A field notebook on the platters, cells, and file systems holding up the AI build-out, for the people who build and buy that infrastructure, and the leaders they report to.
One email a week. Unsubscribe link in every issue.
SK hynix says DRAM and HBM are sold out through 2029 on a record 80% operating margin, Samsung confirms a 783% profit jump and sells out 2027 HBM4, Kioxia warns NAND price hikes are fading as spot prices drop 9.19% in a week, Seagate's record quarter gets sold off anyway, SanDisk bets a third of future NAND demand on KV cache, and the long-term agreement explained.
Read Issue 14 →A running log of smaller developments between issues, each with a source. Updated a few times a week.
OpenAI CEO Sam Altman met with about 50 presidents from Samsung Electronics and its financial and bio affiliates at Samsung's Human Resources Development Institute in Yongin, South Korea, to discuss AI chip and data center cooperation, including memory chip supply for OpenAI's Stargate infrastructure project. Source ↗
SanDisk said at the SEMICON Taiwan 2026 Memory Summit that AI infrastructure customers are returning to secure more NAND flash after underestimating their storage requirements, with its pipeline of prospective long-term supply agreements now exceeding planned output through the end of the decade. Source ↗
SanDisk said it expects AI inference and agent-orchestration workloads to require roughly 60 to 90 gigabytes of NAND flash per CPU, citing a shift in how customers deploy memory and storage tiers. Source ↗
HPE said its Alletra Storage MP X10000, tested by Kamiwaza on Nvidia H200 GPUs running a 70-billion-parameter Nemotron model with more than 30 concurrent agents, cut KV cache offload time to first token by up to 20 times and raised effective inference throughput by up to 17 times versus running without the offload. Source ↗
SK hynix detailed two memory-tiering technologies for AI inference KV caches, SALT-KV, which weighs each cache item's reuse value and storage cost to place it across HBM, DRAM and SSD, and CMM-Hybrid, a CXL-based design that keeps the full KV cache on SSD and predicts what to move into DRAM before it's needed, and said it plans a future High Bandwidth Flash tier between HBM and SSDs. Source ↗
Scality launched an AI Inference Factory software stack that it said loads a 27-billion-parameter Gemma-3 model in 1.9 seconds and restores a 14,000-token KV cache from storage in 166 milliseconds, about 14 times faster than recomputing it. Source ↗
Seagate said its Mozaic HAMR-based SMR hard drives became the first SMR drives approved on Broadcom's host bus adapter vendor list, with Microsoft Azure Storage validating the configuration at hyperscale. Source ↗
SK hynix said it plans a special dividend of roughly 16 trillion won ($12 billion), about 80% of an estimated 20 trillion won in gains from its Kioxia Holdings stake, with details expected alongside third-quarter earnings on October 28. Source ↗
TrendForce data cited by Notebookcheck showed contract prices for 1-terabit TLC NAND flash components rising 123% in the six months to October 21, while a 256GB TLC SSD's contract price rose 303% year-over-year to $134.20 in the third quarter of 2026. Source ↗
Reuters reported that Samsung plans to convert its Xi'an plant from roughly 200-layer to 286-layer V9 NAND under an internal project called Monami, while SK hynix is shifting its Dalian plant to NAND with at least 240 layers, up from 192 layers currently. Source ↗
Micron laid out a one-generation-per-year HBM roadmap at an October 21 technology webinar — HBM4 in 2026, HBM4E in 2027, HBM5 in 2028, HBM5E in 2029, and HBM6 in 2030 — with its current HBM4 delivering more than 2.8 terabytes per second of bandwidth per stack on a 2,048-bit bus. Source ↗
DeepL, the German neural machine translation company, said it is moving to VAST Data's storage and AI OS platform and expects roughly three times the storage capacity from its existing hardware, with data migration handled by VAST's SyncEngine. Source ↗
Most weeks are a fast read. Occasionally, a topic gets the long treatment instead.
This week's numbers, curated news with a "why it matters" take, one rotating deep-dive story, a quick storage/memory term explained, and what's coming up.
Occasional full-market issues covering every enterprise storage vendor, hyperscaler storage, the HDD capacity race, memory and flash, field reliability data, and an optimization playbook.
A visual walkthrough of the data flow from a stored checkpoint to a generated token, and why the memory footprint keeps growing after the model has already loaded.
Read the guide →An interactive walkthrough of the NeMo Framework repos, from data curation through training, parallelism, post-training, evaluation, export, and serving.
Explore the system design →One email a week on AI storage and memory infrastructure. No fluff, no spam.
One email a week. Unsubscribe link in every issue.