Musk States Current Limiting Factor for AI Expansion is Memory, Not GPU or Power
Musk clearly stated in the SpaceX Q2 2026 earnings call that the current limiting factor for AI expansion is memory, not GPU or power.
He pointed out that memory output is growing at about 20% annually, which is fast for mature industries, but AI demand is growing at about 200% or higher, leading to a supply-demand gap that causes prices to continue rising rather than falling. This statement has been quoted by several media outlets as "limiting factor currently is memory."
Tesla has halved the RAM of the AI5 chip to about 72GB LP5 (later adjusted to 96GB to avoid being the only user at the minimum specification) and reduced AI6 to about 144GB LP6, aiming to ensure the supply needed for Optimus mass production and reduce costs. Musk stated that memory bandwidth is more critical than total capacity, and that maintaining bandwidth has a negligible impact on performance.
Micron, Samsung, and SK Hynix are the main suppliers, with HBM capacity sold out until 2026. The U.S. currently lacks large-scale memory wafer fabs, and Micron's new factory in New York is expected to reach production around 2028. Tesla has stated that it has received a "very significant" allocation from Micron, and the prices are relatively reasonable.
Hyperscale cloud providers and xAI continue to lock in multi-year memory orders, further squeezing consumer and server DRAM supply, leading to a significant increase in contract prices.
Market mechanisms indicate that the event is driven by the explosion of AI data center and inference demand, with funds and capacity prioritizing HBM and high-bandwidth memory. Beneficiaries include Micron and the other two memory manufacturers in terms of pricing power and profits, while those under pressure are AI chip designers and system manufacturers reliant on external memory, as well as the consumer electronics supply chain.
Source: Public Information
ABAB AI Insight
Musk has repeatedly warned about the "chip wall" since early 2026, prioritizing memory as a constraint in Tesla and SpaceX earnings reports. Both Tesla and SpaceX are heavily procuring AI chips and memory for autonomous driving, Optimus, and data centers, while Musk is also promoting the Terafab plan to integrate logic and memory manufacturing in Texas to reduce dependence on external supply.
On the capital front, the company is locking in supply by signing long-term allocation agreements with Micron while reducing its own chip memory specifications to match available capacity. xAI and hyperscale cloud providers are seizing HBM quotas through prepayments and multi-year contracts, transforming memory from a general commodity into a strategic resource.
A comparable case is the 2021 automotive chip shortage that forced automakers to reduce features, and Nvidia's concentrated procurement of HBM led to Samsung and SK Hynix's capacity shifting towards AI. Currently, AI is at the "memory bandwidth and capacity becoming the actual expansion bottleneck stage"; after GPU supply improves, memory becomes the new scarce link.
Structurally, this represents a restructuring of the industry chain and a transfer of pricing power. AI workloads are upgrading memory from an auxiliary component to a core bottleneck, and the capacity expansion speed of the three manufacturers determines the overall computing power release pace of the industry. The mechanism is that DRAM and HBM production line construction cycles are long and yield requirements are high; when demand growth far exceeds supply, pricing power shifts from design companies to memory manufacturers, and any entity attempting to rapidly scale AI must first resolve memory allocation issues.
ABAB News · Cognitive Laws
When demand growth is ten times that of supply, prices only rise.
Computing power bottlenecks will shift from chips to memory.
Long-term allocation agreements equal locking in capacity in advance.