Two statements crossed paths in the first week of October, and together they say more about the next three years of humanoid robotics than every capability demo of the past six months combined.
The first came from Micron’s fiscal Q4 2026 earnings call on September 30. CEO Sanjay Mehrotra told investors that each humanoid robot will eventually carry at least around 200GB of DRAM and several terabytes of NAND storage, and that “physical AI can become a significant driver of memory and storage demand by the end of this decade.” He had already gone further in earlier remarks, predicting a “sustained, substantial multi-decade memory demand cycle” as robots move server-grade silicon into millions of standalone bodies.
The second came from Elon Musk on October 1. “We cut our RAM in half for the Tesla AI5 chip (now 72GB of LP5) and 1/3 for AI6 (now 144GB of LP6),” he posted. “This was the only way to get enough volume for Optimus production and greatly reduces cost.” Hours later he nudged the number back up, settling AI5 at 96GB of LPDDR5, with AI6 unchanged at 144GB of LPDDR6.
Read those two numbers against each other. The world’s largest memory manufacturer says a serious humanoid needs 200GB or more. The company planning the largest humanoid production run just designed its robot brain down to 96GB, and the reason was not architecture, it was allocation. Tesla, with more purchasing leverage than any robotics company on earth, could not secure the memory it originally wanted at the volume it wants to build. That is the whole story of the humanoid industry’s next 24 months in one de-spec.
What Micron said, and what Micron earned
The Micron numbers around that robot projection deserve attention first, because they explain why a chip customer with Tesla’s balance sheet is being rationed at all. As reported earnings coverage documents, Micron’s fiscal Q4 2026 revenue came in at $54.23 billion against a $51.07 billion consensus, up 379 percent year on year, at an adjusted gross margin of 87.0 percent. Q1 FY27 guidance calls for $61.5 billion in revenue, and management guided sequential revenue growth in every quarter of FY27.
Those are not the numbers of a cyclical commodity business. They are the numbers of a structurally short market, and Micron’s own language confirms it. The company expects memory and storage “supply-demand conditions to be much tighter in calendar 2027 and 2028 than they were in 2026,” and stated flatly: “Even with additional industry DRAM clean room space plans, with robust demand trends including new upside requests from customers, we do NOT have line-of-sight to when supply and demand will return to balance.”
Mehrotra quantified the shortage in an August CNBC interview: demand for memory currently runs about 50 percent above supply. New capacity is coming, $25 billion of Micron capex in the first half of FY27 alone, with more in the second half, but the majority of that increase is construction spending that accelerates cleanroom availability “in late calendar 2028 and beyond.”
The arithmetic is brutal for anyone building physical AI in 2027. Tesla’s AI5 chip enters volume production next year at Samsung’s Taylor, Texas fab, with trial production already underway, and the dedicated Optimus factory at Giga Texas has gone from raw earth to a nearly complete steel skeleton in roughly six months, targeting 2027 mass production. Every robot that line produces will need its LPDDR5 during the exact window Micron says will be the tightest in the market’s history, with no relief capacity until late 2028. The first mass-production wave of humanoids is scheduled to hit the wall at full speed.
The take-or-pay era has arrived
The most consequential detail in Micron’s disclosures is not the robot forecast. It is this: Micron now holds 26 multi-year take-or-pay agreements covering more than 35 percent of its expected revenue through 2030, and most of its FY2027 supply is locked in at already-elevated prices.
Take-or-pay is the contract structure that liquefied natural gas markets are built on. Buyers commit to fixed volumes for years, pay whether or not they take delivery, and in exchange secure allocation in a scarce market. Its arrival in memory means the industry’s biggest customers, the hyperscalers and AI labs above all, are locking up supply years forward. Whatever is left over is the spot market where everyone else lives, including every humanoid startup that is not Tesla.
This is why Tesla’s memory cut is more significant than a cost optimization. When Musk says halving AI5’s DRAM was “the only way to get enough volume,” he is describing a design decision made against an allocation reality. And when he walked the cut back from 72GB to 96GB, the stated reason was itself a supply-chain fact: Tesla would otherwise be the only company using the minimum-RAM configuration of LPDDR5, which suggests the memory vendors were not enthusiastic about spinning a special low-capacity SKU for a single customer. Even the world’s most valuable automaker has to negotiate with the memory market, not command it.
Bandwidth versus capacity: the engineering defense, tested
Musk’s defense of the de-spec is worth taking seriously, because it is technically literate. “We think this will have a negligible effect on Optimus performance,” he wrote, on the grounds that memory bandwidth matters more than memory capacity for the robot’s workloads. The AI5 architecture backs him up: roughly half of the chip’s TRIP AI compute accelerators are dedicated to SRAM, and calculations resident in that on-chip cache see effective bandwidth an order of magnitude higher than DRAM. A half-reticle part built on TSMC and Samsung 2nm processes, rated at 2000-2500 TOPS, with a single-chip configuration roughly matching an Nvidia H100 and a dual-chip configuration contesting Blackwell-class performance, is not a brain starved of compute.
But bandwidth is not the only thing capacity buys, and three costs of the 96GB robot deserve attention.
First, local model weight. A modern vision-language-action policy, plus perception stacks, plus safety monitoring, plus whatever world-model reasoning runs onboard, is a large memory footprint, and the trend across the VLA literature is toward larger, not smaller, on-device models. Micron’s 200GB figure is not a fantasy of a memory salesman; it is an extrapolation of where on-device AI stacks are heading when they cannot assume a network connection. 96GB forces aggressive quantization, distillation, or a smaller resident policy.
Second, autonomy of the safety case. A humanoid working around people eventually has to degrade gracefully when connectivity drops. Less local memory means the fallback brain is smaller, and the gap between connected and disconnected performance widens. That is a safety-relevant parameter, not a benchmark footnote.
Third, cloud dependence quietly becomes an architecture decision. If the robot’s full capabilities depend on offloading to datacenter inference, then the memory story does not disappear, it moves upstream to the same hyperscalers who signed the take-or-pay contracts. The robot ships with less DRAM and a bigger inference bill. That trade can work economically, and several vendors are explicit about hybrid edge-cloud designs, but it converts a one-time hardware cost into a recurring operating cost that scales with fleet activity, the same token-consumption line item that theme park operators began itemizing publicly last month.
The honest summary: Tesla’s bandwidth argument is probably correct for the workloads Tesla has optimized for, and the 96GB Optimus will very likely do everything Tesla shows it doing. The question is what it cannot yet do, and the answer is a moving target that memory capacity would have absorbed over the chip’s lifetime. Tesla chose the constraint. It did not choose it freely.
The squeeze everyone below Tesla will feel
Tesla can absorb a de-spec because it is vertically integrated into silicon: it designs its own chips, co-fabs at Samsung Taylor, and buys at volumes that make it a priority customer. Now consider the rest of the humanoid industry.
Figure AI, Apptronik, Agility Robotics and their peers buy merchant silicon, mostly Nvidia, and compete for memory allocation against customers with far deeper pockets and signed multi-year commitments. In a market where supply is 50 percent short of demand and new fabs are two-plus years out, their options are worse versions of Tesla’s: pay spot prices that climb every quarter, accept smaller allocations and cap production, or de-spec their own designs without Tesla’s SRAM-heavy custom silicon to cushion the loss. None of these choices existed as constraints when their business plans were written.
The pressure is regressive in a specific way. High memory prices are a margin nuisance for a Tesla or an Nvidia. For a startup burning cash toward a 2027 production commitment, a bill-of-materials line that inflates faster than the fundraising window is an existential risk. Expect the memory squeeze to do what component squeezes always do in young industries: accelerate consolidation, push weaker players toward licensing or acquisition, and reward the companies that locked supply early or designed it away. The 26 take-or-pay contracts are, among other things, a list of who will be eating first for the next four years.
There is also a geopolitical layer. Beijing’s response to the memory shortage has been to build domestic alternatives, and this week brought word of China’s first humanoid robot built on a fully domestic electronic architecture, chips and control systems included. A shortage of American, Korean and Japanese DRAM is, from Beijing’s perspective, a forcing function: every quarter of scarcity strengthens the case for domestic memory in Chinese robots, and Chinese humanoids already dominate global unit shipments. The memory wall does not just ration the industry, it bifurcates it along supply-chain lines.
What the 2x gap is really telling you
Step back and the cleanest way to read this week is as a disagreement between a supplier’s physics and a customer’s procurement, settled in the customer’s favor by scarcity. Micron’s 200GB humanoid is the machine the AI software stack wants to build by 2030. Tesla’s 96GB humanoid is the machine the 2027 supply chain will permit at scale. The distance between them, roughly a factor of two, is the shape of the constraint under which the entire industry’s next hardware generation is being designed.
Morgan Stanley’s near-term estimates, cited in the Micron coverage, put 2030 humanoid production between 250,000 and 1.2 million units per year. At the high end, with Micron’s per-robot figure, that is 240 million GB of DRAM demand, a volume that simply does not exist in any 2027 supply plan. Someone’s forecast is going to be wrong, and the resolution will come from exactly the mechanism Tesla just demonstrated: robots will ship with less memory than their architectures want, and software will be bent to fit.
The humanoid industry has spent two years discovering its physical bottlenecks in sequence. First actuators, then dexterous hands, then rare-earth magnets, now memory. Each choke point looked, from the outside, like a temporary annoyance. Each one, on inspection, determined who could build, how many, and at what cost. Memory will be the largest of them, because unlike magnets, memory demand is competing with the richest customers in the history of computing for the same wafers.
Watch three signals from here. Watch whether other humanoid OEMs disclose de-specs or memory contracts of their own, because Tesla is rarely alone in a constraint for long. Watch the take-or-pay count in Micron’s next filings; every new contract is a chunk of supply that robot makers will never see. And watch what Optimus actually runs locally versus in the cloud through 2027, because the answer will reveal whether 96GB was, as Musk insists, negligible, or the first great compromise of the mass-production humanoid era.