Tae Kim
AI/Semiconductor Journalist, Author of Key Context
AI data-center CPUs will drive a tidal wave of memory demand, contrary to the memory bears' view. Neil Shah cites 1.5 TB of LPDDR5X per Vera CPU and 12 TB across eight CPUs, with growing agent workloads increasing CPU orchestration and state-management needs. This is bad news for Apple, too. x.comÂ
Memory stocks are riding surging AI demand, as Korea's September chip exports jumped 263% year on year to a record $60.3 billion, per Bloomberg. Micron delivered a big beat-and-raise with demand extending into another year, OpenAI is near $70B in ARR, and AI agents are becoming a new secular demand driver. A resolution of the Iran war would let the rally fly. x.com x.comÂ
High-bandwidth memory supply may not keep pace with AI demand, supporting the memory makers. Anthropic's S-1 says SK hynix, Samsung and Micron are expanding capacity, but warns that large buyers or competitors could reserve much of the available output. x.comÂ
Memory-cost pressure, weak volumes and weak services should make Apple's quarter bad, giving the short a chance to work on this quarter alone. Other hardware companies this quarter also point to weakness, though Apple has been a terrible short so far. x.com x.comÂ
Calls from AI lab leaders to pace frontier training are a negative for semiconductor and memory stocks, because training restraint cuts accelerator demand and the infrastructure multiple takes the hit before any lab's revenue does. Sam Altman said progress should be slower than it otherwise could be. Seoul sold off first, with SK hynix down more than 6% and Samsung more than 4%, then ASML, Infineon and Micron. Anthropic expects positive adjusted operating income: the margin stays with the labs. … Show more
AI data-center investment needs outside AI spending to rise 33-fold by 2030 to break even. Hyperscalers hold $4.5 trillion in debt and future liabilities, but corporate and SaaS token spending outside circular deals is only $60 billion annualized against a $2 trillion break-even ARR. SK hynix says memory shortages may last until 2031, raising capex and CoreWeave default risk; Mitsubishi warns turbines are booked through 2031 and cost 400% more, delaying CoreWeave and Oracle data centers. x.com … Show more
Nvidia can capture more AI infrastructure spending by supplying Vera CPUs and networking alongside GPUs. Oracle outlined an initial supply of 150,000 Vera CPUs, and OpenAI says it is putting them into production for agent tool execution; reducing CPU waits could improve cost per task and GPU-memory use. Benefits from independently scaling CPUs and GPUs still need validation in utilization and completion times. x.comÂ
Chip stocks face an oversupply, and oversupplied chips get dumped: AMD sold restricted chips into China at 20% of the US retail price, per a report cited in the post. The same pressure threatens Nvidia, SanDisk, Micron, Broadcom, SK hynix and semiconductor and memory ETFs. x.comÂ
Inference's move toward edge devices and dense, memory-hungry racks should drive demand for specialized chips, links, power delivery and cooling. McKinsey projects inference at 60% of AI computing demand by 2030. Ambarella makes on-device AI chips; Astera Labs links GPUs, memory and networks; Credo supplies short copper connections; Rambus earns from memory interfaces; Navitas reduces power waste; Modine cools dense racks; and Super Micro builds inference-specific server racks at volume. x.com … Show more
US sovereign AI and defense demand for domestically made compute will bring Intel most, if not all, US semiconductor designers as customers. PC weakness also lets Intel shift wafers to higher-margin server CPUs, but that takes a few quarters; capacity to ship more server CPUs is the limiter. The $240 billion 2030 server CPU market forecast is a stretch case, not the base case, versus roughly $200 billion consensus. x.com x.com x.com x.com x.comÂ
Memory stocks are at the peak of the cycle, and FOMO buying into Micron and SanDisk now could be one of the dumbest moves. Chinese factories are coming online at lightning speed, which is why Micron is signing multi-year deals, and pricing power is peaking right now. The market is forward-looking, so a low forward P/E matters little if net income drops 60-90% a few years out. Per Kalshi, the memory ETF just passed $25 billion in assets for the first time. x.com x.com x.comÂ
Agentic AI's repeated context favors cache memory and storage demand for Micron, SK hynix and Sandisk over fresh GPU computation. Per @demian_ai, about 95% of input tokens for DeepSeek V4.1 Flash on OpenRouter were cache hits; storage matters unless routing and caching keep almost everything in GPU memory. New products such as 3D DRAM and HBM-on-logic could be the final trigger for a full memory-stock re-rating. x.com x.comÂ
