CXMT hits 10% DRAM share; report: Samsung to use TSMC for NVIDIA custom-HBM base dies
Silicon Analysts Weekly for the week of 2026-10-12 to 2026-10-19: CXMT hits 10% DRAM share; report: Samsung to use TSMC for NVIDIA custom-HBM base dies. A sourced, provenance-marked briefing on AI-chip supply-chain and pricing movements — HBM pricing, wafer pricing, foundry and CoWoS capacity, per-chip bill-of-materials cost, and HBM qualification status — derived from public data. Part of the Silicon Analysts semiconductor data layer.
Correction: Last week's subject line, "HBM3E price −90.5% WoW", was wrong. Our data agent read a price quoted per gigabit as a price per gigabyte (an 8× unit error) and filed it under the wrong period. HBM3E prices did not fall that week. Our first item, "Samsung × NVIDIA Vera Rubin (HBM3E 12-Hi) now Volume shipping", was also wrong: Vera Rubin uses HBM4, not HBM3E, and no independent source tied those shipments to NVIDIA. Issue #16's "RunPod H200 +27.9% WoW" compared two different RunPod price tiers; the price did not rise. We have fixed all three and added checks to hold items like these for review. (correction log)
The take
One unconfirmed report (Seoul Economic Daily, Oct 18) says NVIDIA wants all three NVHBM suppliers—Samsung, SK Hynix, and Micron—to fab the custom-HBM base die on TSMC 3nm; Samsung reportedly agreed, though it still makes standard HBM base dies in-house. If confirmed, that's a meaningful concentration risk for next-gen accelerator supply. CXMT's jump to 10% global DRAM share from 5% is real, but readers tracking NVIDIA GPU market share should note export-control exposure makes that trajectory fragile.
What moved this week
COREWEAVE H200 spot $/GPU-hr — +1.2% WoW · full history
DDR5 16GB SPOT spot $/unit — −0.9% WoW · full history
Supply-chain & pricing signals
Samsung Outsources NVHBM Base Dies to TSMC as SF2 Yields Fall 50% Short — Samsung's foundry yields on SF2 process (50-60%) lag TSMC N3 (90%+), forcing all three HBM vendors (Samsung, SK Hynix, Micron) to rely on TSMC for Nvidia's NVHBM base dies. This marks a critical erosion of Samsung's vertical integration in premium AI memory and consolidates TSMC's foundry dominance for next-gen accelerator architectures. (Techtimes) · _In our data:_ SK Hynix — 2 developments tracked this quarter · Market Intelligence · HBM price by generation · HBM market share by vendor
China's CXMT Captures 10% Global DRAM Share; Profit Swing Masks Export Reliance — CXMT forecasts 131-132.5B yuan profit (Jan-Sep 2026) vs. 5.28B yuan loss year-ago, driven by AI DRAM demand surge and price appreciation. Market share jumped from 5% to 10%, yet stock fell 6.5% as investors question sustainability post-cycle and tariff/export control risks from Western chip competition. (CNBC) · _In our data:_ Market Intelligence
Microsoft Exits Nscale UK AI Data Center; Power Grid Delays Until 2030s — Microsoft withdrawal from £2B Loughton AI data center due to UK grid supply delays from end-2026 to 2030+ signals critical infrastructure bottleneck constraining accelerator deployment in Europe. Power grid modernization deficits now explicitly blocking hyperscaler expansion, rippling through semiconductor demand forecasts. (BBC) · _In our data:_ Market Intelligence
HBM Price per GB by Generation — HBM3E (headlined in issue #18) — was $1.55/GB (H2 2025) → now $16.4/GB (2025); the H2 2025 point is removed (Not a 90.5% fall: a July 2026 per-gigabit asking price ($1.5-1.6/Gb, about $12.4/GB) was stored as $/GB and filed under H2 2025.)
RUNPOD H200 on-demand $/GPU-hr (headlined in issue #16) — was $4.59 / $4.00 / $4.31 (other RunPod tiers and pages) → now $3.59/GPU-hr, Community Cloud (RunPod lists H200 on demand at $3.59 (Community Cloud) and $4.59 (Secure Cloud). Our check switched tiers; we show Community.)
Samsung × NVIDIA Vera Rubin (HBM3E 12-Hi) (headlined in issue #18) — was Volume shipping @ 2026-01-01 → now withdrawn (The cited article (Sisa Journal-e, 2026-01-08) reports that Samsung had begun supplying 12-high HBM3E without naming a customer, and that…)
capacity readings (customer allocation) — was 26 published readings in 10 series whose only sources were social posts, forum threads, this site or a search page → now all withdrawn; 8 re-published as 5 new readings citing the original publisher (Withdrawn: these readings cited only social posts, forums, our own pages or a search page. Verified originals are now cited as new readings.)
capacity readings (ramp) — was 14 published readings in 10 series whose only sources were social posts, forum threads, this site or a search page → now all withdrawn; 5 re-published as 4 new readings citing the original publisher (Withdrawn: these readings cited only social posts, forums, our own pages or a search page. Verified originals are now cited as new readings.)
capacity readings (utilization) — was 28 published readings in 11 series whose only sources were social posts, forum threads, this site or a search page → now all withdrawn; 12 re-published as 4 new readings citing the original publisher (Withdrawn: these readings cited only social posts, forums, our own pages or a search page. Verified originals are now cited as new readings.)
Citation & licensing. Figures from this issue are free to cite — in articles, research, internal decks, and AI-generated answers — with attribution to Silicon Analysts and a link to this page. Commercial redistribution and derived data products require a license. See the Citation & Data License.
Get the next issue in your inbox
One short, sourced email a week on the AI-chip supply chain.
One short email a week — sourced and dated. No ads. One-click unsubscribe.
Sourced public intelligence, not insider data. Qualification status changes publish automatically once corroborated, not after human review. Full timelines on the HBM Qualification Tracker.