VLSI KoreaDaily Silicon: onsemi $1.604B: NVIDIA HBM4e and AI Chip Design
A practicing semiconductor engineer's daily read on global silicon news - onsemi $1.604B: NVIDIA HBM4e and AI Chip Design
🆓 무료 글 · 누구나 무료로 읽을 수 있는 글입니다
A practicing semiconductor engineer's daily read on global silicon news:
onsemi's $1,604 million Q2 and $1,650-$1,750 million Q3 guide set the earnings baseline, while NVIDIA's four HBM options, $886.7 billion in CSP capex, and LG's three-tier AI-home SoC plan tightened memory, packaging, and design qualification.
Chase's Take - I care less about onsemi's 9% revenue growth than PSG's 19% growth and the Q3 non-GAAP gross-margin guide of 40.0%-42.0%. A broader AI rack power tree does not turn design wins into revenue if qualification stays slow.
Next I am watching whether AI data-center revenue tracks the 2026 goal of more than doubling and whether Q3 PSG mix actually lifts gross margin.
1. onsemi Q2 2026: PSG at $829.0 Million Lifted Margin
TL;DR - Revenue was $1,604 million at 38.4% gross margin, while the 2026 AI data-center revenue target remains more than 2x.
Source: onsemi - onsemi Q2 2026: PSG at $829.0 Million Lifted Margin
2. Four NVIDIA Rubin Ultra HBM4/HBM4e Configurations for 2027
TL;DR - NVIDIA is evaluating 8-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4 alongside the 12-Hi HBM4e baseline; the final spec is not set.
Source: TrendForce - Four NVIDIA Rubin Ultra HBM4/HBM4e Configurations for 2027
3. 2026 CSP Capex at $886.7 Billion, AI Server Growth Forecast at 31%
TL;DR - The nine CSPs will spend more than $886.7 billion in 2026, and the AI server shipment growth forecast rose from 28% to about 31%.
Source: TrendForce - 2026 CSP Capex at $886.7 Billion, AI Server Growth Forecast at 31%
4. LG AI-Home SoC Lineup: 30, 10, and 1.5 TOPS
TL;DR - LG Electronics is developing AI-home SoCs with domestic fabless partners at >30 TOPS premium, 10 TOPS midrange, and 1.5 TOPS entry tiers.
Source: ETNews - LG AI-Home SoC Lineup: 30, 10, and 1.5 TOPS
What to Watch Next
- onsemi Q3 revenue of $1,650-$1,750 million and non-GAAP gross margin of 40.0%-42.0%
This is the first test of whether PSG and AI rack power mix convert into margin, not just revenue. - Rubin Ultra's final 8-Hi and 12-Hi HBM4/HBM4e spec and validation schedule
DRAM wafer allocation and stack yield will determine shippable accelerator volume. - Actual AI rack installations versus CSP capex, plus liquid-cooling and power-energization speed
Energized rack count is a more realistic leading indicator for packaging and interconnect demand than purchase orders. - LG AI-home SoC foundry, node, PDK disclosure, and product-level qualification
This will show whether the 30, 10, and 1.5 TOPS tiers reach real tapeouts under a common software stack.
Signal Radar — More Headlines
- SiC Power Semiconductors: KRW 500 Billion Government R&D Hits Roadblocks (ETNews)
- TinyGPU v2.0 Implements Real Silicon with 240,000 Transistors (PC Watch)
- Renesas MRDIMM Update Targets the DDR5 Memory Bottleneck (EE Times)
- Kioxia XL1 Brings CXL Memory Expansion to AI Workloads (TechPowerUp)
- Comparing TSMC, Intel, Samsung, and GlobalFoundries CPO Roadmaps (Tom's Hardware)
- AMD GFX7 GPU Soft Reset Lands in Linux 7.3 (Phoronix)
- Micron New York Fab PFAS Lawsuit Reopens Water-Treatment Risk (Tom's Hardware)
- Lexar Lists CXMT-Based 32GB DDR5 at 3,999 Yuan (Tom's Hardware)
- Kioxia GP1 PCIe 6.0 SSD Claims Up to 10 Million IOPS for AI (GDM)
- DDR5-6400 Enterprise RDIMM Reaches 256GB with SK hynix A-Die (IT Home)
Further Reading on VLSI Korea
- AWS from Inferentia to Trainium3: The Evolution of an AI ASIC with a Price Tag
Background reading from VLSI Korea. - Google TPU from v1 to Ironwood: Ten Years of Designing an AI Factory
Background reading from VLSI Korea.
Frequently Asked Questions
What did onsemi report for Q2 2026?
onsemi reported Q2 revenue of $1,604 million at a 38.4% gross margin, with PSG revenue of $829.0 million. Q3 guidance is $1,650 million to $1,750 million with non-GAAP gross margin of 40.0% to 42.0%.
How much will the nine CSPs spend in 2026?
The nine CSPs will spend more than $886.7 billion in 2026. The AI server shipment growth forecast rose from 28% to about 31%.
What HBM options is NVIDIA evaluating for Rubin Ultra?
NVIDIA is evaluating 8-Hi HBM4e, 12-Hi HBM4, and 8-Hi HBM4 alongside the 12-Hi HBM4e baseline. The final specification is not set.
What are LG's AI-home SoC tiers?
LG Electronics is developing AI-home SoCs at over 30 TOPS for premium, 10 TOPS for midrange, and 1.5 TOPS for entry tiers. The chips are being developed with domestic fabless partners.
Related reading
