a German corporate restructuring at the equivalent of 31 months of salary per employee laid off, compared to just 7 months in the United States. High severance costs create a fundamental incentive for European businesses to “avoid innovative areas and concentrate on safe, unchanging ones“. This is not because Europeans are inherently more risk-averse, but because avoiding innovative bets is the rational response when the downside of failure is so much larger. Fixing it would require labour market reform; Garicano suggests, for example, allowing workers above the 90th percentile of income to opt out of employment protections entirely, which would make German services highly competitive with the American market.

The 716,000 German automotive workers are navigating a double transition—from combustion to electric, and from mechanical engineering to software-defined vehicles—with employers who are themselves becoming integrators of foreign capability rather than developers of their own. The carrier layer is not disappearing, but it is consolidating, automating, and generating less value per unit than the platform layer above it.

Källenius told his podcast audience in September 2025 that Mercedes was “at the front” of driving automation. Four months later, that claim can’t be squared with the evidence. But of course, the problems didn’t emerge over those four months; they’ve been accumulating for thirty years. Germany’s genuine strengths in precision engineering, manufacturing quality, and systems integration are the skills that matter least in a competition decided by data accumulation and iteration speed. If the pattern established in automotive holds elsewhere, and there is evidence to suggest that it will, then Germany has a real problem.

It can rise to the occasion, or it can decide to keep doing what it did so well last century. If it chooses the latter, then Germany’s position in the next economy will be that of an excellent supplier, to platforms it does not own, in markets it does not control.

Respect to Jeff Fong and Mike Riggs for feedback on earlier drafts.