Several dynamics are driving this:

• Data movement is the real constraint. In hyperscale data centers, the energy and latency cost of shuttling data between memory and processors now dictates AI performance more than the processor's peak theoretical throughput .
• Model architectures demand more memory, not more FLOPS. Mixture-of-Experts (MoE) models and long-context processing have sharply increased capacity requirements for model weights and KV caches, making memory capacity—not compute—the bottleneck .
• HBM has become the strategic centerpiece. Industry attention has pivoted from individual-chip compute benchmarks to system-level memory performance. High-Bandwidth Memory is no longer a supporting component; it's the critical path .
• Markets have already priced this in. Memory stocks significantly outperformed AI chip stocks in 2025, signaling that investors understand the supercycle now belongs to memory, not logic .

Inside the shortage: prices, supply-demand gaps, and why consumers are paying

The 2024–present memory shortage—dubbed "RAMmageddon" or the "RAMpocalypse"—is unlike the pandemic-era chip crunch. It's not a short-term supply-chain disruption. It's a structural reallocation of manufacturing capacity toward high-margin AI memory products, starving consumer and enterprise PC markets .

AI data centers are expected to consume roughly 70% of global high-end DRAM production in 2026 . Hyperscalers such as Meta, Alphabet, and Microsoft are each spending $70–93 billion-plus annually on AI infrastructure, securing memory supply years in advance .

That leaves consumer-grade memory—DDR4, DDR5, and NAND flash—scrambling for the remaining 30% of output. The result has been brutal pricing:

• 64GB DRAM module: Rose from approximately $255 in Q3 2025 to over $420 by May 2026. Estimates point toward $700 by early 2027 .
• Consumer DDR4 and DDR5 prices: Jumped roughly 4x between September and November 2025 alone, according to Deloitte's 2026 semiconductor outlook .
• Overall RAM costs: Increased 172% over the trailing year as of early 2026 .
• Standard DRAM contract prices: Rose nearly 60% by the end of 2025, with further double-digit gains expected through 2026 .

Analysts broadly expect prices to keep climbing through at least late 2026. IDC projects the shortage could persist into 2027, with PC shipments potentially declining up to 9% year-on-year in a worst-case scenario .

Consumer electronics: what the shortage means for real prices

The knock-on effects are already visible on store shelves:

• PC prices have risen 15–20% as manufacturers pass through memory cost inflation. Dell, Lenovo, and HP have all warned of planned price increases, some as steep as 15–30% .
• Smartphones and laptops face higher prices and limited upgrade options in 2026. IDC has cautioned that the effects on consumer device pricing and availability could stretch well into 2027 .
• Consumer demand is beginning to buckle. With RAM costs up 172% year-over-year and PC prices climbing, consumers are hesitating on purchases—a classic demand-destruction signal .
• Edge AI devices are being redesigned around memory constraints. Product decisions are now driven by memory availability and cost rather than processor TOPS or sensor specifications . Raspberry Pi raised prices on several models in December 2025, explicitly citing the unprecedented rise in LPDDR4 memory costs .

The structural nature of this shortage makes it hard to resolve quickly. IDC expects 2026 DRAM supply growth to be only 16% year-on-year, below historical norms, while demand continues to outpace expansion .

Sandisk's HBF: a NAND-based answer to the memory wall

Sandisk has been developing an architecture called High-Bandwidth Flash (HBF) specifically to address the memory wall from the capacity side .

HBF is a NAND-based alternative designed to sit in the same package as the GPU, giving the processor fast access to a much larger pool of memory than HBM alone can provide. Where high-end HBM stacks today top out in the tens of gigabytes, HBF aims to deliver 8–16x the capacity at comparable bandwidth and similar total cost .

Key technical points:

• Architecture: HBF uses a 16-die-plus-base-die NAND stack that fits the same physical footprint as HBM, though with a different interface protocol. The goal is to keep large AI model parameters inside the GPU package rather than fetching them from PCIe-attached SSDs, which would be orders of magnitude slower .
• Target workloads: The technology is purpose-built for AI inference across large data centers, small enterprises, and edge applications .
• Differentiator: By trading NAND's higher latency for dramatically higher capacity and non-volatility, HBF is meant to augment HBM—not replace it—in systems where inference models have outgrown available DRAM capacity .

Timeline and standardization progress

Sandisk has been moving HBF from lab to industry standard through a series of deliberate steps:

• August 2025: Sandisk signed a memorandum of understanding with SK hynix—the dominant HBM supplier—to jointly define the HBF specification .
• August 2025 (same period): Sandisk submitted HBF technology to the Open Compute Project (OCP) for open standardization .
• February 2026: The two companies held an "HBF Standardization Kick-Off" at Sandisk's global headquarters, establishing a dedicated OCP workstream to advance the standard .
• Sampling roadmap: Sandisk plans to deliver first HBF memory samples in the second half of 2026, with the first AI inference devices using HBF expected in early 2027 .

HBF is still in the specification and validation phase; there is no announced commercial production timeline beyond the sampling targets. But the partnership with SK hynix and the push through OCP signal an intent to make HBF an open, multi-vendor standard rather than a proprietary Sandisk product .

HBF represents one of the first memory architectures designed from the ground up for the inference era. Whether it can close the memory wall at scale won't be clear until real silicon ships, but the direction is unambiguous: the AI industry is being rebuilt around memory, not just logic.

Several dynamics are driving this:

• Data movement is the real constraint. In hyperscale data centers, the energy and latency cost of shuttling data between memory and processors now dictates AI performance more than the processor's peak theoretical throughput .
• Model architectures demand more memory, not more FLOPS. Mixture-of-Experts (MoE) models and long-context processing have sharply increased capacity requirements for model weights and KV caches, making memory capacity—not compute—the bottleneck .
• HBM has become the strategic centerpiece. Industry attention has pivoted from individual-chip compute benchmarks to system-level memory performance. High-Bandwidth Memory is no longer a supporting component; it's the critical path .
• Markets have already priced this in. Memory stocks significantly outperformed AI chip stocks in 2025, signaling that investors understand the supercycle now belongs to memory, not logic .

Inside the shortage: prices, supply-demand gaps, and why consumers are paying

The 2024–present memory shortage—dubbed "RAMmageddon" or the "RAMpocalypse"—is unlike the pandemic-era chip crunch. It's not a short-term supply-chain disruption. It's a structural reallocation of manufacturing capacity toward high-margin AI memory products, starving consumer and enterprise PC markets .

AI data centers are expected to consume roughly 70% of global high-end DRAM production in 2026 . Hyperscalers such as Meta, Alphabet, and Microsoft are each spending $70–93 billion-plus annually on AI infrastructure, securing memory supply years in advance .

That leaves consumer-grade memory—DDR4, DDR5, and NAND flash—scrambling for the remaining 30% of output. The result has been brutal pricing:

• 64GB DRAM module: Rose from approximately $255 in Q3 2025 to over $420 by May 2026. Estimates point toward $700 by early 2027 .
• Consumer DDR4 and DDR5 prices: Jumped roughly 4x between September and November 2025 alone, according to Deloitte's 2026 semiconductor outlook .
• Overall RAM costs: Increased 172% over the trailing year as of early 2026 .
• Standard DRAM contract prices: Rose nearly 60% by the end of 2025, with further double-digit gains expected through 2026 .

Analysts broadly expect prices to keep climbing through at least late 2026. IDC projects the shortage could persist into 2027, with PC shipments potentially declining up to 9% year-on-year in a worst-case scenario .

Consumer electronics: what the shortage means for real prices

The knock-on effects are already visible on store shelves:

• PC prices have risen 15–20% as manufacturers pass through memory cost inflation. Dell, Lenovo, and HP have all warned of planned price increases, some as steep as 15–30% .
• Smartphones and laptops face higher prices and limited upgrade options in 2026. IDC has cautioned that the effects on consumer device pricing and availability could stretch well into 2027 .
• Consumer demand is beginning to buckle. With RAM costs up 172% year-over-year and PC prices climbing, consumers are hesitating on purchases—a classic demand-destruction signal .
• Edge AI devices are being redesigned around memory constraints. Product decisions are now driven by memory availability and cost rather than processor TOPS or sensor specifications . Raspberry Pi raised prices on several models in December 2025, explicitly citing the unprecedented rise in LPDDR4 memory costs .

The structural nature of this shortage makes it hard to resolve quickly. IDC expects 2026 DRAM supply growth to be only 16% year-on-year, below historical norms, while demand continues to outpace expansion .

Sandisk's HBF: a NAND-based answer to the memory wall

Sandisk has been developing an architecture called High-Bandwidth Flash (HBF) specifically to address the memory wall from the capacity side .

HBF is a NAND-based alternative designed to sit in the same package as the GPU, giving the processor fast access to a much larger pool of memory than HBM alone can provide. Where high-end HBM stacks today top out in the tens of gigabytes, HBF aims to deliver 8–16x the capacity at comparable bandwidth and similar total cost .

Key technical points:

• Architecture: HBF uses a 16-die-plus-base-die NAND stack that fits the same physical footprint as HBM, though with a different interface protocol. The goal is to keep large AI model parameters inside the GPU package rather than fetching them from PCIe-attached SSDs, which would be orders of magnitude slower .
• Target workloads: The technology is purpose-built for AI inference across large data centers, small enterprises, and edge applications .
• Differentiator: By trading NAND's higher latency for dramatically higher capacity and non-volatility, HBF is meant to augment HBM—not replace it—in systems where inference models have outgrown available DRAM capacity .

Timeline and standardization progress

Sandisk has been moving HBF from lab to industry standard through a series of deliberate steps:

• August 2025: Sandisk signed a memorandum of understanding with SK hynix—the dominant HBM supplier—to jointly define the HBF specification .
• August 2025 (same period): Sandisk submitted HBF technology to the Open Compute Project (OCP) for open standardization .
• February 2026: The two companies held an "HBF Standardization Kick-Off" at Sandisk's global headquarters, establishing a dedicated OCP workstream to advance the standard .
• Sampling roadmap: Sandisk plans to deliver first HBF memory samples in the second half of 2026, with the first AI inference devices using HBF expected in early 2027 .

HBF is still in the specification and validation phase; there is no announced commercial production timeline beyond the sampling targets. But the partnership with SK hynix and the push through OCP signal an intent to make HBF an open, multi-vendor standard rather than a proprietary Sandisk product .

HBF represents one of the first memory architectures designed from the ground up for the inference era. Whether it can close the memory wall at scale won't be clear until real silicon ships, but the direction is unambiguous: the AI industry is being rebuilt around memory, not just logic.