Weekly signal for the people building and buying AI infrastructure — what moved, why it matters, what's next.
This is the first weekly issue, and it's a different animal on purpose. Issue 01 tried to cover the whole AI storage market in one sitting: every vendor, every roadmap, every failure-rate table. That's not something worth attempting every seven days, so it becomes an occasional Field Report instead. Most weeks will look like this one: what actually happened, why it matters if you're the one keeping GPUs fed, and what to watch next.
Seven things that happened this week, and why I'd pay attention to each one.
DDN's object storage will sit underneath SCX.ai's inference infrastructure for Australian enterprise and government customers. The deal itself isn't the story, DDN wins deals.
NVHBM claims up to 30% more bandwidth by pulling what used to be separate silicon directly into the stack, freeing up die space on the compute side.
Specifically Intel's EMIB advanced packaging technology, for SK Hynix's next generation of HBM memory chips.
That's CEO Matt Murphy's phrase, from the earnings call, describing deployments across hyperscalers for both memory expansion and AI inferencing.
The recognition is for AI-driven data engineering automation — Git-style branching and versioning applied to datasets.
DuckDB shows up in a lot of local and edge AI data-prep tooling, so the acquisition matters less than Amazon's stated commitment to keep it free and MIT-licensed.
The explicit goal: become the world's largest NAND manufacturer by the end of 2027.
Areal Density reaches people who actually buy and operate AI storage and memory infrastructure, AI DevOps engineers, infra leads, and the executives they report to. If that's your buyer, this slot is available.
A story instead of seven headlines, since this one deserves the room. It's also the reason Issue 01 had a mistake in it.
Pure Storage renamed itself Everpure, effective February 23, 2026. If you searched "Pure Storage" this week and ended up on everpuredata.com, that's not a phishing redirect. That's just where the company lives now. The ticker stays PSTG on the NYSE for the moment, though Barchart has reported plans to shorten it later this year.
The company's own explanation is that it's "gone beyond storage to become a storage and data management company," and that Everpure fits a portfolio that now includes Pure Fusion and an Enterprise Data Cloud pitch, not just arrays. Outside coverage is less generous about it: this reads like a B2B infrastructure vendor renaming itself on the way upmarket, betting that owning "data primacy" — the layer AI systems actually reason over, not just the disks underneath — is a bigger and more defensible business than selling storage hardware. Whether that bet pays off against NetApp's AI Data Engine or Dell's AI-attach push is the thing actually worth watching, not the new name.
One piece of storage or memory vocabulary, explained properly, every week.
In plain English: it's the model's short-term memory of everything it's already "read" in your conversation, kept around so it doesn't have to reprocess that text token by token.
When a model generates text, it computes a "key" and a "value" vector for every token it processes. The KV cache is just those vectors, stored so the model can reuse them instead of recalculating them from scratch for every new token. Model weights are a fixed size. The KV cache is not — it grows with context length and with how many people you're serving at once. Feed a 128K-token context into Llama 3 70B for a single user and the cache alone runs around 40GB, roughly half of one H100's memory, before you've served a second person.
That's the whole reason KV cache management turned into a storage problem instead of a pure GPU problem. When the cache doesn't fit, you have two choices: evict it and recompute later, which costs latency, or push it somewhere else and pull it back when needed — CPU RAM first, then NVMe, then further out if you have to. The second option only pays off if fetching the cache back beats just recomputing it, which comes down to context length, model size, and how fast your fabric actually is.
Nvidia made this official at CES 2026 with ICMSP (Inference Context Memory Storage Platform), a standard for offloading KV cache straight to NVMe over RDMA using GPUDirect Storage, claiming up to 5x better power efficiency and up to 5x more tokens per second under the right conditions. The partner list reads like a rerun of Issue 01's vendor section: Dell, HPE, IBM, VAST Data, WEKA, Everpure, DDN, and Supermicro. In one benchmark, a 128K context over 400Gbps RDMA took time-to-first-token from 11 seconds down to 1.5.
AI Infra Summit 2026, Santa Clara. About as close as this space gets to a dedicated conference. Expect vendors to time announcements around it.
Q3 earnings season for storage and memory vendors (Samsung, SK Hynix, Western Digital, Seagate). Watch for capex guidance updates and anything on HBM4 qualification.
SC26, Chicago. The HPC/storage world's biggest annual gathering. Parallel file system and interconnect announcements tend to cluster here.