I keep coming back to one number that stopped me cold last week. DRAM demand tied to humanoid robots could jump from roughly 8,300 terabits this year to 172,000 terabits by 2030. That is not a gentle rise. That is a twenty-fold explosion. And it has almost nothing to do with the data-center frenzy we have all been watching for the past two years.
Why Physical AI Changes the Memory Equation Completely
For a while it felt like the only story in semiconductors was hyperscalers gobbling up every available high-bandwidth memory chip for their training clusters. Prices climbed. Consumer electronics makers started sweating. Apple even had to nudge up the cost of its latest premium phones. Then something quieter started appearing in analyst notes and executive comments. The next wave of demand might not sit in a climate-controlled server hall at all. It might walk on two legs.
Humanoid robots need to process video, sound, and spatial data at the same time while constantly updating an internal picture of the world around them. That kind of continuous, multi-modal thinking eats memory. Current prototypes already carry about 19 gigabytes of DRAM on average. By the end of the decade that figure is expected to roughly double to 39 gigabytes per unit. Multiply those larger modules by the hundreds of thousands, then millions, of robots analysts now expect to roll off production lines, and the arithmetic becomes hard to ignore.
The Shipment Numbers Are Moving Faster Than Most People Realize
Earlier this year one major investment bank still projected only 51,000 humanoid shipments for 2026 and a modest 256,000 by 2030. Those forecasts have been rewritten. The new base case sits at 75,000 units next year, 890,000 in 2030, and an eye-watering 6.5 million by 2035. That last figure represents a market opportunity measured in the hundreds of billions of dollars. I have watched enough technology cycles to know that early shipment projections often look aggressive until the moment they suddenly look conservative. This one already feels like it could land on the higher side of realistic.
What drives the revision? Better actuators, cheaper sensors, and, crucially, AI models that finally make continuous real-world reasoning feasible outside the lab. Once a robot can anticipate how a package will shift when it is lifted, or how a person will move through a doorway, the commercial case for deployment strengthens quickly. Warehouses, manufacturing floors, even certain retail settings start to look like natural homes for these machines. And every single one of them will need a healthy slice of DRAM.
Autonomous Vehicles Add Another Layer of Pressure
Robots are not the only physical systems hungry for memory. Level-4 and higher autonomous vehicles routinely require more than 200 gigabytes of memory plus several terabytes of storage. The overlap in technology is obvious. Both platforms must fuse camera, lidar, and radar data in real time while maintaining a constantly updated model of their surroundings. Memory executives have started saying out loud what many of us already suspected: humanoid robots will likely need similar levels of memory and storage to those already designed into advanced self-driving platforms.
That dual demand creates a longer-term squeeze that looks different from the current HBM shortage. High-bandwidth memory for GPUs is specialized and still relatively low volume compared with mainstream DRAM. The robots and vehicles coming down the pipeline will pull on more conventional DRAM capacities as well. In other words, the pressure could spread across more product lines than the current AI-server boom has touched.
Supply Side Signals Are Mixed and Worth Watching Closely
Not every voice in the industry is singing the same tight-supply song. One prominent Taiwanese PC maker recently suggested that memory prices could begin easing in the second half of next year. That view sits in quiet tension with the public statements from the three largest memory producers, who continue to signal firm pricing well beyond 2027. Meanwhile a Chinese manufacturer has announced the start of mass production on more advanced memory chips. The market will get a clearer read when the next major Korean earnings call arrives later this month. Until then the narrative remains unsettled.
I tend to side with the cautious camp for now. Capacity expansions take years, not quarters. And once humanoid production ramps, the incremental demand arrives on top of an already elevated baseline from data centers. Even if prices soften temporarily, the longer arc still points higher. The companies that control the most advanced DRAM processes stand to benefit from both the near-term AI server cycle and the longer physical-AI cycle that is only beginning.
What More Sophisticated Models Actually Require
There is a simple reason memory per robot keeps rising. Early systems could get by with relatively modest onboard storage because much of the heavy computation happened in the cloud or in simpler rule-based code. Newer architectures rely on what researchers call world models. These models let a machine simulate possible futures a few seconds ahead so it can choose safer or more efficient actions. Running those simulations in real time demands both speed and capacity. Image streams, audio, proprioceptive feedback, and predicted environmental changes all have to live in memory at once. The result is a steady upward march in the average DRAM load.
Perhaps the most interesting aspect is how quickly those requirements can change. A year ago 19 gigabytes looked ambitious for many prototype designs. Now 39 gigabytes is treated as a reasonable target for production machines later in the decade. If foundation models continue to grow in parameter count and multimodal capability, I would not be surprised to see the average climb higher still. Memory density and bandwidth become limiting factors almost as soon as processing power does.
Level 4 or higher autonomous vehicles typically require over 200 GB of memory and several terabytes of storage. Humanoid robots will likely need similar levels of memory and storage.
That statement from a major memory company chief executive captures the convergence nicely. The same silicon that powers advanced driver-assistance systems is being designed into walking machines. Shared architectural needs mean shared pressure on the same manufacturing capacity.
Investment Implications Beyond the Obvious Chipmakers
Most investors already know the names of the big three memory producers. The more interesting question is how the broader supply chain responds. Companies that make advanced packaging, specialized testing equipment, or high-speed interfaces stand to see secondary demand. So do the firms developing the low-power process technologies that keep a humanoid from draining its battery after only a few hours of operation. Memory that is fast but power-hungry is of limited use in a mobile robot.
On the demand side, the industrial customers matter as much as the pure technology plays. Logistics operators, contract manufacturers, and even some healthcare providers have begun pilot programs. Early adopters who secure long-term memory supply agreements could gain a cost advantage once volumes scale. Those who wait may find themselves competing for the same constrained wafers that the hyperscalers already dominate.
- Average DRAM per humanoid roughly doubles between now and 2030
- Shipment forecasts revised sharply higher in recent months
- Autonomous vehicles create parallel demand for similar memory profiles
- Chinese capacity additions may ease but not eliminate tightness
- World models push continuous multi-modal processing requirements higher
The Timeline for Real Volume Is Shorter Than Many Expect
Deliveries of production humanoids are expected to begin in meaningful numbers as early as next year. That does not mean millions of units overnight. It does mean the first commercial fleets start generating real DRAM orders. By 2027 and 2028 the ramp becomes steeper. The 2030 figure of 890,000 units already implies a compound growth rate that would have looked fantastical only a couple of years ago. And the 2035 projection of 6.5 million units turns the entire segment into a major industrial market.
I have found that technology adoption curves in physical systems often lag the software curves by a few years, then catch up quickly once reliability thresholds are crossed. We appear to be approaching that inflection. Safety certifications, battery life, and cost per hour of operation still need work. But the trajectory is clear enough that memory planners cannot treat humanoids as a distant curiosity any longer.
How This Differs From Previous Memory Cycles
Past DRAM booms were driven largely by personal computers, then smartphones, then servers. Each cycle had its own rhythm of oversupply and shortage. The current AI-driven shortage already feels different because of the specialized nature of high-bandwidth memory. The robot-driven phase will feel different again. It mixes mainstream DRAM volumes with the need for low-power, high-reliability variants that can survive vibration, temperature swings, and continuous operation. Yield and qualification cycles may prove more demanding than pure data-center silicon.
That complexity could favor the largest, most technologically advanced producers. Smaller players may struggle to meet both the volume and the quality requirements at the same time. In my view this sets up a longer period of elevated pricing power for the leaders than a typical commodity memory cycle would allow.
| Year | Projected Humanoid Shipments | Estimated DRAM Demand (terabits) |
| 2025 | Low tens of thousands | Around 8,300 |
| 2026 | 75,000 | Rising sharply |
| 2030 | 890,000 | 172,000 |
| 2035 | 6.5 million | Multiples higher |
These figures are still estimates, of course. Actual numbers will depend on how quickly costs fall and how readily industries adopt the technology. Yet even if the 2030 shipment total lands 30 or 40 percent lower than the latest forecast, the DRAM demand growth remains substantial. A ten-fold increase instead of a twenty-fold increase is still a major structural shift.
Potential Near-Term Relief and Longer-Term Constraints
Some industry voices expect memory prices to ease in the second half of next year as new capacity comes online and the most acute phase of the current shortage passes. That relief, if it materializes, would be welcome for consumer electronics companies that have already absorbed higher component costs. It would not, however, erase the longer-term demand curve now visible from physical AI. The same factories that expand to meet today’s AI server needs will find themselves planning the next expansion for robots and vehicles almost immediately afterward.
Chinese production of more advanced memory chips adds another variable. Greater domestic supply could reduce pressure on global pricing for certain mid-range products. High-performance, low-power DRAM optimized for mobile robots is a different story. The most advanced process nodes and packaging techniques remain concentrated among a handful of producers. That concentration tends to support pricing power even when overall industry utilization is not at absolute peaks.
The Role of World Models in Driving Capacity Needs
It is easy to underestimate how much memory a continuously running world model consumes. The system must hold multiple hypotheses about the near future, update them with every new sensor reading, and still leave enough free capacity for safety-critical processes. That is a very different workload from batch training of a large language model. Latency matters more. Predictability matters more. Power efficiency matters more. All of those constraints push designers toward larger and faster memory subsystems than early prototypes used.
In my experience watching previous hardware transitions, the memory bill of materials often rises faster than the processor bill once systems leave the lab. Designers discover that real-world edge cases require more buffer, more redundancy, and more simultaneous data streams than simulation suggested. Humanoid robots are likely to follow the same pattern. The 39-gigabyte average projected for 2030 may prove to be a floor rather than a ceiling.
Broader Market Consequences Worth Considering
If the robot-driven demand materializes on anything close to the current forecasts, the memory industry will enter a multi-year period of elevated utilization. That environment typically supports stronger free cash flow and higher capital returns for the leading producers. It also tends to accelerate investment in next-generation process technologies. The companies best positioned to convert those investments into yield and cost advantages will widen their competitive lead.
Secondary effects could appear in adjacent markets. Higher DRAM pricing feeds into the cost structure of every device that uses the chips, from servers to smartphones to industrial controllers. OEMs will respond with a mix of price increases, feature adjustments, and more aggressive qualification of alternative suppliers. The net result is rarely a pure win for end customers, at least not in the short run.
At the same time, the existence of a large new end market reduces the risk of prolonged oversupply once the current AI-server boom eventually cools. Memory makers have historically struggled with boom-bust cycles precisely because demand was concentrated in a few cyclical end markets. Adding robotics and autonomy as structural growth drivers could smooth those cycles somewhat. That would be a healthy development for long-term industry stability.
What to Watch in the Coming Months
Several near-term data points will help refine the outlook. The next round of earnings calls from the major memory producers should provide updated commentary on 2026 and 2027 visibility. Any mention of design wins or volume commitments related to humanoid platforms would be especially telling. On the demand side, pilot program announcements from large industrial customers will signal how quickly real deployments are progressing. And the pace of cost reduction in actuators and batteries will determine how soon the economic case for large fleets becomes compelling.
I will also be watching the evolution of average memory content per robot. If early production units already exceed the current 19-gigabyte benchmark by a wide margin, the 2030 demand figures could move higher still. Conversely, aggressive power-management techniques or more efficient model architectures could temper the growth. Either way, the direction of travel seems clear.
A Structural Shift Rather Than Another Temporary Spike
The last few years taught everyone that artificial intelligence can create sudden, intense demand for specialized silicon. The next phase may teach us that the same intelligence, once embodied in physical machines, creates a more durable and diversified demand for memory. Data centers will continue to matter. They will not be the only story.
Humanoid robots and highly autonomous vehicles represent a different kind of customer. Their memory needs are continuous rather than bursty, mobile rather than static, and quality-sensitive rather than purely performance-driven. Meeting those needs will require investment, process innovation, and careful capacity planning. The producers that prepare early stand to capture a meaningful share of a market that barely existed five years ago and could be enormous by the middle of the next decade.
The twenty-fold increase in DRAM demand projected for humanoids alone is only one piece of a larger picture. When you add the parallel requirements from advanced vehicles and the ongoing needs of AI infrastructure, the case for structural tightness in memory markets becomes harder to dismiss. Prices may fluctuate. Capacity will expand. But the underlying demand drivers look more durable than many previous cycles.
For investors, engineers, and industry observers, the message is straightforward. Physical AI is no longer a distant science-fiction concept. It is beginning to show up in shipment forecasts, executive commentary, and, soon enough, actual production lines. The memory that powers those machines will be one of the more interesting hardware stories of the coming years. And the numbers already on the table suggest it will be a large story indeed.
Keeping an eye on average content per unit, shipment trajectories, and the comments from the largest producers will remain useful through the rest of this year and into the next. The shift from pure digital AI to physical AI is underway. Memory sits right at the center of that transition. The scale of the opportunity is only starting to come into focus, and the companies best prepared for it will likely be the ones still shaping the conversation when the decade ends.
