Counterpoint sees humanoid DRAM demand rising 20-fold by 2030. Per robot, it only doubles

Counterpoint Research expects DRAM in the compute units of humanoid robots to grow from about 8,300 terabits (Tb) in 2026 to about 172,000 Tb in 2030, roughly 20 times, according to a note published on October 10, 2026. Average DRAM per robot rises from 19 GB to 39 GB over the same period. Most of the growth comes from more robots, not from more memory in each one.

Why it matters: memory makers are looking for the next source of demand after HBM. On Counterpoint’s numbers, robots will not be that in volume before 2030. Whether they become one depends on how much AI runs on the robot rather than in a data centre.

How much memory is 172,000 terabits?

About 21,500 terabytes. Nvidia’s GB300 NVL72 rack carries 20 TB of GPU memory, so Counterpoint’s entire 2030 humanoid estimate equals the GPU memory in roughly 1,100 of those racks. That is our arithmetic on Nvidia’s published specification, and it puts the 2030 robot market at the scale of a single large AI cluster rather than a new market segment.

The 2026 figure squares with Counterpoint’s own unit count. At 19 GB a robot, 8,300 Tb works out to about 55,000 robots, in line with its forecast of more than 50,000 shipments this year. The same arithmetic on 2030 implies about 550,000 robots. Counterpoint’s note gives no 2030 unit forecast. Some Korean coverage of the note cited about 256,000 units for 2030; we could not find that figure in Counterpoint’s published text, and it would not reconcile with the DRAM totals. Unit forecasts for this market already vary widely, as our forecast comparison shows.

The type of memory matters as much as the amount. Counterpoint expects robots to rely on LPDDR and high-capacity NAND rather than the HBM used in AI servers, because a battery-powered robot is far more sensitive to power draw. That points the opportunity at mobile-style DRAM lines, where Samsung Electronics and SK hynix also compete, rather than at the HBM capacity that has driven memory earnings.

Why does memory per robot matter more than the total?

Because it measures where the model runs. Counterpoint links the rise to robot “brains” moving from layered control stacks to end-to-end models, which take camera, lidar and touch inputs into one representation and output actions directly, and to early use of world models. Those run on the robot only if the robot carries the memory.

The ceiling is already well above the average. Nvidia’s Jetson T5000 module, the top robot computer in its Thor line, has 128 GB of LPDDR5X; the T4000 has 64 GB. A 2030 average of 39 GB is under a third of the larger module, which implies that most units in Counterpoint’s model will not carry flagship robot computers. That fits today’s market, where the volume leaders sell many lower-cost machines; see humanoid shipments by company.

Micron has set out the high end. On its fiscal third-quarter call in late June 2026, chief executive Sanjay Mehrotra said humanoid robots carry 10 times the memory of an average L2+ vehicle. On September 30, reporting fiscal fourth-quarter results, he said humanoids are expected to need memory and storage comparable to Level 4 autonomous vehicles, which typically carry more than 200 GB of memory and several terabytes of storage, and that “physical AI can become a significant driver of memory and storage demand by the end of this decade.” The gap between Micron’s 200 GB-class framing and Counterpoint’s 39 GB average is the gap between a flagship robot and a fleet average. If on-board models win, the average moves toward Micron’s number. If robots offload inference to the cloud, it stays near Counterpoint’s.