Jun 22, 2026, 8:56 PM · InfraCharm

The Future of Memory: Scarcity and Tomorrow’s Technologies

The last few years have been a roller-coaster for memory buyers. In 2024 the market was full with cheap DDR4 and DDR5, but by mid-2025 prices were doubling every few months. By early 2026 many system builders discovered that RAM and NAND flash accounted for about one-third of the cost of a PC. Understanding why memory became so expensive and what lies ahead is critical for IT managers, system administrators, and business owners planning hardware upgrades.

How We Got Here

AI and High-bandwidth Memory (HBM) Displaced Commodity DRAM

The current crisis isn't just a supply blip—it's a structural shift. Generative AI and large-language models consume unprecedented amounts of memory. Each wafer allocated to HBM for data-centre GPUs yields roughly three times less capacity than the same wafer used for DDR5. Because HBM offers ten times the profit per wafer compared with commodity DRAM, manufacturers have prioritized AI customers. AI workloads now account for an estimated 60–70% of total DRAM output, starving the consumer and enterprise markets for standard modules.

Supply Chains Were Already Tight

Three companies—Samsung, SK Hynix, and Micron—produce 93% of global DRAM. Production ramps cannot happen overnight; new cleanrooms take years to build and qualify. Analysts note that new DRAM capacity won't come online until 2027 or later. Meanwhile, DDR4 production is being wound down to free up equipment for DDR5 and HBM, causing older modules to become scarce. Lead times for DDR4 and DDR5 modules now exceed 20–30 weeks in industrial markets.

Prices Have Soared

The combination of AI demand and limited supply has driven extreme price inflation:

DDR5 chip prices quadrupled between September 2025 and December 2025, climbing from roughly US$6.84 to US$27.20. DDR4 spot prices even traded above DDR5 because of scarcity.
NAND flash isn't immune. Kingston reported that NAND prices surged 246% from early 2025 to late 2025, with 70% of the increase occurring in just 60 days.
By early 2026, memory prices rose another 80–90% quarter-over-quarter and remain higher than contract rates. TrendForce's spot-price index shows DDR4 chips selling for more than US$33 and NAND wafers above US$20.
On the buyer side, HP's CFO noted that memory and storage jumped from 15–18% of a PC's bill of materials in 2025 to about 35% by early 2026. Industrial PC makers report that memory accounts for 15–25% of system cost.

Inventories Are Exhausted

DRAM inventories at suppliers shrank from 31 weeks in early 2023 to about 8 weeks by September 2025. Counterpoint Research observed that by Q1 2026 suppliers could only fulfill 55–60% of core customer demand. Some vendors issue quotes valid for only 72 hours and advise customers to plan for 10–20% monthly price increases through the end of 2026. Data centres, cloud providers, and PC vendors are stockpiling memory, exacerbating the shortage.

Market Outlook: Don't Expect Relief Before 2027

Multiple analyst firms converge on a similar forecast: memory prices will remain elevated throughout 2026 and may not normalize until late 2027 or 2028.

Forecast source	Key projection	Evidence
Gartner	Predicts a 130% surge in combined DRAM and SSD prices by the end of 2026, driving PC prices up 17%. Gartner doesn't foresee meaningful relief until Q1 2027 and expects entry-level PCs to disappear by 2028.	Gartner's February 2026 research notes that hyperscale data centres will consume over 40% of DRAM output.
TrendForce / Winbond	Reports that spot prices exceed contracts. Winbond expects DDR prices to quadruple by mid-2026, and TrendForce predicts 30–50% price hikes each quarter through H1 2026.	Spot prices for DDR4 chips reached US$33 and NAND wafers US$20.586.
IDC	Anticipates DRAM and NAND supply growth of just 16–17% in 2026, far below demand. IDC warns PC vendors to expect 15–20% price hikes, with PC selling prices rising 4–8% and smartphone BOMs seeing memory costs reach 10–20%.	IDC attributes the shortage to AI workloads diverting wafer capacity to HBM and expects supply constraints to persist through 2027.
Unisystem / industrial markets	Suggests data centres could consume up to 70% of total DRAM by 2026. Lead times exceed 20–30 weeks and supply may not stabilize until 2028.	Industrial PC makers are being asked to lock in orders months ahead and prepare for memory representing 15–25% of system BOM.

Together, these forecasts paint a bleak short-term picture. Even if new fabs ramp up in 2027, most capacity is earmarked for HBM and AI customers. Analysts widely agree that mid-2025 pricing is unlikely to return.

Beyond Price: The Next Generation of Memory

While budgets are squeezed today, future technologies promise bigger, faster, and more efficient memory.

DDR6 and LPDDR6

JEDEC started standardizing DDR6 in 2024, completing the main draft by late 2024 and releasing a low-power LPDDR6 draft in mid-2025. The new standard will start at 8,800 MT/s and scale to 17,600 MT/s, roughly 2–3× the bandwidth of DDR5. DDR6 improves energy efficiency by 15–20%, uses a 4×24-bit sub-channel architecture and a CAMM2 interface, and introduces a 16n prefetch mechanism. These changes allow more parallelism, higher density, and slimmer form factors—ideal for laptops and servers when the technology matures later this decade.

Compute Express Link (CXL) Memory Pooling

Traditional servers have fixed memory attached to a single CPU. CXL is a high-speed interconnect that allows memory to be disaggregated and pooled, with latencies of 200–500 ns—far faster than accessing NVMe storage. In demonstrations using two servers with NVIDIA H100 GPUs, CXL memory pooling delivered 3.8× the performance of 200 Gb/s RDMA and 6.5× the performance of 100 Gb/s RDMA for large language model inference. Platforms like Liqid can provide up to 100 TiB of pooled memory. For AI workloads where memory is a bottleneck, CXL enables dynamic scaling without physically adding DIMMs, reducing costs and improving utilization.

Emerging Packaging and Form Factors

New form factors such as CAMM2 and LPCAMM2 (Low-Power Compression Attached Memory Module) aim to replace traditional SO-DIMMs in laptops. They reduce height and improve thermal performance. HBM4 is expected later this decade, promising higher stacks and improved yields. Meanwhile, non-volatile memory technologies like MRAM and PCM are under research for future storage-class memory.

Practical Guidance for System Administrators

Given the current environment, IT professionals should plan carefully:

Audit and optimize existing workloads. Before buying new RAM at peak prices, reclaim unused memory by tuning operating systems, disabling unnecessary services, and right-sizing virtual machines. This can delay upgrades and free capacity.
Plan for longer lead times. Expect 20–30 week lead times. Place orders early, maintain safety stock, and consider multi-sourcing to hedge against supplier delays.
Evaluate DDR4 vs DDR5 platforms. For many workloads the performance gap is modest. DDR4 platforms may be more cost-effective until prices normalize. Conversely, high-performance workloads may justify DDR5 or early DDR6 adoption when available.
Consider modular and disaggregated architectures. Explore CXL-enabled servers or memory pooling solutions that allow memory resources to be shared across nodes. These technologies can deliver more memory per node without buying more DIMMs.
Monitor vendor roadmaps. Track announcements from Samsung, SK Hynix, and Micron for new capacity expansions. Understand when your preferred supplier plans to ramp up DDR6 or HBM4 and adjust procurement accordingly.
Stay informed about security advisories. High-profile vulnerabilities sometimes drive emergency upgrades that coincide with memory shortages. Incorporate security patch planning into your memory strategy.

Conclusion

Memory markets are undergoing a once-in-a-generation transformation. AI and data-centre demand have created a structural shortage, driving DRAM and NAND prices up several hundred percent since 2025. Supply constraints, concentrated production, and a shift toward HBM mean the shortage will likely persist through 2027. Meanwhile, future technologies like DDR6 and CXL memory pooling promise dramatic improvements in bandwidth, efficiency, and flexibility, enabling new levels of performance once they reach maturity. For now, the smartest strategy is to optimize existing resources, plan procurement carefully, and keep abreast of emerging innovations.