Straight Answers — No Spin from the desk of Andrea Klein
I’ve been in this industry long enough to know when a market is different. This one is. Over the past few weeks, I’ve heard the same five or six questions from customers, partners, and our own team. They’re good questions. They deserve honest answers, not the kind that make you feel better in the short term, but the kind that help you make better decisions.
Here they are.
“Is hyperscaler demand a bubble that’s going to burst — and when it does, won’t parts become available?”
The infrastructure being built right now is physically real. Data centers are under construction. Power grids are being commissioned. Chips are being installed. Even if AI revenue models disappoint, even if there’s a correction in AI stocks, that infrastructure doesn’t get unbuilt, and the components consumed to build it don’t come back.
More importantly, the world’s largest buyers, the US Hyperscalers, are currently receiving only 50–70% of the memory they actually need. They are running inventory levels as lean as six weeks. After prices increased more than 225% year over year, they have not reduced a single order. That is not bubble behavior. That is infrastructure urgency. A bubble bursts when demand is artificial. This demand is real, contracted, and attached to buildings in the ground.
“Everyone is double and triple ordering — won’t that inventory shake out and relieve supply in the second half?”
This was exactly the right call in 2021–2022. It is the wrong call in 2026.
In the last cycle, demand was consumer-driven and finite. Customers over-ordered to hedge, supply caught up, phantom demand evaporated, and the correction was fast. That mechanism worked because the excess demand was artificial.
This time, the demand at the top of the market is contracted infrastructure. When a hyperscaler places an order for server memory, it is tied to a data center under construction. That is not a hedge. And critically, the bottlenecks in this cycle are process inputs: packaging capacity, substrate laminate, and specialty gases. These are consumed in production. They don’t pile up in a warehouse and flood back into the channel. There is no phantom inventory waiting to shake out. The customers waiting for that correction are waiting for a mechanism that doesn’t exist in this cycle.
“Will die shrinks at the big three suppliers increase supply enough to matter?”
Die shrinks, moving to a smaller manufacturing process, which increases the number of chips you can produce from a single wafer. This is real, and it matters over time. But it is not a near-term solution for three reasons.
First, the transition takes a minimum of 12–18 months to qualify and ramp to volume. Second, the most advanced fab capacity is already fully allocated to AI accelerators; there is no spare leading-edge capacity available to be repurposed for memory-die shrinks. Third, and most importantly, demand growth is outpacing supply improvements. Even with die shrinks factored in, production growth is running at upper-teens percent year over year, while demand growth is tracking in the mid-twenties percent. The gap is not closing in 2026. Die shrinks help, but they are not a fix.
“Won’t these high prices eventually break the market and force the big three to lower prices?”
For consumer, PC and mobile, yes, this risk is real and building. PC and mobile OEMs are already planning price increases of 20% or more on end products and cutting SKUs for the second half of the year. A consumer demand correction is likely. When it comes, suppliers serving that segment will face pressure.
But here is what matters for your planning: the big three are actively remixing their wafer production away from consumer and toward server. When PC demand softens, suppliers don’t lower prices — they reduce consumer allocation and redirect capacity to hyperscalers who are paying more and still can’t get enough. The pressure on one end of the market becomes the relief valve for the other. Server memory pricing is effectively insulated from consumer demand destruction. Don’t confuse a PC/mobile correction with broad memory market relief.
“Will a 10% drop in consumer demand free up enough supply for everyone?”
No. Not even close, and the math is straightforward.
A 10% decline in PC and mobile DRAM frees up roughly 4–5% of total industry supply. Hyperscalers need 30–50% more than they’re currently getting. That consumer decline covers maybe one-seventh of the gap — on a good day. On the NAND side, it’s worse: hyperscalers want 75%+ growth in enterprise storage this year. Suppliers can deliver less than 50%. A 10% consumer decline releases 2–3% of total supply against a 25%+ structural gap.
There’s also a compatibility problem. The memory freed up by consumer softening — mobile LPDDR5, PC SO-DIMMs — is not the same product as what hyperscalers consume. Remixing wafer production toward server configurations takes quarters, not weeks. The bits don’t just move.
“So when does supply actually even out?”
Honestly, not in 2026 for server memory and enterprise storage. The new fab capacity coming online (TSMC Arizona, Samsung Taylor, Micron Idaho) is already substantially allocated and will come online gradually through 2027–2028. Die shrinks will add incremental supply, but won’t close the structural gap this year. HBM production will continue to cannibalize standard DRAM capacity as long as demand for AI GPUs remains strong.
The most likely scenario: selective improvement in specific categories through late 2026 into 2027, with genuine broad normalization in 2027–2028 at the earliest. Consumer NAND and PC DRAM may see some H2 relief if demand destruction materializes. Server DRAM and enterprise storage will remain structurally tight well into 2027.
Customers who plan around a 2026 normalization will be in a difficult position. Those planning for a 2027–2028 normalization — with extended horizons, allocation agreements, and qualified alternatives in place — will be in a fundamentally different and stronger position.
“Should I just buy from Chinese suppliers to get around the shortage?”
Chinese NAND producers, particularly YMTC, have become meaningful producers at competitive pricing. For consumer applications, it’s a real option. For most of our customers, enterprise, industrial, automotive, and networking, it is not a simple substitute. Western enterprise and industrial qualification processes, export control compliance, and supply chain risk policies effectively close off Chinese-sourced memory for most of the use cases we serve. The supply is real. For most of your programs, it isn’t accessible.
The bottom line:
The questions above all share a common assumption: that a correction is coming that will return this market to the conditions of 2019 or 2023. That assumption is worth examining carefully. The supply chain has been permanently restructured around AI infrastructure. The companies that act on that understanding now will be in a materially stronger position than the ones that don’t.
