Cohere is a model provider that sells its models as software the enterprise runs, and the map shows the consequence. Serving and agents read strong at 2B, the value plane reads strong at Layer 3 on North, retrieval and orchestration read moderate on frontier retrieval models with a managed bundle and on Model Vault's capacity surface, and four layers are gaps everywhere the instrument asks for a substrate or a governance plane the customer administers. The private path (a data center or the customer's own cloud account) is where most of the revenue comes from, so the gaps at Layer 0, 1A, and 1C read Absent rather than Ceded-invisible: the enterprise brought its own GPUs, its own Kubernetes, and its own data governance, and Cohere shipped containers into them.
The capture is coupled and visible, with two open doors. Embed 4, Rerank 4, Compass, and Parse are captive under the embedding carve-out wherever the container runs; North's agents, workflows, and first-party connectors are Cohere's; Model Vault and the private containers are proprietary software on the customer's hardware, and self-deployable isn't Retained. The doors are Command A+ and North Mini Code: Apache 2.0 open weights with no field-of-use restriction, the models the agents run on, operable without Cohere. Add the Compatibility API and the cloud marketplaces on the Delegated side, with the Model Context Protocol (MCP) seam at three layers, and the row reads Retained 3, Delegated 4, Ceded 11 across 18 components, an authority profile Mistral shares and OpenAI and Anthropic don't.
The buyer's trade: the broadest deployment menu on the instrument (metered, dedicated, marketplace, private, air-gapped, turnkey on Dell PowerEdge) and a GA agent workspace, in exchange for a retrieval stack and an application whose opinions rebuild on exit, and an agent runtime whose approval gates are still in beta. The reasoning plane is absent: autonomy policies, inherited access controls, and the North agent directory are properties of one product's runtime, not a plane over agents the enterprise runs elsewhere. What would move a cell: Automations and Guardrails reaching GA in the docs (2B authority to Delegated, 1C to moderate), a customer-administered index and query architecture (1B), or an agent registry or gateway on Cohere's paper (2C). The Aleph Alpha merger, if it closes in the second half of 2026, adds a model family and a platform but no infrastructure.
Layer-by-layer status: Layer 0 (Enterprise or Cloud Responsibility; Cohere Runs on Rented Capacity), Layer 1A (Grounding Connectors, Not a Data Foundation), Layer 1B (Frontier Embed and Rerank Models + a Managed Retrieval Bundle, Deployable Anywhere), Layer 1C (Parsing and Ingest for Retrieval, Not Pipelines), Layer 2A (Model Vault Capacity Controls; Private Deployments Run on the Enterprise's Orchestrator), Layer 2B (Frontier Serving Anywhere + Two Apache 2.0 Exits + North Agents (GA)), Layer 2C (Autonomy Policies, Inherited Access, and an Agent Directory, Not a Governance Plane), Layer 3 (+1) (First-Party Value Plane: North, Private and Sovereign).
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: September 5, 2026. Version: v1.0 - 4+1 v2: Authority Split.
Raw compute, networking, and acceleration fabric
Private deployments require NVIDIA drivers, the container toolkit, and the device plugin; Command A+ is sized for one B200 or two H100s; the Encrypted Vault beta runs on confidential-computing NVIDIA GPUs; the Canadian sovereign stacks with CoreWeave and with Bell, Hypertec, and BUZZ HPC are NVIDIA-accelerated. The carve-out is Cohere's own platform, which has trained and served on Google Cloud TPUs as well. NVIDIA is also a Cohere investor. Cohere sells none of it.
Cohere sells models and a platform, not infrastructure, and the buyer's Layer 0 is whoever's accelerators the models run on. Private deployments come in two documented types: on premises, on the customer's own hardware, and in the customer's own cloud account, typically a virtual private cloud (VPC) on AWS, Azure, GCP, or OCI. In both the customer manages the deployment infrastructure with Cohere's guidance: Cohere ships containers for Command, Embed, Rerank, Parse, and North that run on Docker or Kubernetes with NVIDIA drivers, air-gapped if required, and Command A+ fits on one B200 or two H100s. Roughly 85% of Cohere's revenue comes from those private deployments (CNBC, February 2026). The Cohere Platform runs the same models as a metered API on rented capacity (Oracle Cloud Infrastructure and Google Cloud TPUs over the years), and Model Vault (January 28, 2026) is a Cohere-managed, dedicated, single-tenant serving environment. Dell has offered North on PowerEdge as a turnkey on-premises stack since May 2025, and the Canadian sovereign facilities with CoreWeave and with Bell, Hypertec, and BUZZ HPC are infrastructure Cohere consumes rather than sells. The exposure test governs, as it did for Mistral and Anthropic: a fleet the vendor rents and operates behind its API isn't a customer-administered substrate. Model Vault is dedicated and isolated but Cohere-operated, so it fails the same test at this layer (its capacity controls score at 2A). On the private path the substrate is the enterprise's own, in its building or its cloud account. Calibration: Mistral, Anthropic, and OpenAI read gap with authority Ceded because their customers consume a fleet they can't see; Elastic reads gap with authority Absent because half its base self-manages. Cohere's base mostly self-manages, so the Elastic reading applies: gap, Absent, with the platform and Vault paths named as Ceded-invisible for the customers who take them.
None to borrow on the private path: the enterprise supplies and operates the infrastructure, on its own hardware or in its own cloud account. On the Cohere Platform and Model Vault paths the substrate is Cohere's rented capacity, inherited invisibly. The reading is Absent, the Elastic precedent for a vendor whose customers mostly self-host.
Watch-list, notes only: the Encrypted Vault (beta; supported families exclude Command) on confidential-computing NVIDIA GPUs; the multibillion-dollar Canadian AI data center with CoreWeave (federal backing announced; no operational date); the Bell AI Fabric, Hypertec, and BUZZ HPC sovereign stack (June 18, 2026); the Aleph Alpha merger (announced April 24, 2026, closing expected second half 2026, subject to German and EU clearance), which would fold the STACKIT sovereign perimeter into Cohere's paper. Public evidence that moves the cell: a Cohere-owned or Cohere-sold compute product.
Durable, governed data foundation — the Governance Catalog that Layer 2C queries
Nothing at this layer exists to carry a dependency.
Cohere provides no system of record and no governed data foundation; it grounds on the enterprise's data where it lives, and keeps a derived retrieval index of its own. North's documented integrations are Google Drive, SharePoint, Slack, Salesforce, and any Model Context Protocol (MCP) server, linked to single sign-on, with agents inheriting the identity provider's access controls. Compass indexes documents into a Cohere-managed store for search, scored at 1B. Cohere adds workspace, role, and document-level controls over its own surfaces (role-based access control is generally available in North), but it authors none of the enterprise's data-governance catalog, lineage, classification, or source-system authorization model. The enterprise keeps its data plane and its governance, and Cohere reads through it. Calibration: Anthropic, OpenAI, and Mistral read gap with authority Absent for the same shape, a model provider that respects the customer's governance and ships none. Cohere is that shape with a stronger connector story. Gap, Absent.
None to borrow. The data stays where it was, governed by whoever governed it before; Cohere inherits the source system's access controls and adds workspace-level controls over its own product, none over the enterprise's data. Absent: nothing offered at the layer's function, nothing inherited.
Public evidence that moves the cell: a governed store or catalog on Cohere's paper. The Compass index is a retrieval artifact and is scored at 1B. Gmail and Outlook appear in North marketing imagery but not in the documented integration list; unscored.
Low-latency retrieval for RAG — vector/hybrid search, context windows
Multimodal, multilingual embeddings for text, images, and business documents, served on the Cohere Platform, Amazon Bedrock, SageMaker JumpStart, Azure AI Foundry, OCI Generative AI (on-demand since February 26, 2026 per Oracle's release notes; Cohere's own table still says coming soon), and as a private container on the customer's GPUs. Embedding carve-out: the vectors are useless without Embed at query time and no second vendor serves the space. Ceded, wherever it runs.
Cross-encoder reranking over any retriever's candidates, released December 2025 on the Cohere Platform and on OCI Generative AI since May 9, 2026 (on-demand and dedicated clusters by region), with Rerank 3.5 through Bedrock's Rerank API and private containers for both. Proprietary model; the relevance opinions rebuild on exit.
Parse (parse-v5.0) converts documents to structured Markdown through its own API, Model Vault, SageMaker, and Foundry; Compass bundles Parse, Embed, and Rerank into a managed index with no vector database for the customer to provision, in a VPC, on premises, or in Model Vault. Compass search is GA inside North; standalone Compass is a waitlisted Model Vault bundle with no published pricing. Cohere's index and pipeline: Ceded.
Cohere's connector framework over the documented sources, with agents inheriting the identity provider's access controls and answers carrying citations. Cohere's framework; the source permissions are the customer's. Ceded.
North reads sources through any Model Context Protocol server. A multi-vendor standard interface: the servers lift to another MCP-capable host. Delegated, matching the Anthropic and Mistral 1B chips.
Private deployments of the retrieval models run on NVIDIA GPUs with the NVIDIA driver, container toolkit, and device plugin as prerequisites. Cohere's own platform serving is on rented NVIDIA and TPU capacity.
This is where Cohere's models are most widely consumed inside other people's stacks. Embed 4 is a multimodal embedding model (text, images, and business documents, multilingual), Rerank 4 (December 2025; Pro and Fast variants) is a cross-encoder that re-scores candidates from any retriever, and Parse (parse-v5.0) is a vision model that turns documents into Markdown with tables, bounding boxes, and image descriptions. All three ship on the Cohere Platform, on SageMaker JumpStart and Azure AI Foundry, on OCI Generative AI (Rerank 4 Pro and Fast since May 2026), on Amazon Bedrock for Embed 4 and Rerank 3.5, and as private containers on the customer's own GPUs. Compass bundles them into an end-to-end retrieval system: parsing, automatic chunking and indexing without the customer provisioning a vector database, Embed-based semantic search, Rerank-based reranking, deployable in a VPC, on premises, or in Model Vault. Compass search is generally available inside North; standalone Compass is a Model Vault bundle that may require a waitlist, with no published pricing. The buyer gets the embedding and reranking models that peers embed in their own retrieval products, plus a managed bundle when they don't want to assemble one. The architect's concern is the embedding carve-out in its purest form: every vector Embed produces is useless without Embed at query time, no second vendor serves the same space, and the private container doesn't change that (self-deployable isn't Retained). The grade question is coverage. Rule 4 asks for general coverage of the layer, and what Cohere ships is retrieval models plus a fixed-function managed bundle, not an index and query architecture the enterprise administers. Calibration: Elastic and Snowflake read strong on retrieval estates the customer configures end to end; OpenAI reads moderate on a hosted parse-chunk-embed-search pipeline; Mistral moderate on hosted libraries consumed in-agent; Anthropic moderate with federated search and no estate. Cohere's models are the ones Elastic, Snowflake, and OCI expose inside their own 1B cells, which is the strongest model-side signal on the map, and Compass is the OpenAI shape with better models inside. Moderate, with the escalation to Keith recorded: whether frontier retrieval models the strong peers embed lift a fixed-function bundle to strong.
Low for the models and total for the space: Embed, Rerank, Parse, and Compass are Cohere intellectual property (IP) and Ceded, and the vectors are model-captive under the embedding carve-out wherever the container runs. The MCP connector interface is the one Delegated seam, as on the Anthropic and Mistral rows. The retrieval intelligence inside Compass is Cohere's and opaque: parsing, chunking defaults, the embedding space, the reranker's relevance opinions, inherited and uninspectable, and the enterprise's per-request candidate sets don't change whose judgment ranks them: vendor decides, not visible, not overridable, Ceded, the Mistral, Anthropic, OpenAI, and Perplexity reading for hosted retrieval.
Escalated: the grade (moderate written; strong argued from rule 6, score the function the architect shops for). Standalone Compass is gated in the docs (Model Vault bundle, waitlist) while Compass search inside North is GA; the component is scored on the North surface. Public evidence that moves the cell: a customer-administered index and query architecture, or a published Compass product surface with pricing.
Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering
Nothing at this layer exists to carry a dependency.
Cohere transforms documents for retrieval and moves nothing. Parse is a standalone, generally available document-to-Markdown model with its own API, and Compass chunks and indexes what Parse produces into Cohere's own store; North's connectors read sources in place. There's no extract-transform-load engine, no change data capture, no lineage, no cost-aware movement, and no KV-cache tiering. Applying rule 4: Parse is a single transform primitive with no destination but the caller, the Mistral Document AI shape the instrument holds at gap, and Compass ingest is fixed-function and terminates in Cohere's own retrieval surface, the NetApp AIDE shape without a general movement product beside it. The enterprise's pipelines are its own. Calibration: Anthropic, OpenAI, and Mistral read gap with authority Absent for intake that isn't movement; Mistral holds Document AI, a stronger transform primitive than Parse, at gap. Cohere is the same. Gap, Absent.
None to borrow. Absent: nothing offered at the layer's function, nothing inherited.
Watch-list, notes only, and this is the item that moves the cell: North Automations (launched to all North customers July 27, 2026; listed as Beta in the North user guide's feature-state table, and the docs badge governs) is a graph-based builder for multi-agent workflows with scheduling, API triggers, loops and branching, versioning, and a sandbox. At GA it's the Mistral Workflows shape: an Airflow-shaped orchestration surface for AI steps that reads moderate here and scores at 2B independently. Public evidence that moves the cell: the Automations badge reaching GA, or a pipeline product with choosable destinations beyond Cohere's own index.
GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization
Cohere-managed dedicated serving sold by the instance hour, with customer-set Fixed or Flex capacity, minimum and maximum replicas, a maximum instance cap under autoscaling, pause and resume, and GPU duty-cycle monitoring. Cohere's platform on Cohere's capacity; the bounds are configuration. Ceded.
Private deployments require NVIDIA drivers, the container toolkit, and the device plugin on the customer's cluster; scheduling those GPUs is the customer's Kubernetes' job. Model Vault instances are GPU-sized by Cohere (Small through XL tiers) on capacity the customer never sees.
One orchestration surface the customer administers, and it's Cohere's. Model Vault Standard (January 2026) is a dedicated, isolated, single-tenant serving environment sold by the instance hour: the customer creates a vault, picks the model and performance tier, chooses Fixed capacity (a set instance count) or Flex capacity (a baseline plus autoscaling on request rate and agreed latency objectives, with a maximum instance cap), edits minimum and maximum replicas, pauses and resumes models, and watches GPU duty cycle. Pricing is published per instance hour for Embed, Rerank, and Command (generative models may require a waitlist). That's a capacity and autoscaling surface the enterprise sets bounds on, the exposure test passes, and the layer moves off gap. Everywhere else Cohere orchestrates nothing the customer touches: the Cohere Platform, which Cohere's docs call the most common access path, meters tokens against rate limits on an invisible fleet, and private deployments, the majority of revenue, hand the enterprise container images (through a Replicated registry, copyable into air-gapped environments) and setup guides for Docker, Kubernetes, and AWS EC2 and EKS, with the enterprise's own Kubernetes scheduling the GPUs, setting the quotas, and scaling the replicas. The architect's concern: the Vault's bounds are configuration, not a per-request override, so what the enterprise sets is a ceiling on Cohere's judgment about placement and scale, and none of it lifts. Compare Elastic Cloud Enterprise, a purchasable orchestrator the customer runs; Cohere ships none. Calibration: Elastic reads moderate with authority Ceded on Elastic Cloud Hosted and Serverless autoscaling, a customer-bounded surface on the vendor's capacity, and Model Vault is that shape. OpenAI, Anthropic, and Mistral read gap with authority Ceded on token metering alone; Kamiwaza gap with authority Absent because the enterprise's own orchestration runs beneath it. Moderate, Ceded.
Bounded on the Vault path, none on the private path. Model Vault's replica bounds and Flex autoscaling are configuration the engine honors within Cohere's placement judgment: vendor decides, visible, not overridable, Ceded, the Elastic Hosted reading. On the private path the enterprise's Kubernetes holds the scheduling judgment; on the platform path Cohere's capacity management is inherited invisibly.
Grade moved during review: drafted gap on the exposure test, moved to moderate when Model Vault's documented replica, capacity, and autoscaling controls were read (a customer-administered surface passes the test). Watch-list, notes only: the Encrypted Vault (beta). Public evidence that moves the cell: a region, placement, or quota surface the customer administers beyond replica bounds, or a Cohere-shipped orchestrator for private deployments.
Model serving, agent execution, inference APIs, distributed inference
A 218B-parameter sparse mixture-of-experts frontier model (25B active, 128K context, vision, reasoning, 48 languages) released under a full Apache 2.0 license, quantized to run on one B200 or two H100s. Open-source software the enterprise operates itself with no license gate, and fine-tunes of it are the enterprise's: Retained on the open-source seam, matching Mistral's Apache 2.0 weights.
A mixture-of-experts model trained for agentic coding, released under Apache 2.0 with open weights on Hugging Face, served through the Cohere API and Model Vault or run locally. A model, not an application: Retained on the open-source seam.
Cohere models through the OpenAI SDK at the compatibility endpoint (chat completions with function calling and structured outputs) and through the clouds' native services per the availability table. Managed services behind standard-shaped interfaces the enterprise can switch: Delegated on the inference-interface ruling. Embeddings through the same endpoint are excluded and Ceded at 1B under the carve-out.
Cohere-managed, isolated serving with no noisy neighbors or rate limits, sold by the instance hour (capacity controls scored at 2A). Cohere's platform: Ceded.
Licensed container images pulled from Cohere's registry (copyable into air-gapped environments) and run on Docker or Kubernetes with NVIDIA prerequisites, per Cohere's setup, AWS EC2 and EKS, and Kubernetes guides; North on as few as two GPUs; turnkey on Dell PowerEdge. Proprietary software on the customer's hardware: self-deployable isn't Retained. Ceded.
Platform fine-tuning where Cohere trains and hosts the custom model; private-environment fine-tuning producing merged weights the enterprise holds; a documented path to run merged Command R fine-tunes on SageMaker under a Cohere marketplace license. The artifact leaves, the base license follows it: Ceded, the Mistral revenue-capped-weights ruling. Fine-tunes of Apache 2.0 bases are covered by the Retained chips above.
Custom agents (GA) built with tools from first-party connectors and MCP servers, coordinating with each other, under product-level autonomy policies and RBAC; Automations and Guardrails are Beta and unscored. Cohere's runtime and agent objects: Ceded.
North agents call tools over the Model Context Protocol. A multi-vendor standard interface: Delegated, matching the Anthropic and Mistral 2B chips.
External MCP servers the customer builds and hosts (Cohere publishes an SDK for creating them) and calls from North agents. The logic runs outside the vendor: Retained under the narrowed customer-tools ruling (September 5, 2026).
Command A+ serves on one B200 or two H100s in W4A4; every private container carries the NVIDIA driver, toolkit, and device-plugin prerequisites. No customer-side NVIDIA dependency on the platform API or marketplace paths.
Both halves of the layer are generally available, and the serving half runs anywhere the buyer wants. Serving: Command A+ (May 20, 2026), a 218-billion-parameter sparse mixture-of-experts model with 25 billion active parameters, 128K context, vision, reasoning, translation, and agentic tool use, 48 languages, released under a full Apache 2.0 license with W4A4 quantization that fits it on one B200 or two H100s; Command A and the Embed, Rerank, and Parse models beside it; served through the Cohere Platform API and the Compatibility API (Cohere models through the OpenAI SDK at the compatibility endpoint, with function calling under strict schemas and structured outputs), through the cloud marketplaces per model (Command A on SageMaker, Azure AI Foundry, and OCI Generative AI with Bedrock listed as coming soon; Command A+ on Azure AI Foundry; Embed 4 and Rerank 3.5 on Bedrock), through Model Vault Standard (dedicated, single-tenant), and as private containers on the customer's own Kubernetes, air-gapped if required. Fine-tuning runs on the platform and in private deployments, and merged fine-tuned weights deploy on the customer's own infrastructure. Agents: North (generally available August 6, 2025) is an agentic workspace with grounded chat, document creation, and custom agents (GA) built in Agent Studio with tools from first-party connectors and any Model Context Protocol (MCP) server, with agent-to-agent coordination, on as few as two GPUs on premises, in a VPC, or air-gapped, turnkey on Dell PowerEdge since May 2025 and as a Dell agentic AI platform blueprint since March 2026. North Mini Code, a 30-billion-parameter (3 billion active) coding model under Apache 2.0, serves through the same paths. The buyer gets frontier serving with a real weights exit and an agent runtime that lives inside their perimeter. The architect's concern splits. On the model, the exit is unusually clean: Command A+ and North Mini Code are Apache 2.0 with no field-of-use restriction, so the models the agents run on can be operated without Cohere. On the platform, North's agents and first-party connectors are Cohere's, and the private container for North is still Cohere's software: self-deployable isn't Retained. The approval story is the one to watch: the pre-execution gates (Automations' built-in approvals, Guardrails' block-on-tool-call through HiddenLayer) are both Beta in the docs, so what's GA is autonomy policy at the product level and access control, not a documented per-action gate. Calibration: Mistral reads strong on frontier serving plus an owned-weights exit with the agent API pending; Anthropic and OpenAI strong on serving plus a hosted agent runtime; SAP and ServiceNow strong on constructible agent runtimes. Cohere has Mistral's weights exit twice over, the broadest deployment menu on the instrument (metered, dedicated, marketplace, private, air-gapped), and a GA agent workspace. Strong.
Low for the runtime and unusually low for the model. Command A+ and North Mini Code under Apache 2.0 are open-source software the enterprise can operate on its own hardware with no license gate: Retained on the open-source seam, the Mistral open-weights reading. Model access through the Compatibility API and the marketplaces is Delegated on the inference-interface ruling (embeddings excepted, Ceded at 1B under the carve-out); the MCP interface is Delegated; external MCP servers the customer hosts are Retained under the narrowed customer-tools ruling. Model Vault, the private containers, proprietary-base fine-tunes, and North's agents are Cohere's wherever they run, Ceded. The model decides which tool to call, the call is visible, and no GA surface documents a per-action gate before the effect: model decides, visible, not overridable, Ceded, the ServiceNow reading; it flips to Delegated when Automations approvals or Guardrails reach GA in the docs.
Watch-list, notes only: North Automations (Beta in the user guide; launched to all North customers July 27, 2026) with built-in approvals, scheduling, and API triggers; Guardrails (Beta, select customers) with HiddenLayer runtime detection and ALLOW, BLOCK, and REDACT on tool calls and MCP results; the Encrypted Vault (beta; Embed, Rerank, Parse, Compass, North Mini Code, Transcribe, and customized models, not Command); Command A on Bedrock (listed coming soon); the Aleph Alpha merger (announced April 24, 2026; closing expected second half 2026), which would fold the Pharia model family into Command. Authority applied by the ServiceNow precedent and recorded on the judgment sheet. Public evidence that moves the cell: nothing upward from strong; the authority reading moves on a GA badge for Automations approvals or Guardrails.
Policy-driven placement and resource coordination — the Autonomy Layer
No reasoning-plane product exists to carry a dependency.
Cohere's governance is a property of North, not a plane over agents. Inside North, autonomy policies define which decisions an agent takes alone and which require human review, agents inherit the user's access through single sign-on and the identity provider's controls, outputs carry citations and reasoning chains, role-based access control is generally available, and the Admin Panel lists every default and custom agent in the instance with search, filtering, and user, group, and all-users access. Applying the test: Intelligence-2C exists only as the runtime's permission inheritance and access governance, properties of the agents scored at 2B. The agent directory is an admin console over one product's own agents, with no lifecycle, risk classification, cross-system discovery, or approval-before-activation, the same altitude as Cowork's RBAC on the Anthropic row and workspace RBAC on the OpenAI row, both gaps. There's no request-time gateway with policy over tool calls, no first-class agent identity distinct from the user, no cross-agent orchestration beyond North's own coordination, and no observability of agents that aren't North's. Infrastructure-2C, live placement across model, cost, and compliance tiers, is absent as everywhere. Calibration: Anthropic reads gap with runtime permissions that aren't a reasoning plane; Mistral gap with guardrails and approvals that aren't a governance plane; Qlik gap on permission inheritance and confirmation prompts. North's autonomy policies and agent directory are the same plumbing with better names. ServiceNow, Salesforce, and Snowflake read moderate on productized multi-leg planes that govern other vendors' agents. Gap, Absent.
None to borrow, and that's the finding: the approval and permission behavior belongs to the 2B runtime. To the extent a reasoning plane exists in a Cohere deployment, the enterprise built it. Absent: nothing offered at the layer's function, nothing inherited.
Watch-list, notes only, by feature state in the North docs: Automations (Beta) with built-in approvals and audit-ready logs; Guardrails (Beta, select customers) scanning tool calls and MCP results through HiddenLayer with ALLOW, BLOCK, and REDACT; Audits (Alpha) recording admin events such as MCP server created, user added to agent, and tool source runtime configuration updated; product analytics (Beta). Reviewed for awareness: whether a single-instance agent directory with RBAC counts as a registry leg (ruled subthreshold on the Anthropic and OpenAI precedent). Public evidence that moves the cell: an agent registry with lifecycle and risk state, a request-time policy gateway, or an agent principal distinct from the user, productized and GA.
AI-powered business capabilities — business logic, workflow automation
The first-party application, generally available since August 6, 2025, private-deployable on as few as two GPUs, with chat, custom agents, document creation, Compass search, and RBAC all GA in the feature-state table. Cohere IP.
The agents the enterprise builds in Agent Studio. They run only in North: Ceded under the narrowed customer-tools ruling (September 5, 2026), with customer-hosted MCP servers scored Retained at 2B.
The MCP servers the enterprise connects to North lift to any MCP-capable host at menu altitude, though North-specific installation and configuration is recreated. Delegated, matching the Anthropic, Mistral, OpenAI, and Perplexity Layer 3 chips.
The application runs on whatever GPUs serve the models beneath it.
North is the application, and it's the reason the models get bought. A collaborative agentic workspace with grounded chat over the enterprise's own sources, document creation and refinement, and custom agents, sold to run inside the customer's perimeter on as few as two GPUs, with RBC, Dell, LG CNS, Ensemble Health Partners, and Second Front as named deployers and Fujitsu, Oracle, and Notion among multi-year customers. The distinguishing property is where it runs: the peers' value planes are hosted; North's is deliverable to a data center, a VPC, or an air-gapped room, turnkey on Dell PowerEdge, distributed to the public sector through Carahsoft and to Canadian government and enterprise through Bell. Automations, the workflow surface, is Beta and unscored. The capture is coupled and visible, as on every model-provider row. The workspace, the agents, and the first-party connectors are Cohere's, and the private container doesn't change that; what accumulates inside North is the record of how the enterprise works, its agents, its grounded sources, its delegation habits, per user, per day, and swapping the model underneath (which Cohere makes unusually easy at 2B) moves none of it. The MCP seam is the one exception: servers the enterprise connects lift to another MCP-capable host. Calibration: Anthropic, OpenAI, Mistral, and Perplexity read strong on first-party value planes at assistant altitude; ServiceNow and Salesforce on workflow and business applications. Cohere is the Mistral case with the private-deployment emphasis turned up: a first-party agentic workspace that ships as software the customer hosts. Strong.
High and compounding at the application, architecturally separable beneath it. North, its agents, and its connectors are Cohere IP and Ceded, and the workspace state rebuilds nowhere; the models beneath are brokered or, for Command A+ and North Mini Code, the enterprise's to keep (2B), and MCP-served sources are Delegated. Vendor decides, visible, not overridable, Ceded, the first-party value-plane reading.
Channels, notes only: Carahsoft (July 30, 2026) as public-sector distributor; Bell Canada (June 2026) for sovereign services; Dell for turnkey on-premises delivery and agentic AI platform blueprints (March 2026). Watch-list: North Automations (Beta); PhariaAI joining the portfolio if the Aleph Alpha merger closes (expected second half 2026). Public evidence that moves the cell: nothing upward from strong.
Cohere is a model provider that sells its models as software the enterprise runs, and the map shows the consequence. Serving and agents read strong at 2B, the value plane reads strong at Layer 3 on North, retrieval and orchestration read moderate on frontier retrieval models with a managed bundle and on Model Vault's capacity surface, and four layers are gaps everywhere the instrument asks for a substrate or a governance plane the customer administers. The private path (a data center or the customer's own cloud account) is where most of the revenue comes from, so the gaps at Layer 0, 1A, and 1C read Absent rather than Ceded-invisible: the enterprise brought its own GPUs, its own Kubernetes, and its own data governance, and Cohere shipped containers into them.
The capture is coupled and visible, with two open doors. Embed 4, Rerank 4, Compass, and Parse are captive under the embedding carve-out wherever the container runs; North's agents, workflows, and first-party connectors are Cohere's; Model Vault and the private containers are proprietary software on the customer's hardware, and self-deployable isn't Retained. The doors are Command A+ and North Mini Code: Apache 2.0 open weights with no field-of-use restriction, the models the agents run on, operable without Cohere. Add the Compatibility API and the cloud marketplaces on the Delegated side, with the Model Context Protocol (MCP) seam at three layers, and the row reads Retained 3, Delegated 4, Ceded 11 across 18 components, an authority profile Mistral shares and OpenAI and Anthropic don't.
The buyer's trade: the broadest deployment menu on the instrument (metered, dedicated, marketplace, private, air-gapped, turnkey on Dell PowerEdge) and a GA agent workspace, in exchange for a retrieval stack and an application whose opinions rebuild on exit, and an agent runtime whose approval gates are still in beta. The reasoning plane is absent: autonomy policies, inherited access controls, and the North agent directory are properties of one product's runtime, not a plane over agents the enterprise runs elsewhere. What would move a cell: Automations and Guardrails reaching GA in the docs (2B authority to Delegated, 1C to moderate), a customer-administered index and query architecture (1B), or an agent registry or gateway on Cohere's paper (2C). The Aleph Alpha merger, if it closes in the second half of 2026, adds a model family and a platform but no infrastructure.