If you’ve quoted a laptop, a server, or a workstation in the last quarter, you’ve probably noticed that memory has gotten expensive. DRAM contract prices roughly doubled in the first quarter of 2026 alone. TrendForce had forecast a 50 to 60 percent jump and then revised it up to around 90 percent quarter over quarter, and retail has moved even harder: a 32GB DDR5 kit that was $80 to $110 in mid-2025 has been selling for $300 and up. Even older DDR4 has more than doubled, because production of it is being wound down to free capacity for the parts that pay better. This isn’t a rumor and it isn’t a short-term blip. Here’s the short version of why, written for operators who have to make purchasing decisions through it.
The supply side
Three manufacturers, Samsung, SK Hynix, and Micron, produce effectively all of the world’s DRAM. They’re all running at full capacity. The problem is that capacity is finite and can’t be added quickly. A new memory fab is a multi-year, multi-billion-dollar project, and none of the ones currently under construction come online before 2027.
The demand side
AI datacenters consume a specific kind of memory called High Bandwidth Memory (HBM), which sits directly on GPUs like NVIDIA’s H100 and B200. HBM commands much higher margins than conventional DRAM, so the manufacturers have shifted production toward it. Because HBM dies are stacked and use more silicon per bit, shifting one bit’s worth of capacity from conventional DRAM to HBM reduces the total output of conventional DRAM meaningfully.
SK Hynix has publicly stated its 2026 output (DRAM, HBM, and NAND) is effectively sold out, much of it to NVIDIA, and Micron has said the same about its HBM. AI buyers have locked up supply well into 2027. Samsung is in a similar position.
That leaves laptop RAM, server RAM, embedded systems, cars, and phones competing for the remaining capacity at higher prices.
What this means in practice
For end-user hardware. Consumer RAM has roughly tripled at retail. A 16GB DDR5 kit that was $40 is north of $100; a 32GB kit that was $90 is $300-plus. Pre-built PC and laptop prices haven’t risen by that much yet, but they’re climbing: Dell raised prices 15 to 20 percent in December 2025, Lenovo followed in January 2026, and OEMs are quietly shrinking default configs. A machine that shipped with 16GB and a 1TB SSD a year ago might ship with 8GB and 512GB now at the same price.
For servers and workstations. On-prem hardware costs are up. Many of our clients who were planning to refresh workstations or add storage in 2026 are seeing 20 to 40 percent total cost increases, more if the build is memory-heavy.
For cloud pricing. The major cloud providers buy at scale and insulate customers from short-term price spikes, but over the course of 2026 expect GPU-instance prices to creep up. Spot pricing is already volatile.
For embedded and industrial. Automotive, medical devices, and industrial automation all use meaningful amounts of memory and are affected. Vendors in those markets are quietly raising prices.
What to do if you’re buying
A short, practical list:
- Lock in pricing early if you can. Quotes that were valid for 30 days are now often valid for a week. Get purchase orders signed quickly.
- Budget higher. For 2026 capex, add a 20 to 40 percent buffer to memory and server line items.
- Reconsider cloud vs. on-prem. If you were on the fence, the higher cost of on-prem hardware shifts the math toward cloud for the next 12 to 18 months.
- Optimize what you already have. Quantization, smaller models, and batching reduce memory pressure. If you can run a 7B quantized model where you were using a 13B full-precision one, do.
- Don’t panic-buy. Prices are high but supply isn’t zero. If a vendor is pressuring you into a bigger order than you need, that’s a vendor problem.
When it resolves
Not quickly. The new fabs under construction complete in 2027 and ramp through 2028, so the relief everyone is counting on for 2027 is starting to look more like 2028. Until then, AI demand for HBM is expected to keep growing. A reasonable planning assumption is that memory prices stay elevated through 2027 and gradually normalize in 2028.
That’s a long enough window that it’s worth factoring into ongoing budgets rather than treating it as a one-time surcharge.
Update, August 2026. It kept climbing. A 32GB DDR5 kit that was $300 and up when we first wrote this is now regularly north of $400, and some DDR5-6000 kits that were under $90 a year ago are selling for around $529. Samsung, SK Hynix, and Micron have reportedly filled their 2027 production slots too, so the “relief in 2027” planning assumption is looking optimistic. If you’re still budgeting off the numbers earlier in this post, budget higher.
It’s not just memory, either. The same dynamic has hit storage; we wrote a follow-up on why SSDs got expensive too.
Update, September 11, 2026: the “normalizes in 2028” line in this post is wrong, and the correction comes from the supplier. That is the sentence in the When it resolves section above, and we are leaving it visible rather than quietly editing it, because how the estimate moved is more useful than the estimate.
SK hynix chief executive Kwak Noh-Jung now expects the memory shortage to run through the end of 2030. The company’s previous public guidance was tight supply through 2028, which is where our planning assumption came from. SK Group’s chairman has put the underlying arithmetic more bluntly: wafer supply is trailing demand by about 20 percent, and that is a gap fabs cannot close on a two-year build cycle. Kwak’s framing of the eventual exit is worth noting too, because it cuts against how memory markets have historically behaved: he expects a gradual easing rather than the sharp price crash that has ended every previous DRAM cycle, on the reasoning that AI customers are buying customised DRAM and HBM configurations that make demand far easier to forecast, which removes the overbuild that normally causes the crash.
The market data through 2026 is consistent with that. TrendForce put first-quarter DRAM industry revenue up 81 percent quarter over quarter, and forecast server DRAM contract prices rising another 13 to 18 percent quarter over quarter in the third quarter, with quarterly increases expected to continue from the second half of 2026 through the second half of 2027 at a moderating pace. The reason is supply, not enthusiasm: RDIMM bit supply is growing only 15 to 20 percent year over year while server CPU shipments grow faster. One detail worth knowing if you buy cloud rather than hardware, and it explains why your cloud bill has moved less than a retail RAM price: several large US cloud providers hold multi-year agreements that bar suppliers from raising prices to them at all. The increases land on everyone without such an agreement, which is everyone reading this.
On the retail side, a 32GB DDR5 kit is running roughly $375 to $400 as of early September, against $90 to $191 a year ago, so the August figures in the paragraph above have held rather than eased.
And there is a second-order effect now showing up that is worth watching. On September 10, Reuters reported that Chinese AI chipmakers are raising accelerator prices specifically because of HBM costs. Huawei has lifted the indicated price of its Ascend 950DT card above ¥250,000, roughly US$37,000, a jump of 20 to 50 percent from quotes given two months earlier, and Cambricon has repriced its next-generation part 20 to 30 percent higher, with MetaX and Iluvatar CoreX moving similarly. The mechanism is that HBM is a large share of an accelerator’s build cost, and since the December 2024 US export controls, Chinese buyers have been sourcing it through grey-market channels at several times the price paid elsewhere. That does not change what you pay for a DDR5 kit in Toronto. What it tells you is that the constraint has propagated all the way down the chain into finished compute, in the market with the most incentive to engineer around it, which is not what you would expect to see if this were a cycle rather than a structural gap.
The revised planning assumption. Treat elevated memory pricing as the baseline through 2027 at minimum, and budget on the supplier’s own number rather than ours: no meaningful normalisation before the end of the decade, easing gradually rather than collapsing. Practically, that changes two of the recommendations above. “Reconsider cloud vs. on-prem” is no longer a 12 to 18 month judgement call, it is a standing one. And “optimize what you already have” moves from a stopgap to the actual strategy, because the hardware you are waiting to get cheap is not going to.
If you’re sizing hardware or infrastructure for a project and want a second set of eyes on the numbers, we’re happy to help.
