Storage Industry Encounters the Physical AI Frontier: The Divergence Between Counterpoint's Forecast and Musk

After two lively years in the AI PC race, the Windows camp has finally brought out a flagship chip that genuinely rivals the Mac ecosystem.

Microsoft has teamed up with Nvidia to launch the Surface Laptop Ultra, powered by the RTX Spark superchip. This is no longer a routine iteration of Microsoft going it alone in hardware; it is the first time the Windows ecosystem has secured Nvidia's integrated top-tier SoC, using on-device compute to fill a long-standing gap.

Pavan Davuluri, Microsoft's corporate vice president for Windows and Devices, said at a company event in San Francisco on Wednesday that the Surface Laptop Ultra is now available for preorder and will begin shipping on October 16. The computer features a magnetic USB-C port, similar to Apple's MagSafe charging connector.

Counterpoint Research previously projected that storage demand for humanoid robots would increase 20-fold over five years. Musk directly dismissed that as "absurd," saying the memory capacity robots require will exceed the total of all other devices on Earth.

Because robots must rely on a physical substrate to achieve atomic-level shaping. The compute demands of superintelligence will far exceed that scale, but almost all superintelligence compute will be deployed at deep-space compute nodes.

Yet the disagreement between the two sides has never been a numerical contest over DRAM demand: one side is a mature research framework making linear extrapolations along the growth path of already-commercialized devices; the other defines the endpoint of embodied intelligence as material reconstruction, then works backward from that endpoint to the upper limit of infrastructure demand.

Clearly, the two frameworks are not discussing the same time scale or the same type of task morphology.

Two estimation frameworks, misaligned underlying assumptions

The analysts' calculation framework is rooted in an estimation paradigm that the semiconductor industry has used for years and that is widely accepted.

Its logical chain is not complicated: it centers on per-unit storage specifications and device shipments, adds penetration-rate assumptions, and estimates overall demand for the sector.

Autonomous driving is the most valuable reference sample for this framework. The task scope of mass-produced assisted driving is relatively controllable, typical scenarios are highly structured, and sufficient deployment data has accumulated on compute and storage configurations for cross-reference.

English text is being prepared.

But this reasoning rests on several implicit premises, which are rarely spelled out in research reports.

First is the constraint on task scope. In the short term, commercial humanoid robot deployment scenarios are concentrated on standardized operations such as handling, inspection, simple assembly, and human-machine collaboration. The accompanying world models tend toward lightweight and localized designs, triggered for updates as needed, without the need to continuously maintain a global, high-precision physical state representation.

Second is the layered design of compute scheduling: perception and basic planning are completed on the device side, while complex model inference and long-cycle task scheduling are sent back to the cloud. Storage load mainly carries images, point clouds, short-term state caches, and policy weights, and the load composition is relatively close to that of automotive systems.

Third, the growth path is one of gradual release: demand rises steadily with yield optimization, cost reduction, and scenario penetration. It is an incremental addition on top of the overall semiconductor market, not a paradigm shift that overturns the existing industry framework.

This reasoning is not built on thin air; it is highly consistent with the current mass-production approach of Optimus and the vast majority of humanoid robot makers. To control BOM costs and advance commercialization, engineers tend to compress onboard memory and prioritize bandwidth, using model distillation to reduce hardware compute pressure.

This also exposes the most common confusion in this discussion: many viewpoints compare short-term hardware solutions aimed at mass production with the long-term endgame described by Musk on the same dimension.

Mass-production engineering must answer how to deliver products stably and at low cost, while that "absurd" assessment points to a vision of intelligence whose time scale and underlying assumptions are entirely different.

In the end, what Counterpoint provides is an incremental estimate for the medium-term commercialization cycle.

It presupposes that humanoid robots will be embedded into the existing production system as a new type of device, with controllable task complexity and a rigid cost red line, and an iteration cadence that follows the proven iteration paths of consumer electronics and automotive electronics.

The report does not predict whether silicon-based intelligence can transform the material world; it answers only one question: if humanoid robots are deployed along the current path, how much incremental demand will the storage market see.

Musk's challenge jumps directly outside these presuppositions. He no longer uses existing device forms as the yardstick, but redefines the ultimate goal of embodied intelligence: atomic-level shaping.

Under this vision, an agent needs to continuously maintain a high-fidelity, strongly coupled physical world model, tracking physical processes such as deformation, mechanical feedback, friction, and phase transitions. Every grasp, contact, or intervention in matter will drive synchronous updates of the model state.

On this premise, what determines the scale of storage demand is no longer the number of large-model parameters, but this continuously evolving and mutually coupled physical state space.

Behind the vision, hard engineering constraints that are hard to avoid

Atomic-level shaping is not a scaled-up version of 3D printing.

It points to a closed-loop and sustainable capability for manipulating matter: an agent needs to maintain a high-fidelity, strongly coupled physical world model and track dynamic processes in real time, including solid deformation, contact mechanics, friction effects, heat conduction, and material phase transitions.

Every grasp, assembly, or rewriting of material structure requires synchronous updates to a vast state space.

Traditional large models process symbolic information such as text and images, and their data can be sharded, sampled, and compressed, with a certain degree of distortion still acceptable; by contrast, embodied intelligence aimed at atomic-scale manipulation involves strong causal correlations among states, and small local perturbations continuously propagate and diffuse along physical chains.

The phrase "memory capacity exceeding the sum of all other devices on Earth" is also often simplified and misread as meaning that a single robot needs to carry massive storage.

An interpretation closer to the logic of the original statement points to cluster collaboration scenarios: a large number of robots no longer independently execute scattered instructions, but instead collaboratively complete material reshaping, sharing and synchronizing a single global physical world model, with any physical interaction driving a synchronized refresh of the global state.

Of course, the preconditions for this conjecture to hold are extremely stringent: does the global state space support sharding and hierarchical caching? If high-frequency local states are kept on the device side and low-frequency global inference is handed to remote computing power, can storage pressure be effectively distributed? Around these questions, the industry still has no mature engineering implementation path, and it remains merely a hypothesis awaiting verification.

The concept of deep-space computing nodes should also not be simply labeled as science fiction.

It is essentially a breakout approach that attempts to break through the ceiling of ground-based computing power. Ground-based supercomputing clusters have long been constrained by multiple rigid limits: heat dissipation limits, energy density, gravitational perturbations, and land and electricity costs.

Once the goal of embodied intelligence shifts from information generation to large-scale material transformation, computing demand may expand to a degree that surface infrastructure struggles to accommodate, and seeking new computing carriers outward is logically a direction worth exploring.

The migration trajectory of computing site selection has a traceable pattern: from dispersed server rooms to large data centers and then to undersea and polar computing clusters, the common logic behind this is the search for deployment environments with better temperature, energy, and electromagnetic interference conditions.

Deep-space computing nodes can be seen as a conjecture in which this evolutionary thread is pushed to its limit. But the speed-of-light latency is a constraint this concept can hardly avoid, and it is also the most core engineering contradiction.

Force control, tactile interaction, and precision assembly for surface humanoid robots rely heavily on millisecond-level closed-loop feedback. There is a fixed optical-path delay between surface and deep-space nodes, and remote computing power can hardly undertake such physical manipulation tasks with stringent real-time requirements.

Extrapolating along this constraint, a compromise division-of-labor framework may gradually take shape: on-device robots retain local physical models, real-time perception, and closed-loop control, undertaking the rigid low-latency needs; deep-space super nodes handle long-cycle task planning, global material system optimization, ultra-large-scale world model inference, and long-term path search.

This division-of-labor model is expected to lower the threshold for deploying deep-space computing power, but it may also simultaneously weaken its irreplaceability.

If the vast majority of high-value real-time closed-loop control remains on the surface device side, deep-space nodes are closer to a long-term supplementary computing pool rather than an indispensable core of this embodied intelligence system.

In addition, in-orbit chips also face a series of high-threshold challenges such as radiation hardening, high launch costs, difficult in-orbit maintenance, stability of inter-satellite laser communication, and lifespan degradation of storage media in the space environment.

It is not an alternative route that can be activated at any time; only after ground-based solutions truly hit substantial bottlenecks is it likely to enter serious route evaluation.

Examining these two layers of constraints together, it is not hard to see: the picture Musk depicts is a logically self-consistent thought experiment, but its validity depends on a series of stringent assumptions being fulfilled simultaneously.

As long as any single engineering bottleneck cannot be overcome, this inferential logic regarding storage and computing demand may converge substantially.

It can become a direction for long-term technological exploration, but it is hard to use to revise the industry demand curve for the next five years.

Two parallel narratives, a game of chip allocation

English text is being prepared.

The first is the script of moderate volume growth, consistent with Counterpoint's calculation logic.

LPDRAM, high-reliability storage, and high-bandwidth interfaces will bring steady demand growth alongside the commercialization penetration of humanoid robots. The increment on this route is quantifiable, capacity can be planned in an orderly manner, and the secondary market can also price it using mature valuation frameworks.

Its risks are concentrated in commercialization falling short of expectations, slow scenario expansion, and terminal demand failing to take off for a long time. This is also the main line of current industry resource investment: storage manufacturers' capacity expansion plans, iteration of storage interface standards, and development of supporting test solutions mostly use this narrative as the baseline.

This route is conservative enough, with clear risk boundaries, and is more suitable for leading large manufacturers and public market funds to bet on.

The second is the script of paradigm shift, corresponding to the long-term hypothesis proposed by Musk.

Demand no longer expands linearly along the existing curve; instead, with the switch in the underlying logic of intelligent tasks, it may force a reconstruction of the entire storage architecture. Near-memory computing, high-bandwidth stacking, persistent memory, and distributed shared state pools are expected to welcome entirely new application scenarios. This route has a low probability of materializing, but once it works, it will reshape the value distribution pattern within the storage track.

At present, the participants are mostly frontier laboratories, small R&D teams, and primary-market venture capital. Although they cannot yet dominate current capacity planning, they have imperceptibly raised the R&D priority of some frontier technology routes.

Industrial capital's layout strategies have also diverged accordingly: betting most chips on mid-term increments with greater certainty, while allocating a small portion of resources to advance research on new storage architectures adapted to ultra-large physical state spaces.

This dual-track layout is itself not contradictory.

The Optimus team continues to compress per-unit memory configuration and prioritize clearing obstacles to mass production, while its long-term research group is also inferring the next-generation hardware form for physical world models. Short-term engineering implementation optimization and long-term technological conjecture can absolutely coexist within the same entity.

Analysts' forecasts are essentially linear extrapolation of existing mature paradigms. This method is good at estimating the hardware increment brought by phones, servers, and automotive electronics: samples are sufficient, variables converge, and the correlation of historical data can support predictions.

Musk's judgment is not an industry demand forecast in the standard sense. It is more like a goal statement: first anchor the endgame that embodied intelligence wants to reach, then work backward to what capability ceiling the underlying infrastructure needs to possess.

One narrative looks backward, inferring increments based on historical patterns; the other anchors the endpoint forward, working backward to the extreme boundaries of the hardware system.

For a long time to come, the storage industry will hover between the two narratives, and no one can be certain which path will necessarily take shape. The industry can only step steadily along visible stairs to advance commercialization, while quietly laying technological groundwork for those distant visions that may even ultimately fail to materialize.

The core of the disagreement has never been 20x in 5 years or a higher demand magnitude, but what capability boundaries we are willing to set for silicon-based intelligence.

This article comes from the WeChat public account "Silicon-Carbon Variables," author: Hu Yidao, published by 36Kr with authorization.