Best of LinkedIn: Smart Manufacturing CW 05/ 06

00:00:00: This episode is provided by Thomas Allgaier and Frenus, based on the most relevant LinkedIn posts about smart manufacturing in calendar weeks O five and O six.

00:00:09: Frenus is a B to B market research company that supports enterprises across the smart manufacturing industry.

00:00:14: with the market, customer and competitive insights they need to navigate dynamic markets and drive customer centric product development.

00:00:23: Welcome to the deep dive.

00:00:25: We are looking at weeks O five and O six of twenty twenty six.

00:00:29: And you know, if I had to put a bumper sticker on what we're seeing, it'd be something like the end of pilot purgatory.

00:00:35: That's

00:00:36: a good way to put it.

00:00:36: Because we aren't just seeing press releases about future possibilities anymore.

00:00:40: We're seeing these massive structural changes.

00:00:43: Yeah, it's a fascinating, almost polarized landscape right now.

00:00:47: Now so.

00:00:48: Well, on one side, you have the futurists building what they call world models, which is mind-bending stuff we should get into.

00:00:53: Okay.

00:00:54: And on the other, you have the pragmatists, you know, practically screaming that our data is garbage and we need to stop buying tools we don't understand.

00:01:01: It's that tension, right?

00:01:02: The Star Trek future... versus the rusty clipboard reality.

00:01:06: Exactly.

00:01:07: So let's start with the Star Trek side, because the NVIDIA factor is just impossible to ignore in this batch of updates.

00:01:14: It feels like they're partnering with everyone, but the Siemens one, that's the one that really stood out.

00:01:19: Right,

00:01:19: and David Kau highlighted this in his analysis.

00:01:22: It's not just about Siemens buying chips from NVIDIA.

00:01:25: They are building what they call an industrial AI operating system.

00:01:30: And that phrase, industrial AI operating system, It just stops you in your tracks.

00:01:35: It does.

00:01:35: I mean, we think of an OS as Windows or iOS.

00:01:39: It manages files, opens apps.

00:01:41: What does an OS do for a factory?

00:01:42: Well, think about the fragmentation.

00:01:44: You've got design software over here, simulation software over there, production scheduling, the actual machines.

00:01:50: And

00:01:50: none of them really talk to each other fluently.

00:01:52: Right.

00:01:52: So David Cowb points out that this partnership aims to build a unified execution layer.

00:01:58: something that spans from the very first sketch of a product all the way to its operation.

00:02:03: Okay,

00:02:03: I think I follow.

00:02:04: But the key difference, and Cal was very specific on this, is moving from passive models to active decision engines.

00:02:12: Okay, so a digital twin.

00:02:14: right now, it usually just shows you data.

00:02:16: Yeah.

00:02:16: It's a dashboard that says, hey, this machine is getting hot.

00:02:19: Exactly, it reports the problem.

00:02:22: You're saying an industrial AIOS changes that.

00:02:25: Ideally, yeah.

00:02:26: Instead of just reporting, I am hot, the system actively manages the thermal load, adjusts the production schedule, maybe even orders a maintenance part.

00:02:34: It's

00:02:34: closing the loop.

00:02:35: It's the difference between a weather app telling you it's raining and a smart home system that automatically closes your windows.

00:02:40: I see.

00:02:41: But that requires a level of trust in the AI that, well, I'm not sure we have yet.

00:02:45: And that brings up a crucial point from Alisa Prisoner-Levine and Tobias Berenstrow.

00:02:50: About the Dessault Systems Partnership.

00:02:51: Yes, also with NVIDIA.

00:02:53: They use a very specific phrase, science-based AI.

00:02:57: And this is probably the most important technical concept in this whole batch of research.

00:03:01: We're all used to things like Chad GPT, right?

00:03:04: You ask it a question, and it predicts the next likely word.

00:03:08: It's basically guessing what sounds right.

00:03:10: Which is why it sometimes hallucinates and tells you two plus two five if it's having a bad day.

00:03:15: Exactly.

00:03:16: And that is fine for a marketing email.

00:03:18: It is not fine if you're managing pressure valves in a chemical plant.

00:03:22: You can't read a factory on probabilistic guesses.

00:03:24: You

00:03:24: can't.

00:03:25: And that's what prisoner Levine and Barron Strouch are arguing for with industry world models.

00:03:31: These aren't just trained on text.

00:03:32: They are trained on the laws of physics, fluid dynamics, material.

00:03:38: Okay, so let me see if I have this right.

00:03:40: A standard AI looks at a sentence and asks, does this look like good English?

00:03:45: A science-based AI looks at a pipe design and asks, does this violate the laws of physics?

00:03:51: That's the perfect way to put it.

00:03:52: It constrains the AI with reality.

00:03:54: So if the model generates a solution that would cause a structural failure, the physics engine just rejects it instantly.

00:04:01: So you get the speed of AI, but the safety of physics.

00:04:04: That's the goal.

00:04:05: That sounds incredibly powerful, but also... incredibly expensive, computer-power-wise.

00:04:09: Oh, it requires massive compute, which brings us right to Janina Ballner's update on what's happening in Munich.

00:04:16: Right.

00:04:16: Germany isn't just watching this happen.

00:04:18: No,

00:04:18: they're making a sovereignty play.

00:04:20: They've launched the first industrial AI factory.

00:04:23: I saw the numbers in Ballner's post.

00:04:25: Up to ten thousand GPUs.

00:04:27: That's basically a supercomputer for industry.

00:04:30: It is.

00:04:31: But the why is more interesting.

00:04:33: Ballner points out that for German industries automotive, Chemical data is the crown jewel.

00:04:39: Mercedes isn't going to upload its deepest trade secrets to a public cloud in California.

00:04:43: So this meeting facility is like a walled garden for AI development.

00:04:47: It's a sovereign bunker.

00:04:49: It lets European companies access that Silicon Valley level compute power.

00:04:53: To train these world bottles we just talked about.

00:04:55: But without the data ever leaving their control.

00:04:58: It lowers the barrier to entry while keeping the IP lawyers happy.

00:05:01: So the infrastructure is finally catching up to the vision.

00:05:04: But having the compute as one thing, having a valid model to run on it is another.

00:05:09: Which leads us directly into our second theme, virtual factories.

00:05:12: And I'm so glad Steven Spasic said what he said in his post.

00:05:16: Stop calling your three D model a digital twin.

00:05:19: It needed to be said.

00:05:20: I feel like every marketing department just labeled their CAD files digital twins and called it a day.

00:05:26: Spasic is drawing a hard line.

00:05:28: He argues a three-D model is just geometry.

00:05:30: It looks nice, but it's static.

00:05:33: A true digital twin has to include machine kinematics, physics, and real data.

00:05:38: It's about constraints again, isn't it?

00:05:40: A three-D model doesn't know that a robot arm has weight.

00:05:43: Precisely.

00:05:44: Or that if you swing it too fast, momentum will make it overshoot.

00:05:47: If your digital model lets you do things that are physically impossible, it's not a twin.

00:05:52: It's

00:05:52: a video game.

00:05:52: It's a video game, and it's useless for real operations.

00:05:56: And when you get it right, When it's not a video game, the results are just staggering.

00:06:01: Roland Locker shared that case study on the Siemens Nanjing plant, a digital native factory.

00:06:05: Meaning it was born digitally before it was born physically.

00:06:08: They simulated everything, production lines, material flow before laying a wringled brick.

00:06:13: The numbers Locker quoted were, I mean, a seventy-eight percent reduction in lead time.

00:06:17: You usually fight for two percent or three percent.

00:06:20: Seventy-eight percent is a different business model entirely.

00:06:24: Think about what that means.

00:06:25: You get a product to market in a fraction of the time.

00:06:28: Lausher also noted a twenty-eight percent reduction in CO-II.

00:06:32: That's

00:06:33: the power of debugging reality before you build

00:06:35: it.

00:06:35: You don't have to tear down a wall because a forklift can't fit.

00:06:38: You fixed all that in the software.

00:06:40: And it scales down too.

00:06:42: It's not just for billion-dollar plants.

00:06:44: I saw Cervando Cedillo's post about... bros, the auto supplier.

00:06:48: Right, they're using plant simulation for buffer sizes.

00:06:51: It sounds boring, but it's so critical.

00:06:53: It sounds boring until you run out of cash.

00:06:55: Inventory is money sitting on a shelf.

00:06:57: Too big, you waste capital.

00:06:58: Too small, the whole line stops.

00:07:00: So Cedillo shows how Bros runs what-if scenarios.

00:07:04: They crash the system virtually to see if the buffers hold up.

00:07:07: They're stress testing the business logic.

00:07:09: Then

00:07:09: you have Bozana Eminin and Armand Berenclaw taking it down to the absolute micro level.

00:07:14: CNC digital twins.

00:07:15: This is where you run the specific machine code on a PC, right?

00:07:19: Exactly.

00:07:20: High-end CNC machines are incredibly expensive.

00:07:22: The materials are pricey.

00:07:24: You do not want to press start and watch the tool crash.

00:07:27: enabling operators to run the exact production code on a virtual controller.

00:07:32: You catch the collision in the software.

00:07:34: Not with a fifty thousand dollar spindle replacement.

00:07:36: Right.

00:07:37: So we have the God view of the factory, the traffic control view of the flow, the pilot view of the machine.

00:07:44: The tech stack is incredible.

00:07:45: Right.

00:07:46: But, and there's always a but.

00:07:48: Here comes the reality check.

00:07:50: The reality check.

00:07:51: Not everyone is drinking the Kool-Aid.

00:07:53: Chacobo Lorette-Casal had a post that went absolutely viral.

00:07:56: He called industry four point oh the biggest fraud of the decade.

00:08:00: It

00:08:00: was provocative for sure.

00:08:02: But if you strip away the anger, was he wrong?

00:08:06: I don't know.

00:08:06: What do you think?

00:08:07: I don't think he was.

00:08:08: He wasn't saying the technology doesn't work.

00:08:09: He was saying the implementation is often fraudulent.

00:08:12: His argument was that companies are spending millions on software and sensors that nobody looks at.

00:08:18: It's shelf wear.

00:08:19: It's the shiny toy syndrome.

00:08:21: Executives have FOMO fear of missing out.

00:08:24: They want an AI strategy.

00:08:26: But Casal argues that if your underlying process is inefficient, digitizing it just makes the inefficiency faster.

00:08:32: You're buying a Ferrari to deliver pizza when a moped would have done the job better.

00:08:37: And cheaper.

00:08:38: This connects perfectly to that story.

00:08:40: Muserat Hussein shared, a salesperson who refused to sell.

00:08:44: I love that.

00:08:45: It's a man bites dog story in the corporate world.

00:08:48: Totally.

00:08:49: Hussein had a client begging for an AI module.

00:08:52: Give us the AI.

00:08:53: But he looked at their data infrastructure and realized he

00:08:56: was a mess, not standardized, not clean.

00:08:59: Right.

00:09:00: So he told them, if I sell you this, it will fail.

00:09:03: It's the classic garbage and garbage out.

00:09:05: You feed an AI bad data, it gives you bad decisions with high confidence, a dangerous combination.

00:09:10: So

00:09:11: he walked away from the revenue.

00:09:12: That highlights the real barrier.

00:09:14: It's not compute power.

00:09:15: We have ten thousand GPUs in Munich.

00:09:17: It's the basic hygiene of data management.

00:09:19: And the basic hygiene of people management.

00:09:21: Jeff Winter brought some Deloitte data to the table that was pretty damning on this.

00:09:26: This is the idea that we're over-investing in tech and under-investing in people.

00:09:29: the data supports it.

00:09:31: Winters show that while investment in operational systems is sky-high, the maturity of human capital is the lowest part of the stack.

00:09:38: We're asking workers to bridge the gap.

00:09:41: When the smart system fails, we rely on the heroic efforts of an operator to figure it out.

00:09:46: We're asking people to work around broken systems.

00:09:50: Winters' point is that you can't have twenty-first century tech and twentieth century training.

00:09:55: Sean C. had a great insight on this for the auto industry, moving from a component mindset to a learning organization.

00:10:02: It's a huge cultural shift.

00:10:04: In the old days, if you wanted a better car, you bolted on a better part, a better carburetor.

00:10:09: The muffler guy didn't need to talk to the carburetor guy.

00:10:12: But with software-defined vehicles.

00:10:13: Everything is integrated.

00:10:15: A change to the battery software affects cooling, range, torque, everything.

00:10:19: You can't just bolt on a chain.

00:10:21: So

00:10:21: C. is saying the organizations are still structured like the old cars?

00:10:25: in silos.

00:10:25: And in many cases, the org chart is the bottleneck, not the technology.

00:10:29: Speaking of technology, I want to pivot.

00:10:32: We've talked software, data, culture.

00:10:35: But there was some hardware innovation that was just literally reshaping how we make things.

00:10:40: It's interesting.

00:10:41: Hardware is usually slow to evolve.

00:10:43: Tobias Claus highlighted that MESH technology.

00:10:46: This was the construction tech.

00:10:47: I saw the video.

00:10:48: It looked like a robot weaving in midair.

00:10:50: Essentially, yeah.

00:10:52: Rebar concrete reinforcement is heavy, manual work.

00:10:55: MESH uses robotic arms to create three-D reinforcement cages automatically.

00:11:00: And because it's a robot, it doesn't just have to make a square grid, right?

00:11:03: Right.

00:11:03: It allows for complex organic geometries, curved walls, twisting pillars, shapes that would be astronomically expensive to do by hand.

00:11:13: It's mass customization for concrete.

00:11:15: Yeah.

00:11:15: And Daniel Kieber had a similar example with metal forming, Makina Labs.

00:11:19: This was one of my favorites, dialess forming.

00:11:21: Okay, for anyone who isn't a metallurgist, why is dialess such a big deal?

00:11:25: Well,

00:11:25: normally, to make a car fender, you machine a massive block of steel into a molded die.

00:11:29: It costs a fortune.

00:11:31: If you want to change the fender shape by an inch, you throw the die away and start over.

00:11:35: Great for a million cars, terrible for fifty.

00:11:37: Exactly.

00:11:38: Makina Labs uses two robots, one on top, one on bottom, to physically shape the metal sheet.

00:11:44: Like a Potter-shaping clay.

00:11:46: So there's no mold at all.

00:11:47: No mold.

00:11:48: It's software-defined shaping.

00:11:49: You want a new fender, you just change the code.

00:11:51: It eliminates fixed tooling costs.

00:11:54: That is the definition of agility.

00:11:56: But the hardware story that really blew my mind, and connects back to our first topic, was from Robert Little about Boom Supersonic.

00:12:04: Ah, the return of Supersonic flight.

00:12:06: Well, yes, a fast jet is cool.

00:12:08: Yeah.

00:12:09: But it was what they're doing with the engine that was so interesting.

00:12:13: The symphony engine, it's a brilliant pivot.

00:12:15: They realize a jet engine is just an incredibly efficient power generator.

00:12:19: And who needs massive amounts of stable power right now?

00:12:21: Data centers, the AI factories we talked about at the start.

00:12:24: Exactly.

00:12:25: Boom is monetizing their engine tech to power AI data centers before the plane even carries a single passenger.

00:12:31: That really connects everything, doesn't it?

00:12:32: The AI needs the data centers, the data centers need the power, the aerospace manufacturers provide it.

00:12:38: and they probably used a digital twin to design the engine.

00:12:41: It's a closed loop.

00:12:43: And just to cap it off, we have to mention humanoids.

00:12:45: Alexander Dronovic shared how Siemens is approaching

00:12:48: this.

00:12:49: Are we finally seeing C-IIIPO on the assembly line?

00:12:52: We

00:12:52: are seeing the simulation of C-IIIPO.

00:12:56: Dronovic's point was crucial.

00:12:58: Don't just buy a walking robot.

00:13:00: Simulate it first.

00:13:01: Back to the digital twin again.

00:13:02: Always back to the twin.

00:13:04: Humanoids are bipeds.

00:13:06: They have to walk on uneven floors, step over cables.

00:13:10: You need to simulate that to see if the robot will actually be useful.

00:13:13: Or if we'll just trip over a pallet and break something expensive.

00:13:16: Exactly.

00:13:17: Simulate the robot in the virtual factory before you let it loosen the real one.

00:13:21: You know, looking at all these themes, a picture starts to form of where we are in early twenty-twenty-six.

00:13:26: We are definitely maturing.

00:13:27: Moving past that pilot purgatory, the tools, like the industrial AI operating system, are becoming foundational.

00:13:34: But, as Jacobo Loret-Casal and Museratu Sein reminded us, the tech is useless without the unglamorous work of cleaning data and fixing processes.

00:13:44: And changing how we manage people, that's the lingering question from Jeff Winter's data.

00:13:49: We can build the perfect virtual factory, but we still need to staff it with real people who need to be empowered.

00:13:56: That's the big takeaway for me.

00:13:58: The tech is ready.

00:13:59: The question is whether our organizations and our data are ready to handle it.

00:14:03: I think that's the challenge for the rest of twenty twenty six.

00:14:06: Closing that gap.

00:14:06: A lot

00:14:07: to chew on.

00:14:07: It's going to be an interesting year.

00:14:09: Indeed.

00:14:10: If you enjoy this episode, new episodes drop every two weeks.

00:14:13: Also check out our other editions on digital construction and digital power tools.

00:14:17: Thank you for listening.

00:14:18: Make sure to subscribe so you don't miss the next one.

00:14:20: We'll see you then.