Best of LinkedIn: Software-Defined & AI-Defined Vehicles CW 41/ 42

00:00:00: Provided by Thomas Allgaier and Frneus, based on the most relevant LinkedIn posts about software-defined and AI defined vehicles in calendar weeks twenty nine and thirty.

00:00:09: Frneius supports tier one automotive suppliers with early stage market validation for their R&D efforts by combining secondary research direct OEM expert interviews and facilitated customer meetings.

00:00:21: you can find more info in the description.

00:00:23: Welcome to The Deep Dive!

00:00:25: And just to set this stage right away we are You know, we're really looking past the superficial headlines today.

00:00:32: We are pulling apart the underlying mechanics of software-defined and AI defined vehicles And we're pulling from the absolute most critical conversations happening across LinkedIn right now.

00:00:42: Yeah because when you think about the history of the automobile it has always been this Masterpiece of purely physical engineering.

00:00:51: I mean for over a century innovation just meant better thermodynamics stronger steel tighter gear ratios.

00:00:57: Right, it was entirely a mechanical ecosystem.

00:01:00: Exactly!

00:01:01: You could open the hood point to a physical component and just understand what it did by looking at its shape.

00:01:07: but today we're completely unpacking how that physical architecture is shifting to AI orchestration.

00:01:14: We'll get into the honestly billion dollar economics of middleware.

00:01:18: And why human organizational culture not The technology itself Is the biggest bottleneck right now because the value of the vehicle used to be entirely bound in metal and grease, you know.

00:01:29: But this transition to a software-centric model is just tearing that historical foundation apart.

00:01:34: The physical hardware Is now frequently at war with the digital architecture.

00:01:39: That's trying to control it

00:01:40: Right.

00:01:41: And To really grasp the scale of This software revolution You first have to understand The physical brains Of the car And how they are fundamentally changing.

00:01:50: For decades automakers operated on This totally distributed Model

00:01:54: Yeah...the ECU model.

00:01:56: Every time they invented a new feature, They just added an isolated microcomputer to the car.

00:02:01: These electronic control units are ECUs.

00:02:04: You had specific computer for anti-luck brakes A totally different one for climate control Another one just rolled down windows

00:02:10: And over time you ended up with vehicle hauling around.

00:02:13: what?

00:02:14: Seventy or eighty different computers?

00:02:16: Oh easily All made by different suppliers speaking entirely different digital languages Over thousands of feet heavy copper wire

00:02:25: Which, I mean that is an absolute nightmare for latency.

00:02:28: When you have a vehicle traveling at highway speeds attempting to process data from cameras radar or LiDAR all at once It just cannot afford to send data packets across this slow fragmented network of eighty separate nodes

00:02:40: Right!

00:02:41: That delays literal life-or death safety issue

00:02:43: Exactly And Ben Rubin highlighted how the industry is actively dismantling This old system.

00:02:49: He pointed specifically to General Motors and the Cadillac Escalade IQ.

00:02:55: So GM is utilizing a centralized NVIDIA Thor processor to basically replace dozens of those independent legacy computers

00:03:03: and the performance leap there is just immense.

00:03:05: I mean Ruben noted that this centralized approach gives the vehicle up to thirty five times more AI computing power

00:03:12: And like a thousand times more bandwidth.

00:03:14: Yeah, a

00:03:14: thousand time Moving from a fragmented bus network to a single centralized brain Fundamentally changes how fast the car can react to its environment.

00:03:24: But okay, let's unpack this.

00:03:27: It's like going from a committee of fifty people shouting at each other to one hyper-intelligent conductor.

00:03:32: Right but wasn't that committee created for reason?

00:03:35: Weren't those fifty separate ECUs actually a good idea at the time?

00:03:39: Well yes Olaf Locke provided this brilliant analysis on this.

00:03:43: Actually he pointed directly to Conway law.

00:03:47: So conways law states that organizations inevitably design systems that mirror their own internal communication structures.

00:03:54: Oh, that makes so much sense.

00:03:55: Because automakers were heavily siloed

00:03:57: completely.

00:03:58: you had the powertrain division sitting in one building The interior division and another and chassis somewhere else entirely.

00:04:05: And because the VP of Powertrain didn't really talk to the VP have interiors their engineering teams Didn't talked either.

00:04:11: So the power train team goes to one supplier and buys a dedicated computer for engine timing right?

00:04:17: In the Interior Team goes to a completely different supplier for a computer to manage the seed heaters.

00:04:23: The physical layout of the corporation literally dictated the physical wiring harness of the vehicle.

00:04:28: Oh,

00:04:29: for a long time that complexity was rational.

00:04:31: it let departments work independently.

00:04:33: but today That organizational legacy is the main barrier to consolidating compute power.

00:04:40: You have decades of procurement habits fighting a unified digital architecture.

00:04:44: Okay.

00:04:45: But I want to push back on the idea of total centralization here.

00:04:48: if we consolidate all this power into one massive supercomputer in the dashboard.

00:04:54: Don't we lose the fail safes?

00:04:55: We used to have

00:04:56: heavy mean,

00:04:57: well if this centralized AI crashes because it's say overloaded processing a complex three D navigation route I still need my brakes To function instantly.

00:05:07: we can't have a generative AI model hallucinating when a pedestrian steps into the crosswalk.

00:05:12: No absolutely not.

00:05:13: and Thomas Bohm addressed This exact tension.

00:05:16: he argues that The term ai defined vehicle is actually a bit misleading

00:05:21: because it implies AI is doing literally everything.

00:05:24: Exactly, in reality you have to build a strict firewall between systems.

00:05:28: safe deterministic control.

00:05:30: so the underlying physics of steering accelerating braking that must remain at the lower layers.

00:05:36: those systems execute rigid code that guarantees a specific physical reaction in milliseconds every single time.

00:05:43: So the AI acts as the high-level strategist but the deterministic code

00:05:50: Central compute, it has to be paired with highly efficient localized controllers at the edge.

00:05:56: Alfredo Barata expands on this emphasizing cross-domain zone controllers

00:06:00: instead of one giant computer or eighty tiny ones.

00:06:03: Yeah you create regional hubs in the car.

00:06:05: It's about finding the absolute perfect location within the vehicles architecture for software Safety and real time control to execute without stepping on each other

00:06:14: which brings us To the operational cost of running all this intelligence because here is where it gets really interesting.

00:06:20: We are moving away from standard generative AI, the kind that just answers a prompt to change the radio toward agentic AI.

00:06:28: Right systems that think and act on their

00:06:29: own Exactly.

00:06:31: Agentic AI holds context, uses digital tools takes proactive action.

00:06:36: it might monitor battery degradation check upcoming elevation changes and then proactively negotiate a charging slot twenty miles away based on live grid pricing

00:06:44: without you ever asking to.

00:06:46: yeah but if we are pumping these cars full of agentic ai that thinks in acts like what the computing cost just explode off-the charts.

00:06:54: oh they do.

00:06:55: computational weight is staggering.

00:06:58: Marcus Wambach broke down the math on this, noting that agentic AI requires twenty to fifty times more inference compute than a simple generative AI request.

00:07:09: And just to clarify for anyone who might not be super deep into machine learning, inference is the actual real time thinking process.

00:07:16: it's NOT THE INITIAL LAB TRAINING IT'S THE CONTINUOUS PROCESSING THE AI HAS TO DO ON THE FLY WHILE YOU'RE DRIVING.

00:07:23: Yeah and because it has hold memory in constantly run parallel scenarios devours processing cycles.

00:07:30: Wampock points out a harsh reality, falling AI token prices will not save automakers.

00:07:36: the sheer volume of data is too vast.

00:07:39: so what's the solution?

00:07:40: The only real cost lever automakers possess Is the orchestration layer!

00:07:45: The vehicle's architecture must dynamically constantly decide What Data is processed locally on-the-edge hardware and what gets sent to cloud.

00:07:54: Right.

00:07:54: Because if you try to process everything locally, You need a fifty thousand dollar supercomputer in every sedan which destroys your margin.

00:08:01: But If you send it all the cloud Your cellular data costs explode and lose function In dead zone.

00:08:08: That orchestration layer is tightrope.

00:08:10: automakers are walking right now.

00:08:12: And because building that intelligent orchestration Layer Is so incredibly expensive Doing from scratch in house is basically financial suicide.

00:08:21: This immense pressure is forcing former rivals to collaborate on foundational software.

00:08:26: But really quickly, before we dive into the billion-dollar economics of these global ecosystems.

00:08:30: I just want to casually remind you to subscribe to The Steep Dive so that you catch all our future editions

00:08:47: Middleware

00:08:49: being the digital plumbing, right?

00:08:50: Exactly.

00:08:51: The invisible layer translating high-level commands into a language that physical hardware can actually execute

00:08:57: exactly!

00:08:58: The consumer never sees it they don't interact with it and They definitely don't base their buying decision on them.

00:09:03: yet long in Walter notes It costs roughly one hundred million euros to develop a robust middleware platform from scratch.

00:09:10: Wow One hundred million just for the plumbing

00:09:12: Right.

00:09:13: so if seven different automakers let there pride dictate their strategy and build our own independently, it is seven hundred million euros burned across the industry.

00:09:23: And over a ten-year vehicle lifecycle with maintenance security patches updates that balloons to one point.

00:09:29: seventy five billion euros

00:09:31: tied up in invisible software That offers zero competitive differentiation In the showroom.

00:09:36: exactly

00:09:37: It's insane which Is exactly why The autosar adaptive platform is now in production Across what

00:09:42: sixteen

00:09:43: different OEMs they're finally getting.

00:09:45: The goal is to build that common foundation together, and then compete on the user interface in the applications the customer actually cares about.

00:09:53: But how are they managing higher level stuff?

00:09:56: The advanced driver assistance systems or ADAS?

00:09:59: Augustine Friedl had a great breakdown of this.

00:10:01: he outlined four specific integration archetypes companies using to navigate these costs.

00:10:07: Right different strategic camps.

00:10:09: Yeah

00:10:10: First you have vertical integration.

00:10:11: Think Tesla or Rivian where try to own the entire stack from Silicon to UI.

00:10:17: Then there's the OEM orchestrated model, which is like BMW and Toyota acting as general contractors plugging in external partners for specific

00:10:25: domains.".

00:10:26: And Friedl also points out the AI application-led approach.

00:10:30: so Nissan partnering with Wave basically handing the core driving intelligence over do a dedicated AI firm.

00:10:37: And finally The platform led archetype where players pre-packaged chunk of the hardware and software ready to drop in.

00:10:46: And because there's no consensus on which model wins, we're seeing some wild strategic moves.

00:10:52: Yeah so what does this all mean for Europe?

00:10:54: Because Pedro Pacheco pointed out that some European automakers are literally buying Chinese vehicle architectures now.

00:11:01: Isn't it just a shortcut to catching up?

00:11:03: I mean It sounds like buying us store bought cake and putting your own icing on it.

00:11:07: Well pachego gave very specific warning about exactly.

00:11:11: He says it's a short-term fix with massive long term brand risk.

00:11:15: European cars already have this growing image problem in China, they're seen as expensive but digitally old fashioned compared to local EVs.

00:11:24: Right because Chinese consumers view the car is rolling smartphone.

00:11:28: The software IS the luxury feature?

00:11:30: Exactly So.

00:11:32: if a european brand licenses chinese tech drops into their chassis and tries sell at premium just based on historical brand prestige Consumers see right through it.

00:11:43: Selling a car loaded with Chinese tech at higher price point won't fix their problems, consumers would just buy the domestic Chinese car instead.

00:11:50: Okay so if licensing is dangerous shortcut how do European legacy brands accelerate their own software?

00:11:56: Kishore Patil shared a vision for this, detailing the rise of a Europe-India mobility corridor.

00:12:01: Yes and this is a massive shift!

00:12:04: Europe brings deep R&D in manufacturing mastery while India brings unparalleled software scale and AI capability.

00:12:11: But it's not old outsourcing model right?

00:12:13: No none at all.

00:12:15: Patil emphasizes shifting India from being just an extended workbench where they get handed low level coding tasks to be strategic coengineering partner with end-to-end responsibility.

00:12:26: That makes total sense.

00:12:28: But here is the thing, we have the centralized compute architecture figured out conceptually.

00:12:33: We see these global economic partnerships forming but none of this matters if The internal corporate structures are still stuck in the mechanical era.

00:12:42: Oh!

00:12:42: This is the biggest issue...the clash between modern code and legacy bureaucracy Is just

00:12:47: staggering.

00:12:47: I have to share this story from Peter Virk.

00:12:49: It perfectly captures the absurdity of us.

00:12:52: back in twenty thirteen he was at a UK OEM trying to push an over-the-air software update.

00:12:58: keep in mind, This is purely digital code.

00:13:00: right

00:13:01: but The legacy process required him To input physical part number submit a CAD drawing and Input await

00:13:08: await for software

00:13:09: exactly Software doesn't weigh anything But the procurement database was built for break calipers, so it demanded a weight in grams.

00:13:18: The software could be written in a week but navigating the bureaucracy just to get a meaningless part number took six-to eight

00:13:25: weeks.".

00:13:25: That is incredible!

00:13:27: The engineering wasn't the bottleneck at all...the human bureaucracy was and John Francois Alice shared a staggering statistic on this kind of legacy complexity.

00:13:36: In twenty-twenty A VW Passat Could Be Configured different combinations.

00:13:43: Half a billion?

00:13:44: Yes,

00:13:45: compared to around two hundred total combinations for Tesla that same year.

00:13:49: That physical diversity must just crush the electronic architecture validation process.

00:13:53: It creates a massive deficit in collective speed.

00:13:57: QA teams are trying to validate code across millions of physical wire routing setups.

00:14:01: So the technical cure for that has to be virtualization.

00:14:05: Tiago Rodriguez talked about this.

00:14:07: Google's AAOS-SDV runs on these digital twins like Cuttlefish and Cloud Horizon.

00:14:12: Which is such a game changer,

00:14:14: right?

00:14:14: Because it allows teams to architect design and validate vehicle components in the cloud entirely before the hardware even exists.

00:14:22: Jens Beckman completely agrees with this noting that The old Hardware Tide box model Is completely broken.

00:14:28: now

00:14:28: you can simulate tens of thousands Of driving scenarios overnight without stamping A single piece of metal.

00:14:34: Okay, but I have a question here.

00:14:36: We have digital twins we have cloud testing...we've all these amazing tools.

00:14:40: But as Dr Tim Sturm and Prashant Dorswamy both point out that technology is moving much faster than the actual organizations.

00:14:48: Are we just trying to cram twenty-first century software into a twentieth century org chart?

00:14:52: Yes That's exactly what happening!

00:14:55: DorsWami emphasizes that real software defined vehicle transformation requires massive cross functional collaboration.

00:15:02: The mechanical, electronic and software teams can't be siloed anymore.

00:15:06: Software has to become part of the organization's DNA

00:15:10: Not just an IT department.

00:15:11: bolt it onto a factory Precisely

00:15:13: And Dr.

00:15:14: Sturm echoes this.

00:15:15: He says until you have software defined organizations that are capable of fast decisions An end-to-end ownership AI is only going to accelerate complexity not solve it.

00:15:24: Right because if your corporate structure Is a tangled mess AI will help generate a tangled mass at light speed.

00:15:31: Exactly It always comes back to Conway's law.

00:15:34: So we've spent this whole time looking at the individual software-defined vehicle, but I want to leave you with a final concept to mull over based on a post from Rupali Mondgei.

00:15:45: She brought up V two X or Vehicle To Everything communication.

00:15:49: Oh This is where it gets really sci fi

00:15:52: right because as we focus so much on a single car's architecture What happens when cars start continuously sharing hazard events, braking status and speed beyond line of sight?

00:16:04: It changes everything.

00:16:05: If your car becomes just one node in a distributed network of moving computers the whole concept of standalone AI vehicle might become obsolete.

00:16:14: We won't be engineering cars anymore.

00:16:16: we will be engineering synchronized intelligence.

00:16:19: So think about this next time you drive what is very first piece data that should share with the network.

00:16:26: If you enjoyed this episode, new episodes drop every two weeks.

00:16:29: Also check out our other editions on charging and battery tech commercial fleets in autonomous and mobility markets.

00:16:35: absolutely thank you so much for joining us For the steep dive.

00:16:39: remember to hit subscribe So you don't miss out.

00:16:42: as we saw today upgrading The software is honestly the easy part.

00:16:45: it's Upgrading the organizational mindset.

00:16:47: that's the real challenge.

00:16:49: see ya next time.