The compute services, embedded platforms and memory that develop and run machine intelligence.
AS OF 2026-09-13DATED · NON-EXHAUSTIVE
Research reviewed 2026-09-13. Available means documented access or production, not assured spare capacity for every specification. Tight refers to the scope identified in the node, not every robot component. Supply unclear means the reviewed evidence does not settle availability at the selected boundary. Expansion describes how additional usable supply can be delivered, not a promised lead time. Pricing power describes our assessment of the ability to defend price or terms; it is not inferred from technical importance alone.
Learning at one scale, acting at another
A training cluster is not the computer in the robot
Training can use shared data-centre infrastructure. Running a machine adds different constraints: power, heat, size, response time, sensor interfaces and product lifetime. A robot does not need a miniature copy of the cluster that trained its model. It needs a computing system that performs its assigned work within the machine's limits.
This branch therefore separates rented training capacity from on-machine platforms. It also keeps working memory and persistent storage distinct. They have different functions and manufacturing economics even when the same supplier sells both or the customer receives them inside one module. Folding them into a generic electronics node would hide an important source of upstream exposure.
The AI compute map already follows fabrication, packaging and data-centre infrastructure. This page connects Physical AI to those markets rather than drawing a second copy of the entire semiconductor industry. Fast motion-control loops and the software that coordinates them belong in Control & runtime.
More machines, more capability, less compute per task
Compute demand can increase through more training programmes, larger experiments, more deployed machines or more demanding behaviour on each machine. Better models, compression and hardware efficiency can reduce the resources needed for a given task. None of those mechanisms should be replaced by a single robot-unit multiplier.
The most direct shared exposure may sit upstream. In its June 24, 2026 prepared remarks, Micron reported that DRAM and NAND demand exceeded supply. That is a supplier's industry assessment, not evidence that a particular robot builder cannot obtain its chosen memory. The distinction matters when translating an AI infrastructure constraint into a Physical AI investment claim.
Supply-chain map
Developing the capability
Training services sell access to infrastructure. A builder can also own its training system; this is the purchased-service route.
The output is a development resource, not a finished robot policy. The upstream chips, networking, memory and energy are mapped in the AI compute theme.
Training compute services
Supply unclear
Role
Provides execution capacity for model development
Expansion
Provision existing capacity, or add servers, networking and power
Pricing power
Depends on usable capacity and commitments; no robotics-specific premium established
Why it exists
A development team needs processors, memory, storage and interconnect to execute training and evaluation. Buying access as a service transfers infrastructure ownership and operations to the provider. It is a different commercial boundary from buying accelerators, even though both routes ultimately need the same physical work.
Azure, a Microsoft business: ND H100 v5 virtual machines provide GPU-based capacity for demanding AI workloads.
Google Cloud / Alphabet: Cloud TPU access provides a training-infrastructure route alongside rented GPU systems; it is not a robot-specific compute allocation.
Worth understanding
These are established offerings serving many AI customers, not capacity reserved for robotics. The relevant availability question is whether a team can obtain the required configuration, region, duration and interconnect at an acceptable price. A published instance specification does not answer that question; this read has not measured allocation queues, hence the unclear supply label.
Expansion has two clocks. A provider may allocate an existing cluster relatively quickly. Adding genuinely new capacity requires hardware and a powered, connected site. The bottleneck can sit outside the cloud service's own software layer, as the AI branch pages explain.
Our inference is that effective cost per completed experiment matters more than the advertised price of one GPU-hour. Faster communication or better utilization can offset a higher hourly charge. Conversely, model reuse can reduce purchased training work. Do not count the cloud provider's full revenue and the underlying equipment sale as two independent measures of final Physical AI spending.
The arrows express input requirements, not verified supplier contracts. Platforms include processor and software support at varying boundaries. Vehicle qualification remains a distinct route; these are not two successive stages.
Working memory
Tight
Role
Holds active models, observations and intermediate results
Expansion
Process transitions and additional qualified DRAM capacity
Pricing power
Current industry scarcity supports suppliers; robot-specific terms are not disclosed
Why it exists
The processor must repeatedly access instructions, model parameters and current observations. Working memory provides this active workspace. Capacity and bandwidth can limit what the machine can execute within its power budget even when its nominal arithmetic performance is high.
Relevant businesses
Micron: low-power DRAM for edge workloads, distinguished from the persistent storage used for logs and other retained data.
Samsung Electronics: automotive DRAM, including low-power memory products, alongside separate storage offerings.
SK hynix: automotive LPDDR memory alongside its other AI-memory products. The March 2026 portfolio display establishes product scope, not available capacity for every grade.
Worth understanding
Low-power DRAM on an embedded platform is not interchangeable with the HBM stack on a training accelerator. The shared economics reside in memory manufacturing capability and capacity allocation, not in every robot purchasing the same package.
Micron's June 2026 supply discussion links constrained growth to complex fab expansion and technology transitions, and describes pressure from HBM on other DRAM supply. The tight label refers to this broader industry assessment. It is not a measured lead time for every LPDDR grade.
Our inference is that qualification and lifecycle requirements can make an apparently available substitute costly to adopt. However, a module buyer may have no direct memory-supplier choice. Memory revenue then sits inside the purchased platform's price; adding both amounts would double-count that content. The separate node survives because merging it would hide a manufacturing constraint with a different expansion clock.
NAND output plus controller, packaging and product qualification
Pricing power
Industry supply supports pricing; endurance and qualification differentiate products
Why it exists
A machine has to retain its operating software and model files without power, and may record observations for later analysis. Embedded storage supplies that persistent record. It differs from working memory in use, technology and failure considerations, even when a single vendor supplies both.
Relevant businesses
Micron: NAND-based storage for retained drone data, alongside a separate DRAM workspace.
Samsung Electronics: automotive SSD products for vehicle data storage; other form factors are available within its automotive portfolio.
Kioxia: automotive UFS storage combining flash and controller functions in a supplied device, illustrating a storage-specific product boundary.
Worth understanding
A logging policy can change storage demand without changing the number of machines. Keeping more sensor history increases capacity and write requirements; selecting useful events or compressing records can reduce both. The economically relevant unit is not simply bytes generated by sensors.
Micron's June remarks also describe NAND supply constraints and capacity allocation toward DRAM. The tight rating is at this industry level, not proof of a shortage in every automotive or industrial storage product.
Our inference is that endurance, firmware, data integrity and validated operation can support differentiation beyond raw capacity. But a specification advantage is not automatically a durable margin: suppliers still face competing products and customer redesigns. Separating storage from DRAM lets the reader follow different production and usage mechanisms without turning the controller, flash die and package into compulsory additional nodes.
Runs perception and learned behaviour on the machine
Expansion
Qualified processor/module production plus carrier, cooling and software integration
Pricing power
Software ecosystem and redesign costs can matter; spare production capacity unverified
Why it exists
A machine builder needs a supported way to run its models and connect sensors within a size, power and heat envelope. The purchase may be a processor platform, a production module or a more complete computer. This node keeps the platform and its essential development support together rather than presenting each processor function as a separate market.
Relevant businesses
NVIDIA: Jetson production modules with published lifecycle dates; developer kits are separately identified as prototyping products.
Qualcomm: Dragonwing IQ9 compute platforms for demanding edge-AI and industrial applications.
Worth understanding
The delivered boundaries are not identical: a processor platform leaves different board-level work with the customer than a module. They belong in the same commercial comparison because builders evaluate these routes to obtain an embedded computing platform. Carrier boards and cooling remain real inputs, but are not assumed to be separate mandatory merchant markets in every design.
NVIDIA's lifecycle table distinguishes production modules from developer kits, for which it does not specify lifecycle support. It also shows different end dates across module generations. A working prototype therefore says little about the supply plan for years of production. No consistent programme-level inventory or lead-time evidence was established here, so the supply label remains unclear rather than inferring abundance from a catalogue.
Our inference is that model tooling, drivers and integration knowledge can raise switching costs. Higher arithmetic capability alone does not establish higher customer value if bandwidth, latency or cooling limit the task. A successful new platform can expand supplier content while a simpler model or more efficient implementation can reduce the compute needed in each machine.
Provides in-vehicle compute within an automotive programme
Expansion
Semiconductor supply plus vehicle-specific integration and qualification
Pricing power
Programme commitments and software integration can support retention
Why it exists
A road vehicle's computing platform must fit the vehicle architecture, interfaces and operating requirements. The commercial relationship may include hardware, development software and substantial programme support. Treating this as a generic robot computer would conceal the buyer's qualification process and the possibility of combining several vehicle functions on one platform.
Relevant businesses
NVIDIA: DRIVE in-vehicle computing platforms and associated development technology.
Qualcomm: Snapdragon Ride platforms; Ride Flex can support cockpit and driver-assistance workloads on a shared architecture.
Worth understanding
Combining functions changes the purchased scope. It can increase the value of a platform while eliminating separate computers elsewhere in the vehicle. A design win should therefore be examined as a specific programme architecture, not multiplied by a generic number of AI chips per car.
The reviewed platform material establishes supplier participation, not accepted production capacity for every announced customer. Development-kit availability is particularly weak evidence for this purpose. The unclear supply rating means the programme-level queue is not established; it does not mean that the node is an insignificant market.
Our inference is that qualification, software integration and long programme lives can make an incumbent difficult to replace mid-cycle. That is different from unlimited pricing power at the next platform selection. Large vehicle customers can negotiate architecture and sourcing across generations. The vehicle platform remains separate because those obligations differ materially from an embedded module bought for a smaller robot programme.
The strongest current supply signal in this read is upstream memory tightness, not a demonstrated universal shortage of robot computers. Training services, production modules and automotive programmes have different allocation and integration questions. Their labels should not be borrowed from one another.
For a supplier comparison, examine revenue per successful task alongside qualification costs, platform reuse and replacement risk. For the theme, distinguish more machines from more compute per machine, and distinguish either from the amount of training used to create a reusable capability.
This dated map selects economically useful boundaries. Company participation does not establish a trade relationship between the named suppliers, and a product announcement is not evidence of accepted machine output.