Nebius is the European NVIDIA cloud with a model service on top. The map reads it strong at three layers, moderate at three, and gap at two. Layer 0 is strong at all three sub-layers: GPU virtual machines and clusters at 0A, InfiniBand at 0B, and object and shared-filesystem storage at 0C, with GPU virtual machines and InfiniBand GPU clusters from H100 to B300 in nine public regions across Finland, Spain, France, Israel, the UK, and the US (one private region, Iceland, serves existing customers), with partition-key isolation and topology labels, virtual networks, and security groups; the capacity is Nebius's own and, by contract, partners' too. Layer 2A is strong on Managed Kubernetes with NVIDIA's operators and topology-aware scheduling, Managed Soperator (its open-source Slurm-on-Kubernetes operator) with ephemeral nodes that move capacity between training and inference, Serverless AI, and quotas, the CoreWeave scheduler pair. Layer 2B is strong on Token Factory: open models behind OpenAI-compatible chat, responses, vision, embeddings, and rerank endpoints, serverless per token or on dedicated endpoints with replica bounds, full or Low-Rank Adaptation (LoRA) fine-tuning and custom speculator training with Data Lab and export to Hugging Face format, and open-source agent tooling as managed applications; custom weights and Sandboxes are beta. Layer 1A is moderate on bucket policies by key and prefix and a tenant-project identity model with Security Assertion Markup Language (SAML) single sign-on and OpenID Connect (OIDC) workload credentials, with audit logs in preview and no catalog; the object storage and virtiofs filesystems read at 0C. Layer 1B is moderate on embedding endpoints, Qdrant as a managed app, and Managed PostgreSQL with pgvector. Layer 1C is moderate on Object Storage data transfers (a managed sync between buckets, third-party S3 stores, and Azure Blob) and Data Lab's captive dataset workbench. Layer 3 is a gap under the marketplace ruling: a marketplace of other people's applications, one NVIDIA blueprint under its own license, and Nebius Echo, a console assistant with approval-gated actions, none of them a business application. Layer 2C is a gap: identity for people and workloads, nothing generally available over agents.
The capture is at Nebius's control planes and nowhere at the interfaces. Kubernetes, Slurm, S3, virtiofs, SAML, OIDC, PostgreSQL, and the OpenAI-compatible endpoints are standards; Soperator, Qdrant, Flowise, Open WebUI, JupyterLab, and ComfyUI are open source with export pages, and the same apps self-run on a VM or the enterprise's cluster are the enterprise's own; the models are their owners' open weights and fine-tunes export as the enterprise's artifact; that's why ten of the nineteen components read Delegated or Retained. What is Nebius's and stays Nebius's: the instances, fabrics, networks, and disks; the account model, keys, and audit; the data-transfer engine; node-group provisioning, capacity blocks, Serverless AI, quotas, and the marketplace deployer; the fine-tuning service with Data Lab; Echo. NVIDIA is everywhere underneath (every GPU, the operators, the interconnect, the one blueprint under its own license) and is the dependency the row carries at 0A, 0B, 2A and 2B.
The buyer's trade: NVIDIA capacity by the second in Europe, with Kubernetes and Slurm it already knows, open models it can call, tune, and take away, a managed sync between object stores, and a shelf of open-source apps, in exchange for a rented substrate Nebius sometimes rents itself, a model service whose catalogue is visible only by API and whose burst isn't guaranteed, a modality retired in April, and an identity and audit layer still partly in preview. The decision-authority readings follow: vendor / Ceded at all three Layer 0 sub-layers (a fabric and a platform are a menu), code / Retained at 1A, where the access permits execute as the enterprise wrote them, vendor / Delegated at 1B on per-request controls, vendor / Ceded at 1C (transfer flags are configuration), vendor / Delegated at 2A on two documented pins (topology labels the enterprise's scheduler places on, Slurm nodes released and repowered by name), model / Delegated at 2B (the loop is the enterprise's), Absent at 2C and Layer 3. What would move cells: a first-party business application (Layer 3), Sandboxes leaving beta with a gateway or registry (2C), custom weights leaving beta (a 2B chip), audit logs leaving preview (1A), a catalog or policy surface over data (1A).
Layer-by-layer status: Layer 0A (NVIDIA GPU VMs by the Second, H100 to B300), Layer 0B (InfiniBand GPU Cluster Fabrics With Partition-Key Isolation), Layer 0C (Shared Filesystems Documented at AI-Training Throughput), Layer 1A (Access Permits Over Nebius's Data Services; No Data Catalog), Layer 1B (Embedding Endpoints on Open Models in Token Factory, Qdrant as a Managed Application, and Managed PostgreSQL With pgvector; No Retrieval Engine of Nebius's Own), Layer 1C (Object Storage Data Transfers: Continuous or One-Shot Movement Between Buckets, Third-Party S3 Stores, and Azure Blob With Overwrite and Delete Semantics; Data Lab Prepares Fine-Tuning Datasets by SQL Over Inference Logs and S3 Sources; No ETL, Lineage, or Cache Tiering), Layer 2A (Managed Kubernetes With GPU Node Groups, Operators, a Cluster Autoscaler, and Topology-Aware Scheduling; Managed Soperator (Open-Source Slurm on Kubernetes) With Ephemeral Nodes That Move Capacity Between Training and Inference; Serverless AI Jobs and Endpoints; Quotas, Preemptible Capacity, and Commitments), Layer 2B (Token Factory: Open Models Behind OpenAI-Compatible Chat, Responses, Embeddings, Vision, and Rerank Endpoints, Serverless per Token or on Dedicated Endpoints With Replica Bounds, Plus Full or LoRA Fine-Tuning and Custom Speculator Training With Data Lab; Custom Weights and Sandboxes in Beta; Agents Through Managed Open-Source Applications), Layer 2C (Tenant, Project, and Service-Account Identity for People and Automation, Audit Logs in Preview; Nothing Over Agents: Nebius Echo Is a Console Assistant, Sandboxes Are Beta, the Agents Blueprint Is a Reference Architecture), Layer 3 (+1) (No Business Application: an Applications Marketplace of Others' Open-Source Apps Deployed as Managed Services, Onto the Enterprise's Kubernetes, or Onto a Root-Access Virtual Machine, One NVIDIA Blueprint Under a Bring-Your-Own License, and Nebius Echo, a Console Assistant).
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: v1.22 - 4+1 v3: Layer 0 Sub-Layers.
Nebius is the European NVIDIA cloud with a model service on top. The map reads it strong at three layers, moderate at three, and gap at two. Layer 0 is strong at all three sub-layers: GPU virtual machines and clusters at 0A, InfiniBand at 0B, and object and shared-filesystem storage at 0C, with GPU virtual machines and InfiniBand GPU clusters from H100 to B300 in nine public regions across Finland, Spain, France, Israel, the UK, and the US (one private region, Iceland, serves existing customers), with partition-key isolation and topology labels, virtual networks, and security groups; the capacity is Nebius's own and, by contract, partners' too. Layer 2A is strong on Managed Kubernetes with NVIDIA's operators and topology-aware scheduling, Managed Soperator (its open-source Slurm-on-Kubernetes operator) with ephemeral nodes that move capacity between training and inference, Serverless AI, and quotas, the CoreWeave scheduler pair. Layer 2B is strong on Token Factory: open models behind OpenAI-compatible chat, responses, vision, embeddings, and rerank endpoints, serverless per token or on dedicated endpoints with replica bounds, full or Low-Rank Adaptation (LoRA) fine-tuning and custom speculator training with Data Lab and export to Hugging Face format, and open-source agent tooling as managed applications; custom weights and Sandboxes are beta. Layer 1A is moderate on bucket policies by key and prefix and a tenant-project identity model with Security Assertion Markup Language (SAML) single sign-on and OpenID Connect (OIDC) workload credentials, with audit logs in preview and no catalog; the object storage and virtiofs filesystems read at 0C. Layer 1B is moderate on embedding endpoints, Qdrant as a managed app, and Managed PostgreSQL with pgvector. Layer 1C is moderate on Object Storage data transfers (a managed sync between buckets, third-party S3 stores, and Azure Blob) and Data Lab's captive dataset workbench. Layer 3 is a gap under the marketplace ruling: a marketplace of other people's applications, one NVIDIA blueprint under its own license, and Nebius Echo, a console assistant with approval-gated actions, none of them a business application. Layer 2C is a gap: identity for people and workloads, nothing generally available over agents.
The capture is at Nebius's control planes and nowhere at the interfaces. Kubernetes, Slurm, S3, virtiofs, SAML, OIDC, PostgreSQL, and the OpenAI-compatible endpoints are standards; Soperator, Qdrant, Flowise, Open WebUI, JupyterLab, and ComfyUI are open source with export pages, and the same apps self-run on a VM or the enterprise's cluster are the enterprise's own; the models are their owners' open weights and fine-tunes export as the enterprise's artifact; that's why ten of the nineteen components read Delegated or Retained. What is Nebius's and stays Nebius's: the instances, fabrics, networks, and disks; the account model, keys, and audit; the data-transfer engine; node-group provisioning, capacity blocks, Serverless AI, quotas, and the marketplace deployer; the fine-tuning service with Data Lab; Echo. NVIDIA is everywhere underneath (every GPU, the operators, the interconnect, the one blueprint under its own license) and is the dependency the row carries at 0A, 0B, 2A and 2B.
The buyer's trade: NVIDIA capacity by the second in Europe, with Kubernetes and Slurm it already knows, open models it can call, tune, and take away, a managed sync between object stores, and a shelf of open-source apps, in exchange for a rented substrate Nebius sometimes rents itself, a model service whose catalogue is visible only by API and whose burst isn't guaranteed, a modality retired in April, and an identity and audit layer still partly in preview. The decision-authority readings follow: vendor / Ceded at all three Layer 0 sub-layers (a fabric and a platform are a menu), code / Retained at 1A, where the access permits execute as the enterprise wrote them, vendor / Delegated at 1B on per-request controls, vendor / Ceded at 1C (transfer flags are configuration), vendor / Delegated at 2A on two documented pins (topology labels the enterprise's scheduler places on, Slurm nodes released and repowered by name), model / Delegated at 2B (the loop is the enterprise's), Absent at 2C and Layer 3. What would move cells: a first-party business application (Layer 3), Sandboxes leaving beta with a gateway or registry (2C), custom weights leaving beta (a 2B chip), audit logs leaving preview (1A), a catalog or policy surface over data (1A).
CPUs, accelerators, servers, integrated systems, OS and hypervisor, physical plant
Instances the enterprise rents on Nebius's implementation of NVIDIA's design; one private region, Iceland, serves existing customers and is named, not scored. Standard NVIDIA instances: a CUDA or PyTorch workload moves to another cloud's NVIDIA instances or an OEM server without a rebuild. Retained under the Layer 0 runtime rule, read from the customer's seat; the CUDA dependency is NVIDIA's and sits in the NVIDIA column, the CoreWeave GPU compute reading.
GB200 and GB300 NVL72 racks as NVLink instance groups, which Nebius's Terraform provider documents, with GB300 NVL72 "deployed in production" per its product page and sold through the sales team. Ceded on the integrated-system rule, the Lenovo NVL-class reading: a rack-scale system built around a captive NVLink fabric can't be lifted as deployed.
Non-GPU platforms the enterprise rents: cpu-d3 on AMD EPYC 9654 in every region, presets from 2vcpu-8gb to 256vcpu-1024gb (the four largest outside eu-north1 and eu-south1), and cpu-e2 on Intel Xeon Gold 6338 in eu-north1 only, up to 80vcpu-320gb.
H100, H200, B200, B300, L40S and RTX PRO 6000. Managed Kubernetes requires the NVIDIA GPU Operator. The CUDA dependency is read on NVIDIA's row.
Nebius sells the GPU cloud itself. Virtual machines with one to eight NVIDIA GPUs (H100, H200, B200, B300, L40S, RTX PRO 6000) come regular or preemptible and billed per second. GB200 and GB300 NVL72 racks come as NVLink instance groups through the sales team. They run in nine public regions across Finland, Spain, France, Israel, the UK and the US, with GPU platforms varying by region. CPU-only VMs on AMD EPYC and Intel Xeon sit beside them. The buyer gets NVIDIA capacity by the second, in Europe first. The architect's concern is that the substrate is rented from a company that also rents it. Some capacity sits with third-party GPU clouds the sub-processor list names, and Nebius owns no silicon. The accelerators are NVIDIA's. Calibration: CoreWeave reads strong on the same shape, and Nebius is CoreWeave's shape in Europe; AWS and Azure strong on rented fleets with silicon of their own.
Rented, so the runtime call is Nebius's. Nebius places each VM on a host in the fabric the enterprise chose or the preselected one, and preempts preemptible VMs; the enterprise picks region, platform and count and reverses nothing below it: vendor decides, visible, not overridable, Ceded. Portability: the GPU and CPU VMs are standard instances a CUDA or PyTorch workload leaves without a rebuild, Retained, the CoreWeave reading; the NVL72 instance groups read Ceded on the integrated-system rule.
Named, not scored: the private Iceland region (existing customers only), the Meta supply agreement (March 16, 2026), the NVIDIA Groq 3 LPX adoption for Vera Rubin NVL72 (blog, August 24, 2026; no product documentation) and the third-party capacity sub-processors. Public evidence that moves the cell: a bare-metal tier would change the authority reading.
Interconnects, switching and routing, network virtualization
The InfiniBand fabric Nebius builds GPU clusters on. Ceded, the CoreWeave network fabric reading.
The captive NVLink fabric inside the NVL72 racks Nebius sells as NVLink instance groups, a 72-GPU NVLink domain. Ceded with the system on the integrated-system rule.
Nebius's network primitives in Nebius's regions. Ceded, the CoreWeave network fabric reading.
GPU clusters run on InfiniBand fabrics, one platform per fabric. For B200 or InfiniBand, Managed Kubernetes requires the NVIDIA Network Operator, with GPUDirect RDMA on by default.
GPU clusters live in named InfiniBand fabrics at 3.2 Tbps per eight-GPU node, and Nebius isolates traffic between clusters with InfiniBand partition keys. GB200 and GB300 NVL72 racks carry a captive NVLink fabric inside. Topology comes back as node labels so the enterprise's scheduler can place jobs close. Virtual Networks add IPv4 subnets, pools, routing tables and stateful or stateless security groups with priority-ordered rules. The buyer gets an NVIDIA reference fabric already built. The architect's concern is choice. Nebius says most customers won't need to change the preselected fabric, and its topology tiers aren't guaranteed to match its examples. Calibration: CoreWeave reads strong on the same InfiniBand shape; AWS's EFA and GCP's RoCE GPU network strong on fabrics the cloud runs.
Ceded where the grade rests. Security group rules and routing tables are enterprise policy, but the InfiniBand fabric, preselected in most cases and isolated by Nebius's partition keys, makes the transport calls: vendor decides, visible, not overridable, Ceded. Portability: InfiniBand is a standard, but the enterprise can't take a fabric with it, Virtual Networks are Nebius's, and the NVLink fabric goes with the rack, Ceded.
Arrays, storage servers and software, front-end protocols: where the data sits
Block devices attached to Nebius VMs in Nebius's formats. Ceded; the object and file stores are the Layer 1A chips.
Buckets and objects behind a multi-vendor interface. The S3 API is its own chip. The service, its storage classes and bucket policies, is Nebius's and doesn't lift: Ceded.
File and image contents that copy out through standard tools, on Nebius's mounts. The WEKA and VAST types are named, not scored. The virtiofs mount is the guest-to-hypervisor attachment, not a protocol the VM can point at another provider, so it gets no door, as with cloud block devices. The service, its bucket attachment and encryption, is Nebius's: Ceded. Not scored within this chip under the GA evidence table: Bucket Attachment ("Filesystem buckets in Object Storage are in preview.").
The S3 API, which independent object stores implement. Workloads confined to the supported common S3 operations move by changing endpoints, credentials and provider-specific configuration; the data, policies and unsupported service features migrate or adapt. Delegated. The service is the Object Storage chip, Ceded.
Assessment pending
Nebius's shared filesystems mount to many VMs in a project as virtiofs devices, encrypted by default, and can expose buckets through the mount. Nebius documents tested aggregate read bandwidth up to 1.9 TiB/s and write up to 1.5 TiB/s, and it published a 987 GB/s read on a 302-host production cluster. Object Storage is S3-compatible in every region, and network SSD, non-replicated, IO-M3 and local disks handle block. The buyer gets storage documented to feed a training cluster. The architect's concern is the handle. Nebius's engines decide placement and replication, and the disk type and storage class are configuration. No per-file placement handle is documented on the filesystems or the object store. Calibration: CoreWeave reads strong on AI object storage the cloud places; AWS and Azure strong on managed Lustre with a per-file layout handle.
Ceded where the grade rests. Nebius's engines decide placement and replication for disks, buckets and shared filesystems, and the enterprise's disk types and storage classes are configuration, not overrides: vendor decides, visible, not overridable, Ceded. Portability: buckets behind S3 and filesystem contents that copy out with standard tools read Delegated; compute storage in Nebius's formats reads Ceded.
Named, not scored: the WEKA and VAST filesystem types, partners' engines behind a Nebius menu; the owners' rows score the engines. Out of the Shared Filesystems chip under the GA evidence table (October 24, 2026): Bucket Attachment ("Filesystem buckets in Object Storage are in preview.").
The data itself, organized and governed: the Governance Catalog that Layer 2C queries
Nebius's account model over Nebius's resources; SAML and OIDC are standard doors onto it. Ceded, the CoreWeave dedicated-tier reading; audit logs are named, not scored. Not scored within this chip under the GA evidence table: KMS With Data and Key-Encryption Keys, Audit Logs ("The service is in preview, and available in all Nebius AI Cloud regions.").
Buckets and objects behind a multi-vendor interface. Delegated, the CoreWeave CAIOS and Cloudflare R2 reading.
Object storage, filesystems, identity, and keys run on Nebius's CPUs and storage.
Nebius's 1A function is access governance. Bucket policies by key and prefix on Object Storage, and tenants and projects, groups with access permits, and service accounts with access keys for the S3-compatible service, decide who may use which data, with Security Assertion Markup Language (SAML) and OpenID Connect (OIDC) federation and Key Management Service (KMS) keys. The stores themselves hold the bytes and read at 0C. Nothing on Nebius's paper answers what the data is: classification, sensitivity or lineage. That's governance over data without a queryable classification catalog: moderate. Calibration: CoreWeave reads moderate on a Dedicated-tier catalog; Red Hat moderate on bucket policies.
Nebius enforces the groups and access permits the enterprise defined, as written: code decides, visible, overridable, Retained. Portability: the account model, access permits and KMS are Nebius's, Ceded; SAML and OIDC are standard doors onto them.
Unscored on its badge: Audit Logs (preview). Public evidence that moves the cell: a catalog over the data. Shared Filesystems and the Container Registry are storage, so they're not scored at 1A. Out of the Identity, Keys, and Audit chip under the GA evidence table (October 24, 2026): KMS With Data and Key-Encryption Keys, Audit Logs ("The service is in preview, and available in all Nebius AI Cloud regions.").
Low-latency retrieval for RAG — vector/hybrid search, context windows
Open weights behind standard doors on Nebius's GPUs. The OpenAI-compatible API surface is its own chip. The service behind it, its operations, management and configuration, is Nebius' and doesn't lift: Ceded.
An open-source vector database Nebius operates for the enterprise, with a documented export. The Qdrant API surface is its own chip. The service behind it, its operations, management and configuration, is Nebius' and doesn't lift: Ceded.
Standard PostgreSQL with an open extension, operated by Nebius. The PostgreSQL with pgvector surface is its own chip. The service behind it, its operations, management and configuration, is Nebius' and doesn't lift: Ceded.
The OpenAI-compatible API has independent implementations: code and definitions written to it move to another compatible implementation within the features both support, the data moves by export, replication or rewrite, and the provider's control-plane configuration and operational state don't lift. Delegated. The service behind it is its own chip, Ceded.
The Qdrant API has independent implementations: code and definitions written to it move to another compatible implementation within the features both support, the data moves by export, replication or rewrite, and the provider's control-plane configuration and operational state don't lift. Delegated. The service behind it is its own chip, Ceded.
PostgreSQL with pgvector has independent implementations: code and definitions written to it move to another compatible implementation within the features both support, the data moves by export, replication or rewrite, and the provider's control-plane configuration and operational state don't lift. Delegated. The service behind it is its own chip, Ceded.
Token Factory's embedding and rerank models run on the same NVIDIA fleet as everything else; the models are open weights.
Nebius gives the retrieval layer its parts. Token Factory serves an OpenAI-compatible /v1/embeddings endpoint (the example model is BAAI/bge-en-icl) and a /v1/rerank endpoint ('Reranks documents based on their relevance to a query') per token on the serverless tier; the Applications marketplace deploys Qdrant, the open-source vector database, as a standalone managed application ('Managed infrastructure with zero setup time') in eu-north1 and us-central1 with an export-data page, and points the customer at Qdrant's own API for the rest (hybrid and full-text search included); Managed Service for PostgreSQL ships 'the pg_vector extension' to 'store and index vector embeddings required for retrieval-augmented generation (RAG)-based systems', and the RAG tutorial builds on it; Data Lab runs SQL over inference logs into datasets, which is data preparation rather than retrieval. The buyer gets open embedders, a hosted vector store, and a vector-capable Postgres from one account. The architect's concern is that there's no engine of Nebius's own: the embedders are open weights, the stores are Qdrant's and PostgreSQL's, and nothing Nebius wrote does retrieval over the enterprise's data as a service. Calibration: Cloudflare reads moderate on a managed index plus open embedding and rerank models; Mistral moderate on an embeddings API and libraries; Hugging Face moderate on open embedders and an engine; Cohere moderate on frontier embed and rerank; Elastic and MongoDB strong on native engines. Nebius is Cloudflare's shape with partners' indexes. Moderate.
Delegated on open parts. The embedding and rerank endpoints serve open weights the enterprise could run itself, behind an OpenAI-shaped embeddings door: the door reads Delegated on the open-weights embeddings ruling and the Token Factory service behind it Ceded (the Cloudflare and Hugging Face reading), not the carve-out, because the weights aren't Nebius's. Qdrant as a managed application is Apache 2.0 software Nebius operates, with an export path, and Managed PostgreSQL with pgvector is standard Postgres whose schema and queries lift to any Postgres: both services read Ceded, with the Qdrant API and PostgreSQL as Delegated doors. The runtime call is an endpoint embedding or ranking what the enterprise sent, with per-request model choice: vendor decides, visible, overridable, Delegated, the Elastic reading.
The Token Factory model catalogue is enumerated only through the API (GET /v1/models); which embedding and rerank models are live isn't in a static page in the source set. Public evidence that moves the cell: a Nebius-operated retrieval or hybrid-search engine of its own; nothing describes one. GA evidence table (October 24, 2026): the /v1/rerank endpoint is documented but the public model catalog lists no reranker model, so rerank comes out of the Token Factory chip until a model is offered.
Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering
Nebius's managed sync engine over standard object stores. Ceded, the Cloudflare Super Slurper and Sippy reading; the objects themselves stay behind S3. Not scored within this chip under the GA evidence table: Filesystem Buckets as Destinations ("Filesystem buckets in Object Storage are in preview.").
Nebius's workbench on Nebius's engine for Nebius's trainer, with the datasets exportable. Ceded, the NetApp AIDE captive-pipeline reading.
Data transfers and Data Lab run on Nebius's CPUs and storage.
Nebius moves objects and shapes training data. Data transfers 'allow you to move data between buckets' for replication, backup, 'or when you want to migrate from a different region or third-party storage service', with 'both Object Storage in Nebius AI Cloud and S3-compatible third-party services as sources and destinations' (Azure Blob containers too; an external destination only from a Nebius source), run as consecutive iterations that list, compare, and copy (multipart above 100 MB), with documented overwrite and delete rules for destination objects, key prefixes, stop conditions (one iteration, a run of empty ones, or indefinitely), and management from the console, CLI, or Terraform; a transfer can land in a shared filesystem through a filesystem bucket. Data Lab in Token Factory is 'a unified workspace ... for working with inference logs (chat completions) and datasets': auto-collected chat completions (unless zero data retention is on), SQL-filtered datasets, uploads in JSON Lines (JSONL), S3 connections for JSONL and Parquet 'without creating an additional raw-data copy', batch and fine-tuning outputs, all processed 'exclusively in the EU-North1 (Finland) data center'. Soperator's ephemeral nodes move compute, not data (scored at 2A). The buyer gets a managed sync between object stores and a dataset workbench for fine-tuning. The architect's concern is that it's movement without a pipeline: no transformation, no change data capture, no lineage, no cost-aware orchestration beyond same-region egress being free, no KV-cache tiering; Data Lab prepares data for Nebius's own trainer and nothing else. Calibration: Cloudflare reads moderate with Super Slurper and Sippy (bulk and on-read migration into R2) as scored chips beside Queues and Workflows; CoreWeave gap on LOTA, a transport accelerator with no movement product to buy (its cell reads data movement as a throughput property, not a pipeline); Nutanix gap on snapshot replication; Mistral and Cerebras gap; Hugging Face moderate on a hosted runner plus libraries. Nebius has a movement product the enterprise configures and holds Nebius to, and a captive-destination workbench, each a fixed-function slice (Data Lab on the September 11, 2026 ruling that trainer-input preparation reads on the NetApp AIDE line). Moderate (ruled September 15, 2026): a managed, configurable, continuous object sync between clouds is the purpose line's cost-aware movement as a fixed-function slice; the CoreWeave gap is a transport accelerator with nothing to configure and the Nutanix gap is disaster recovery, so neither is the same shape.
Ceded at the transfer, with the data on standard doors. The transfer resource, its iteration engine, and its overwrite and delete semantics are Nebius's, though both ends are S3 and the enterprise could substitute its own sync tooling: Ceded, the Cloudflare Super Slurper reading. Data Lab is a captive workbench for Token Factory's fine-tuning, on Nebius's SQL engine in Finland: Ceded, the AIDE reading. The runtime call is Nebius's transfer service deciding, each iteration, which objects to copy, overwrite, or delete inside the flags and prefixes the enterprise set; those are configuration, not per-object overrides: vendor decides, visible (iteration status), not overridable, Ceded.
Ruled September 15, 2026: moderate stands on the fixed-function reading and the Cloudflare migration-chip reading; gap had been argued on the CoreWeave movement-not-pipeline line. Data Lab settled September 11, 2026: a captive-destination preparation surface is a fixed-function slice (the Cerebras and NetApp AIDE line) and is scored as a Ceded chip. Named, not scored: the Batch API (referenced in rate limits and the Files API; no standalone page in the Token Factory index as of September 10, 2026). Public evidence that moves the cell: a transformation, connector, or lineage product. Out of the Object Storage Data Transfers chip under the GA evidence table (October 24, 2026): Filesystem Buckets as Destinations ("Filesystem buckets in Object Storage are in preview.").
GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization
Standard Kubernetes with NVIDIA's operators on Nebius's nodes; workloads and manifests lift to any cluster. Delegated, the CoreWeave CKS and Nutanix NKP reading; node-group provisioning is the control-plane chip.
An open operator Nebius manages; Slurm jobs and the operator itself lift to any Kubernetes, untested elsewhere by Nebius's own note. Delegated, the CoreWeave SUNK reading; the capacity-block machinery is the control-plane chip.
Nebius's provisioning, placement, scaling, and bounds over Nebius's capacity. Ceded, the CoreWeave Mission Control reading.
GPU node groups without Nebius's boot image 'must have NVIDIA GPU Operator installed'; B200 and InfiniBand node groups require the NVIDIA Network Operator; GPUDirect RDMA is on by default. The schedulers (Kubernetes, Slurm, Kueue, Volcano) are open.
Nebius orchestrates GPUs three ways and lets the enterprise move them. Managed Service for Kubernetes (free control plane, nodes billed as VMs) runs GPU node groups on a platform and preset ('You cannot change the VM platform and preset or the GPU cluster of an existing node group'), installs the NVIDIA operators or ships its own boot image, scales with 'the cluster autoscaler [that] seamlessly adds nodes when there are unschedulable Pods', exposes InfiniBand topology as tier labels for Kueue or Volcano, and attaches security groups per node group. Managed Service for Soperator runs Slurm as Kubernetes pods ('an open-source solution from Nebius', github.com/nebius/soperator, Apache 2.0) with a shared root filesystem, and, from version 3.0 with Nebius enabling it, ephemeral nodes the enterprise powers down with scontrol to 'reuse the same capacity block group for other workloads, such as inference'. Serverless AI provisions VMs for jobs and endpoints since July 28, 2026; Token Factory dedicated endpoints scale between minimum and maximum replicas of a chosen GPU type in a chosen region ('Minimum replicas are reserved and non-preemptible ... Maximum replicas scaling depends on available burst capacity and is not guaranteed'); quotas bound each region and platform, preemptible VMs are cheaper, and commitment discounts are 'only available if you have signed up ... as a company'. The Applications marketplace deploys onto Kubernetes or as standalone managed apps. The buyer gets Kubernetes and Slurm on NVIDIA fabrics with the levers to move capacity between training and inference. The architect's concern is where the plane stops: topology tiers whose 'exact physical meaning ... is not guaranteed', a preselected fabric, node groups that can't change platform, dedicated-endpoint burst that isn't reserved, and an ephemeral-node feature Nebius must switch on. Calibration: CoreWeave reads strong on CKS plus SUNK (Slurm on Kubernetes) plus Mission Control; Nutanix strong on Prism, ADS, and a Kubernetes platform; NetApp, VAST, and Everpure moderate on data-infrastructure control planes; Mistral and Cerebras gap on metering edges. Nebius is CoreWeave's shape on the schedulers (Kubernetes plus Slurm on Kubernetes) without a Mission Control: no documented fleet-health, straggler, or node-lifecycle plane across both, and Soperator 'has not been tested on platforms other than Nebius AI Cloud'. Strong (ruled September 15, 2026): the strong 2A cohort reads on managed Kubernetes and its scheduling (AWS on EKS plus capacity management, GCP on GKE, Azure on AKS plus Arc, Nutanix on Prism, ADS, and NKP) and none of it requires a fleet-operations plane; CoreWeave's Mission Control is the exception, not the criterion.
Delegated at the schedulers and two documented pins, Ceded at the control plane. Kubernetes and Slurm are open interfaces the enterprise's workloads already speak, Soperator is Apache 2.0, and the topology labels feed the enterprise's own Kueue or Volcano: Delegated, the CoreWeave CKS and SUNK reading; node-group provisioning (platform, preset, fabric, boot image), the capacity-block machinery behind ephemeral nodes, Serverless AI, the autoscaler's bounds, quotas, and the marketplace deployer are Nebius's: Ceded. The runtime call is placement. A fabric per cluster and a platform and preset per node group are menus, and autoscaler minimums, maximums, and quotas are bounds, so none of them is an override under the standing rulings; two controls are: the topology labels the enterprise's own scheduler places pods on (the WEKA node-selector reading), and scontrol power down and up on named Slurm nodes the enterprise chooses to release and repower (a per-node pin). Nebius picks the host inside the fabric and the burst it can honor: vendor decides the defaults, visible (topology labels, quotas, statuses), overridable on those two documented controls, Delegated, the Dataiku and WEKA pin reading. Cross-row: CoreWeave's 2A reads Ceded on the same controls, scored before the override rule.
Ruled September 15, 2026: strong stands on the cohort calibration (managed Kubernetes plus Slurm on Kubernetes plus capacity moves exceeds the AWS and GCP basis for strong); moderate had been argued on the missing fleet-operations plane and Soperator being untested off Nebius. Cross-row: CoreWeave 2A (Ceded, September 2) against this row's Delegated on the two documented pins; the override-rule column item. Soperator ephemeral nodes require version 3.0, Nebius enabling the feature, and a shared capacity block group on one GPU platform; moving capacity means resizing a Managed Kubernetes node group on the same block. Serverless AI is documented by an overview and pricing page only. Public evidence that moves the cell: nothing upward from strong.
Model serving, agent execution, inference APIs, distributed inference
Open-weight models their owners publish, behind multi-vendor interfaces. Delegated, the inference-interface ruling; embeddings and rerank are the 1B chips; image generation was retired April 13, 2026.
Tool logic the enterprise writes and runs. Retained, the customer-tools ruling.
Nebius's provisioned serving of the same open weights behind the same OpenAI-compatible API, with replica bounds as configuration. Delegated, the inference-interface ruling; custom weights are named, not scored. Not scored within this chip under the GA evidence table: Custom Weights ("Working with custom model weights is currently in beta and available on request.").
Nebius's training service and workbench; the weights it produces export to Hugging Face format and are the enterprise's. Ceded for the service, with the artifact exit named, the Databricks Mosaic AI Model Training reading.
Others' open-source applications Nebius operates for the enterprise with documented exports. Delegated, the Nutanix managed-open-engine reading. Not scored within this chip under the GA evidence table: Flowise Agent Builder, Open WebUI ("Deploy JupyterLab on a Kubernetes cluster").
Dedicated endpoints are provisioned by NVIDIA GPU type (L40S, H100, H200, B200, B300); serverless runs on the same fleet; Nebius calls itself 'the first AI cloud to adopt NVIDIA Groq 3 LPX' for Vera Rubin NVL72 in Token Factory (undated). Clarifai's core team joined and Nebius 'licenses inference IP' from it (undated).
Nebius's runtime is a European-hosted open-model service with the training side attached. Token Factory 'offers an OpenAI-compatible API for inference and fine-tuning': chat completions (streaming), a Responses API (no badge), completions, embeddings, vision through image_url, and rerank (text-to-image models and Low-Rank Adaptation (LoRA) per-token endpoints were retired on April 13, 2026, 'as we streamline supported modalities'); function calling with 'Connect to Model Context Protocol (MCP) servers' among its named uses; JSON schema and JSON object modes where a model carries the JSON mode tag; two flavors per model (base and fast, 'identical model outputs' at different price and latency); a catalogue of third-party open weights enumerated through GET /v1/models (Llama, Qwen, DeepSeek, and 'Kimi K3 is now available', 'Qwen3.8-Max ... with Nebius Token Factory as a Day 0 launch partner'), with serverless deprecation waves on June 22 and August 31, 2026 that 'will be unaffected' on dedicated endpoints. Dedicated endpoints ('a private, provisioned instance' shape) let the enterprise 'control: region, GPU type and count per replica, minimum and maximum replicas' in eu-north1, eu-west1, or us-central1, billed per GPU-hour by the minute with no standard rate limits, with observability (traffic, latency percentiles, autoscaling, errors) and Prometheus and Grafana federation for dedicated endpoints only; custom weights 'is currently in beta and available on request'. Fine-tuning jobs run supervised training (full by default, LoRA optional with rank up to 128) or custom speculator training ('a drafter tuned to your own base model and workload') on listed Hugging Face base models, with Weights and Biases and Hugging Face export integrations on the guide page and MLflow in the API reference, over datasets from Data Lab (reinforcement fine-tuning is a limited professional service by request); rate limits scale dynamically to twenty times a base allocation before an Enterprise plan; zero data retention is an organization switch and the chat endpoint is HIPAA-eligible under a BAA with it on (fine-tuning and batch 'excluded from BAA scope'). The Applications marketplace deploys Flowise (an open-source agent builder), Open WebUI, JupyterLab, and ComfyUI as standalone managed apps with export pages. The buyer gets open models it can call, tune, and export, on European floors, with the agent tooling as hosted open-source apps. The architect's concern is what Nebius itself adds above the fleet. No model of its own (a licensed Clarifai inference stack inside), a catalogue visible only by API, a burst that 'is not guaranteed', custom weights in beta, batch documented by reference, a modality just retired, and no generally available agent runtime beyond function calling and someone else's app (Sandboxes, a branching code-execution API for agents with a Contree MCP server, 'are currently in Beta'). Calibration: CoreWeave reads strong on open serving under KServe plus its own inference product and sandboxes; Cloudflare strong on serverless open-model inference plus a gateway plus agents; Hugging Face moderate on open serving under maturity gates; Cerebras strong on its own engine at the frontier; Databricks strong on serving plus training plus agents. Nebius has open serving with replica control, generally available full and LoRA fine-tuning and speculator training with export, and a marketplace of open agent tools: the CoreWeave shape with training on top. Strong.
Delegated at the interfaces and the weights, Ceded at Nebius's surfaces. Serverless model access through OpenAI-compatible endpoints is the inference-interface reading, and the models are their owners' open weights: Delegated. Fine-tuned weights export to Hugging Face format and are the enterprise's artifact: Retained for the weights, the customer fine-tunes reading. Dedicated endpoints serve the same open weights behind the same interface on provisioned replicas: Delegated, the inference-interface ruling. Data Lab, the fine-tuning service, and the flavor catalogue are Nebius's control surfaces on Nebius's GPUs: Ceded, the Databricks Mosaic AI Model Training and Hugging Face flavor readings. Flowise, Open WebUI, JupyterLab, and ComfyUI are open-source applications Nebius operates with documented exports: Delegated. Tools the enterprise writes are Retained. The decision to call a tool is the model's, executed in the enterprise's own process (Token Factory 'does not execute the functions itself'): model decides, visible, overridable, Delegated, the OpenAI and Anthropic reading.
Badges: custom weights 'currently in beta and available on request', unscored; Sandboxes 'currently in Beta' (Contree SDK, CLI, and MCP server), unscored; the Responses API carries no badge, and its OpenAPI schema lists MCP, code-interpreter, shell, and web-search tool types with approval-request items that no guide documents as executing, unscored. Retired April 13, 2026: text-to-image models (FLUX) and LoRA per-token serverless endpoints. Named, not scored: Reinforcement Fine-tuning ('upcoming', a limited professional service by request, no GA date), the Batch API (referenced, no standalone page), the Nebius Agents Blueprint (blog, June 10, 2026: a reference architecture with recipes naming LangChain Deep Agents, LangSmith, Pinecone Nexus, Tavily, and Snowglobe), the Clarifai inference IP license and the Groq 3 LPX adoption (August 24, 2026). Data Lab processes in Finland regardless of where inference ran. Public evidence that moves the cell: nothing upward from strong; custom weights or Sandboxes leaving beta would add a chip. Ruled October 15, 2026: Dedicated Endpoints read Delegated under the inference-interface ruling, since the same client can point at another provider serving the same open weights and replica bounds are configuration; the Cohere Model Vault precedent no longer applies, as it serves Cohere's own models. Out of the Dedicated Endpoints chip under the GA evidence table (October 24, 2026): Custom Weights ("Working with custom model weights is currently in beta and available on request."). Out of the Managed Open-Source Applications chip under the GA evidence table (October 24, 2026): Flowise Agent Builder, Open WebUI ("Deploy JupyterLab on a Kubernetes cluster").
Policy-driven placement and resource coordination — the Autonomy Layer
Nothing at this layer is a scored capability.
Read against the five legs, Nebius has none as a plane over agents. Identity: tenants, projects, groups, access permits, service accounts, and OIDC federated credentials identify people and workloads to the cloud and to Token Factory; no agent principal. Gateway: none; Token Factory's function calling hands tool calls back to the enterprise, and the Responses API's schema lists MCP and hosted-tool types with approval-request items that no guide documents as executing. Registry and orchestration: none generally available. Nebius Echo is documented ('an AI assistant that helps you navigate Nebius documentation, explore and manage project resources', executing state changes 'with explicit approval'), and it's one agent Nebius runs over its own console, scored at Layer 3; the Nebius Agents Blueprint (blog, June 10, 2026) is 'an open reference architecture ... and runnable recipes with cloneable code' over Token Factory, LangChain Deep Agents, LangSmith, Pinecone Nexus, Tavily, and Snowglobe, a design, not a service; Sandboxes, a branching code-execution API for agents with a Contree MCP server, 'are currently in Beta'. Observability: dedicated-endpoint metrics, Monitoring and Logging in preview, and Audit Logs in preview watch the cloud, not agents; Sandboxes' logs and artifacts are beta with the product. The buyer brings the plane. The architect's concern is nil at this layer: there's nothing generally available to be captured by. Calibration: CoreWeave, OpenAI, Anthropic, and Cerebras read gap; Cloudflare moderate on a gateway of its own; WEKA gap on application-scoped legs. Nebius is CoreWeave's shape. Gap, authority Absent; Echo is the one-agent case the Confluent, Hugging Face, and WEKA sheets read as a leg, not a plane.
Nothing offered as a plane. Absent.
Named, not scored: Nebius Echo (documented; scored at Layer 3 as a console assistant), the Nebius Agents Blueprint (blog, June 10, 2026; no product documentation and no GA date, so not watch-listed), Sandboxes and Contree MCP (Beta), the Responses API tool schema (undocumented beyond the OpenAPI listing), Audit Logs, Monitoring, and Logging (preview). Public evidence that moves the cell: Sandboxes leaving beta with a gateway or registry over the enterprise's agents, or documentation for the Blueprint as a service.
AI-powered business capabilities — business logic, workflow automation
The documented Kubernetes application is the NVIDIA Blueprint for virtual screening, which 'orchestrates a suite of GPU-accelerated NIM microservices' under a bring-your-own NVIDIA AI Enterprise license; the standalone apps run on Nebius's NVIDIA VMs. Nothing at this layer is a scored capability.
Nebius's application layer is a marketplace with one assistant of its own. Applications deploy three ways ('Virtual machine: ... root access and custom setup'; 'Standalone: ... infrastructure managed by Nebius AI Cloud ... zero setup time'; 'Kubernetes: ... on Managed Service for Kubernetes clusters'), once per project per option and not every option for every app, with documented per-app pages (connect, export data) for JupyterLab, Open WebUI, ComfyUI, Flowise, and Qdrant, standalone pricing that charges for the compute and storage under an app and nothing for the app, and the NVIDIA Blueprint for virtual screening on Kubernetes, which 'orchestrates a suite of GPU-accelerated NVIDIA Inference Microservices (NIM)' and requires 'a valid and sufficient NVIDIA AI Enterprise License' the enterprise brings; standalone apps run in eu-north1 and us-central1. Nebius Echo, 'an AI assistant that helps you navigate Nebius documentation, explore and manage project resources', lists resources and 'executes actions like creating Nebius resources, or starting and stopping virtual machines (VMs) with explicit approval' from the console. The buyer gets a hosted chat UI, an agent builder, notebooks, an image pipeline, and a drug-discovery blueprint in a few clicks, none of them Nebius's, plus a console assistant that is. The architect's concern is that Nebius writes none of the applications and runs a marketplace of open-source projects and one NVIDIA blueprint; Echo is an operations assistant over Nebius's own console, not a business application; the value above the console is the enterprise's or the projects'. Calibration: under the Layer 3 marketplace ruling (September 11, 2026), a marketplace of other people's applications earns no capability grade, a console assistant over the vendor's own platform doesn't either, and the grade rests on first-party business applications; WEKA, NetApp, Everpure, Cerebras, and Groq read gap; Databricks and Snowflake strong on first-party applications. Nebius has none. Gap, authority Absent.
Nothing offered as a business application. Absent. The relationship boundary records on DAPM where it exists, and here it doesn't: the marketplace bills the compute and storage under an app to the enterprise's Nebius account and carries no commercial relationship for the applications themselves (open source, or NVIDIA's under NVIDIA's license).
Ruled September 11, 2026 (the Layer 3 marketplace ruling): the first draft read moderate, escalated, on the Nutanix ISV-catalogue reading; a marketplace of others' applications isn't creditworthy at Layer 3, and Echo is an operations assistant rather than a business application. Named, not scored: the managed standalone apps (JupyterLab, Open WebUI, ComfyUI, Flowise, Qdrant; the Qdrant retrieval facet is scored at 1B and the agent tooling at 2B), the self-run VM and Kubernetes deployment paths, the NVIDIA Blueprint for virtual screening (bring-your-own NVIDIA AI Enterprise license), Nebius Echo (no availability badge on its page). Public evidence that moves the cell: a first-party business application.
Nebius AI Cloud documentation at docs.nebius.com (overview, services, regions, quotas; Compute: index, GPU clusters, topology, virtual machines, storage types, pricing, quotas; Managed Service for Kubernetes: index, components, GPU set-up, topology-aware scheduling, storage over CSI, networking and security groups; Slurm and Soperator: index, overview, ephemeral nodes and capacity moves; Object Storage: overview, quickstart, pricing, storage classes, bucket policies, data transfers (overview, launch); Managed Service for PostgreSQL and the RAG tutorial; HIPAA Implementation Guideline; Nebius Echo; Container Registry; IAM: overview, authorization, access tokens, access keys, static keys, federated credentials, roles, federations and SAML SSO; KMS; Audit Logs; Virtual Networks: overview, security groups; security and encryption; Applications: index, types, deploy, standalone quotas and pricing, the NVIDIA Blueprint for virtual screening and its licenses page; MLflow; PostgreSQL; Serverless AI; llms.txt); Nebius Token Factory documentation at docs.tokenfactory.nebius.com (API reference for inference, embeddings, rerank, images, responses, files, datasets, fine-tuning, dedicated endpoints; dedicated endpoints overview, capacity and scaling, control and data plane, lifecycle, billing, custom weights, FAQ; function calling, JSON, rate limits, observability, playground, Data Lab; post-training overview and fine-tuning guide; sandboxes overview; deprecation notices of March (effective April 13), June 22, and August 31, 2026; legal quick guide, HIPAA guideline, sub-processors, terms, privacy; llms.txt); nebius.com pages (AI Cloud, Token Factory, prices, services, newsroom through September 8, 2026, the Meta agreement of March 16, 2026, the Agents Blueprint post of June 10, 2026, the Groq 3 LPX post of August 24, 2026); GitHub license API for nebius/soperator. Peer-reviewed cell by cell through the labs claims ledger (nebius-<layer>-chatgpt, ChatGPT gpt-5.5) and as a whole row by Antigravity (nebius-row-agy); totals and escalated items in reviews/nebius-judgment.md. /reconcile (October 5, 2026): the settled 2A escalation (ruled September 15, 2026) cleared from the published notes. No grade, component, or authority change. /reconcile (October 5, 2026): Layer 0 runtime rule (ruled September 28 and October 4, 2026). GPU VMs and clusters split: VMs Ceded to Retained from the customer's seat (the CoreWeave GPU compute reading), cluster fabrics Ceded on the CoreWeave network fabric reading. Grade unchanged. v1.4 (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. 0A strong, 0B strong, 0C strong; summary edits (summary worksheet, Keith, October 16, 2026). v1.5 (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. v1.6 (October 22, 2026, /reconcile): the 0B calibration line cites GCP's RoCE GPU network in place of Virgo, which GCP v1.26 moved to a named, unscored note. Grade and authority unchanged. v1.7 (October 22, 2026, /reconcile): GB200 and GB300 NVL72 instance groups added at 0A (system) and 0B (NVLink facet), both Ceded on the integrated-system rule. Nebius's Terraform provider documents the resource ("Represents an NVLink InstanceGroup," GB200 or GB300), and its GB300 page says the system is deployed in production, sold through the sales team. Found by the 0B fabric research. Grades and authority unchanged. v1.8 (October 22, 2026, /reconcile): 0A region counts follow Nebius's regions doc (nine public, adding Madrid and Minnesota; one private, Iceland), and the CPU chip's range runs to 256 vCPUs on cpu-d3 and 80 on cpu-e2, per the VM types doc; the summary carries the same region counts. Found by the 0A compute research. Grade and authority unchanged. v1.9 (October 23, 2026, /reconcile): Layer 1 research pass. 1A: "Shared Filesystems (virtiofs Mounts to Many VMs in a Project" removed (a shared filesystem and a container registry are storage). Found by the Layer 1 slot cut. Grade and cell authority unchanged. v1.10 (October 23, 2026, storage door pass): The S3 API is a door; the object and shared-filesystem services read Ceded. Grade unchanged. v1.11 (October 23, 2026, Layer 1 managed-services pass): Token Factory Embedding and Rerank Endpoints splits; Qdrant as a Managed Standalone Application splits; Managed Service for PostgreSQL With pgvector splits. Grade unchanged. v1.12 (October 24, 2026, /reconcile): GA evidence recorded on 2 of 6 Layer 1B chips under the GA evidence table. Grade unchanged. v1.13 (October 24, 2026, /reconcile): Layer 1B under the GA evidence table version 2: GA evidence on every chip, preview items and unaffirmed chips to notes. Grade unchanged. v1.14 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 2 Layer 1A chips under the GA evidence table version 2. Grade unchanged. v1.15 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 2 Layer 1C chips under the GA evidence table version 2. Grade unchanged. v1.16 (October 24, 2026, /reconcile): GA evidence recorded on 3 of 3 Layer 0B chips under the GA evidence table version 2. Grade unchanged. v1.17 (October 24, 2026, /reconcile): GA evidence recorded on 3 of 3 Layer 0A chips under the GA evidence table version 2. Grade unchanged. v1.18 (October 24, 2026, /reconcile): GA evidence recorded on 3 of 4 Layer 0C chips under the GA evidence table version 2. Grade unchanged. v1.19 (October 24, 2026, /reconcile): GA evidence recorded on 3 of 3 Layer 2A chips under the GA evidence table version 2. Grade unchanged. v1.20 (October 24, 2026, /reconcile): GA evidence recorded on 3 of 5 Layer 2B chips under the GA evidence table version 2. Grade unchanged. v1.21 (October 24, 2026, /reconcile): GA evidence table version 2.1 applied: 5 chips recorded or narrowed. Grade unchanged. v1.22 (October 27, 2026): each component now carries a durable offering id; nothing else changed.