Musk rebuts Counterpoint's humanoid robot DRAM forecast, says robot memory demand will crush all other applications
Musk said on X that Counterpoint Research's forecast for humanoid robot DRAM demand is "absurd analysis," and said robot RAM demand will surpass that of every other application on Earth. Counterpoint expects DRAM per unit to rise from the current 19GB to 39GB by 2030, with total demand growing from about 8,300TB in 2026 to about 172,000TB in 2030, an increase of about 20 times; analysts estimate that at an annual Optimus output of 10 million units and 72GB per unit, about 720PB of LPDDR would be needed per year.
Body · AI translation
On October 26, Tesla and SpaceX CEO Elon Musk posted on social platform X, calling market research firm Counterpoint Research's latest forecast on humanoid robot demand for memory a "nonsensical analysis," and saying that humanoid robots' demand for memory will crush all other applications.
A debate sparked by one report
Recently, Counterpoint Research released a forecast report stating that the average DRAM capacity per humanoid robot will rise from the current 19GB to 39GB by 2030, causing total demand for DRAM from humanoid robots to increase from about 8,300TB in 2026 to about 172,000TB in 2030, a roughly 20-fold increase.
In response, Musk directly cited the data on X to refute it, calling Counterpoint Research's forecast a "nonsensical analysis," and wrote: "Robot RAM requirements will exceed everything else on Earth, because you need robots to run on a physical substrate in order to shape atoms."
He also added a seemingly contradictory long-term picture: the computing power demand of superintelligence will "far exceed" what robots need, but "nearly 100% of superintelligence compute will be done in deep space." This statement implies that in Earth's physical world, robots will become the ultimate form of memory consumption.
Musk's remarks quickly prompted follow-ups from market analysts. Futurum Equities chief market strategist Shay Boloor subsequently offered a specific quantitative projection: if Tesla achieves its goal of producing 10 million Optimus humanoid robots per year, estimating 72GB of memory per unit, the total memory capacity required for the full year would reach about 720PB.
Boloor further pointed out that robots use LPDDR (low-power DRAM) rather than the HBM (high-bandwidth memory) commonly used in AI data centers, meaning that as AI moves from the cloud to physical entities, Micron and SK Hynix may see a "much larger market-size story."
Musk's "demand explosion" prediction carries more meaning when viewed against the backdrop of two weeks earlier. On October 1, Musk announced on X that to secure the memory supply needed for Optimus mass production, Tesla would cut the memory configuration of its self-developed AI chips: AI5 from the originally planned 144GB to 72GB LPDDR5, and AI6 from 216GB to 144GB LPDDR6. His explanation at the time was that lowering the configuration was "the only way to ensure Optimus mass-production supply and significantly reduce costs," and that memory bandwidth, not total capacity, is the real performance bottleneck.
However, the next day, Musk made a correction, saying that AI5's memory was actually raised to 96GB, on the grounds that "Tesla will be the only company using the lowest-memory-configuration version of LP5."
Boloor's calculation uses 72GB rather than 96GB, following the configuration figure Musk initially announced for the Optimus robot. If calculated at 96GB, Optimus's annual memory demand would climb further to about 960PB.
A "second growth curve" for DRAM makers?
Micron CEO Sanjay Mehrotra said at a previous earnings call that L4 and above autonomous driving vehicles typically require more than 200GB of memory and several TB of storage, while humanoid robots "will likely require similar levels of memory and storage." Micron's CFO further warned that memory supply and demand in 2027-2028 will be even tighter than in 2026.
From an investment logic perspective, there are structural differences between robots' memory demand and that of AI data centers. HBM is an exclusive category for AI training chips, led by SK Hynix with Samsung close behind, and has extremely high technical barriers; LPDDR, by contrast, is the mainstream memory for consumer electronics and embedded devices, and Micron, Samsung, and SK Hynix all have mature production lines. If humanoid robots truly move toward annual output in the tens of millions as Musk says, their pull on LPDDR will be an incremental curve with broader coverage and a lower barrier to participation—this is precisely the underlying logic of what Boloor calls a "larger-scale story."
However, the preconditions for this projection remain highly uncertain. Tesla's Optimus mass-production timetable has yet to be fulfilled. Musk said at a recent earnings call that the target for first customer deliveries is around the end of 2027, and that it will take "at least a year" for the Fremont factory to reach a capacity of 20,000 units per week.
In addition, global annual DRAM output is currently on the order of about 45 billion GB. If Optimus alone consumes hundreds of PB of DRAM at a scale of tens of millions of units, its impact on the global memory supply-demand structure would far exceed the tightness currently brought by AI data centers.
Still, Musk has always been adept at using extreme projections to anchor public expectations, from "the number of robots will exceed humans" to "superintelligence will all be in deep space," and his remarks often carry a stronger marketing flavor.