Everpure is Pure Storage under a new name (February 23, 2026; ticker P since April 17; S&P 500 on September 21, 2026) with a data-management thesis bolted onto a storage company: the Enterprise Data Cloud, which the company itself calls 'an architectural approach ... not a product'. The map reads it moderate at six layers and gap at two, with no strong cell. Storage is the deepest cell, strong at 0C: Purity's data services across block, file, and object, SafeMode immutability, and ActiveCluster and ActiveDR, with Fusion and Pure1 managing the fleet at 2A. Layer 1A is a gap: Data Intelligence (the acquired 1touch), announced 'available now' on June 17, 2026 as a catalog across clouds, SaaS, mainframes, and third-party storage, has no documentation on Everpure's support portal and is named, not scored, because without product documentation it isn't GA, so nothing on Everpure's paper answers what the data is. 0C is strong on a software-defined all-flash substrate the enterprise owns or subscribes to; 0A and 0B are gaps, because compute and fabric come only through Cisco's FlashStack and NVIDIA's DGX designs on partners' paper. Layer 1B is moderate, on Data Stream 1.0.0 (August 12, 2026): an on-appliance ingestion, embedding, hybrid-search, and chat appliance on FlashBlade with eight NVIDIA GPUs, held to moderate by its first-release limits (one node, 750,000 pages). Layer 1C is moderate on fleet replication and mobility plus Data Stream's preparation pipeline. Layer 2A is moderate on Fusion presets, Pure1 capacity management, and Portworx on the enterprise's Kubernetes, with no compute scheduling. Layer 2B is moderate on Data Stream's OpenAI-compatible chat endpoint over a sealed on-appliance model plus two MCP servers for storage operations; the first draft read it gap on the NetApp precedent. Layer 2C and Layer 3 are gaps: governance of storage actions isn't a plane over agents, and there's no application.
The capture is total below the protocol and standard at it. Purity, DirectFlash, Fusion, Pure1, Data Stream, Portworx Enterprise, and SafeMode are Everpure's with no second implementer, and Evergreen//One sells the same arrays on Everpure's paper by the gibibyte; that's why 13 of the 17 components read Ceded. What lifts: the object store behind S3, the front-end protocols (NFS, SMB, NVMe over Fabrics, Fibre Channel) the enterprise's hosts speak, PX-CSI's Kubernetes door, and the open-source Fusion MCP Server the enterprise runs itself. NVIDIA NIM inside Data Stream is Ceded to NVIDIA under the channel-substitution rule, GPUDirect Storage is NVIDIA's stack rather than a standard, and the two hyperscaler DirectFlash design wins move nothing on the customer's seat.
The buyer's trade: a storage fleet that governs and protects itself, and a sealed RAG appliance from the storage vendor, in exchange for every format, policy, and control plane being Everpure's and every model being NVIDIA's. The decision-authority readings follow: vendor / Ceded at 0C, 1C and 2A; Absent at 0A, 0B and 1A (Purity and SafeMode enforce by design; choosing an array, a pod, or a replication target is configuration of a relationship, not an override of a decision Everpure makes at run time, and nothing Everpure ships schedules compute); vendor / Delegated at 1B on Data Stream's documented per-request search parameters; model / Ceded at 2B, where the appliance's undisclosed model answers over its own index with no swap and no gate; Absent at 2C and Layer 3. What would move cells: documentation for Data Intelligence (1A, from gap), multi-node Data Stream (1B), Workload Rebalance and Mobility shipping in Q4 2026 (1C and 2A), a model catalog or agent runtime on Everpure's paper (2B), and a first-party application or a formal ecosystem program (Layer 3).
Layer-by-layer status: Layer 0A (No Compute on Everpure's Paper), Layer 0B (No Fabric on Everpure's Paper), Layer 0C (All-Flash Arrays on DirectFlash Under Purity, Owned or on Subscription, and Purity in Your Cloud Account), Layer 1A (No Scored Catalog or Data Governance; Data Intelligence Announced, Undocumented), Layer 1B (Data Stream 1.0.0 (August 12, 2026): an On-Appliance Ingest, Embedding, Hybrid Search, and Chat Appliance on FlashBlade With NVIDIA GPUs, Single-Node and 750,000 Pages in Its First Release; Data Intelligence Context Over MCP Announced, Undocumented), Layer 1C (Fleet-Wide Replication and Mobility (ActiveCluster for Block, ActiveDR, CloudSnap, Portworx Disaster Recovery and Migration) and Data Stream's Ingest, Curation, Embedding, and Indexing Pipeline; KV Cache Accelerator, File ActiveCluster, and Workload Rebalance Undocumented or Pending), Layer 2A (Fusion Fleet Automation With Presets, Workloads, and Placement Recommendations; Pure1 Capacity Management; Portworx for Kubernetes and KubeVirt VMs on the Enterprise's Clusters With a Standard CSI Door; No Compute or GPU Scheduling), Layer 2B (Data Stream's On-Appliance Language Model Behind an OpenAI-Compatible Chat Endpoint, Answering Only Over Its Own Index; the Open-Source Fusion MCP Server for Storage Operations (Pure1 Manage MCP Undocumented); No General Serving, No Agent Runtime), Layer 2C (Attribute-Based Data Policy for Agents Over MCP and a Human-in-the-Loop Gate on Storage Actions; Not a Plane Over Agents), Layer 3 (+1) (No Value Plane: Storage Beneath Others' Applications; Copilot Workflow Execution Is a Storage-Operations Assistant With an Unconfirmed Q2 2026 Date; Validated Designs Are Reference Architectures).
Assessment framework: 4+1 Layer AI Infrastructure Model. Scoring model: Decision Authority Placement Model (DAPM) — Retained, Delegated, or Ceded. Published by The CTO Advisor LLC (DBA The Advisor Bench). Author: Keith Townsend. Date assessed: October 27, 2026. Version: v1.18 - 4+1 v3: Layer 0 Sub-Layers.
Everpure is Pure Storage under a new name (February 23, 2026; ticker P since April 17; S&P 500 on September 21, 2026) with a data-management thesis bolted onto a storage company: the Enterprise Data Cloud, which the company itself calls 'an architectural approach ... not a product'. The map reads it moderate at six layers and gap at two, with no strong cell. Storage is the deepest cell, strong at 0C: Purity's data services across block, file, and object, SafeMode immutability, and ActiveCluster and ActiveDR, with Fusion and Pure1 managing the fleet at 2A. Layer 1A is a gap: Data Intelligence (the acquired 1touch), announced 'available now' on June 17, 2026 as a catalog across clouds, SaaS, mainframes, and third-party storage, has no documentation on Everpure's support portal and is named, not scored, because without product documentation it isn't GA, so nothing on Everpure's paper answers what the data is. 0C is strong on a software-defined all-flash substrate the enterprise owns or subscribes to; 0A and 0B are gaps, because compute and fabric come only through Cisco's FlashStack and NVIDIA's DGX designs on partners' paper. Layer 1B is moderate, on Data Stream 1.0.0 (August 12, 2026): an on-appliance ingestion, embedding, hybrid-search, and chat appliance on FlashBlade with eight NVIDIA GPUs, held to moderate by its first-release limits (one node, 750,000 pages). Layer 1C is moderate on fleet replication and mobility plus Data Stream's preparation pipeline. Layer 2A is moderate on Fusion presets, Pure1 capacity management, and Portworx on the enterprise's Kubernetes, with no compute scheduling. Layer 2B is moderate on Data Stream's OpenAI-compatible chat endpoint over a sealed on-appliance model plus two MCP servers for storage operations; the first draft read it gap on the NetApp precedent. Layer 2C and Layer 3 are gaps: governance of storage actions isn't a plane over agents, and there's no application.
The capture is total below the protocol and standard at it. Purity, DirectFlash, Fusion, Pure1, Data Stream, Portworx Enterprise, and SafeMode are Everpure's with no second implementer, and Evergreen//One sells the same arrays on Everpure's paper by the gibibyte; that's why 13 of the 17 components read Ceded. What lifts: the object store behind S3, the front-end protocols (NFS, SMB, NVMe over Fabrics, Fibre Channel) the enterprise's hosts speak, PX-CSI's Kubernetes door, and the open-source Fusion MCP Server the enterprise runs itself. NVIDIA NIM inside Data Stream is Ceded to NVIDIA under the channel-substitution rule, GPUDirect Storage is NVIDIA's stack rather than a standard, and the two hyperscaler DirectFlash design wins move nothing on the customer's seat.
The buyer's trade: a storage fleet that governs and protects itself, and a sealed RAG appliance from the storage vendor, in exchange for every format, policy, and control plane being Everpure's and every model being NVIDIA's. The decision-authority readings follow: vendor / Ceded at 0C, 1C and 2A; Absent at 0A, 0B and 1A (Purity and SafeMode enforce by design; choosing an array, a pod, or a replication target is configuration of a relationship, not an override of a decision Everpure makes at run time, and nothing Everpure ships schedules compute); vendor / Delegated at 1B on Data Stream's documented per-request search parameters; model / Ceded at 2B, where the appliance's undisclosed model answers over its own index with no swap and no gate; Absent at 2C and Layer 3. What would move cells: documentation for Data Intelligence (1A, from gap), multi-node Data Stream (1B), Workload Rebalance and Mobility shipping in Q4 2026 (1C and 2A), a model catalog or agent runtime on Everpure's paper (2B), and a first-party application or a formal ecosystem program (Layer 3).
CPUs, accelerators, servers, integrated systems, OS and hypervisor, physical plant
Assessment pending
Everpure sells no compute. FlashStack and AIRI//S sit under Systems on its products page, but the compute in them is Cisco's UCS and NVIDIA's DGX, sold through channel partners, the NetApp FlexPod shape. Cloud Block Store's instances read on the clouds' rows. Calibration: WEKA, DDN, NetApp and VAST read gap here too; Dell and HPE sell storage and compute together.
Nothing offered and nothing inherited from Everpure at 0A: Absent.
Named, not scored: FlashStack for AI Factories (Cisco UCS C845a, Nexus 9000, NVIDIA AI Enterprise, FlashBlade//S; GA Q1 2026 through channel partners) and AIRI//S, both on Cisco's and NVIDIA's paper; the DirectFlash design wins with two top-five hyperscalers, supply agreements inside those clouds and no surface the enterprise buys.
Interconnects, switching and routing, network virtualization
Assessment pending
Everpure sells no switches. The Nexus switches in FlashStack are Cisco's paper, and the enterprise's own switches carry NFS, SMB, S3, NVMe over Fabrics and Fibre Channel. Calibration: VAST and NetApp read moderate here because they sell cluster switches; WEKA and DDN read gap.
Absent: the fabric is the enterprise's own purchase.
Arrays, storage servers and software, front-end protocols: where the data sits
Everpure's "family of scale-up, unified block and file storage systems," serving "NFS and SMB • Fibre Channel, iSCSI, and NVMe-oF" per the family datasheet, racked and administered by the enterprise. Self-deployable isn't Retained; the modules, the array designs, and Purity are Everpure's: Ceded, the NetApp AFF and VAST reading.
Everpure's software on the hyperscaler's substrate in the enterprise's account; the substrate is the cloud's row, the storage is Everpure's. Ceded. Hosts attach over iSCSI or NVMe/TCP: "storage connectivity uses iSCSI or NVMe/TCP depending on the host integration pattern."
Standards the enterprise's hosts and switches speak, pNFS among them for FlashBlade//EXA; the array behind them is Everpure's. Delegated, the NetApp storage-networking reading. GPUDirect Storage is NVIDIA's cuFile stack, named in the NVIDIA note, not here.
The storage half of Purity Data Services: serving, snapshots, replication and data reduction on Everpure's arrays, in Everpure's formats. Ceded, the NetApp ONTAP reading.
Buckets and objects behind a multi-vendor interface; the data lifts to any S3 store. Delegated, the NetApp StorageGRID and Dell ObjectScale reading. Not scored within this chip under the GA evidence table: FlashArray Object Storage ("it's the 1.0 product that's just coming out that's gonna be GA, in a couple of months.").
The protection half of Purity Data Services: SafeMode immutable snapshots and encryption at rest on Everpure's arrays. Ceded, as the chip reads.
Immutable recovery points no administrator can delete, Everpure's software. Ceded, the Dell cyber-resilience reading. The Evergreen//One Cyber Recovery SLA (the clean-array guarantee with on-site services, the per-TiB uplift) is contract terms, not a component (Keith, October 16, 2026), and is named, not scored. Everpure Protect and the Human-in-the-Loop mandate are release-only and named in notes.
Carries volume and replica placement across nodes, pools and failure domains, and replication factors of 1, 2 or 3
Fusion's presets and storage classes set the protection, replication and placement the arrays apply: storage policy, a 0C function. The fleet plane itself is 2A.
The FlashArray//ST, whose datasheet lists "Connectivity NVMe-oF (RoCE)" on "2-port 100/200Gb Ethernet" and no Fibre Channel, iSCSI or file protocols. Purity and the array design are Everpure's: Ceded.
Everpure's scale-out file and object systems: FlashBlade//S "provides native multiprotocol access for NFS, S3, and SMB," and FlashBlade//E the same protocols. Purity and the DirectFlash designs are Everpure's: Ceded.
Everpure's parallel file system for AI: "Leveraging pNFS, the metadata core communicates with the compute cluster using NFSv4.1 over TCP," with data nodes on off-the-shelf servers serving RDMA. The metadata core and Purity are Everpure's: Ceded.
iSCSI and NVMe/TCP, the host connectivity Everpure documents, which independent implementations share: hosts and applications re-point to another server or target speaking the same protocols, and the data and provider-specific policies migrate with it. Delegated. Purity and Pure1's management stay in the Cloud Block Store chip, Ceded.
The Container Storage Interface, through which Portworx "implements a CSI driver that integrates with the Kubernetes storage framework," and NFS for sharedv4 volumes. Kubernetes' standard storage interface, which independent storage engines implement: workload manifests keep the Kubernetes storage contract, though changing drivers means reprovisioning and migrating the data. Delegated. The Portworx driver, and volume and replica placement, stay in the engine chip, Ceded.
FlashBlade//S is certified for DGX BasePOD and DGX B300 SuperPOD and integrates GPUDirect Storage, NVIDIA's cuFile and nvidia-fs stack. FlashBlade//EXA carries NVIDIA-Certified Storage validation and is being aligned with the STX reference architecture. The arrays themselves have no accelerator dependency.
Everpure sells the storage substrate. FlashArray (//XL, //X, //C, //E, //ST) and FlashBlade (//S, //E, //EXA) are all-flash systems on Everpure's DirectFlash modules under the Purity operating environment, racked in the enterprise's data center. On //EXA the metadata core and data nodes scale independently. Evergreen//One sells the same systems by consumption, and Cloud Block Store runs Purity in the enterprise's own AWS and Azure accounts. Purity serves block, file and object, takes snapshots, replicates with ActiveCluster and ActiveDR, and keeps SafeMode snapshots no administrator can delete. Portworx Enterprise places volumes and replicas on the enterprise's Kubernetes nodes. The buyer gets the flash tier of an AI factory as hardware, as a subscription, or as software in its cloud. The architect's concern is that the substrate is Everpure-defined at every layer below the protocol: Purity, DirectFlash and the array designs. Calibration: WEKA, DDN, VAST and NetApp read strong on the same function.
Ceded where the grade rests. Purity places, reduces and protects data inside the array, and the enterprise sets bounds through presets and storage classes, which is configuration: vendor decides, visible through Pure1, not overridable, Ceded. Portability: the arrays, Purity, SafeMode, Cloud Block Store and Portworx read Ceded; the front-end protocols and the object store behind S3 read Delegated.
Evergreen//One, //Flex and //Forever are consumption contracts, not components (Keith, October 16, 2026); Overdrive and the Cyber Recovery SLA are contract terms, named, not scored. Watch-listed: ActiveCluster for file (stated GA Q2 2026, unconfirmed). Named, not scored: NVIDIA STX and Vera BlueField-4 work, 'in development' (June 17, 2026). Out of the FlashBlade and FlashArray Object Storage Behind S3 chip under the GA evidence table (October 24, 2026): FlashArray Object Storage ("it's the 1.0 product that's just coming out that's gonna be GA, in a couple of months.").
The data itself, organized and governed: the Governance Catalog that Layer 2C queries
Purity, Fusion, Pure1, and Data Intelligence run on the arrays and Everpure's management plane; Data Intelligence's classification runs on CPUs.
No Everpure chip documents a 1A function. SafeMode, encryption and the Purity data services hold and protect the bytes, 0C functions; Fusion's presets are storage policy at 0C; Fusion and Pure1 as fleet planes are 2A. The catalog that would answer what the data is, Data Intelligence, was the former 1touch and went 'available now' on June 17, 2026. It's described as classifying data, finding personally identifiable information (PII) and protected health information (PHI), and tracking lineage. No documentation for it exists on Everpure's support portal, so it's named, not scored. Calibration: Red Hat reads gap on the same ground; WEKA and Dell read moderate on metadata-only indexes.
Absent: Everpure ships no scored catalog or data governance at 1A, and nothing here is inherited from it.
Ruled September 15, 2026: 'available now' isn't GA. Named, not scored, from the May 19, 2026 release only: Everpure Protect threat correlation and the Human-in-the-Loop mandate for sensitive data actions. Public evidence that moves the cell: Data Intelligence documentation, which on its described scope would read strong.
Low-latency retrieval for RAG — vector/hybrid search, context windows
Everpure's index, search service, and rules on Everpure's storage, with NVIDIA's microservices inside. Ceded to Everpure for the appliance and to NVIDIA for NIM under the channel-substitution rule; the chat door is scored at 2B.
The supported first release is 'a single-node deployment (an 8-GPU offering)' on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs; the design 'extends upon the NVIDIA AI Data Platform reference design' and uses NVIDIA NIM microservices for model functions, co-engineered with Supermicro. The pipeline, streams, rules, and API are Everpure's; the GPU dependency is the enterprise's to carry.
Everpure's retrieval story is an appliance. Data Stream 1.0.0 (release notes dated August 12, 2026, after the June 17 'availability' announcement and the March 16 beta statement) is, in its own terms of use, Everpure's 'data ingestion, indexing, search, and generative AI software appliance': it ingests from NFS shares and S3-compatible buckets, runs configurable embedding rules with concurrent text and vision-language models, auto-classifies through a classifier model (the classification rules tab is a beta feature), indexes, and serves 'hybrid semantic and keyword search (BM25) including document previews and optional LLM summarization' (LLM: large language model) through a REST API whose search endpoint takes query, stream selection, search mode, top_k, vector_db_top_k, and vision processing per request, and whose chat endpoint is 'OpenAI-compatible' against an on-appliance large language model; it honors each data source's authorization so users retrieve only what they may read; everything stays on the appliance and the FlashBlade; licensing is token-based for search queries and generative requests per purchased system. The supported release is a single node with eight NVIDIA GPUs, no multi-node clustering, 200,000 pages of initial ingest per stream, 750,000 pages of total platform capacity, 30 MB per document, on FlashBlade//S200, //S500, or //E with Portworx's Container Storage Interface (CSI) driver provisioning the volumes. Data Intelligence's knowledge graph as agent context over MCP is announced (June 17, 2026) and undocumented, named at 1A. The buyer gets retrieval-augmented generation (RAG) over its files as a sealed appliance from the storage vendor. The architect's concern is scale and youth. One node, three quarters of a million pages, a beta classification tab, an undisclosed on-appliance model, and a first release four weeks old; the retrieval quality and the model choice are Everpure's, and the record of the product's status ran from 'beta launching later in 2026' (March) to 'availability' (June) to a 1.0.0 release note (August). Calibration: NetApp reads moderate on storage-integrated vectorization and RAG endpoints; Dell moderate on an Elastic-powered search engine; HPE moderate on a reference retrieval stack; VAST strong on native vector search at scale with permission-aware retrieval; Elastic and MongoDB strong on native engines. Everpure now has an index and a search service of its own, the VAST shape in miniature, held to moderate by first-release scope gates. Moderate (ruled September 15, 2026): the scope-gate test names first-release maturity as a documented reason a real engine reads moderate, and the single-node, 750,000-page scope is a second gate; strong had been argued on the documented SearchService and EmbeddingService (the VAST reading).
Ceded at the appliance, Delegated at the request. Data Stream's streams, embedding and curation rules, index, search service, and chat surface are Everpure's software on Everpure's storage: Ceded; the NIM microservices inside are NVIDIA's on an NVIDIA license: Ceded to NVIDIA under the channel-substitution rule, the SUSE and Dataiku NIM reading; vectors made on the appliance belong to the model that made them: Ceded under the carve-out. The OpenAI-compatible chat endpoint is a standard door onto a captive index and a fixed model, not a model the enterprise can point elsewhere, so it doesn't earn the inference-interface Delegated reading. The runtime call is the appliance retrieving and ranking per the request's search mode, top_k, vector_db_top_k, and stream selection, documented per-request parameters that change the retrieval outcome: vendor decides, visible (results and alerts), overridable per request, Delegated, the Elastic reading under the override rule.
Ruled September 15, 2026: moderate stands on maturity and scope gates (single node, 750,000 pages, beta classification tab, release 1.0.0 dated August 12, 2026, the Elastic first-release case); strong had been argued on Everpure owning the index and search service (the VAST reading); multi-node clustering or a second release would move it. Availability record: March 16, 2026 'Data Stream Beta, launching later in 2026'; June 17, 2026 'announced the availability'; product page 'now available' with early-access language; release notes 1.0.0 August 12, 2026; the release notes govern. Trade coverage's 'billions of documents' and 'doesn't need an Everpure array' are contradicted by the limits and architecture documentation and dropped. Public evidence that moves the cell: multi-node clustering and capacity beyond the first release would argue strong; a documented on-appliance model choice would inform the DAPM.
Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering
Everpure's replication protocols between Everpure's endpoints and the enterprise's cloud account. Ceded, the NetApp SnapMirror reading.
The pipeline function of Data Stream, scored here for what it moves and prepares; the search it serves is the 1B chip and the chat the 2B chip. Everpure's orchestration with NVIDIA's model stages: Ceded. Not scored within this chip under the GA evidence table: Classification Rules ("The feature is marked as Beta because we are continuing to improve its robustness").
The ingest, embedding, and classification stages run on the appliance's eight NVIDIA GPUs with NIM microservices for the model functions; Everpure owns the pipeline. The replication surfaces carry no accelerator dependency.
Everpure moves data between its own systems and, with Data Stream, from files into a model. Movement: ActiveCluster synchronous replication across FlashArrays (and, announced March 11, 2026 for Q2 general availability, across file systems with fleet-level, policy-driven mobility integrated with Fusion), ActiveDR continuous replication, asynchronous replication and CloudSnap snapshot offload to object storage and the public cloud, Cloud Block Store as a cloud landing zone with a common data plane, and Portworx's migration and disaster recovery for Kubernetes applications and KubeVirt virtual machines across clusters, sites, and clouds. Transformation: Data Stream is a fixed-function pipeline, ingestion from NFS and S3 sources through parsing, curation and classification rules, embedding, and indexing on the appliance's GPUs, with stream owners defining sources and rules and running or pausing streams, collapsing 'two to four months of data preparation into minutes' by Everpure's account (the search and chat it serves are scored at 1B and 2B). Tiering: the KV Cache Accelerator on FlashBlade, which keeps inference key-value caches on flash and injects them back into GPU memory over remote direct memory access (RDMA), is described in the June 17 release and an Everpure technical blog with no product documentation, so it's named, not scored. The buyer gets its data wherever the fleet is and an AI-ready copy of its files without writing the pipeline. The architect's concern is fixity. Every mover here moves Everpure's data between Everpure's systems or the enterprise's cloud account; Data Stream is a pipeline whose stages are NVIDIA's blueprint, with no arbitrary code and no connector catalog; Workload Rebalance and Mobility (automatic movement across the fleet) is 'Available Q4 2026'; and nothing here does lineage across systems (Data Intelligence's lineage is scored at 1A). Calibration: NetApp reads moderate on a fixed AI-ingest pipeline plus SnapMirror, FlexCache, and FabricPool; Dell moderate on a data orchestration engine with a marketplace; VAST strong on a general DataEngine with user functions; Confluent and Cloudera strong on movement platforms. Everpure is NetApp's shape exactly. Moderate.
Ceded at every mover, and at the decision. ActiveCluster, ActiveDR, CloudSnap, and Portworx migration are Everpure's protocols between Everpure's endpoints: Ceded, the NetApp SnapMirror reading; Data Stream's pipeline is Everpure's orchestration with NVIDIA's model stages: Ceded (NIM to NVIDIA). The runtime call is replication and the pipeline running the policy and rules the enterprise configured; choosing a target array, pod, or replication relationship is configuration of the relationship, not an override of a movement decision Everpure makes at run time, and the documentation shows none: vendor decides, visible, not overridable, Ceded, the NetApp reading.
Named, not scored: the KV Cache Accelerator (release and blog only, no documentation), ActiveCluster for file (stated Q2 2026, unconfirmed), Workload Rebalance and Mobility (Q4 2026); all watch-listed where dated. Authority written Ceded on the NetApp reading after the ChatGPT pass argued the per-object replication choice is configuration; the same line the Hugging Face row drew at 2A. Public evidence that moves the cell: a general pipeline engine that runs enterprise code (the VAST DataEngine shape) would argue strong; KV Cache Accelerator documentation would add a scored chip. Out of the Data Stream as the Ingest and Preparation Pipeline chip under the GA evidence table (October 24, 2026): Classification Rules ("The feature is marked as Beta because we are continuing to improve its robustness").
GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization
Everpure's fleet automation over Everpure's arrays. Ceded, the NetApp Console and VAST Polaris reading.
Everpure's hosted operations plane sizing Everpure's service inside the enterprise's subscription. Ceded.
Proprietary storage and data-management software the enterprise runs on its own clusters. Ceded; the Kubernetes underneath is the enterprise's, the Cloudera customer-run OpenShift reading applies to it, not to Portworx.
A standard interface in front of Everpure volumes; the claims and classes lift to any CSI driver. Delegated, the NetApp Trident and Dell CSI Operator reading.
Fusion and Portworx orchestrate storage and stateful workloads; the GPUs in an Everpure-fed factory are scheduled by the enterprise's Kubernetes or NVIDIA's software, scored on their rows.
Everpure orchestrates storage across a fleet and stateful workloads across the enterprise's Kubernetes. Fusion, in Purity, manages arrays as one fleet: presets template a workload's storage (volumes, protection, replication, allowed storage classes, dynamic inputs such as a billing identifier), workloads are provisioned from them with the administrator choosing the array within the preset's allowed class or asking Pure1 for a placement recommendation, and policy for performance, protection, compliance, and cost is enforced through an API-first control plane; Compliance and Agentic Triage (drift detection with agentic root-cause suggestions) and Workload Rebalance and Mobility (automatic movement across the fleet) are 'Available Q4 2026'. Pure1 adds capacity automatically at 80 percent of allocation on Evergreen//One, holds 25 percent or 100 TiB of headroom, and, per the June 17, 2026 release, Copilot Workflow Execution lets administrators 'use natural language to plan, validate, and trigger secure, end-to-end operations' ('Available Q2 2026'). Portworx Enterprise (version 3.7 documentation; 3.6.0 in March 2026) is a software-defined storage overlay on the enterprise's Kubernetes (OpenShift, EKS, AKS, GKE, Rancher, upstream) with PX-CSI to FlashArray and FlashBlade, Autopilot, Stork, backup and disaster recovery across clusters, KubeVirt virtual machines ('over a hundred thousand VM Volumes on Kubernetes'), and, since May 11, 2026, Plugin 2.2 bringing storage, protection, and disaster recovery into the OpenShift console with Advanced Cluster Management. The buyer gets storage that provisions and protects itself by policy and a Kubernetes data layer that follows its applications and VMs. The architect's concern is that nothing here schedules compute. Fusion orchestrates arrays, Portworx orchestrates volumes, and the placement decisions that matter to an AI factory (which GPU, which node, which job) are made elsewhere; the automatic movers are Q4 2026. Calibration: NetApp reads moderate on a data-infrastructure control plane with no GPU scheduling; VAST moderate on Polaris; Dell moderate on rack-level management; Nutanix strong on Prism, ADS, and a Kubernetes platform; HPE strong on GreenLake. Everpure is NetApp's shape with a stronger Kubernetes story through Portworx. Moderate.
Ceded at the planes, Delegated at the CSI door. Fusion, Pure1, and Portworx Enterprise are Everpure's software with no second implementer: Ceded; PX-CSI speaks the Kubernetes CSI standard (PersistentVolumeClaims and StorageClasses) in front of Everpure volumes: Delegated at that interface, the NetApp Trident and Dell CSI Operator reading; the Kubernetes underneath is the enterprise's and is scored on its own row. The runtime call the layer asks about is scheduling: which node, which GPU, which job, what share. Nothing Everpure ships makes or overrides that decision; Fusion's array choice is storage provisioning, and Pure1's recommendation is advice. Vendor decides, visible, not overridable, Ceded, the NetApp Console and VAST Polaris reading.
Unscored on their dates: Copilot Workflow Execution and Enhanced Cyber Anomaly Detection ('Available Q2 2026', stated June 17, 2026, no confirming document since), Fusion Compliance and Agentic Triage and Workload Rebalance and Mobility (Q4 2026), Evergreen//One Overdrive (Q3 2026); all watch-listed. Portworx Data Services (database-as-a-service on Kubernetes) is named on the documentation index and not detailed in the source set. Authority written Ceded after the ChatGPT pass argued the Fusion array choice is storage provisioning, not the Layer 2A scheduling decision; the override-rule column item (September 5 sheet) stays open for the day Workload Rebalance ships. Public evidence that moves the cell: GPU or compute scheduling from Everpure, which nothing describes.
Model serving, agent execution, inference APIs, distributed inference
A fixed model on one box behind a standard door the enterprise can't redirect. Ceded to Everpure for the appliance and to NVIDIA for NIM; the interface shape doesn't lift the reading.
Open-source code the enterprise runs on its own machine with its own API tokens; the tools act on Everpure resources. Retained on portability, the Confluent open-source MCP reading.
Data Stream's generation runs on its eight NVIDIA GPUs with NIM microservices; the GenAI Pod validated designs (November 2024), AIRI with DGX BasePOD, and the FlashStack CVD for AI Factories (NVIDIA AI Enterprise on Cisco UCS, GA Q1 2026 through channel partners) put NVIDIA's serving stack on Everpure storage. Ceded to NVIDIA under the channel-substitution rule wherever it appears.
Everpure serves one model, sealed inside an appliance, and two tool servers over its own fleet. Data Stream 1.0.0's terms of use call it a 'generative AI software appliance': its ChatService is a 'Drop-in replacement for OpenAI with enterprise context utilizing on-appliance LLMs' with multi-turn chat and streaming, POST /chat/responses is 'OpenAI-compatible completion/responses', and 'Generative Output is produced entirely on End User's appliance using the search results available to Data Stream at the time of the request', licensed by tokens for 'search queries and generative output requests' per purchased system; the model is undisclosed, single-node, and answers only over the appliance's own index. Around the fleet: the Everpure Fusion MCP Server (open source, Apache 2.0, with a published user guide) gives an assistant a read-oriented tool surface over FlashArray and FlashBlade with optional, hidden-by-default supervised write tools (placement recommendation, preset create and update, workload deployment) that the enterprise's agent must approve; a Pure1 Manage MCP Server exposing read-only Pure1 APIs under Pure1 roles is 'generally available now' per the June 25, 2026 Everpure post but has no documentation, so it's named, not scored; Pure1 AI Copilot 'moves from a conversational assistant to an active operator' for anomaly triage inside Pure1. The validated designs (GenAI Pod, AIRI, FlashStack for AI Factories) put NVIDIA's runtime on Everpure storage on partners' paper. The buyer gets a RAG chat appliance and a way to let its assistants read and, with approval, act on its storage. The architect's concern is that there's no serving platform in any of it. The chat endpoint's standard shape hides a fixed model on one box answering only what Data Stream indexed; there's no model catalog, no bring-your-own model, no agent runtime, and Fusion Compliance and Agentic Triage is 'Available Q4 2026'. Calibration: NetApp reads gap because its AI Data Engine retrieves without generating and the runtime is NVIDIA's; Dell moderate on its own Agentic AI Platform blueprints plus services; HPE moderate on Private Cloud AI as a packaged platform; VAST strong on AgentEngine; Nutanix moderate on platform-native serving with agents pre-GA. Everpure sits between NetApp and Nutanix: a fixed-function generative appliance on its own paper (a fixed-function slice caps at moderate) plus MCP tool servers. Moderate (ruled September 15, 2026): a fixed-function generative appliance on the vendor's paper is a fixed-function slice; gap had been the first draft's reading on the NetApp precedent.
Ceded at the appliance, Retained at the open tool server. Data Stream's chat is a fixed on-appliance model over a captive index behind an OpenAI-compatible door the enterprise can't point elsewhere: Ceded (NIM to NVIDIA under channel substitution); the standard shape doesn't earn the inference-interface Delegated reading because the model and the context are the appliance's. The Fusion MCP Server is Apache 2.0 code the enterprise runs against its own arrays: Retained, the Confluent open-source MCP reading. The decision to answer is the appliance's model over the appliance's retrieval, with no model swap and no pre-effect gate documented; the write tools on Fusion MCP are gated in the enterprise's own agent, a client-side gate on storage operations rather than on the model: model decides, visible (results, alerts, audit), not overridable, Ceded, the Cohere and Perplexity reading.
Ruled September 15, 2026: moderate stands on Data Stream 1.0.0's documented generative endpoint (a fixed-function slice); gap had been argued on the NetApp reading. The Pure1 post (June 25, 2026) says both MCP servers are 'read-only by design'; the Fusion MCP Server user guide documents optional supervised write tools hidden by default; the documentation governs. The Pure1 Manage MCP Server (post only, no documentation) is named, not scored. Ruled September 15, 2026 (component): the ChatService is Ceded; Antigravity had argued Delegated at the interface on the Dataiku NIM reading, and the corollary now under the inference-interface ruling says a standard-shaped door with nothing substitutable behind it earns no Delegated. Named, not scored: GenAI Pod validated designs (2024; no current status document), AIRI//S, FlashStack CVD for AI Factories (Cisco's and NVIDIA's paper), Fusion Compliance and Agentic Triage (Q4 2026), Pure1 AI Copilot's agentic triage (post only). Public evidence that moves the cell: a model catalog or bring-your-own model on Data Stream, or an agent runtime on Everpure's paper, would argue strong; the gap reading is closed.
Policy-driven placement and resource coordination — the Autonomy Layer
Nothing at this layer is a scored capability.
Read against the five legs, Everpure has none as a plane over agents. Identity: Pure1 and Purity identify administrators and API tokens, and an MCP client inherits the token's Pure1 role; Data Intelligence's announced attribute-based access controls 'as AI systems and agents interact directly with data' would govern what an agent may read, not who the agent is, and are undocumented. Gateway: the Fusion MCP Server (and the undocumented Pure1 Manage MCP Server) sit in front of an agent's calls to Everpure storage (read-only, or supervised writes the enterprise's agent approves), scored or named at 2B; nothing Everpure ships sits in front of an agent's model calls or its tool calls in general. Registry and orchestration: none. Observability: Pure1 and Everpure Protect watch the storage estate, and Enhanced Cyber Anomaly Detection ('Available Q2 2026') watches login patterns and behavioral drift across arrays, storage telemetry rather than agent traces. What Everpure does give the reasoning plane is a governance substrate: SafeMode recovery points no administrator can delete, and a Human-in-the-Loop mandate requiring multi-party, out-of-band authorization for sensitive data actions, which is a deterministic gate on storage operations that an agent's action would also hit. The architect's concern is the same as NetApp's: data guardrails aren't agent governance, and every control here is on Everpure's resources. Calibration: NetApp reads gap ('data guardrails ≠ agent governance'); Dell and VAST gap; HPE moderate on GreenLake Intelligence as an IT-operations reasoning plane; Nutanix moderate on an agent gateway. Everpure is NetApp's shape with a stronger recovery gate and a tool server. Gap, authority Absent; the ChatGPT pass argued a thin moderate on the MCP servers, rejected on the Confluent and Snowflake precedent that vendor MCP servers are 2B tool surfaces.
Nothing offered as a plane over agents. The enterprise that wants agent identity, a request gateway, a registry, or orchestration brings its own; Data Intelligence's policy travels with the data, not with the agent. Absent.
Named, not scored: Data Intelligence attribute-based access for agents over MCP (announced, undocumented), the Human-in-the-Loop mandate (release only), Enhanced Cyber Anomaly Detection (stated Q2 2026, unconfirmed), Fusion Compliance and Agentic Triage (Q4 2026), Pure1 AI Copilot's agentic triage. The Fusion MCP Server is scored at 2B; the Pure1 Manage MCP Server is named there. Public evidence that moves the cell: an agent principal, a gateway over agent requests generally, or a registry of the enterprise's agents; nothing in the documentation describes one.
AI-powered business capabilities — business logic, workflow automation
The GenAI Pod validated designs automate NVIDIA NeMo and NIM deployments through NVIDIA AI Enterprise, the drug-discovery solution white paper is built on NVIDIA BioNeMo, and the FlashStack CVD runs NVIDIA AI Enterprise; nothing is scored, and every application pattern Everpure names runs on NVIDIA software.
Everpure sells no application. Copilot Workflow Execution, a natural-language assistant for storage administrators to 'plan, validate, and trigger secure, end-to-end operations across the entire global infrastructure estate', was stated 'Available Q2 2026' in the June 17, 2026 release and has no confirming document since; even shipped, it's an infrastructure-operations assistant of the kind the HPE row scores at 2C and the Confluent row carries in notes, not a business capability. The GenAI Pod validated designs (November 2024: drug discovery, trade research and investment analysis with KX, and RAG-based knowledge management and chatbots, with Meta, NVIDIA, Red Hat, Supermicro, WWT, Cisco, and Arista; 'expected to be generally available in the first half of 2025'), the life-sciences page's 'Everpure GenAI Solution for Drug Discovery' on NVIDIA BioNeMo (a white paper with deployment guidance), AIRI, and Cisco FlashStack for AI Factories are reference architectures and validated designs the enterprise builds applications on, none with product documentation or a program page of its own, and the partner ecosystem (Cisco, NVIDIA, Supermicro, Commvault, Kyndryl, Red Hat, Microsoft) is a go-to-market, not a program that ships applications on Everpure's paper. The buyer gets nothing to log into at this layer. The architect's concern is nil: there's no application to be captured by. Calibration: NetApp reads gap ('no value plane, data foundation beneath others' apps'); HPE reads partner because customers buy ISV applications from HPE on the GreenLake Marketplace; Dell, VAST and Nutanix read gap, their AI ecosystem programs certification rather than procurement and platform enablement earning no Layer 3 grade. Everpure is NetApp's shape. Gap, authority Absent. Partner had been argued on the Dell and HPE reading if the GenAI Pod applications, the BioNeMo drug-discovery solution, and the FlashStack designs count as an ecosystem program; under the Layer 3 marketplace ruling (September 11, 2026) a partner shelf earns no grade and the evidence stays unscored on the GA-gate, not absent; the MongoDB and Redis developer-assistant question doesn't reach an operations copilot.
Nothing offered, nothing inherited. Absent.
Ruled September 11, 2026: gap stands on the Layer 3 marketplace ruling; partner had been argued on the Dell and HPE ecosystem-program reading. Named, not scored: Copilot Workflow Execution (stated Q2 2026, unconfirmed; watch-listed), the GenAI Pod validated designs and their named applications (2024, no current status document), the GenAI Solution for Drug Discovery on NVIDIA BioNeMo (white paper), AIRI//S, FlashStack for AI Factories. Public evidence that moves the cell: a first-party application or a named ecosystem program shipping ISV applications on Everpure's paper.
Everpure press releases (Pure Storage Becomes Everpure and Announces Intent to Acquire 1touch, February 23, 2026, with the May 11, 2026 completion note; Everpure Extends ActiveCluster Support for File, March 11, 2026; Everpure Simplifies Enterprise AI with Evergreen//One for AI and Data Stream Beta, March 16, 2026; Portworx by Everpure Is Redefining Modern Virtualization, March 24, 2026; Everpure to Change Ticker Symbol, April 7, 2026; Portworx Makes Data Management Native to Red Hat OpenShift, May 11, 2026; Everpure Redefines Cyber Resilience, May 19, 2026; Everpure Announces Data Stream and Everpure Unveils Data-Primacy Architecture, June 17, 2026; Everpure Lands New Design Win with Second Top-Five Hyperscaler, August 10, 2026; Gartner Magic Quadrant, August 21, 2026; Everpure to Join the S&P 500, September 8, 2026; Pure Storage Introduces New GenAI Pod, November 18, 2024; next-generation AIRI, 2022; Purity 6.0, 2020; Cloud Block Store on Azure, 2021); everpuredata.com product pages (products, block, file, and object storage, unstructured data storage, enterprise file, Evergreen storage as a service with Flex and Forever, Kubernetes data management, Data Intelligence, automation and orchestration, Enterprise Data Cloud, AI, NVIDIA, Cisco FlashStack, cloud, HPC, virtualization; What is Evergreen//One; the Fusion presets and workloads demonstration); pure.ai (Data Stream, FlashBlade//EXA); docs.portworx.com (index, Portworx Enterprise 3.7); NVIDIA AI Data Platform page; trade coverage (SiliconANGLE June 10, June 17, and September 4, 2026; TechTarget on Data Stream; StorageNewsletter). Peer-reviewed cell by cell through the labs claims ledger (everpure-<layer>-chatgpt, ChatGPT gpt-5.5) and as a whole row by Antigravity (everpure-row-agy); totals and escalated items in reviews/everpure-judgment.md. /reconcile (October 5, 2026): Layer 3 calibration citations updated for the Layer 3 re-reads (Nutanix, VAST, VMware to partner; CoreWeave, NVIDIA, AMD to gap) and the MongoDB BI ruling; no grade or authority change on this row. /reconcile (October 5, 2026): settled escalations cleared from the published text and ruling attributions neutralized; open escalations stay named as pending a ruling. No grade, component, or authority change. /ga-check (October 5, 2026): Evergreen//One Overdrive added to the Evergreen//One chip on the binding terms of use (everpuredata.com, updated July 24, 2026; Keith approved the terms as evidence). Grade unchanged. v1.4 (October 16, 2026, hardware-management scope ruling): a consumption contract (pay-per-use, storage as a service) isn't a component, so the control-plane software under it is what gets scored (Keith, October 16, 2026). The Evergreen//One chip moved from layer0 components to notes: the arrays under it are the FlashArray and FlashBlade chip, and the software that manages capacity to the SLA is Pure1, already scored on the Pure1 Operations chip at layer2a. Overdrive moves to notes with it. The SafeMode chip at layer1a now scores SafeMode and names the Cyber Recovery SLA as contract terms. Summary count 14 of 18 Ceded to 13 of 17. Layer 0 grade and authority unchanged. v1.5 (Layer 0 sub-layers, October 16, 2026): Layer 0 split into 0A Compute, 0B Network Fabric and 0C Storage Substrate by runtime decision (Keith, October 16, 2026); storage left 1A for 0C, and 1A, now Data Foundation & Governance, is graded on the catalog that classifies the data. 1A moderate to gap; 0A gap, 0B gap, 0C strong; summary edits (summary worksheet, Keith, October 16, 2026). v1.6 (October 19, 2026, /reconcile partner ruling): calibration prose updated to the new Layer 3 readings of peer rows; no grade or reading on this row changed. (Same pass) Plain-language pass: methodology rule numbers in the prose and audit notes replaced with the reasoning they stood for; no grade, chip or authority reading changed. v1.7 (October 23, 2026, /reconcile): 0C: Block arrays, file and object systems, and a parallel file system were one chip. 0C: The map had no protocol for Cloud Block Store. Found by the Layer 0 data pass after the sniff read. Grade and cell authority unchanged. v1.8 (October 23, 2026, storage door pass): The front-end door names pNFS; Cloud Block Store and Portworx get their own doors. Grade unchanged. v1.9 (October 24, 2026, /reconcile): 1B calibration updated: VAST reads moderate at 1B, InsightEngine unscored without a product guide. Found in the 1B dependency audit follow-ups. Grade unchanged. v1.10 (October 24, 2026, /reconcile): 1B calibration restored: VAST reads strong at 1B again, InsightEngine GA on VAST's own announcement. Grade unchanged. v1.11 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 1 Layer 1B chips under the GA evidence table. Grade unchanged. v1.12 (October 24, 2026, /reconcile): Layer 1B GA evidence migrated to the GA evidence table version 2. Grade unchanged. v1.13 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 2 Layer 1C chips under the GA evidence table version 2. Grade unchanged. v1.14 (October 24, 2026, /reconcile): GA evidence recorded on 13 of 14 Layer 0C chips under the GA evidence table version 2. Grade unchanged. v1.15 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 4 Layer 2A chips under the GA evidence table version 2. Grade unchanged. v1.16 (October 24, 2026, /reconcile): GA evidence recorded on 2 of 2 Layer 2B chips under the GA evidence table version 2. Grade unchanged. v1.17 (October 24, 2026, /reconcile): GA evidence table version 2.1 applied: 2 chips recorded or narrowed. Grade unchanged. v1.18 (October 27, 2026): each component now carries a durable offering id; nothing else changed.