Silicon Analysts Weekly · Issue #16 · week of 2026-09-22 → 2026-09-29

RunPod H200 +27.9% WoW; HBM4 qual slips to late-Q1 2026

Silicon Analysts Weekly for the week of 2026-09-22 to 2026-09-29: RunPod H200 +27.9% WoW; HBM4 qual slips to late-Q1 2026. 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

RunPod H200 spot pricing jumped 27.9% week-on-week while AWS H100 fell 12.3%, a split that reflects real-time GPU-tier rotation as buyers chase available Blackwell-adjacent capacity. Separately, TSMC N5 wafer prices rose 8.1% WoW, adding another cost layer to an already tight CoWoS stack. With HBM pricing the most-watched topic this week, the HBM4 qual slippage to late Q1–Q2 2026 is the cleaner story: stricter NVIDIA specs are the bottleneck, not fab capacity.

What moved this week

MediaTek CoWoS customer allocation — ▼ now ~20,000 CoWoS wafers in 2026 (was 180,000 wafers in 2027) (source)
TSMC CoWoS lead time — now 52–78 weeks (one source cites 36–52 weeks) (was 2026 52–78 weeks) (source)
Samsung HBM bit share — now 17% (Q2 2025) vs ~33% (reportedly by late 2026) — two different time periods reported (was ~33% (doubled)) (source)
SK Hynix HBM bit share — ▲ now 62% (was Q1 2026 ~56–58%) (source)
TSMC N3 ramp — now ~150k–180k wafers/month currently, with ramp targets of 180k–210k by mid-2027 (was 2026 ~180k wafers per month) (source)
TSMC CoWoS capacity — ~120,000–140,000 wafers/month by end-2026 (central estimate ~130,000), up from ~70,000 in… (source)
RUNPOD H200 on-demand $/GPU-hr — +27.9% WoW · full history
GOOGLE GEMINI 3 1 PRO PREVIEW output $/Mtok — +14.3% WoW · full history
AWS H100 SXM5 on-demand $/GPU-hr — −12.3% WoW · full history
GOOGLE GEMINI 3 1 PRO PREVIEW input $/Mtok — +9.1% WoW · full history
IOT EDGE gross margin — +8.5pp WoW · full history
TSMC N5 wafer price (avg) — +8.1% WoW · full history

Supply-chain & pricing signals

US-China trade truce extended only 2 months to Jan 10; rare-earth, chip export controls unresolved — Short 60-day extension leaves semiconductor supply chains exposed to tariff escalation and export-control reinstatement. With 22 restricted minerals and 0.1% threshold for licensing, full implementation of Chinese export controls could suppress $6.5 trillion in downstream production globally, creating acute supply-chain risk for advanced chip packaging and rare-earth magnets. (Z2data / Vocal Media) · _In our data:_ packaging tech storyline
AMD joins $1T market cap club; chip rally driven by sustained AI data-center capex — AMD's 187% YTD gain reflects investor confidence in AI accelerator demand and GPU design wins. However, 12 S&P 500 firms now account for 44% of index value, creating systemic concentration risk if AI capex cycles soften. (Investor's Business Daily) · _In our data:_ AMD — 3 developments tracked this quarter · AMD storyline
Japan-US alliance reaffirmed on semiconductors, critical minerals amid China tensions — Japan-US trilateral coordination on supply-chain resilience and critical material sourcing (rare earths, tungsten, semiconductor precursors) strengthens non-China semiconductor ecosystem. Strategic emphasis on advanced materials reduces single-source dependency risks. (News On Japan) · _In our data:_ supply chain storyline

HBM watch

SK Hynix × NVIDIA Rubin (HBM4): Sampling → In qualification — TrendForce reports SK Hynix resubmitted samples and is refining designs against stricter NVIDIA specifications; mass production projected no earlier than late Q1–Q2 2026. (source · timeline)
Micron × NVIDIA Rubin (HBM4): Now in qualification — TrendForce reports Micron resubmitted HBM4 samples and is continuously refining designs to meet NVIDIA's stricter specifications. (source · timeline)

Catch up from the archive

SK Hynix GDDR Memory Components: Suppliers, Channel Structure, and Market Position — Who actually supplies SK Hynix GDDR memory components, how the supply chain is structured, and what procurement teams need to know about sourcing in 2026.
The 750W Floor: Why Power Infrastructure Is Now a First-Order Cost Variable in AI Accelerator Deployments — As AI accelerator TDPs push past 700W per GPU and rack densities climb toward multi-kilowatt norms, power supply and thermal management have moved from procurement afterthoughts to strategic capex variables. This analysis quantifies the economics.
The $5,000 Graphics Card: How AI Demand and GDDR7 Scarcity Are Restructuring the Discrete GPU Market — RTX 5090 pricing has surged to more than $5,000 — 2.5x its $1,999 MSRP — as AI-driven GDDR7 memory competition and structural supply constraints reshape discrete GPU market economics for corporate procurement teams and strategic planners.

AI engines are citing

AMD vs NVIDIA: The AI GPU War in Numbers
SK Hynix GDDR Memory Distributors: How the Channel Actually Works
Analysis

Tool spotlight

Price/Performance Frontier — TFLOPS-per-dollar and TCO across the accelerator landscape.

Sources this week

Z2data / Vocal Media
Investor's Business Daily
News On Japan
yic-electronics.com
io-fund.com
valueaddvc.com
astutegroup.com
investing.com
exponentialindustry.com

Corrections

HBM4 capacity — was 48 GB → now 36 GB (The HBM4 row of the HBM generation table carried the 16-Hi stack capacity (48 GB, a 2027 part) labelled "Future Gen", while every HBM4 figur)

Full correction log

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.

Sourced public intelligence, not insider data; status changes are human-reviewed before publication. Full timelines on the HBM Qualification Tracker.