Curating the latest in AI, semiconductors, EVs & global tech breakthroughs | Not affiliated with any brand
Earth
Joined January 2022
- Tweets13.2K
- Following59
- Followers27.3K
- Likes7.1K
Long-term supply agreements locked in by the top memory makers:
• SK hynix: >50% of 2027–28 capacity
• Samsung: 60–70% of production capacity (target)
• Micron: >35% of revenue through 2030
AI demand is forcing multi-year contracts. Denominators differ - not directly comparable.
#Memory #HBM #AI #Semiconductors #SKhynix #Samsung #Micron
NAND spot prices under pressure this week 📉TLC
512Gb: −9.19% to $16.88
MicroSD 64GB: −8.13%
eMMC 64GB: −1.48%
Only MLC 64Gb edged up +1.61%.
Chinese smartphone brands cut prices for Double 11 to clear inventory → aggressive spot selling. Consumer demand remains weak while enterprise/AI SSD demand stays firm.
#NAND #Memory #TrendForce #Semiconductors #SpotPrice
SK hynix just posted record 3Q26 results.Revenue crossed 100 trillion KRW for the first time → 104.6T (+32% QoQ)
Operating profit hit 83.3T with an 80% marginDriven by server DDR5, data center SSDs, and HBM4 shipments.
LTAs already cover over 50% of 2027–2028 capacity.
Company expects memory demand to keep outpacing supply through 2030.AI memory supercycle is still accelerating.#SKhynix #HBM #AI #Semiconductors #memory
Tesla’s Optimus AI5/AI6 already carry 2–4× more DRAM than Counterpoint’s 2030 humanoid average.CP: 19 GB → 39 GB
Tesla: 72 GB (AI5) / 144 GB (AI6)Humanoid DRAM demand: 8,300 → 172,000 terabits by 2030 (~21.5 PB). Still tiny vs total capacity - for now.
Anthropic’s contracted compute mix (est.):NVIDIA GPU ~4.7 GW
Google TPU ~4.5 GW
AWS Trainium ~1.2 GW
AMD GPU ~0.8 GW~11 GW total. NVIDIA is large but not dominant - multi-chip strategy is real.
#Anthropic #AI #NVIDIA #Semiconductors
TSMC is reportedly preparing to raise prices for its CoWoS advanced packaging services by as much as 20%, according to Taiwanese media reports.
The technology is critical for high-performance AI chips, connecting GPUs or other processors with high-bandwidth memory (HBM).
Demand for CoWoS capacity is also expanding beyond Nvidia, with AMD and Broadcom among companies developing AI accelerators and custom chips that require advanced packaging.
Supply-chain sources reportedly say customers are now prioritizing delivery dates over price negotiations.
Meanwhile, lead times for some packaging equipment have reached up to eight months, highlighting the severe capacity constraints surrounding AI infrastructure. TSMC has not officially confirmed the reported price increase.
$TSM
From SEMICON West 2026: Nvidia’s OOI prototype brings optical engines closer to the compute chip by placing them directly on the silicon interposer (100 mm → 8 mm span). Paired with a DWDM demo delivering 256 Gb/s per fiber.
Samsung & SK Hynix are targeting Q2 2027 for mass production and volume shipments of 8-layer HBM4, specifically for NVIDIA’s Rubin Ultra rack-scale AI systems.
While 12-layer HBM4 remains the primary focus through 2026, the share of 8-layer is expected to rise significantly from mid-2027 - driven by yield and thermal constraints on taller stacks.
NVIDIA is adjusting its strategy and evaluating a mix of configurations.
#HBM #HBM4 #NVIDIA #Samsung #SKHynix #AI #Semiconductors
NVIDIA just locked in lower HBM4 bandwidth for Vera Rubin.Earlier targets → Official specs (Oct 2026):
• Rubin GPU: 22 → 19.2 TB/s
• Superchip: ~44 → 38.5 TB/s
• NVL72 Rack: ~1,580 → 1,400 TB/sCapacity stays 288 GB/GPU. Pin speed cut (~11.7 → ~9.5 Gbps) to improve yields and multi-supplier volume.Still a big jump over Blackwell’s HBM3E.#NVIDIA #VeraRubin #HBM4 #AIHardware
Amkor’s CapEx stayed flat ~$0.75–0.91B from 2021–2025.2026 guide jumps to $2.5–3.0B.NVIDIA is prepaying $2.8B ($1.5B Arizona + $1.3B Korea) to lock capacity.
