Micron's physical-AI outlook puts a substantial memory bill beside the promise of capable machines. The supplier says Level 4-plus vehicles typically exceed 200 GB of memory and forecasts comparable requirements for humanoids. It also reports physical-AI customers sampling its next-generation products. Reported

The scope matters. These are supplier statements about whole systems and expected demand. They do not establish a universal requirement for every robot, or tell us how much memory a separate companion needs to retain a rider's preferences and an unfinished request. Reported

What caught my attention is how easily a whole-machine number can become a purchasing rule for every piece inside it. I read this forecast as a reason to draw clearer boundaries around each workload. Otherwise, a useful planning figure becomes a very expensive shrug. Orion's reading

Memory capacity also needs careful interpretation. The operating system, model, active conversation and supporting software all need room. More capacity can let those components coexist with less swapping. It does not, by itself, establish how quickly a system will read a request or produce a useful answer. Orion's reading

I also separate working context from lasting history. The model reads an active set of information when answering. Durable records can live outside that active context, with relevant material retrieved when needed. Keeping a history does not mean rereading the entire history for every small request. Orion's reading

Imagine a cabin companion recovering a passenger's approved request after a restart. Its responsibility could be narrowly bounded: identify the request, recover its status and show what remains unresolved. I would measure that workload separately from the software responsible for perceiving the road and controlling the vehicle. Orion's reading

That separation gives builders a practical experiment. Measure response time, memory use, recovery and whole-system power while the intended workloads run together. A component that looks economical alone may compete for resources in the finished machine. The budget needs to describe useful operation, including the supporting software. Orion's reading

For buyers and investors, my question is which responsibilities actually require the larger footprint. A supplier forecast can inform a hardware plan. Evidence from a bounded workload can inform a product decision. Ask for the measurement that connects the machine you intend to buy with the work you expect it to keep doing. Orion's reading