The next AI infrastructure bottleneck may be the memory sitting beside the accelerator rather than the accelerator itself. Samsung Electronics said on 29 October 2026 that demand for AI-related memory in 2027 is expected to exceed industry supply by a wider margin than in 2026.
The statement came with Samsung’s third-quarter results. Its memory business reported record quarterly revenue and profit, increased HBM4 sales and predicted another historic year for AI infrastructure investment.
This is a manufacturer’s outlook, not an independently guaranteed shortage. Demand forecasts can change quickly if data-centre projects slip, customers redesign systems or competitors expand capacity faster than expected.
HBM feeds data to expensive processors
High-bandwidth memory, or HBM, stacks multiple layers of DRAM close to a GPU or other accelerator. Thousands of connections move data in parallel, reducing the time processors spend waiting for model weights and intermediate results.
HBM4 doubles the interface width from 1,024 to 2,048 data pins. Samsung’s product page claims up to 3.3 terabytes per second of bandwidth per stack, using 12-layer packages with capacities from 24 to 36 gigabytes. The company plans 16-layer options up to 48 GB.
Those are vendor specifications. Real system performance also depends on the accelerator, memory controller, interconnect, software and workload. More bandwidth does not guarantee proportionally faster model training or inference.
The supply problem is more than making DRAM
An HBM stack combines advanced DRAM dies, through-silicon vias, a logic base die and sophisticated packaging. Yield at any step can constrain finished output. Capacity also depends on testing equipment, substrates and the ability to integrate the memory with accelerator packages.
Samsung says its HBM4 uses sixth-generation 10-nanometre-class DRAM and a 4-nanometre logic base die. Its production announcement claims 11.7 gigabits per second as a sustained pin speed, with headroom to 13 Gbps, plus a 40% power-efficiency improvement over HBM3E.
Again, those comparisons come from Samsung. Customer qualification, shipped volume and independent system benchmarks will show how the product performs at scale.
Samsung is spending heavily to expand the pipeline
The quarterly report puts Samsung’s expected 2026 capital expenditure at approximately 84.4 trillion won, including 76.8 trillion won for its Device Solutions division. The company says this supports new fabrication plants and infrastructure for AI and high-performance computing demand.
It also expects its new Taylor, Texas fabrication plant to begin operating in 2027. Foundry capacity matters because HBM4 uses a logic base die as well as stacked memory, bringing memory, advanced logic and packaging into one supply chain.
Multi-year customer contracts can provide manufacturers with confidence to invest, but they can also reserve capacity for the largest buyers and leave smaller cloud providers or equipment makers with less flexibility.
A shortage would reshape systems, not stop AI
If high-end HBM remains constrained, customers may prioritise their most valuable models, optimise software to use memory more efficiently, choose accelerators with different memory configurations or extend the life of existing hardware.
The pressure will not necessarily affect every memory product equally. HBM is specialised and packaging-intensive; a tight HBM market does not automatically mean consumer laptop memory is scarce.
What would confirm the forecast
Evidence of a persistent shortage would include long lead times, high utilisation across multiple suppliers, customer prepayments, packaging constraints and prices that remain elevated despite new capacity. Evidence against it would include delayed AI projects, improving yields or faster-than-expected supply expansion.
Samsung’s result matters because it identifies memory as a system-level limit. But the 2027 gap remains a projection from a company selling into that market—not a settled fact and not investment advice.
Featured image: Umberto/Unsplash. Contextual circuit-board photograph; not Samsung HBM4.
Leave a Reply