Korean HBM is showing up in Malaysia. That does not mean CoWoS just moved. Customs data, the “composite-structure chip IC” bucket that includes HBM, put August shipments to Malaysia at $1.625 billion. That is up 466.6% from a year earlier. Weight rose from 1.4 tons to 3.8 tons. Taiwan still took $3.32 billion that month, up 32.2%. Malaysia went from 11.4% of Taiwan’s August haul last year to 48.9% this year. The longer series is the same bend. Malaysia was $93.7 million in 2022, 11th among destinations. It hit $3.00 billion last year, fifth. January-August this year is already $6.31 billion, fourth. August alone ranked third after Hong Kong and Taiwan. The pull is packaging, not a new Korean fab. Malaysia already does about 13% of world assembly and test. ASE is expanding Penang from 1 million to 3.4 million square feet. TF-AMD is adding bumping and advanced packaging in Batu Kawan, about 5 billion ringgit, with 3 billion aimed at FOED and 2.5D. The government is funding a local HBM4 test-chip consortium for 2027 process checks. Hong Kong is still a re-export hub. Taiwan is still the CoWoS core. Economic Daily has said Intel’s Malaysia plant is taking some overflow back-end work as TSMC packaging stays tight. A Korean industry source told Chosun Biz Intel volume may be part of the spike, but Intel’s book is not large enough to call this a stolen CoWoS socket. It reads more like extra AI packaging demand finding a second OSAT geography. Destination is not consignee. Watch packaged units, not just the export line. If Malaysia is already half of Taiwan’s monthly HBM haul, how much of that do you assign to Intel versus ASE and inventory?

Is EMIB really chipping away at CoWoS? CoWoS has owned advanced packaging for years. TSMC keeps adding capacity aggressively but it still isn't enough to satisfy customer demand. Intel stepped in with a compelling alternative with EMIB — and customers are responding. (1/3)🧵

Samsung Electronics and SK Hynix will significantly increase High Bandwidth Memory (HBM) production next year. Still expect 2027 to run short. The Fact says both firms have pulled expansion forward because a new line takes two to three years from dirt to wafers. Customers are already shopping for 2027 bits. Samsung is said to aim for twice this year’s HBM4 and HBM4E output next year. Wafer starts would rise from about 180,000 a month to 250,000. HBM4-class mix would go from 40% of HBM shipments to 80%. Glass-carrier clean work jumping from 20,000 wafers a month to 50,000 is read as another tell. About half of its 4nm foundry, some 15,000 wafers a month, is said to be reserved for HBM4 base dies. Pyeongtaek P5 may drop floors to finish sooner. Counterpoint put its second-quarter HBM sales share at 33%, up 12 points from the prior quarter. $SKHY SK Hynix is lifting Cheongju M15X from about 10,000 wafer starts a month toward 80,000. The first Yongin fab is being pulled toward a 2027 start. 2027 volume and price talks with big buyers are still open. Last year’s deals slipped to October. Samsung said in Q1 that supply versus demand was the worst on record and the gap would widen in 2027. Kiwoom sees 2027 HBM supply up 50% and demand up 56%, with HBM4 prices up about 65%. Jensen Huang $NVDA said last week he wants to sell twice as many chips next year. Rubin Ultra is slated for 16 HBM4E stacks per package. Micron $MU has said 2027 HBM is sold out. CXMT is reported to be sampling HBM3E for 2027 volume.

Not sure it’s only Muse driving demand. It’s inference workloads in general. The ratio of inference to training was 2:1, projected to skew higher. Barrons writes reason for CPU bump is - Intel, Arm Holdings and AMD shares rose Monday on investor excitement about CPU growth driven by Meta Platforms’ Muse AI agent.

Why is everyone suddenly talking about CPU demand and supply? For two years, the AI story was simple: GPUs sold out, memory exploded, everything else waited. That script just changed. Before focusing on anything else, read these points: 1. Intel’s $INTC CEO said it can meet only about half of customer CPU demand. 2. AMD $AMD doubled its server CPU TAM outlook. 3. NVIDIA $NVDA started selling Vera as a standalone CPU built for agents. 4. Arm $ARM said the data-center CPU is now central, not a sidekick. 5. Bank of America lifted its 2030 server CPU market view toward $170-210 billion. That is the supply chain, not a retail thread. The reason is agentic AI. Training a model is a GPU job. Running an agent is a system job. An agent does not answer once and stop. It plans, calls tools, writes code, hits databases, moves data, checks the result, and loops. Orchestration, scheduling, I/O, sandboxes, and memory management sit on CPUs. Chat inference can live at one CPU per four to eight GPUs. Agent clusters are being designed closer to 1:1. Some swarm designs want even more host cores per accelerator. That is why the talk sounds like 2023 GPUs. Demand showed up faster than wafer plans. Server parts stay tight into 2026-27. Intel and AMD have room to raise prices. Hyperscalers are signing multi-year CPU deals. Client PCs may get the leftovers, the same way gamers waited while data centers ate GPUs. GPU-like outcomes are possible, with limits. Allocation fights. Longer lead times. Higher ASPs. A second “who has supply” trade in AMD and Intel. NVIDIA and Arm also win if they own the host CPU next to the accelerator. Memory is still the more violent shortage. Power and packaging still cap how fast anyone ramps.

Why is everyone suddenly talking about CPU demand and supply? For two years, the AI story was simple: GPUs sold out, memory exploded, everything else waited. That script just changed. Before focusing on anything else, read these points: 1. Intel’s $INTC CEO said it can meet only about half of customer CPU demand. 2. AMD $AMD doubled its server CPU TAM outlook. 3. NVIDIA $NVDA started selling Vera as a standalone CPU built for agents. 4. Arm $ARM said the data-center CPU is now central, not a sidekick. 5. Bank of America lifted its 2030 server CPU market view toward $170-210 billion. That is the supply chain, not a retail thread. The reason is agentic AI. Training a model is a GPU job. Running an agent is a system job. An agent does not answer once and stop. It plans, calls tools, writes code, hits databases, moves data, checks the result, and loops. Orchestration, scheduling, I/O, sandboxes, and memory management sit on CPUs. Chat inference can live at one CPU per four to eight GPUs. Agent clusters are being designed closer to 1:1. Some swarm designs want even more host cores per accelerator. That is why the talk sounds like 2023 GPUs. Demand showed up faster than wafer plans. Server parts stay tight into 2026-27. Intel and AMD have room to raise prices. Hyperscalers are signing multi-year CPU deals. Client PCs may get the leftovers, the same way gamers waited while data centers ate GPUs. GPU-like outcomes are possible, with limits. Allocation fights. Longer lead times. Higher ASPs. A second “who has supply” trade in AMD and Intel. NVIDIA and Arm also win if they own the host CPU next to the accelerator. Memory is still the more violent shortage. Power and packaging still cap how fast anyone ramps.

KOSPI crossed 7,000 and then slipped back. The tape is still a chip tape. At 10:10 a.m. on Sept. 21, had the index at 6,993, up 99 points or 1.45%. It opened at 6,938.34 and briefly hit 7,007.1. US futures were mixed Friday. The Philadelphia semiconductor index jumped 2.78%. Nvidia, Micron, Broadcom, and AMD all rose. The 10-year Treasury yield sat back above 5%. On KOSPI, Samsung Electronics was up 3.45% at 270,000 won. SK Hynix was up 0.92% at 1.874 million. Samsung preferred, SK Square, Samsung C&T, and Samsung Electro-Mechanics also rose. LG Energy Solution and Hyundai Motor lagged. KOSDAQ was at 834.13, up 0.85%, after dipping to 823. Retail bought that market. Foreigners and institutions sold.