Every few weeks another memory maker announces new capacity, and the market reads it as a date on a calendar. Build the plant, ease the shortage, move on.
Steve Neamtz thinks that reflex is where investors are getting 2027 wrong. Neamtz is chief executive of Yorkville America, which launched a passive memory and storage ETF on 10 September, so he has a position in this.
His argument is still the most specific thing anyone has said about the next eighteen months.
“The market may be underestimating the gap between a new factory announcement and memory that customers can actually use,” he told Fathom.news in an interview. “A factory may be able to make more chips, but HBM must also be stacked, packaged, tested, and approved by customers.”
That gap matters because the headline numbers have become genuinely strange. A 32GB DDR5 kit that cost under $100 a year ago now sits near $375. SK Hynix, which holds around 56% of the high-bandwidth memory market, had sold out its 2026 output before 2026 began, and its chief executive has said publicly that the squeeze will probably run past 2030.
The company’s own new capacity is not expected at volume until late 2027 at the earliest, and it has said even that will not be enough.
Why A Wafer Is Not A Chip
The reason extra capacity converts so slowly is partly arithmetic. High-bandwidth memory consumes roughly three times the wafer area per gigabyte of standard DDR5, and by TrendForce’s estimate a single gigabyte of HBM takes up the equivalent of four gigabytes of ordinary DRAM.
HBM now accounts for something like 30% of DRAM revenue while making up about 8% of output, and margins on it run several times higher than on consumer memory.
That explains the price of your next laptop. It also explains why a new plant does not translate into AI memory at the speed the announcement implies. Wafers are the beginning of the process, not the end of it, and the stacking, packaging and customer qualification that follow are where the real queue forms.
Neamtz expects that queue to sort unevenly. Ordinary DRAM and NAND flash, he said, could become easier to buy while the newest HBM stays tight, which means relief in one part of the market will be read as relief across all of it. The two are not the same trade.
2027 Gets More Expensive, Not Cancelled
The obvious question is whether memory ends up capping how fast AI actually gets built. Neamtz thinks the effect shows up in budgets rather than in cancelled projects.
“Memory could slow AI growth in 2027, but it is more likely to change the cost and timing of projects than stop them,” he said. If HBM and the advanced packaging that connects it to AI chips stay tight, fewer finished chips ship, system prices rise, and some spending slides to a later year, which pushes back the revenue those data centres were supposed to produce.
He also argues the industry has started judging whole systems rather than individual chips. Training a model leans on raw processing power, but generating answers one token at a time depends far more on memory and on how fast data can move.
Also Read: Ripple Wires XRP Into Stripe’s New Standard for AI Agent Payments
With inference projected to account for more than half of AI computing by 2030, the parts of the stack that were treated as plumbing start setting the ceiling.
The Squeeze Lands On Whoever Cannot Negotiate
There is no price at which memory suddenly becomes a crisis, in his reading. It becomes a problem when higher costs lower the expected return on a data centre, delay a project, or force a provider to raise what it charges.
That is already happening for systems built around large amounts of HBM, and the burden is not evenly spread. The largest cloud companies can negotiate better terms or design more efficient systems around the constraint. Smaller buyers have fewer options and less leverage.
Follow that through and the memory shortage stops being a supply story and becomes a competitive one, quietly concentrating AI capacity among the handful of buyers who can absorb the bill.
Where His Own Thesis Breaks
To his credit, Neamtz is explicit about how the whole argument comes apart. The industry could overbuild if plants are planned around today’s demand and finish after growth has slowed, and the trigger he names is uncomfortable for anyone positioned for a long shortage.
“The biggest risk would come if large technology companies cut spending just as major new factories start producing memory,” he said. Rising inventories, fabs running below capacity, shorter delivery times and falling contract prices are the early tells.
The concentration that supports pricing power today is its own risk. HBM supply sits with three manufacturers, so a single fab problem, a run of poor yields or a trade restriction could slow AI chip supply across the industry.
Memory is also a deeply cyclical business, and he is careful not to claim that today’s pricing power is permanent or that every memory company is an automatic winner because of it.
What Is Actually Worth Watching
The useful part of his answer is the list. Shipment volumes, yield rates, packaging capacity, contract prices and the capital spending plans of the largest technology buyers.
“The important signal is not how much companies say they will spend,” he said. “We believe it is how quickly that spending produces approved HBM.”
That is a harder thing to track than a press release, which is rather the point. For the next two years the memory story will be told in announcements, and the announcements will be running a year or more ahead of anything a customer can plug in.
Read Next: Google, Nvidia Back New Alliance to Fix AI Power Crunch