Quick favor: If you like this edition, forward it to one person who builds real things. It helps us grow. 💪
Welcome to Issue #008 of Neural Newton
This edition is a full shop tour. NVIDIA’s Apollo puts AI physics on center stage at SC25, Supermicro wheels out rack scale AI factories, and Rockwell greenlights a million square foot smart plant to prove it is not just slideware.
On the data spine, PTC and Garrett plug CAD, PLM, and ALM into a single SaaS stack so copilots can find the truth. In the tool crib, Simcenter Prescan, NVIDIA Warp, and PhysicsX bring fresh options for simulation and AI native engineering. The rumor mill says CAM teams are prepping Apollo assisted toolpath trials. The money trail points to Lambda’s mega round, PhysicsX’s top up, Cavela’s seed, and NcodiN’s photonics play for optical interposers. We wrap with Blake Moret’s factory as a demo floor and a robot baseball clip to keep everyone humble.
🏭 The Retrofit - NVIDIA’s Apollo gambit puts AI physics into your CAD, CAE, and fab flows
SC25 was not just another GPU parade. NVIDIA introduced Apollo, an open family of AI physics models designed for real-time-ish predictions across CFD, FEA, EM, climate, lithography, and more. Translation: surrogate models that behave like your expensive solvers, only faster, so you can explore 100 design variants before the espresso cools. Apollo arrived as an official, named release at Supercomputing 2025 in St. Louis, not a backroom demo, which tells you where this is headed. NVIDIA Blog
Early users span real industrial trenches. NVIDIA called out Applied Materials, Cadence, KLA, Luminary Cloud, PhysicsX, Rescale, Siemens, and Synopsys as adopters or partners. That reads like your CAE and EDA vendor list, not a research poster wall. The big bet is simple: wire Apollo into design workflows, reserve your high-fidelity runs for validation, and let surrogates prune the bad ideas first. NVIDIA Blog
What this means if you build things: your bottleneck shifts from solver hours to data hygiene. Apollo will happily learn your modeling sins if your geometry is dirty, your mesh is a tet soup, or your boundary conditions live in a Slack thread from 2019. Treat Apollo like turbo on your existing stack, not a replacement for physics. Validate on your geometry and your operating envelope, then unleash it to shrink search space. Engineers get cycles back, and program managers stop asking why “one more mesh” takes all week. NVIDIA Blog
Main image suggestion: split screen with NVIDIA’s Apollo graphic next to a residual heatmap comparing solver vs. surrogate on an airfoil. NVIDIA Blog
⚙️ Gearbox - Rapid News Roundup (post Nov 16)
Supermicro shows off “AI factory” clusters at SC25. Rack-scale B300 and NVL72 based builds so your pilot model does not live on a pizza box. Facilities will love the liquid cooling slide, probably. Super Micro Computer+1
Rockwell plans a new greenfield plant in Wisconsin. Over 1 million square feet with advanced automation and digital systems, and part of a 2 billion dollar US expansion plan. A living demo line for customers, finally. Manufacturing Dive+1
NVIDIA makes Apollo official at SC25. Open AI physics models pointed straight at engineering workloads, from semiconductors to structures. You will see this show up as buttons in the tools you already use. NVIDIA Blog+1
🧱 Blueprints IRL: Garrett Motion standardizes on PTC’s SaaS spine
Problem: Global teams building turbos, EV thermal systems, and hydrogen compressors were dealing with scattered data and slow change control.
AI approach: Move to a unified SaaS backbone - Onshape for CAD and PDM, Windchill+ for PLM, and Codebeamer+ for ALM - so design, software, and compliance live on one data model that AI copilots and surrogates can actually find. PTC
Result so far: Garrett and PTC say the expansion extends traceability from requirements to release, improves collaboration, and sets the stage for AI-driven checks and optimization across the lifecycle. The message is not “AI first,” it is “clean product data first.” Metrics will follow as programs roll through the new stack. Engineering
Takeaway: If your AI roadmap is just “add a copilot,” you skipped leg day. Converged CAD, PLM, and ALM makes your org AI ready by default. PTC
📉 Signal Drop - AI physics will not fix bad engineering hygiene
Some of you will feed Apollo a STEP with non-manifold edges, a mesh with element quality that screams, and boundary conditions copied from email. When results look weird, you will blame “AI hallucination.” Please do not. Surrogates are fast approximators that mirror your inputs with gusto. If your CAE vault is a junk drawer, the model is a mirror, not a miracle.
🛠️ Toolbench - Three to kick the tires on
Simcenter Prescan 2511 - Scenario and sensor simulation for ADAS and AV, now with a smoother Blender import path that speeds up scene building for digital twins. Great, now test engineers can complain about materials and shaders faster. Siemens Blog Network
NVIDIA Warp 1.10 - Open source Python framework that JITs physics kernels to the GPU. New release improves JAX interoperability and performance, so you can prototype physics ops without writing raw CUDA. Your ML team will act like it was their idea. NVIDIA Developer Forums+1
PhysicsX platform - Startup tooling for AI native engineering that moves heavy numerical physics from solve time to inference. Fresh Series B extension plus a new Deutsche Telekom partnership point to growth, and yes, this plugs into real manufacturing and energy use cases. physicsx.ai+1
🛰️ Shop Floor Rumors - plausibly true whispers
CAM teams are planning “Apollo-assisted” toolpath scoring trials. Surrogates will rank strategies first, then traditional engines validate, which means less solver queue thrash when you are chasing chatter. NVIDIA Blog
Expect Q1 bake offs for “NC programming copilots,” with at least two OEMs comparing vendor copilots vs. in-house LLMs on nasty legacy parts. The metric will not be BLEU score. It will be scrap. Manufacturing Dive
💸 Grease Money - follow the cash
Lambda - Series E - 1.5 billion dollars plus. AI infrastructure for training and inference at industrial scale, with capital earmarked for GPUs and vertically integrated data centers. If you plan to train physics surrogates on the clock, this is one of the clouds that will take your money. Business Wire+1
Cavela - Seed - 6.6 million dollars. Startup claims AI powered supplier sourcing and pre tariff cost modeling for brands and manufacturers. If your sourcing team is spreadsheet surfing at 2 a.m., they will at least be curious. TechCrunch
PhysicsX - Series B extension. Fresh capital added to the UK startup building an AI native engineering stack, right alongside new go to market moves with Deutsche Telekom’s Industrial AI Cloud. That is a signal that enterprise distribution is coming, not just cool demos. physicsx.ai+1
👤 Machine Whisperer - Blake Moret, Rockwell Automation
With a new greenfield Wisconsin plant exceeding a million square feet, Rockwell’s CEO is not just selling The Connected Enterprise, he is building one people can walk through. Expect a very public proving ground for robotics, edge AI, and digital twin assisted commissioning. Customers will not need a slide deck. They can watch the lines. Manufacturing Dive
🧠 Dumb Things Smart People Will Say in the AI Era
“We do not need PLM cleanup, the copilot can read between the BOM lines.”
“Our surrogate is 98 percent accurate, just not on Tuesdays or stainless.”
“The digital twin is off by 5 percent, so we adjusted reality.”
🧊 Coolant Break - Robots play ball
IEEE Spectrum’s latest Video Friday shows robots throwing, catching, and even batting baseballs. Somewhere a plant manager is wondering if the same arm can catch falling OEE. Until then, enjoy the uncanny swing mechanics and those custom robo gloves. IEEE Spectrum+1
Until next time, keep your meshes manifold and your twins out of PowerPoint.
Neural Newton