Good for ↑
Samsung forecasts memory shortage extending to 2028, driven by AI accelerator demand. Market focuses on DRAM producers. The real trade is upstream in advanced packaging: HBM memory requires complex CoWoS/SoIC interposer work, and that substrate/packaging capacity is the true bottleneck. Long US-listed advanced packaging and substrate suppliers.
Source: TheGamer ↗
The first-order read — 'memory shortage is good for Samsung and SK Hynix' — is already priced into Korean chipmakers and widely discussed. The consensus is also long Nvidia because HBM scarcity constrains GPU supply. That is the obvious chain and it is crowded. The non-obvious chain: HBM3e and HBM4 are not just memory dies — they are complex 3D-stacked assemblies that require CoWoS (Chip-on-Wafer-on-Substrate) interposer integration at TSMC, and advanced organic substrates from a small number of suppliers. Every additional HBM unit Samsung ships into the AI supply chain must pass through this packaging bottleneck. US-listed companies providing advanced packaging services, substrate materials, and test equipment for HBM integration are capacity-constrained and under-discussed. These are mid-cap US semiconductor equipment and materials companies — names in advanced substrate supply (think ABF substrate materials), chip-packaging test systems, and thermal interface materials for stacked memory — that see direct volume uplift when HBM shipments grow. The market context supports the setup: XLK is up only 0.6% this week despite the memory shortage headline, suggesting the tech sector hasn't broadly re-rated on this news. VIX at 16 (down 13.9% on the week) signals declining fear, which supports risk-on positioning in growth-adjacent industrials. The 10Y yield at 4.74% is a headwind for long-duration growth names, but mid-cap capital equipment and materials companies with near-term earnings visibility are less rate-sensitive than pure software. The falling dollar (DXY -1.6% on the week) also helps US companies with foreign revenue from Asian customers paying in stronger local currencies.
The market believes the memory shortage is primarily a tailwind for HBM producers (Samsung, SK Hynix) and GPU designers (Nvidia), with the trade being long those names. Advanced packaging is seen as a TSMC story, already priced into TSMC's premium valuation.
Why that may be wrong
Slow diffusion of information: the HBM packaging supply chain is technically complex and covered primarily by semiconductor equipment specialists, not the generalist tech analysts who dominate the narrative. Most buy-side models treat packaging as a cost line for TSMC rather than a standalone capacity constraint. The result is that substrate and advanced packaging suppliers trade at a discount to their actual throughput leverage on the HBM shortage cycle.
If AI accelerator demand softens materially in H2 2026 — signaled by hyperscaler capex guidance cuts in Q3 2026 earnings — HBM order books would thin and the packaging bottleneck would ease before the thesis plays out. A specific observable signal: if Microsoft, Google, or Amazon trim 2027 datacenter capex in their next earnings calls, the entire HBM demand stack deflates.
The strongest bear case is that Samsung's shortage forecast is self-serving guidance designed to lock in long-term supply contracts at favorable pricing — a classic oligopoly signaling move, not a genuine supply analysis. If Samsung is overstating the shortage to justify price increases, actual HBM supply could normalize faster than expected, particularly as SK Hynix and Micron both aggressively expand HBM capacity. The specific assumption most likely to be wrong is that packaging remains the binding constraint: TSMC has been investing heavily in CoWoS expansion, and if that capacity comes online faster than the market expects (observable via TSMC capacity announcements in Q3/Q4 2026), the packaging bottleneck thesis collapses before the trade pays.
The 2016-2018 DRAM supercycle offers the closest analog. As DRAM shortages deepened in 2017, the market initially focused on Samsung and Micron. But the second-order winner was companies like Entegris and Cabot Microelectronics (now CMC Materials), which supplied process chemicals and CMP slurries — their stocks doubled over 18 months while the memory names saw more volatile, commodity-like returns. The pattern: shortage narratives draw capital to the most visible producers, but the durable margin expansion lives in the picks-and-shovels layer where switching costs are high and capacity additions take 2-3 years. HBM packaging is structurally more complex than standard DRAM, making the bottleneck even stickier this cycle.
Samsung forecasts memory shortage extending to 2028, validating sustained AI-driven HBM demand. Memory producer stocks and Nvidia benefit from the supply-scarcity narrative.
Every incremental HBM unit must be assembled via advanced CoWoS-style interposer packaging before it reaches an AI accelerator. TSMC's CoWoS capacity is already cited as a bottleneck; substrate suppliers (ABF substrates, organic interposers) and packaging test equipment vendors face the same constraint with less capital to expand quickly.
US-listed mid-cap semiconductor materials, substrate, and packaging test companies — those supplying advanced organic substrates, thermal interface materials for stacked dies, and burn-in/test systems for HBM modules — see multi-year volume commitments from HBM assemblers. Their revenue visibility extends to 2028 alongside the shortage forecast, but their valuations have not re-rated because generalist analysts model them as cyclical equipment names rather than structural AI infrastructure plays. The falling dollar adds a currency tailwind as Asian customer payments translate back at favorable rates.
No comments yet.
Sign in to comment
Earlier connections in the same sector that share a theme with this one.