Silicon Analysts Weekly for the week of 2026-09-28 to 2026-10-05: HBM4 contract price +117% WoW; Samsung enters NVIDIA qualification. A sourced, provenance-marked briefing on AI-chip supply-chain and pricing movements — HBM spot prices, wafer pricing, foundry and CoWoS capacity, per-chip bill-of-materials cost, and HBM qualification status — derived from public data and human-reviewed. Part of the Silicon Analysts semiconductor data layer.
The take
HBM4 contract pricing nearly doubled week-on-week — a signal that buyers are repricing scarcity risk as Micron confirms 75%+ of FY2027 supply is already committed and tightness runs through 2028. Samsung entering NVIDIA HBM4 qualification adds a second credible supplier to the race, which is where GPU market-share watchers should focus their attention. CoWoS capacity and HBM supply are now the two binding constraints on how fast AI infrastructure actually scales.
What moved this week
TSMC CoWoS pct locked — 2026 >85% of CoWoS capacity locked by Nvidia, Broadcom, and AMD combined… (source)
TSMC CoWoS lead time — now 52–78 weeks (was 52–78 weeks (one source cites 36–52 weeks)) (source)
Broadcom CoWoS customer allocation — now ~231,000–240,000 wafers/year (2026 allocation) (was 425,000 wafers in 2027) (source)
TSMC CoWoS capacity — ▲ now Q4 2026 ~120,000–140,000 wafers/month by end of 2026 (was ~70–80K wafers/month currently (2025); targeting ~127,000–130,000 wafers/month by end-2026) (source)
TSMC CoWoS CoWoS-L pct locked — now Q1 2026 70% (was no new bookings accepted beyond Q1 2027) (source)
HBM Price per GB by Generation — HBM4 (Contract — per-stack basis) — +116.6% WoW · full history
GOOGLE GEMINI 3 1 PRO PREVIEW output $/Mtok — −25.0% WoW · full history
ARIZONA industrial ¢/kWh — +21.0% WoW · full history
GEORGIA industrial ¢/kWh — +13.2% WoW · full history
Supply-chain & pricing signals
Micron Guides Tighter DRAM/HBM Supply Through 2028; No Equilibrium Visibility — Micron CEO Mehrotra stated supply-demand imbalance will persist through 2028 with HBM consuming disproportionate wafer capacity. 75%+ of FY2027 HBM supply already committed at higher prices signals structural tightness and margin expansion for memory oligopoly. (CHOSUNBIZ / Micron Earnings) · _In our data:_ Micron storyline · HBM price by generation · HBM market share by vendor
Micron posts record $54B Q4; locks $100B hyperscaler contracts for future supply — Structural shift from spot contracts to multi-year prepayments ($22B in advance commitments) fundamentally breaks the traditional memory cycle. This creates pricing power and supply certainty for hyperscalers but locks out smaller players and sustains elevated memory costs through 2028. (Techtimes) · _In our data:_ Micron storyline
HBM supply shortage to persist through 2028; capacity limited by packaging yields — CEO commentary confirms HBM will consume 30% of DRAM fab capacity by 2027 (vs 20% currently), severely constraining consumer DRAM supply. Production ramp delays and GAA node transitions limit wafer productivity gains, keeping enterprise memory margins elevated and consumer devices underfunded. (Ars Technica) · _In our data:_ HBM storyline · HBM price by generation · HBM market share by vendor
HBM watch
Samsung × NVIDIA (HBM4): now In qualification — Samsung reportedly passed NVIDIA's HBM4 prototype test; NVIDIA is requiring quality validation on both HBM3E and HBM4 before finalizing pricing with SK Hynix. (source · timeline)
Who Is Prepaying for AI Chips? The AI Hardware Bottleneck Map, October 2026 — The clearest sign of a supply squeeze is money paid in advance. In the second quarter of 2026, customer deposits rose 56% at TSMC and at Samsung, and Micron's rose about 23-fold the quarter after. SK hynix's barely moved. We graded eight AI hardware bottlenecks against filings and earnings calls, and three widely repeated claims did not survive the check.
groq-lpu perf_fp16_tflops — was 900 → now 188 (Spec correction, not a market move: Groq rates the Gen1 LPU at 188 TFLOPS FP16 (published specs, Feb 2024). We listed 900.)
mi325x perf_fp8_tflops — was 2614 → now 5230 (Basis correction, not a market move: FP8 here is the sparse rating, as for MI300X (5,230). AMD rates MI325X at 5,230 sparse; we listed the d)
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 — only when the data moves.
One short email a week — only when the data moves. No marketing, ever. One-click unsubscribe.