Notes

Short form, between the posts.

Notes are the working margin of AutomotiveCloudWatch: a signal worth flagging, a correction, an argument that is not yet a post. Newest first, straight from Substack.

Newest pieceOct 19, 2026

Zonal Architecture [Explained]

The shift that lets software run the car is, underneath, a story about copper, fuses and who builds the boxes.

20 notes

Oct 19, 2026
Traditional vehicle electronics work like home appliances hardwired directly into the walls. Replacing an oven means opening drywall and rewiring the circuit. A software-defined vehicle introduces the automotive equivalent of a standard wall socket. An abstraction layer sits between physical hardware and application software. Engineers can build, validate, and update features for years without redesigning underlying control units. Toyota structured its Arene platform around this model. They designed Arene as a software development platform with dedicated tools, an SDK, and a data pipeline.…
- Note
0 reactionsOpen on Substack
Oct 14, 2026
An AI agent took five automotive test requirements to a Simulink model, a compiled FMU, five ecu.test runs and a green 5 of 5 report in 8 minutes and 37 seconds. The reasoning cost about 28 cents. Then I laid the requirements beside the executed tests. "At t=0" was checked at t=1s. "At all sample points" was checked once. In the published demo, the FMU that actually ran came from a hand-written C script, and the agent confirmed the match by reading equations. The team fixed that two weeks later, but the summary looked identical either way. On paper, the suite catches one of six plausible…
When an Agent Runs the V-Cycle, the Answer Key Is the Product
- Note
1 reactionOpen on Substack
Oct 13, 2026
A robot that finishes the demo can still leave your operators holding every exception that decides real throughput. That is why I think the useful factory metric is the cost of an accepted part across the full production mix, including the intervention, inspection, and rework that never show up in a monthly robot payment. RobCo's Alfie makes this timely because it targets changing parts, containers, and layouts rather than one repeated motion, but adaptability only has value where the process needs it and the economics hold once the easy cases disappear. When your plant trials physical AI,…
The Robot Economics Start at the Accepted Part
- Note
2 reactionsOpen on Substack
Oct 8, 2026
The most important architectural idea in NVIDIA's Open Agent Safety Platform is the independent observer. Placing Sentry on a BlueField-4 DPU separates the watchdog from the host OS. An agent cannot modify the monitor judging its execution. How should enterprise architectures isolate observers from the agent runtimes they monitor?
- Note
1 reactionOpen on Substack
Oct 8, 2026
Most Physical AI updates report one number, and it is usually about the model. A humanoid has five layers, and each fails for its own reasons. The brain picks the action. The body has to execute it under load. The parts have to repeat it thousands of times a shift, and dexterous hands alone run to roughly 31 percent of a humanoid's bill of materials. The fuel sets the working day, with most humanoids still good for two to four hours of real industrial work before a swap or a recharge. The proving ground is the customer's floor, where Figure 02 spent eleven months at BMW Spartanburg to log…
The Physical AI Stack, Explained
- Note
1 reactionOpen on Substack

Every other note

Page links need JavaScript. Without it, every other note is listed here, newest first.

Oct 7, 2026
Honda opened Marysville in 1982 and Nissan opened Smyrna in 1983. Both were called transplants. Over the next decade suppliers followed them across the Pacific, American engineers were trained, and the process knowledge moved with the tooling. I was a general manager at Nissan, and I saw what that localization took. It worked because nothing stopped the knowledge from travelling. That condition no longer holds for every investor. Since 2025 Beijing has required export licenses for battery cathode and lithium processing technology. A Chinese-owned plant in America could assemble packs without…
The Leverage Isn’t in the Room
- Note
2 reactionsOpen on Substack
Oct 7, 2026
A robot reducer is a set of gears in a housing. You can buy one, cut it open and measure every surface. Competitors have been doing that for decades. Nabtesco still estimates it holds about 60 percent of the world market for the precision reduction gears in industrial robot joints. Harmonic Drive Systems leads in strain wave gears, the type many humanoid joints rely on. Chinese makers such as Leaderdrive have taken a growing share of their home market, so the geometry was never the secret. What the drawing leaves out is the process. Tolerances held across thousands of units, surface finish,…
Japan’s Quiet Monopoly Inside Every Robot Joint
- Note
4 reactionsOpen on Substack
Oct 7, 2026
Figure ran its Figure 02 humanoids at BMW Spartanburg for eleven months. They logged more than 1,250 hours, loaded over 90,000 sheet metal parts and helped build 30,000 X3s. The most useful line in Figure's own report was about what broke. The top hardware failure point was the forearm, where three degrees of freedom, wiring and heat were packed into a human-sized limb. It carried a circuit board passing signals between the main computer and the wrist actuators. No model update fixes that. Figure 03 removed the board. This is why I start a robot review at the first failed task. A missed…
The Physical AI Stack, Explained
- Note
2 reactionsOpen on Substack
Oct 7, 2026
In May 2025 Ford idled the Explorer line in Chicago. The cause was a rare earth magnet, a part that barely registers on the bill of materials and sits inside seats, wipers, speakers and motors. China's export licensing had slowed the supply, and nobody could swap the part quickly. A substitute magnet changes how the motor behaves under heat, which reopens validation, which takes months the program does not have. The purchase price said nothing about that exposure. BMW's Neue Klasse shows how far the digital side of the car has come, with four central computers, four wiring zones and 600…
The Software-Defined Vehicle Has a Second Architecture Nobody Draws
- Note
2 reactionsOpen on Substack
Oct 6, 2026
The microkernel idea behind QNX is neither secret nor unique to QNX. The design is well documented, and a capable team can build a working demonstration of it. That still leaves them a long way from something an automaker can put in a car. QNX keeps its kernel small. Drivers, file systems and network stacks run as separate processes and talk through explicit message passing. A fault stays inside the process that caused it, and timing stays predictable when something goes wrong. That matters in a vehicle because safety software and infotainment now share the same compute, and QNX Hypervisor…
How QNX Works and Why It Is Difficult to Replicate
- Note
5 reactionsOpen on Substack
Oct 6, 2026
NVIDIA is treating the vehicle as another node on its broader computing platform. The car is no longer treated as a collection of isolated microcontrollers running vendor firmware. Upstream training clusters, factory digital twins, and in-vehicle drive silicon share a single development environment. That alignment lets an engineering team distill large autonomy models in the data center and deploy deterministic runtimes to the car without translating across incompatible toolchains.
NVIDIA Is Turning the Car Into a Node on Its AI Platform
- Note
1 reactionOpen on Substack
Oct 5, 2026
NVIDIA's industrial position now reaches well past the GPU. An independent map of 80 publicly documented entities across eight functional layers puts automakers beside CAE vendors, factory robotics firms, cloud platforms and systems integrators. NVIDIA gains influence wherever those tools come to depend on a common accelerated computing and simulation layer. The value sits at the boundaries between tools. Geometry loses context as it moves from one system to the next, simulation models fall behind the current product configuration, and factory planners work from assumptions engineering has…
NVIDIA Is Building the Operating Layer for Industrial AI
- Note
3 reactionsOpen on Substack
Oct 2, 2026
Everyone is scoring the robot race by counting machines. I mapped 80 companies in NVIDIA's robotics ecosystem, and the machine count turned out to be the least interesting number on the page. NVIDIA's pitch is three computers, one for training, one for physically grounded simulation and one on the robot itself. FANUC, ABB Robotics, YASKAWA and KUKA are integrating Isaac simulation into virtual commissioning, and together they represent an installed base of more than two million robots. That is real distribution. It is not proof. A foundation model anyone can adapt does little for a team that…
NVIDIA Is Building the Common Stack for Physical AI
- Note
1 reactionOpen on Substack
Sep 29, 2026
NVIDIA announced its Open Agent Safety Platform on September 28, and the detail worth reading closely is where it draws the boundary. OpenShell confines an agent at the runtime level, restricting which files it can write, which binaries can make outbound calls and which hosts it can reach. Sentry, a reference design running on BlueField-4, adds a watchdog the host agent cannot edit. That moves a real share of agent safety out of the model and into infrastructure an operator can inspect. For automotive and manufacturing teams, the harder problem sits one step past the sandbox. A well-scoped…
NVIDIA’s Open Agent Safety Platform Moves the AI Boundary Out of the Model
- Note
3 reactionsOpen on Substack
Sep 28, 2026
Check out my new short article! Jev is TypeSafe AI's first System One model, and automotive and manufacturing leaders should understand what it actually does. It never writes text. You give it a situation and a set of allowed answers, and it returns a typed decision with a calibrated confidence score in milliseconds, at a fraction of the cost of a generative model. In fleets, that means triaging fault codes and warranty claims. In factories, it means routing quality alarms and maintenance tickets, with uncertain calls escalated to a person. In robots, it means choosing the next task while…
Jev Explained for Automotive and Manufacturing Leaders
- Note
4 reactionsOpen on Substack
Sep 23, 2026
Trump says letting China build here is just like the Japanese transplants. He is reaching for the wrong analogy. I was a general manager at Nissan, and I watched how the transplants localized. Honda and Nissan pulled suppliers across the Pacific and moved real process knowledge into American plants in the 1980s. They could do that because nothing stopped them. Tokyo had no export controls forbidding it. The politics were friendly. China inverts both conditions. Beijing now licenses its best battery know-how so it cannot leave the country. Washington treats a China-owned maker as a security…
The Leverage Isn’t in the Room
- Note
4 reactionsOpen on Substack
Sep 23, 2026
I was recently asked a deceptively simple question: If QNX can isolate critical workloads inside a vehicle, could the same architecture make public cloud infrastructure more secure? The interesting answer has almost nothing to do with kernel size. A 50,000-line hypervisor does not create a 50,000-line trust boundary if a much larger management domain beside it can map tenant memory or control DMA-capable devices. The real trusted computing base is every component capable of breaking isolation. That is what makes QNX Hypervisor 8.0 interesting. Its microkernel architecture offers a…
A Microkernel Will Not Save Your Cloud: QNX Hypervisor 8.0, Xen, and the Trusted Computing Base Nobody Publishes
- Note
3 reactionsOpen on Substack
Sep 23, 2026
A car assembled in Ohio is not the same as a car that originates in Ohio. One is a location. The other is a question of where the value comes from. A plant can bolt together a vehicle that is overwhelmingly Chinese in its battery, motors, magnets and software, then call it American-made because the last screws were turned in the Midwest. USMCA’s rule of origin is what separates the two. Today it requires 75 percent North American content, up from 62.5 under NAFTA. In the current review the US is asking for 82 percent, with 50 percent required from the US specifically. That review runs year…
The Leverage Isn’t in the Room
- Note
0 reactionsOpen on Substack
Sep 22, 2026
Here is the part of the Chinese car panic almost no one mentions. The technology transfer everyone fears cannot legally happen. In 2025 Beijing restricted its core battery technologies, the LFP and LMFP cathode work and lithium processing. Any transfer abroad, through trade, investment or a joint venture, now needs a license from the Chinese government. China holds more than 95 percent of global cathode capacity and about 70 percent of the battery market. That is the crown jewel, and Beijing has made it hard to move offshore. So a BYD plant in America could bolt battery packs together. It…
The Leverage Isn’t in the Room
- Note
1 reactionOpen on Substack
Sep 22, 2026
Xi lands in Washington tomorrow, and much of Detroit has decided Trump will settle the Chinese car question over dinner. He mostly cannot. Three rules already on the books have decided it for him. A China-built EV lands at roughly 127.5 percent in stacked tariffs. The connected vehicle rule bars a China-owned maker from selling a connected car here from model year 2027, even one built in America. And China’s own export controls stop its firms from bringing their best battery technology to a US plant. The real protection is quieter. It is the USMCA rules of origin review, running to 2036,…
The Leverage Isn’t in the Room
- Note
1 reactionOpen on Substack