Joby Aviation puts Blackwell class AI in the flight loop

This week, aviation straps in and the shop floor takes notes. Joby Aviation wires NVIDIA IGX Thor into Superpilot so real AI sits in the flight loop, not the sizzle reel. In factories, Rockwell previews Nemotron at the edge so operators get on-prem copilots that speak alarms and SOPs. Meanwhile, DOE’s Grid Deployment Office reminds everyone that the grid is the new constraint and your twin’s appetite still counts against your substation.

On the twin side, IEB tracks European plants moving to AI-enhanced twins that predict and act instead of narrating. Over in compute land, GMI Cloud spins up a Blackwell powered AI center in Taiwan while SC25 and Ansys showcase how simulation plus AI belongs on clusters, not laptops. Design isn’t sitting out either. Siemens NX X on Azure hints at CAM copilots, and Autodesk Fusion drops TetMet lattices so lightweight doesn’t mean unmanufacturable.

Money check. Foxglove grabs a fresh Series B for physical AI data plumbing, Azumuta raises to turn work instructions into living documents with AI, and Bone AI lands seed cash for robotics that looks suspiciously ready for plant duty. Theme of the week: fewer speeches, more closed loops.

Joby Aviation just hit a new waypoint for autonomy: it’s the launch partner for NVIDIA’s IGX Thor edge platform.

Think safety-critical, Blackwell-era AI doing perception on-aircraft for Superpilot.

In plain English: the eVTOL guys are moving from hand-crafted rules to learned models that can “see,” segment, and decide in real time - under certification constraints that would make your quality team blush.

Aviation Week reports Joby is using IGX Thor specifically in its perception stack. If regulators sign off on this paradigm, avionics stops being “logic plus lookup tables” and starts looking like model lifecycle management with flight-test-grade MLOps.

For factory folks: the same pattern (validated AI at the edge) spills into vision cells, AMR navigation, and safety-rated PLC companions - less scripted logic, more models with evidence. The faster you treat model updates like part revisions (requirements, verification, traceability), the less rework you’ll be doing at 10,000 feet - or under your press line. Aviation Week. Aviation Week

Problem: Long commissioning cycles and on-site adjustments were killing schedules and ROI.
AI approach: A plant-level digital twin (MELSOFT Gemini 3D) front-loads layout, logic, and robot pathing - now with AI agents providing predictive support and anomaly triggers inside the twin (their words, not ours).
Result: At Nagoya Works, sim-first changes cut cycle time on a five-station line from 128 to 92 minutes (≈30% improvement); broader projects report weeks shaved off timelines.
Takeaway: Twins stop being Zoom props when you embed AI for prediction and diagnostics - and wire the resulting decisions to the actual PLC/robot code. Then commissioning moves from “cross your fingers” to “run the playbook.” Industrial Ethernet Book (Nov 16). IEB Media

Execs keep waving around “AI strategy” like a Costco membership. Fortune just ran back-to-back reality checks: CEOs admitting ROI stalls without redesigned work, and a post-mortem on why AI programs face-plant (data chaos, pilot purgatory, no process owner).

Industrial translation: if you can’t sustain layered process control, you can’t sustain model lifecycle control. Pick one value stream, define data contracts, push decisions to the edge (where latency matters), then add AI. Otherwise, congrats - you’ve automated your demo environment. Fortune TechCrunch+1
Mic drop: If your “AI center of excellence” doesn’t own a breaker bar or a firewall rule, it’s a book club.

Autodesk’s AM lead just introduced TetMet’s Lattice Design Suite inside Fusion - positioning AI-assisted lattice generation and robotic welding for lightweight parts that don’t crumble under CAM reality.

The subtext: AI in design has to handshake with manufacturability, not just make pretty Voronoi.

Alexander Oster’s drumbeat matters because AM is where “AI-assisted geometry” meets heat maps, distortion, and post-machining.

Useful heresy: “Efficient lattices are only useful if your process control can actually print them.” Autodesk Fusion Blog.

At XPeng’s AI Day in Guangzhou, the company’s humanoid robot moved so naturally that people insisted a human was inside. To shut that down, the CEO literally cut into the robot on stage and showed the internals. Yes, that happened this week, and yes, there is video. Welcome to product demos in the age of physical AI. Business Insider • South China Morning Post • Silicon UK