Nokia's AI infrastructure is the network, with a service-provider platform beside it. Layer 0B is strong: 7220 IXR switches on Broadcom Tomahawk 5 and 6, with an April 2026 validated design that documents RDMA over Converged Ethernet (RoCEv2) congestion control, with Explicit Congestion Notification (ECN) firing before Priority Flow Control (PFC), against NVIDIA H200 endpoints. Three cells read moderate on Nokia's telco cloud stack on Red Hat OpenShift, sold to telecom service providers: compute on Dell and HPE servers (0A), Everpure storage (0C), and orchestration through OpenShift beside Event-Driven Automation (EDA), the fabric's fleet plane (2A). Nokia Mediation reads moderate at 1C as a usage-data pipeline for the same buyers, and Deepfield Defender reads moderate at Layer 3 as distributed denial-of-service (DDoS) mitigation for service providers, cloud builders and large digital enterprises. The data foundation, retrieval, the runtime and the agent plane are gaps. Authority sits in several places: the network operating system makes the fabric's runtime calls inside the enterprise's thresholds, the service provider owns and runs the servers, Everpure places the bytes, OpenShift's scheduler places workloads under overrides the operator can pin, Mediation executes what the operator configured, and Defender's models choose mitigations an operator can override.
The capture is coupled at the switch, with a real exit, and decoupled one layer up. The same hardware runs SR Linux or Community SONiC with Nokia support, an open network operating system (NOS) exit on the same hardware, as Cisco's 8000 offers. SR Linux's published YANG data models, NetOps Development Kit (NDK) bindings, free container image and containerlab make the platform feel open, and they're real conveniences. They don't port the opinions. Those accumulate in EDA's intents, apps and allocation pools, and EDA's multi-vendor support (Arista in beta, Cisco in alpha) is reach, not exit. The NOS choice looks like the lock-in decision. It's the one with an exit. The automation plane is the commitment buyers underprice.
The architect gets a fabric that stands with Cisco, Dell and HPE, pays for day-two operations with a captive automation plane, and brings AI compute and storage from another row: Lenovo in Nokia's own validated designs. The telco cloud stack and Mediation are real but narrow, built for operators running network functions and usage data, not for an enterprise's AI estate; Defender reaches further, to cloud builders and large enterprises. Nokia's industrial edge business is being sold to AURELIUS, and Data Suite stays named until its documentation is held. One piece deserves a reader's attention: EDA ships deterministic dry-run, gated commit and all-node rollback and an agent-facing Model Context Protocol (MCP) door into it, for its own fabric only. Buy Nokia for the fabric, and decide about EDA as deliberately as you'd decide about the switch.
Layer-by-layer status: Layer 0A (CPU Compute Inside the Telco Cloud Stack, Sold to Service Providers), Layer 0B (7220 IXR on Tomahawk 5 and 6, a Documented RoCEv2 Design, and a SONiC Exit), Layer 0C (Everpure Block Storage Inside the Telco Cloud Stack), Layer 1A (Data Suite Is CSP-Facing, Its Docs Gated), Layer 1B (Enterprise Responsibility), Layer 1C (Nokia Mediation, a Usage-Data Pipeline for Service Providers), Layer 2A (EDA Over the Fabric and OpenShift in the Telco Cloud Stack; No GPU Scheduling), Layer 2B (No Runtime; Ask EDA Is a Console), Layer 2C (An MCP Door Into EDA, Scoped to One Platform), Layer 3 (+1) (Deepfield Defender, a Documented Security Application).
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 19, 2026. Version: v1.3 - 4+1 v3: Layer 0 Sub-Layers.
Nokia's AI infrastructure is the network, with a service-provider platform beside it. Layer 0B is strong: 7220 IXR switches on Broadcom Tomahawk 5 and 6, with an April 2026 validated design that documents RDMA over Converged Ethernet (RoCEv2) congestion control, with Explicit Congestion Notification (ECN) firing before Priority Flow Control (PFC), against NVIDIA H200 endpoints. Three cells read moderate on Nokia's telco cloud stack on Red Hat OpenShift, sold to telecom service providers: compute on Dell and HPE servers (0A), Everpure storage (0C), and orchestration through OpenShift beside Event-Driven Automation (EDA), the fabric's fleet plane (2A). Nokia Mediation reads moderate at 1C as a usage-data pipeline for the same buyers, and Deepfield Defender reads moderate at Layer 3 as distributed denial-of-service (DDoS) mitigation for service providers, cloud builders and large digital enterprises. The data foundation, retrieval, the runtime and the agent plane are gaps. Authority sits in several places: the network operating system makes the fabric's runtime calls inside the enterprise's thresholds, the service provider owns and runs the servers, Everpure places the bytes, OpenShift's scheduler places workloads under overrides the operator can pin, Mediation executes what the operator configured, and Defender's models choose mitigations an operator can override.
The capture is coupled at the switch, with a real exit, and decoupled one layer up. The same hardware runs SR Linux or Community SONiC with Nokia support, an open network operating system (NOS) exit on the same hardware, as Cisco's 8000 offers. SR Linux's published YANG data models, NetOps Development Kit (NDK) bindings, free container image and containerlab make the platform feel open, and they're real conveniences. They don't port the opinions. Those accumulate in EDA's intents, apps and allocation pools, and EDA's multi-vendor support (Arista in beta, Cisco in alpha) is reach, not exit. The NOS choice looks like the lock-in decision. It's the one with an exit. The automation plane is the commitment buyers underprice.
The architect gets a fabric that stands with Cisco, Dell and HPE, pays for day-two operations with a captive automation plane, and brings AI compute and storage from another row: Lenovo in Nokia's own validated designs. The telco cloud stack and Mediation are real but narrow, built for operators running network functions and usage data, not for an enterprise's AI estate; Defender reaches further, to cloud builders and large enterprises. Nokia's industrial edge business is being sold to AURELIUS, and Data Suite stays named until its documentation is held. One piece deserves a reader's attention: EDA ships deterministic dry-run, gated commit and all-node rollback and an agent-facing Model Context Protocol (MCP) door into it, for its own fabric only. Buy Nokia for the fabric, and decide about EDA as deliberately as you'd decide about the switch.
CPUs, accelerators, servers, integrated systems, OS and hypervisor, physical plant
Sold directly by Nokia as part of its complete vertical stack on Red Hat OpenShift, under one SLA, to telecom service providers running CNFs. CPU compute; no accelerators documented. Commodity x86 from two OEMs under a platform that runs on any vendor's servers, so the workloads lift.
Assessment pending
Nokia sells compute in one place: its telco cloud on Red Hat OpenShift, which Nokia describes as available directly from Nokia, the complete vertical stack, under a single-vendor service-level agreement (SLA). The stack runs Nokia's cloud-native network functions (CNFs) and autonomous networks applications on Red Hat OpenShift, and a stack that runs them has to include the servers they run on, so the 0A credit follows from what the stack requires. Nokia's page names its platform partners: chipsets from AMD and Intel, servers from Dell and HPE. The offer is aimed at telecom service providers (the page says telcos and enterprises can choose the vertical stack, but the workloads in it are telecom network functions): general availability here means availability to the service providers the stack is sold to, not a general enterprise or AI compute offer. Nothing documents accelerators in it, so it reads as CPU-only compute, which the 0A rubric grades moderate. Nokia sells no AI compute. Its AI fabric designs bring servers from elsewhere: the NVIDIA validated design (April 2026) was tested on Lenovo SR685a V3 servers with H200 GPUs and BlueField-3 NICs, and the Lenovo-Nokia sovereign blueprint (LP2339) pairs AMD Instinct with Lenovo servers. AirFrame, Nokia's former data center server line, is gone from its site. AI-RAN isn't GA: anyRAN on NVIDIA ARC-Pro runs on Dell PowerEdge servers (Dell's paper, NVIDIA's reference design), with pilots at the end of 2026 and commercial availability stated for 2027. MX Industrial Edge doesn't score: its documented hardware is the discontinued AirFrame Open Edge server and, from 2023, a Dell PowerEdge XR11 with an NVIDIA A2, and a discontinued product disqualifies it; the unit is also moving to AURELIUS with the Enterprise Campus Edge sale. Calibration: Cisco, Dell and HPE read strong on accelerated systems they sell. Nokia's compute is general-purpose servers inside a service-provider platform, the moderate band.
Little. The servers are Dell and HPE x86 under Red Hat OpenShift, which runs on any vendor's servers, so the network functions and the platform lift to another OEM's hardware. The service provider owns and operates the hardware, and it makes no runtime tradeoff on Nokia's behalf. Lifecycle operations for the stack belong to 2A.
Watch item: AI-RAN on NVIDIA ARC-Pro, pilots at the end of 2026 and commercial availability in 2027 (Nokia release, July 15, 2026); the AirScale AI-RAN capacity plug-in unit has no date of its own. Not scored: MX Industrial Edge hardware (discontinued AirFrame base; Dell hardware; moving to AURELIUS, close expected H1 2027). NVIDIA's $1B equity investment in Nokia (October 28, 2025) is a capital fact, not a capability.
Interconnects, switching and routing, network virtualization
Tomahawk 5 and 6 AI back-end and front-end switching under SR Linux, documented in the SR Linux 26.7 hardware overview and the AI-DC validated design (DCQCN, ECN and PFC, DLB, rail-optimized stripes). SR Linux is proprietary; its YANG models and NDK are BSD-licensed and its container image is free to try, which makes the platform open to build against without making the opinions portable. Leaving means rebuilding the fabric's configuration on another NOS.
Upstream Community SONiC on Nokia hardware with Nokia support and a 7220 IXR-H SONiC datasheet; Nokia is a SONiC Foundation Premier member. Configuration built here lifts to any SONiC switch, so Nokia is a substitutable provider behind a multi-vendor NOS. Scoped to the platforms Nokia's SONiC page lists.
FP5-based routing for data center gateways, interconnect and backbone, as in CoreWeave's backbone. SR OS opinions (policy, services, routing configuration) are captive to the Nokia stack.
Line systems for data center interconnect and metro transport. Their management and configuration, where the opinions accumulate, are captive.
ICE-X ZR and ZR+ pluggables are interoperable with any Optical Internetworking Forum (OIF) compliant ZR transceiver and run in other vendors' routers. They swap like commodity hardware and accumulate no opinions of their own.
Programs the fabric from declared intent (Kubernetes custom resources rendered to device configuration) and decides when switch state changes or reverts: dry-run against YANG, all-or-nothing commit across nodes, revert on failure, drift reconciliation, Git-backed history. EDA's core is proprietary and its opinions live in its own app and resource model; support for Arista EOS (beta) and Cisco NX-OS (alpha) is reach, not exit. The same product carries a fleet-plane facet at 2A with the same reading.
Validated against NVIDIA H200 GPUs on Lenovo servers, with BlueField-3 NICs (ConnectX-7 also listed). The NICs make transport decisions: the design configures BlueField-3 QoS and congestion-notification handling, so the NIC-side congestion behavior is an NVIDIA dependency at 0B. The switches carrying the fabric are Nokia's. No Spectrum-X validation: SR Linux for Spectrum-X was announced as a collaboration in October 2025 with no deliverable, and NVIDIA's Spectrum-X stack lists only Cumulus.
This is Nokia's home layer, and it's strong. The 7220 IXR line covers the AI back end and front end: the H5 family on Tomahawk 5 (the H5-64D and H5-64O at 51.2 terabits per second, the H5-32D at 25.6) and the H6-64 on Tomahawk 6 (102.4T, 64 ports of 1.6T, air or liquid cooled), both in the SR Linux 26.7 hardware overview, beside the D-series leaves and the 7250 IXR chassis. The grade rests on product documentation, not press: the AI-DC Nokia Validated Design for NVIDIA (3HE-21918, Issue 3, April 2026) documents Data Center Quantized Congestion Notification (DCQCN) with Explicit Congestion Notification (ECN) and Priority Flow Control (PFC) thresholds (ECN fires before PFC, PFC before tail drop), dynamic load balancing, queue-pair hashing, rail-optimized stripes and multitenancy, tested on SR Linux 25.10.1 and EDA 25.12.2. That's the 0B criterion: a generally available (GA) remote direct memory access (RDMA) scale-out fabric with congestion control for collective traffic, and the switching to run it at cluster scale. Around the fabric, Nokia sells 7750 SR and 7250 routing for data center gateways and backbones (CoreWeave's backbone runs on 7750 SR FP5), and ex-Infinera optical (the acquisition closed February 28, 2025): 1830 and GX line systems for data center interconnect, ICE-D intra-data-center optics, and ICE-X pluggables that interoperate with any Optical Internetworking Forum (OIF) compliant ZR transceiver. Event-Driven Automation (EDA), the fabric's intent and automation plane, carries a facet here. Its transactions decide when switch state changes or reverts (all-or-nothing commit across nodes, revert on failure, drift reconciliation), and its fabric apps carry the partitioning and isolation intent. Its fleet-plane facet is scored at 2A. The two operating systems are the interesting part. The same hardware runs SR Linux, Nokia's own network operating system (NOS), or Community SONiC with Nokia support on listed platforms (7215 IXS-A1/C1, 7220 IXR-D4, H4 and H5, 7250 IXR-X3b and 10e). On SR Linux, opinions accumulate in its configuration model; the BSD-licensed YANG models and NDK don't change that, because published models of a single-vendor NOS don't run on another vendor's switches. On SONiC, the same configuration lifts to any SONiC switch. Calibration: Cisco is strong on Nexus Hyperfabric and Silicon One, Dell on PowerSwitch with Spectrum-X, HPE on Slingshot and Juniper. A Tomahawk 5 and 6 line with a documented RoCEv2 design stands with them; neither grade needs Spectrum-X. Not scored: Ultra Ethernet compliance (press release only) and the Nscale deployment (Nscale names Nokia its preferred networking partner, September 2025), which is deployment evidence the grade doesn't rest on.
The fabric's runtime calls are Nokia's. On the SR Linux path that carries the grade, the NOS decides routing, equal-cost multipath (ECMP) and dynamic load balancing, and ECN and PFC marking, inside thresholds the enterprise sets; setting thresholds is configuration, not override. On the SONiC path the same transport decisions are made by open code the enterprise runs, which is the exit, but the reading follows SR Linux because the validated design and the grade rest on it. EDA's commits and reverts execute intent the enterprise wrote, so that half reads as code; the cell reads with the transport half the grade would fall without. Leaving SR Linux and EDA means rebuilding the fabric's configuration and intents; leaving through SONiC on the same hardware doesn't.
H6 on SONiC: the November 13, 2025 press release says the H6 runs Community SONiC, but Nokia's SONiC page doesn't list it, so the SONiC chip is scoped to the platforms that page lists. H6 availability: stated 1Q26 and present in the SR Linux 26.7 hardware overview; no shipping notice found, and the grade holds on the H5. Customers named in press releases: Nscale (September 25, 2025), CoreWeave backbone (September 16, 2024), Microsoft data center switching on 7250 IXR with SONiC (April 2022), Telefonica edge data centers (February 2026). AI Networking Innovation Lab (Sunnyvale, May 2026) with AMD, Lenovo, Supermicro, Everpure, WEKA and others: a lab, not a product. Supermicro switches running SR Linux and EDA (September 8, 2025) are Supermicro's paper.
Arrays, storage servers and software, front-end protocols: where the data sits
Block storage and File Share for CNFs, sold as part of Nokia's vertical stack to telecom service providers. Proprietary storage operating system and data services (ActiveCluster, ActiveDR, Autopilot capacity automation); Ceded to Everpure through Nokia's paper, as on Everpure's own row.
Multi-vendor standard storage protocols; the volumes and their consumers swap to another provider behind the same protocols. Delegated, as on Everpure's row.
Software-defined storage on the cluster's own nodes at edge sites, per the solution brief. Its replication and placement opinions are Portworx objects; the fleet facet is scored at 2A. Ceded, as on Everpure's row.
Assessment pending
Nokia sells storage in one place: Everpure FlashArray with Portworx CSI as the data layer of its telco cloud on Red Hat OpenShift. Everpure reports that the Nokia-led stack with Everpure reached Ready for Sale status at Nokia in December 2025, with general availability targeted for April 2026, and on September 24, 2026 that Everpure is validated as the data layer in the Red Hat telco network cloud reference architecture. Nokia sells the complete vertical stack directly, under one SLA, so the storage is on Nokia's paper. Service providers can also buy Everpure directly. As at 0A, general availability here means availability to the telecom service providers the stack is sold to. It's block storage for network functions, with persistent volumes through the Container Storage Interface (CSI) over Internet Small Computer Systems Interface (iSCSI), Fibre Channel or NVMe over Fabrics (NVMe-oF) over TCP, plus FlashArray File Share and, at edge sites, a hyperconverged Portworx deployment without FlashArray (solution brief). The 0C rubric grades that moderate; nothing documents it as storage built for accelerator throughput. Nokia's AI fabric designs bring storage from partners: the Lenovo-Nokia blueprint's figures label the storage DDN AI400X2T (its prose says AIX400; the figure governs), and Everpure and WEKA appear in the Sunnyvale AI Networking Innovation Lab. Those are partnerships and lab work, not Nokia paper. Nokia's part is the storage network, which sits at 0B; the backend storage leafs are 7220 IXR-H4 in both design figures, while the blueprint prose names the IXR-D5. Calibration: the chips follow Everpure's own row, where FlashArray's Purity and the Portworx storage engine read Ceded and the standard front-end protocols read Delegated, and the channel rule reads an engine on a channel's paper as it reads on its owner's row. Cisco reads strong because VAST EBOX on its price list is AI-throughput storage; Nokia's channel storage is block storage for a telco platform.
Everpure's. FlashArray places, protects and replicates the bytes (ActiveCluster, ActiveDR, Autopilot capacity automation) inside bounds the operator sets, and the opinions are Everpure's whichever paper it arrives on. Leaving means moving off Everpure, not off Nokia.
Storage partners named in Nokia AI designs and labs: DDN AI400X2T (Lenovo LP2339 blueprint and Nokia white paper figures), Everpure and WEKA (AI Networking Innovation Lab, May 2026).
The data itself, organized and governed: the Governance Catalog that Layer 2C queries
Assessment pending
Nothing on Nokia's paper governs an enterprise's data foundation on documentation we hold. EDA's state store and its query language, EQL, query fabric state, not enterprise data, and Deepfield is traffic analytics and DDoS defense. Nokia does sell data governance, to communications service providers. Nokia Data Suite (formerly AVA Data Suite) is a catalog of ready-to-use, telco-specific data products on a data mesh, with TM Forum-aligned semantics and, per its product page, a data catalog, data lineage, policy controls and data validity engines. It launched on Google Cloud with BigQuery and Vertex AI in 2023, now describes itself as cloud-agnostic, and on September 17, 2026 Nokia announced it connected to Microsoft Fabric, saying the solution is available now. Its product documentation isn't in Nokia's public documentation center, and the whitepaper and eBook sit behind a form. Gated documents count once held, and these weren't, so Data Suite is named, not scored. Even documented, it would top out at moderate: prepackaged network-data products are a fixed-function slice, the product page doesn't say classification is queryable at the object or field level, and where it runs over BigQuery or Microsoft Fabric, the data platform isn't Nokia's. The data-product model, where a service provider's curation opinions would accumulate, would read Ceded. Calibration: Cisco reads strong on VAST's catalog on its price list, Dell and HPE moderate on documented storage-side metadata. Nokia has no data-foundation product an enterprise can buy on documents we hold.
None scored. The enterprise governs its own data. A service provider that adopts Data Suite would accumulate its curation opinions in Nokia's data-product model, which is outside what this cell can score until the documentation is held.
Named, not scored: Nokia Data Suite (product page; whitepaper and eBook form-gated; product docs not in the public documentation center). Data Suite with Microsoft Fabric (press release, September 17, 2026: available now, with no product documentation, so named, not scored). Autonomous Network Fabric data products and chain-of-custody monitoring (press release, June 18, 2025, no GA date).
Low-latency retrieval for RAG — vector/hybrid search, context windows
Assessment pending
Nokia sells nothing an enterprise can use for retrieval. The only retrieval on Nokia's paper is inside its own tools: Ask EDA, the assistant in Event-Driven Automation (EDA), uses an embedding model for its text-based tools, and those embeddings must come from OpenAI whichever chat model the operator picks; the documentation doesn't describe what it indexes. That's a feature of an operations console over Nokia's own platform, not a retrieval product, and it's described at 2A. Autonomous Network Fabric's knowledge engine is described on a product page and in a 2025 white paper, with no product documentation, and runs as software as a service (SaaS) on Google Cloud. The fabric carries retrieval traffic, but transport isn't retrieval: the function an architect shops for here is search and context, and the fabric is scored at 0B. Calibration: Dell reads moderate on Elastic-powered retrieval it sells, HPE moderate on a reference stack, Cisco strong on VAST InsightEngine on its price list.
None. The enterprise builds or buys its own retrieval.
Ask EDA's OpenAI embedding requirement is an EDA dependency, recorded at 2A.
Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering
Usage and network data collection, transformation and routing for service providers: streaming, real-time and batch, JSON and Avro, HDFS and Kafka, and an SDK for custom processing logic. Standard interfaces in and out; the configured streams and processing logic don't lift to another mediation product.
Assessment pending
Nokia sells one documented pipeline product: Nokia Mediation, which collects, transforms and routes usage and network data for service providers. Its datasheet (Release 26.7, June 2026) cites more than 170 service providers, online, streaming, real-time and batch processing, formats including JSON and Avro, destinations including the Hadoop Distributed File System (HDFS) and Kafka, and a software development kit (SDK) for building new processing logic. For a service provider, that's a real data-movement layer under billing, assurance and analytics. It's a slice of the layer. Mediation runs pre-configured data streams for usage mediation, and the buyer is a telecom service provider; the workload has to fit the product, a fixed-function slice that grades moderate. Data Suite would have been the same case at 1A if its documentation had been held. The fabric, the 7750 SR gateways and the ex-Infinera interconnect carry traffic, which is transport and scored at 0B; KV cache tiering and general ETL for an enterprise's AI data aren't on Nokia's paper. Calibration: HPE and Dell read moderate on pipeline engines they sell (Data Fabric, Dataloop), Cisco strong on VAST DataEngine on its price list.
The operator decides. Mediation's streams execute what the operator configured, and the SDK runs logic the operator wrote; no engine judgment is documented, so the runtime decision reads as code. Portability is the other axis: the stream definitions and SDK logic are Nokia Mediation objects, and its standard interfaces carry data in and out without porting the configuration.
Data center interconnect optics are 0B transport, not a pipeline. Data Suite's data-product pipelines are named at 1A.
GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization
Kubernetes-native automation across the fabric's switches: declarative intents as custom resources, installable apps (including AIFabrics), transactions with dry-run, commit and rollback, streamed telemetry, drift detection and a digital twin. Multi-vendor by design, with Arista EOS in beta and Cisco NX-OS in alpha in the AIFabrics app; that's reach, not exit. The opinions are EDA objects and the core is proprietary. Same product and reading as its 0B facet.
Kubernetes orchestration for the service provider's CNFs, sold by Nokia as part of its vertical stack. The consumed interface is the Kubernetes API, so workloads and manifests lift to another conformant platform. Retained, as on Red Hat's row.
Container storage orchestration, replication and placement policy for OpenShift workloads. The policies are Portworx objects. Ceded, as on Everpure's row.
Assessment pending
Event-Driven Automation (EDA) is Nokia's fleet plane for the data center fabric, generally available and documented at 26.8. It's Kubernetes-native: desired state is declared as custom resources, and features install as apps, the AIFabrics app among them. Every change is a transaction (dry-run, all-or-nothing commit, revert on failure, Git-backed history with restore), and telemetry is streamed, with drift detection and a digital twin. For a buyer standing up an AI fabric, this is the day-two story: intent goes in, the fabric is rendered and kept in line, a failed transaction rolls back automatically, and a successful change that turns out wrong can be reversed with a manual revert. EDA also carries Ask EDA, a built-in assistant that brings the operator's own model through any OpenAI-compatible endpoint (store apps exist for OpenAI, Gemini, xAI and Claude), requires an OpenAI key for embeddings in every case, allows one large language model (LLM) app at a time, and runs operations through a NetOps agent. It's an operations console over Nokia's own platform, so it's described here and scored nowhere. Workload orchestration comes from the telco cloud stack. Nokia's page says the complete vertical stack includes its core network functions and autonomous networks applications on Red Hat OpenShift, so OpenShift Container Platform sits on Nokia's paper directly, with Portworx Enterprise beside it for container storage. That's Kubernetes scheduling, quotas and role-based access control (RBAC) for the service provider's CNFs. Nothing in the stack documents GPU scheduling or fair-share for AI jobs: OpenShift AI, Kueue and the GPU operators that carry Red Hat's own 2A aren't in Nokia's offer. An AI buyer still runs a scheduler beside it (Slurm, Kubernetes with Kueue, Run:ai), which is the day-two corollary, so the layer reads moderate. Calibration: Red Hat reads strong on OpenShift with OpenShift AI's GPU queueing; Nokia carries the OpenShift half without the GPU half. Cisco reads moderate with Intersight, HPE strong on GreenLake, Dell moderate on OpenManage. The hardware-management ruling puts a fleet plane run across many systems at 2A, which is what EDA is, so the layer isn't a gap. Not scored: the Network Services Platform (NSP), Nokia's IP and WAN service automation for service providers (documentation not held, and not AI infrastructure orchestration), and Nokia Cloud Operations Manager, the hybrid cloud lifecycle product linked from the telco cloud page (product page only, no documentation held).
The workload half is the one an architect shops for here, and OpenShift's scheduler makes that call: it places the service provider's pods inside the quotas and policies the operator set, and per-object pins (node selectors, affinity rules, taints and tolerations) are overrides the scheduler must honor, so the reading is Delegated, as on Red Hat's row. The network half is the enterprise's code: EDA renders and enforces intent the enterprise declared, and its pool allocation hands out addresses, autonomous system numbers (ASNs) and virtual LANs (VLANs) from ranges the enterprise defined. Portability splits: Kubernetes manifests lift to any conformant cluster, while EDA's intents, apps and pools and Portworx's policies don't lift without a rebuild.
EDA 26.4 (April 30, 2026) shipped the first publicly consumable Ask EDA with root-cause alarms and marked its OpenAI use GA; the remediation half of the AIOps announcement (November 13, 2025) isn't confirmed. Watch item: NSP agentic framework, commercially available by the end of 2026 (June 11, 2026 release). The AIFabrics app's source is listed as coming soon; the EDA core isn't open source. The free Try EDA playground runs digital-twin nodes only.
Model serving, agent execution, inference APIs, distributed inference
Assessment pending
Nokia serves no models for the enterprise's own workloads and hosts no runtime that carries an enterprise's own agent loop on documentation we hold. Two surfaces come close. MX Industrial Edge is an on-premises industrial edge application platform: ecosystem neutral, with click-to-deploy apps from Nokia and partners (Bosch ctrlX OS, Litmus, Crosser) and edge orchestration with lifecycle management, MX Workmate, an on-premises generative AI app for operational-technology workers, runs on it while running parallel LLMs, and MX Grid orchestrates pools of micro-edge devices under it. For an industrial buyer it's a real runtime for edge AI apps, but the evidence is product pages: no runtime documentation, no supported-model or GPU list. It's named, not scored, and it's moving to AURELIUS with the Enterprise Campus Edge sale. Documented, it would read moderate as an edge-scoped runtime behind a scope gate. Ask EDA is GA and documented, but it's an assistant over Nokia's own platform, not a runtime the enterprise deploys its applications on, so it's described at 2A. EDA's MCP server is the door an agent acts through, which makes it a gating question at 2C, not a runtime. Calibration: Dell reads moderate on blueprints and services, HPE moderate on Private Cloud AI, Cisco moderate on AI Defense. Nokia has no AI security product at this layer; Deepfield Defender is DDoS protection, not model or agent security.
None. The enterprise runs its own serving and agents, and Nokia's base doesn't inherit a runtime from it.
Named, not scored: MX Industrial Edge as an application runtime, MX Workmate and MX Grid on it (2023 infographic and brochure, product pages; no runtime guide or supported-model list; ECE sale to AURELIUS signed October 13, 2026, close expected H1 2027). AI-RAN inference co-located with RAN workloads on NVIDIA ARC-Pro: commercial availability stated for 2027.
Policy-driven placement and resource coordination — the Autonomy Layer
Assessment pending
One documented, generally available surface touches this layer: the Model Context Protocol (MCP) server built into the Event-Driven Automation (EDA) API server over Streamable HTTP. The enterprise defines each tool as a custom resource backed by a workflow, an EDA Query Language (EQL) query or an HTTP call, so it decides which actions an agent can see. Clients authenticate with Open Authorization (OAuth) bearer tokens through Keycloak, workflow tools need Create rights on the target resource, queries run with the caller's own permissions, and the docs recommend a least-privilege client for headless agents. That's a real authorization point. Workflow tools are dispatched through EDA's workflow engine, and EDA's transactions offer dry-run, pre-commit pipeline gates, all-or-nothing commit and a Git-audited revert; HTTP-backed tools forward to a configured URL, and pipelines can be skipped. For a buyer pointing its own agents at the fabric, it's close to the deterministic gate before the effect that agent governance needs. It's scoped to one platform. It governs agents acting on EDA and nothing else: no agent identity beyond a service account, no registry, no orchestration across agents, no gateway for agent or model calls generally. Legs scoped inside one product are that product's own governance, and the moderate floor at 2C is a gateway every agent and model call passes through. The docs also don't say MCP workflow writes pass through the transaction pipeline; "dry run, then confirm" appears only as text in a sample prompt, which is a probabilistic lever, not a control. The nearest scored analog is HPE's GreenLake Intelligence, an IT-operations agent mesh read moderate. Nokia's counterpart, the Assurance Center agents built on Google's Agent Development Kit and Gemini (June 22, 2026), had a Google Cloud Marketplace listing planned for September 2026 that couldn't be confirmed, so they're named, not scored. Calibration: Dell reads gap and Absent, Cisco moderate on Duo agentic identity and AI Defense, HPE moderate on GreenLake Intelligence.
None scored. The enterprise brings its own agent governance. The sub-threshold finding is worth stating: Nokia ships deterministic dry-run, gated commit and rollback in EDA and an agent-facing MCP door into it, for its own fabric only; the documentation doesn't say every MCP write passes through the gates.
Named, not scored: Assurance Center agents on Google's Agent Development Kit (Apache-2.0, model-agnostic) and Gemini (Router and Event Triage agents described as operational; Marketplace SaaS listing planned September 2026, unconfirmed; remaining agents late 2026 through 2027). Watch item: NSP agentic framework with MCP to external agents, commercially available by the end of 2026. Autonomous Networks Agent Library: no GA dates. Open question: whether every EDA MCP workflow write commits through EDA transactions and their pipeline gates.
AI-powered business capabilities — business logic, workflow automation
AI-driven DDoS detection and automated mitigation that instructs routers to filter attack traffic, with telemetry, DNS and BGP inputs and the Secure Genome feed. Documented in a public solution sheet and AI/ML application note. Proprietary models and policy; leaving means rebuilding the mitigation tuning.
Assessment pending
Nokia sells a documented first-party security application: Deepfield Defender, AI-driven distributed denial-of-service (DDoS) detection and mitigation for service providers, cloud builders and large digital enterprises. Its solution sheet (26 pages, September 15, 2026) and its AI and machine-learning application note (27 pages) are public, and NL-ix (2024) and Hetzner (March 2026) have announced deployments. Defender combines network telemetry, Domain Name System (DNS) and Border Gateway Protocol (BGP) data with Nokia's Secure Genome feed, calculates a mitigation strategy and instructs routers to apply the filters. It watches the enterprise's traffic, whatever network carries it, so it's a security capability, not a console over Nokia's own platform. The scope is network security, a fixed-function slice, so the grade is moderate. The rest of Nokia's application software doesn't score here. MantaRay AutoPilot and Self-Organizing Networks (SON) act on the radio access network, taking corrective action against an operator's intent, which is network operations, not a business application; NTT DOCOMO announced on June 22, 2026 that it had put AutoPilot into commercial operation on June 19. Assurance Center, the Autonomous Networks suite and NetGuard are named. The enterprise applications (MX Workmate, MX Context, Scene Analytics, Visual Position and Object Detection) sit in Enterprise Campus Edge, are product pages only, and leave with the AURELIUS sale. The MX Industrial Edge portfolio is a catalogue of Nokia and partner applications, which earns no Layer 3 grade. Nokia runs a technology-partner program for independent software vendors (ISVs) and Network as Code partners, but no AI application program procured on Nokia's paper, so partner doesn't apply. Calibration: Arista reads moderate on its documented NDR and AGNI security applications, Cisco credits its security portfolio as first-party application components, and Lenovo reads moderate on first-party applications.
The enterprise borrows Defender's mitigation judgment: its models calculate the mitigation and instruct the routers, and the solution sheet documents full control to monitor or override mitigation settings, which is why the reading is Delegated. It can't take the tuning with it: mitigation policy and the Secure Genome context are Nokia's, and leaving means rebuilding them in another DDoS product.
Named, not scored: MantaRay AutoPilot (network operations on the RAN; customer announcement, NTT DOCOMO, June 22, 2026), Assurance Center, NetGuard, and the ECE applications MX Workmate, MX Context, Scene Analytics and Visual Position and Object Detection (product pages; ECE sale to AURELIUS signed October 13, 2026). Agentic AI for home and broadband networks (press release, May 12, 2026): not verified.
Nokia SR Linux 26.7 product overview and hardware overview (July 16, 2026), SR Linux 26.3 QoS what's-new (ECN slope on 7220 IXR-H5, DCBX ETS), AI-DC Nokia Validated Design - NVIDIA (3HE-21918-AAAA-TQZZA, Issue 3, April 2026), Nokia Community SONiC page and 7220 IXR-H SONiC datasheet, 7220 IXR hardware documentation hub, EDA 26.8 documentation (docs.eda.dev: Try EDA, AIFabrics app, transactions, Ask EDA, MCP server), Nokia press releases (EDA launch September 17, 2024; 7220 IXR-H6 and EDA AIOps November 13, 2025; Infinera close February 28, 2025; Nokia and Google Cloud Gemini agents June 22, 2026; Data Suite with Microsoft Fabric September 17, 2026; Enterprise Campus Edge sale to AURELIUS October 13, 2026; ECE recast as discontinued operations July 23, 2026), Nokia FY2025 annual report and 20-F, Nokia Data Suite, Deepfield Defender and MX Industrial Edge product pages, Lenovo Press LP2339 (Lenovo-Nokia sovereign AI blueprint, November 24, 2025), NVIDIA newsroom (NVIDIA investment, October 28, 2025), customer announcements (NTT DOCOMO MantaRay AutoPilot, June 22, 2026; NL-ix and Hetzner Deepfield deployments), published 4+1 model. Evidence boundary: gated Nokia documents count once held. The Data Suite whitepaper and eBook and the customer-portal documentation for Data Suite and MantaRay weren't held, so those products are named, not scored. The Deepfield Defender solution sheet and AI/ML application note are public and carry Layer 3. v1.0 (October 18, 2026): new row, graded cell by cell with Keith. Reconciliation changed no grade; it added the pluggable-optics note to the 0B optical chip, stated the 2A portability and authority split in borrowed judgment, and named the SONiC path in the 0B authority prose. v1.1 (October 19, 2026): a ChatGPT review of every locked cell (reviews/nokia-judgment.md; 38 claims, 29 upheld, 9 rejected after verification) moved four cells. 0A gap to moderate: Nokia sells the complete vertical telco cloud stack on Red Hat OpenShift directly, so its servers are on Nokia's paper (Keith: a stack sold whole needs no named server maker, as AWS needs none), aimed at telecom service providers. 0C gap to moderate: Everpure FlashArray with Portworx CSI inside that stack (Ready for Sale at Nokia, December 2025; validated as the data layer, Everpure, September 24, 2026). 1C gap to moderate on Nokia Mediation's public datasheet. Layer 3 gap to moderate on Deepfield Defender's public solution sheet and application note, model and Delegated on the documented override. 0B split the optics chip (ICE-X pluggables Retained, line systems Ceded). MX Industrial Edge stays unscored: a discontinued product disqualifies it (Keith). Fact fixes: H5-32D capacity, DDN AI400X2T, the IXR-H4 storage leafs, the Fabric release's availability line, the NIC-side NVIDIA dependency, Ask EDA's indexing, acronym expansion. Two dated items went to the watch list (AI-RAN, 2027; NSP agentic framework, end of 2026). Whole-row review (October 19, 2026, Antigravity on gemini-4-argon-high; reviews/nokia-row-agy.json; 7 claims, all upheld): 2A gained OpenShift Container Platform (Retained) and Portworx Enterprise (Ceded) from the telco cloud stack, and its authority moved to vendor and Delegated on OpenShift's scheduler, as on Red Hat's row (Keith: 0A is credited by what the stack requires; 2A applies directly). 0C split into FlashArray (Ceded), front-end protocols (Delegated) and the Portworx storage engine (Ceded) to match Everpure's row. 0A names the Dell and HPE server partners from Nokia's page. The summary now accounts for all ten cells and scopes Deepfield's buyers. v1.2 (October 19, 2026): the evidence-boundary sentence corrected to say the Deepfield Defender documents are public; no grade or reading changed. v1.3 (October 19, 2026, /reconcile): 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.