Investor Mackenzie Morehead has published a survey of early-stage memory technologies that could eventually outperform high-bandwidth memory, the stacked DRAM that has become the single most expensive component in an Nvidia GPU. Her conclusion: two approaches, magnon-based memory and vertical ferroelectric RAM, look like the most credible long-term candidates. Neither is close to a shippable product.
The stakes are unusual for a hardware niche. Morehead notes that HBM now accounts for roughly 60 percent of the manufacturing cost of an Nvidia GPU, and that SK Hynix’s chief executive has said supply shortages will persist beyond 2030. Memory, not raw compute, is the line item squeezing AI infrastructure budgets right now.
That is worth naming directly, because it cuts against how most people outside chip design talk about AI constraints. The conversation defaults to GPU counts and training FLOPs. But the bottleneck engineers actually hit when serving large models in production is how fast data moves between memory and logic, not how many matrix multiplications a chip can theoretically run. A piece about lab-stage memory physics belongs on the reading list of anyone who thinks they only care about models.
Morehead splits the field into two camps. The first chases access speeds faster than HBM by moving memory physically closer to logic. She points to d-Matrix and Qualcomm as the most advanced entrants here, calling their approaches capable of reaching manufacturing within several years, though with speedups she describes as meaningful but not dramatic. Further out sits magnon-based memory, which stores information in spin waves rather than electrical charge. Morehead says antiferromagnetic magnons in insulating materials are particularly promising precisely because the active material carries no moving charge at all, which is what keeps dissipation low. She cites six years of progress on signal readout, up roughly four orders of magnitude, with commercially useful levels still out of reach.
The second camp targets HBM-class bandwidth at NAND-like density, and Morehead’s pick there is vertical ferroelectric RAM, or FeRAM. Ferroelectric materials have been proposed as a universal memory platform for two to three decades because they are non-volatile and require little energy to switch states. She notes Micron invested $1 billion in a ferroelectric capacitor demonstration that was ultimately shelved: its density scaled the same way DRAM’s does, undermining the entire premise. Morehead describes newer academic work on vertical, stacked capacitor architectures as a possible way around that dead end, while cautioning that any startup pursuing it would face close scrutiny from incumbent memory makers and would need both a decisive technical edge and a defensible patent position.
Morehead is explicit that these remain speculative bets. She writes that commercializing either approach would require founders capable of raising hundreds of millions to over a billion dollars, pointing to Micron’s shelved $1 billion demo and Cerebras’s $3 billion raise as the scale of capital this category demands. She also flags a structural reason to expect faster progress than prior hardware cycles: AI tools for materials simulation and chip design, plus coding agents that let hardware startups run lean engineering teams, could compress the discovery-to-manufacturing timeline she says historically took new memory technologies decades to cross.
None of this changes near-term GPU procurement. Morehead’s own framing is that if magnonics or vertical FeRAM eventually reach production, on a timeline she puts at four years or more if AI-accelerated materials discovery holds, they would reshape the constraints AI systems are designed around rather than merely improve them at the margins. Investors and infrastructure teams betting on the current HBM supercycle should treat that as a multi-year tail risk to the assumption that stacked DRAM defines the ceiling on inference performance, not as a near-term alternative to today’s supply contracts.
Mackenzie Morehead detailed this research on her own site, mackenziemorehead.com, in a post published September 1, 2026.
