Intel’s EMIB-T is scaling. It is not replacing CoWoS. 5 → 20 → 45 kwpm by 2028e. Still ~12% of the 2.5D stack. TSMC remains the base layer. Morgan Stanley · $INTC Data versionMorgan Stanley still has EMIB-T as the primary TPU packaging flow. TSMC CoWoS is overflow, not a takeover. Intel mix: EMIB-T 5 / 20 / 45 kwpm. EMIB-M stays ~95. 2.5D stack: 142 → 250 → 374 kwpm. Intel sliver = 10–15%. Could support ~2.5M TPU units by 2028

Micron capex is not a cycle spike. It is a construction calendar. FY25 actual: $13.8B FY26 outlook: ~$27B FY27: >$45B - 3.3× in two years New wafers still arrive late: Jan 2026 - NY ground break Mid-2027 - ID1 Boise + Tongluo Late 2028 - ID2 Idaho That is why management says industry supply only improves gradually in 2028. Demand is doubling. The fabs that could close the gap cannot show up sooner. $MU reports Wednesday.

🚨 J.P. Morgan HBM Update After accounting for de-spec: Supply Shortage• CY26: -20%• CY27: -19%• CY28: -16% Accumulated shortage widening to 23 weeks by CY28E. AI memory tightness isn’t going away anytime soon. $MU $DRAM $EWY $DISK

US data-center IT power demand is exploding. Morgan Stanley: 9.19 GW in 2025 → 78.57 GW by 2029. That’s 8.5x in four years. 2026: 17.96 GW 2027: 35.46 GW 2028: 52.31 GW Not just more buildings. Denser racks. Vera Rubin ~234 kW. Rubin Ultra ~600 kW. Chips getting more efficient. Total electricity still climbing. The bottleneck is no longer only GPUs. It’s power. Source: Morgan Stanley, Sept. 2026 $NBIS $IREN $CIFR $WULF

SK hynix at TSMC OIP 2026: HBM3E / Rubin HBM4 → 12-hi × 3GB = 36GB HBM4 → 16-hi × 3GB = 48GB HBM4E → 12-hi × 4GB = 48GB The jump is per die, not stack height. 4GB vs 3GB is how 12 layers match a 16-high part. Nvidia Rubin: 288GB HBM4 (8×36GB). Swap those sites for HBM4E already on the booth → 384GB. HBM5 is validated on the CoWoS package it will ride. Memory per GPU keeps climbing through the next two cycles. $SKHY

Component lead times are no longer about GPUs. GPU: 35w = “balanced” DRAM 8 → 20 NAND 8 → 16 HDD 16 → 50 ABF 12 → 52 MLCC 12 → 32 The queue has moved to memory, storage, substrates, and caps. $MU $SNDK $WDC $STX $NVDA $AMD

Samsung vs SK Hynix in HBM gets tight by 2027. JPM sales share: 2025 → SKH 60% / SEC 20% / MU 20% 2027E → SKH 41% / SEC 39% / MU 21% Micron holds ~1/5. The fight is Korea vs Korea. $DRAM $MU $EWY

Bernstein: Data center power demand ~19% CAGR → 1,570 TWh by 2032E. AI servers are the load. 400 TWh (2024) → 1,570 TWh (2032E) $NVDA $DELL $GEV $IREN $CRWV $NBIS

The GPU is the only balanced line on the rack. ABF 4.3× HDD 3.1× MLCC 2.7× DRAM 2.5× NAND 2.0× CPU 1.5× GPU 1.0× Agentic AI adds CPU hours, storage and substrates first. Discrete GPUs are not the scarce SKU this week.

$MU operating margins by unit, last five quarters. Cloud Memory - the unit that sells HBM - started as the high-margin business at 46%. By May 2026 it was second-lowest at 78%. Same year: Mobile & Client 15% → 86% (+71 pts) Cloud / HBM 46% → 78% (+32 pts)

$MU quarterly revenue, FQ4-24 → FQ3-26 $7.8B → $41.5B in eight quarters. FQ4-26 guide: $49–51B Street: $51.24B Reports Wed, 30 Sep after the close.

DRAM’s No.1 and No.2 races in one chart. Samsung flipped SK hynix twice in 2025 and now leads by 14.5 points. Meanwhile SK hynix’s lead over Micron collapsed from 16.7 → 1.6 points. $SKHY $MU

Blended HBM pricing forecast (Jefferies/Fubon): 📈 Rises 45% from $18 (2Q26) → peaks at $26.1/GB in 2Q27 📉 Then enters a shallow correction (not a sharp crash) Still ~38% above 2Q26 base even by 2Q28. $MU $NVDA $SKHynix $DRAM

The "GPUs are the only tollbooth" narrative is dying. Merchant GPU shipments: 6.8M → 13.9M by 2028 Custom ASIC/XPU: 1.5M → 17.0M by 2028 By 2028, custom silicon overtakes GPUs in raw volume. Hyperscalers are taking control of inference economics.

CoreWeave’s growth trajectory according to UBS: $13B (FY26) → $84.8B (FY30) EBIT margin expanding to 27%

Samsung claws back HBM share in 2027e HBM bit share by vendor 2023 → 2027e Samsung: 48% → 38% SK Hynix: 48% → 40% Micron: 4% → 18% CXMT: 0% → 4.4% $SKHY $MU solana:iUSRerdqvY4Si9PxT8e5RZtnvvM4MsV1EVrKPPBpump

Even in UBS’s TSMC stress test, Intel still reaches 12-17% of advanced packaging by 2030. Sometimes the biggest winners come from market growth, not market share gains. $INTC

DRAM shortage deepens into 2027e 📉 Morgan Stanley data: 2023: -4% 2024: -5% 2025: -4% 2026e: -17% 2027e: -15% Absolute gap keeps widening. 2027 is shaping up to be the worst year in memory history. $MU $SKHY $SSNLF

Silicon wafers are repricing hard. GlobalWafers 12" ASP: • 2026: $92 • 2027e: $122 (+32.6%) • 2028e: $166 (+36.9%) 2027 LTA hikes: → Polished (memory): +40% or more → Epi (logic): +15–25% → Some 2nd tier buyers: Over 50% Inference runs on silicon. That’s why. $MU $SKHY

Hyperscaler capex isn’t stopping - it’s just slowing down. Goldman sees $1.2T in 2027 and $1.4T in 2028. Growth drops from ~100% → 54% → 12%. Cash flow is no longer enough. Debt will have to do more of the heavy lifting.