Every AI accelerator on Earth — Nvidia's, AMD's, Google's own — is currently sitting idle some fraction of the time, waiting for data. High-bandwidth memory is sold out through 2026 across all three global suppliers. This letter argues the real scaling constraint on AI was never just compute — it is moving data less, and that shift is quietly making memory makers the more durable winners.
While the world has spent 2025 and 2026 fixated on GPU shortages, a more fundamental constraint has been quietly reshaping the AI hardware industry: high-bandwidth memory, or HBM, is sold out through 2026 across every major supplier, with 2027 allocations already being negotiated. This letter is about why that matters more than the next chip announcement.
Every AI chip — Nvidia's H100, H200, and B200, AMD's MI300X and MI325X, Google's TPU v5p and v7 series — depends on HBM to feed data to its compute cores fast enough to keep them busy. Without sufficient HBM, the most powerful processor in the world sits idle waiting for data to arrive. HBM delivers 5-10x the bandwidth of standard GDDR memory through 3D-stacked DRAM connected via through-silicon vias — genuinely different engineering from the memory in a consumer laptop, not just a faster version of the same thing.
The supply side is a structural oligopoly: SK Hynix, Samsung, and Micron control over 95% of global DRAM production between them, and as of 2026 have collectively reallocated roughly 93% of combined production capacity toward HBM for AI data centers, according to IDC. HBM now consumes 23% of total DRAM wafer output, up from 19% in 2025 — and producing a single bit of HBM requires roughly 300% more wafer capacity than equivalent DDR5, a structural, physics-level bottleneck that no amount of capital spending resolves quickly.
SK Hynix holds roughly 50-62% of HBM market share depending on the measure, having secured early exclusive supply agreements with Nvidia for the H100 and H200 generations. Micron's turnaround has been the most dramatic: fiscal Q1 2026 revenue reached $13.64 billion, up 57% year-over-year, with gross margins climbing above 50% — more than double the roughly 22% margins the company posted in fiscal 2024. This is not cyclical recovery. It is structural transformation in product mix, and Micron has gone so far as to exit the consumer memory and storage market entirely to concentrate on AI and enterprise customers, a decision that would have been unthinkable three years ago.
Bank of America has explicitly labelled 2026 a "supercycle similar to the boom of the 1990s," forecasting global DRAM revenue growth of 51% and NAND growth of 45% year-over-year, with average selling prices up 33% and 26% respectively. HBM demand itself is projected to grow 70% year-over-year in 2026 alone.
The reallocation of capacity toward HBM has a direct, measurable victim: everyone buying consumer electronics. DRAM prices surged roughly 90% in Q1 2026 alone compared to Q4 2025. Samsung raised prices on 32GB DDR5 modules from $149 to $239 — a 60% increase — in a single pricing cycle, and contract DDR5 pricing has more than doubled from around $7 to $19.50 per unit. Smartphone shipments are projected to decline 12.9% in 2026 and the PC market faces an 11.3% contraction, directly attributable to component cost inflation flowing from the AI memory reallocation. This is the AI boom's least-discussed negative externality: it is currently taxing every consumer electronics purchase on the planet to subsidise the compute buildout.
"HBM production consumes triple the wafer capacity of standard DRAM per gigabyte — meaning even as total memory manufacturing capacity grows, the share available for anything other than AI data centers keeps shrinking, not growing."
HBM4, entering mass production in 2026 across all three suppliers, pushes bandwidth past 1.5 terabytes per second per stack with 64GB capacity — roughly double the throughput of the prior generation. But the more structurally interesting shift sits one layer further out: Processing-in-Memory (PIM) architectures, which embed compute logic directly inside the HBM base die rather than shuttling data back and forth to a separate processor. Published engineering analysis shows PIM-HBM configurations achieving a 53% performance gain and 10.4% energy efficiency improvement versus traditional GPU-HBM setups — a genuine architectural shift, not an incremental speed bump, and one that inverts the entire historical logic of computing: instead of moving data to the compute, move the compute to the data.
This is the parallel to the CPU-to-GPU shift of the 2010s that this letter's framing invokes deliberately: an unglamorous infrastructure layer quietly becomes the actual scaling constraint, and the market spends years mispricing the shift before it becomes consensus.
The AI scaling story the market has priced is "more compute wins." The scaling story actually playing out in 2026 is "whoever moves data most efficiently wins" — and that favours the memory makers (SK Hynix, Samsung, Micron) and the advanced-packaging enablers (TSMC's CoWoS platform, the dominant integration vehicle for GPU-HBM assembly) at least as much as it favours the chip designers getting most of the market's attention. Micron's structural margin expansion — doubling gross margins in under two years — is the clearest evidence this repricing is already underway, not a future thesis. The risk: new capacity from Samsung's Taylor, Texas fab and SK Hynix's Indiana packaging plant will not meaningfully relieve supply until 2027-2028 at the earliest, meaning this remains a multi-year, not a multi-quarter, structural position — and any technology that reduces memory-bandwidth requirements per unit of AI compute (efficient model architectures, better quantisation) is the single biggest risk to the thesis holding at current pricing power.
Founder, NextGen Economics · Bangalore, India · July 2026
Sources: Introl (AI Memory Supercycle analysis, Jan 2026) · SK Hynix 2026 Market Outlook · Patsnap (HBM Technology Landscape 2026) · TrendForce & IDC (DRAM wafer allocation data) · Network World / Gartner (Samsung memory shortage pricing, 2026) · DataCenterDynamics (Samsung & SK Hynix capacity expansion) · Bank of America research (memory supercycle forecast) · Micron Technology fiscal Q1 2026 results.
Not investment advice. This letter evaluates a semiconductor sub-sector; it does not constitute a recommendation regarding any security. Semiconductor markets are cyclical and pricing power can compress rapidly if new capacity outpaces demand.