Silicon Analysts Weekly · Issue #6 · week of 2026-07-06 → 2026-07-13
TSMC H1 rev +35.6% YoY; HBM demand 7x oversubscribed
Silicon Analysts Weekly for the week of 2026-07-06 to 2026-07-13: TSMC H1 rev +35.6% YoY; HBM demand 7x oversubscribed. 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
TSMC's 35.6% H1 revenue growth and 67.9% June YoY surge confirm CoWoS-driven packaging demand is pulling the entire advanced supply chain tighter — and with HBM running 7x oversubscribed and SK Hynix holding 56.4% share, anyone sourcing memory for 2027 builds has no slack left. Wafer pricing is the right thing to be watching right now.
What moved this week
TSMC CoWoS ramp — up — CAGR of over 60% until at least 2026; more than quadruple from 2023 levels by end of 2026 vs 56% YoY growth in 2026 (source)
Supply-chain & pricing signals
TSMC H1 2026 Revenue Surges 35.6% YoY on AI/HPC Demand, Signaling Foundry Strength — TSMC's 67.9% YoY June revenue growth and 35.6% H1 growth demonstrate robust foundry demand driven by AI/HPC products. Strong execution supports advanced packaging capacity expansion and validates continued AI infrastructure investment thesis. (ANI News)
Intel reportedly planning Altera divestiture; Marvell seen as likely acquirer — Intel asset sale reflects mounting financial distress and retreat from programmable logic market. Marvell acquisition of Altera would consolidate custom silicon capability for cloud giants (AWS, Microsoft), intensifying competitive pressure on Intel's foundry ambitions. (CNBC)
SK Hynix CEO forecasts worst-ever memory shortage in 2027, demand to exceed supply through 2030 — Sustained memory supply constraints will support elevated pricing and long-term contracting power for HBM and DRAM producers. SK Hynix's dominant position in HBM supply to AI infrastructure players (Nvidia, hyperscalers) ensures strong revenue visibility, but supply scarcity will accelerate capex race and raise customer qualification costs. (Reuters)
SK Hynix claims 56.4% HBM market share as demand runs 7x oversubscribed — SK Hynix's dominant HBM position creates supply concentration risk for Nvidia and datacenter builders. 7x oversubscription confirms severe memory chip shortage persists, supporting sustained pricing power and margin expansion through 2H 2026. (Yahoo Finance)
HBM Supply Constrained Until H2 2027, DRAM Undersupply Persists — Memory supply shortage will extend through 2027 as chipmakers balance HBM demand from AI against commodity DRAM, creating supply chain bottlenecks. Even 2028 recovery projected to leave market undersupplied, constraining AI infrastructure buildout timelines. (CNBC)
SK Hynix IPO Momentum Falters on AI Spending Sustainability Concerns — SK Hynix stock volatility reflects investor uncertainty about AI capex durability post-listing, signaling broader market skepticism about memory chip demand sustainability. Post-IPO decline pressures valuations for other memory chipmakers seeking capital. (Barron's)
Supply-chain & pricing signals
TSMC H1 revenue +35.6% YoY, June alone +67.9% YoY — AI/HPC demand is driving foundry utilization and advanced packaging expansion. (ANI News)
SK Hynix claims 56.4% HBM market share; demand 7x oversubscribed — supply concentration at a single vendor is a systemic risk for datacenter builders. (Yahoo Finance)
SK Hynix CEO forecasts worst-ever memory shortage in 2027, demand to exceed supply through 2030 — long-term contracting windows are narrowing faster than most procurement teams have modeled. (Reuters)
HBM supply constrained through at least H2 2027; 2028 recovery still leaves market undersupplied — AI infrastructure build timelines will be gated by memory, not compute. (CNBC)
Intel reportedly planning Altera divestiture; Marvell seen as likely acquirer — a deal would consolidate custom silicon capability for major cloud buyers and further narrow Intel's foundry addressable market. (CNBC)
HBM watch
Samsung × NVIDIA (HBM4 12-Hi): now Volume shipping — Samsung began world-first HBM4 mass production shipment at 11.7 Gb/s (max 13 Gb/s), 3.3 TB/s bandwidth, securing NVIDIA supply approval. (source · timeline)
SK Hynix × NVIDIA (HBM3E 12-Hi): now Volume shipping — Jensen Huang confirmed all three major Korean memory makers, including SK Hynix, passed NVIDIA qualification and are in production. (source · timeline)
SK Hynix × NVIDIA (HBM4): now In qualification — SK Hynix met NVIDIA's 11.7 Gb/s speed requirement after a partial redesign; top-tier HBM4 supply to NVIDIA may not begin until end of year. (source · timeline)
Samsung × NVIDIA (HBM4): now In qualification — Samsung entered final qualification phase after supplying initial samples in September, per people familiar with the matter. (source · timeline)
SK Hynix × AMD MI400 (HBM4): now In qualification — SK Hynix HBM4 listed in qualification against the projected AMD MI400 series. (source · timeline)
This week's analysis
What You're Actually Paying For in a GPU-Hour: Hardware, Power, and the Scarcity Premium — We decompose H100, H200, and B200 rental prices into hardware, energy, and scarcity premium: 52-88% of every rented GPU-hour is demand premium, not cost.
The Hyperscaler Capex Wall: $434B of Buyer-Side AI Spend, the Depreciation Lag, and Why Big Tech Borrows — Microsoft, Alphabet, Amazon, and Meta spent a combined $434B on property and equipment over the last four quarters — nearly 3x their reported depreciation. The gap is by design, it is temporary, and it explains why cash-rich companies are suddenly issuing record bonds.
The Inference Accelerator Wars: Why Cost-Per-Token Is Now the Defining Metric in AI Silicon — OpenAI's Jalapeño ASIC and the broader custom inference push are reshaping GPU vs ASIC economics. This analysis breaks down total cost of ownership, cost-per-token dynamics, and what the custom silicon wave means for enterprise AI infrastructure strategy.
AI engines are citing
Tool spotlight
Fab Site Explorer — Global fab capacity by node, geography, and outlook.
Sources this week
Corrections
coreweave-h200 on_demand — was 3.89 → now 6.305 ($3.89 is not supported by any CoreWeave page; the published price is $50.44 per 8-GPU HGX H200 instance ($6.305/GPU-hr, archive.org 2026-03-)
coreweave-h100-sxm5 on_demand — was 4.76 → now 6.155 (cloudPricing.ts (Cloud GPU Pricing Tracker / Build-vs-Rent) carried the retired GPU-component rate: since Jan 2025 CoreWeave quotes 8-GPU bu)
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.
More recent issues
Sourced public intelligence, not insider data; status changes are human-reviewed before publication. Full timelines on the HBM Qualification Tracker.