{ "id": "lightning-ai", "name": "Lightning AI (Lightning Studios, GPU Marketplace, Jobs, Pipelines and Deployments + Model APIs + Voltage Park GPU Cloud + GraphN Agent Workflows)", "subtitle": "Mapped to the 4+1 Layer AI Infrastructure Model", "version": "v1.0 - 4+1 v3: Initial Assessment", "date": "November 13, 2026", "source": "Lightning AI platform documentation (lightning.ai/docs/platform, all 270 pages reached from lightning.ai/llms.txt and in-page links: GPU Cloud, GPU marketplace, bring your own cloud, Cluster, Data, Drive, cloud folders, data connections, datasets, move data, Inference, Model APIs, Lightning Inference, autoscaling, clouds and regions, multi-node inference, self-serve vLLM, custom models and model registry, deploy containers, Pipelines, batch jobs, Train, multi-node training, experiment management, Build, Studios, Sandbox, Agents, Security, RBAC, SSO, audit logs, Team Management), the Python SDK reference (lightning.ai/docs/sdk/stable, including Machine, Job, Deployment, Agent, K8sCluster, Teamspace), the CLI reference (lightning.ai/docs/cli/stable), LitData and LitServe documentation, lightning.ai/pricing, the Lightning AI terms and conditions, the lightning-sdk 2026.11.13 package, and PyPI metadata for litserve, litdata, litmodels and litai, read November 13, 2026; Voltage Park help center (support.voltagepark.com: Deploy GPUs, Bare Metal, Virtual Machines, Deploy and Attach Storage, VAST NFS mounting, InfiniBand setup, Managed Kubernetes add-on, Dedicated Reserve, local NVMe) and product pages (voltagepark.com: storage, managed Kubernetes, bare metal, Blackwells, pricing, infrastructure), read November 13, 2026; GraphN (graphn.ai: llms.txt, FAQ, AI workflow platform, workflows, blueprints catalog, pricing, terms, docs for getting started, concepts, inference quickstart, models, LoRA adapters, DSL reference, API reference, Function SDK reference, coding agents, document analysis, and the generated OpenAPI specification v0.39.0 at github.com/voltagepark/graphn-openapi), read November 13, 2026; Lightning AI blog (Voltage Park merger post, January 21, 2026; September 2026 recap); Wayback Machine snapshots of the withdrawn GPU Cloud > Clusters pages (October 25, 2026). Peer-checked cell by cell by ChatGPT (gpt-5.6-sol); ledgers in reviews/lightning-ai-*-chatgpt.json.", "status": "complete", "summary": { "title": "Summary Finding", "paragraphs": [ "Lightning AI is a GPU cloud built from the developer's side in. It merged with Voltage Park on January 21, 2026, which gave the company that built PyTorch Lightning an owned NVIDIA fleet to sit under its Studios, and since then it has added GraphN, an agent-workflow platform at graphn.ai. The map reads it strong at three layers, moderate at five and gap at two. Compute and fabric are strong: H100 bare metal on 3.2 Tbps InfiniBand, self-serve to 1,016 GPUs, plus a GPU marketplace that places work on Lightning's cloud, AWS, Google Cloud, Lambda or Nebius, with B200 documented in the SDK. Storage is moderate, because Voltage Park's VAST-backed volumes are documented only as NFS over TCP. The runtime is strong on two doors that carry the enterprise's own agent loop: managed Deployments running the official vLLM image, and GraphN's OpenAI-compatible inference with tool calling on every built-in chat model. Data governance, retrieval, pipelines and orchestration are moderate. There's no reasoning plane and no packaged business application: GraphN's 72 blueprints are starting points the enterprise copies and finishes.", "The capture is decoupled. The parts the buyer sees are open or standard: NVIDIA machines a CUDA or PyTorch workload leaves without a rebuild, the vLLM container and the OpenAI and Anthropic interfaces at both inference doors, open-weight models, and Kubernetes on Voltage Park. What compounds sits beside them, in Lightning's own surfaces: the Studio, Job and Deployment definitions and their settings, Lightning's placement engine, the proprietary filesystem behind the Drive, LitLogger's experiment history, GraphN's workflow DSL and its knowledge-base indexes. Sixteen of twenty-seven components read Ceded. Lightning's own pitch is portability across clouds, and on compute it's real; the opinions an ML team accumulates over a year still live in Lightning's namespace, the CoreWeave Forge shape.", "Decision authority mostly sits with Lightning where it runs things and with the enterprise where it writes them. Lightning decides which nodes and clouds a workload lands on, and a cloud, region or replica pin is the EC2-region choice, not an override. The enterprise's code decides at 1A and 1C, where Lightning enforces access and runs pipelines exactly as configured, and at 2B, where the model decides to act but the enterprise's own loop or GraphN guards stand before the effect. The trade for the buyer is speed for namespace: GPUs found across clouds, a training-to-serving path with no glue code, and agents in a day, in exchange for accumulating the team's working opinions in a platform that's open at the edges and proprietary in the middle. The one decision to make up front is which of Lightning's surfaces to build on and which to keep in the enterprise's own code, because the workloads leave easily and the platform habits don't." ] }, "layers": [ { "id": "layer0a", "label": "Layer 0A", "shortName": "Compute", "title": "Compute", "purpose": "CPUs, accelerators, servers, integrated systems, OS and hypervisor, physical plant", "status": "strong", "statusLabel": "An Owned H100 Fleet on InfiniBand Plus a GPU Marketplace Reaching B200", "authority": { "decides": "vendor", "visible": true, "overridable": false, "boundary": "vendor", "direction": "Ceded" }, "nvidia": [ { "component": "Every Accelerator Is NVIDIA's", "detail": "Voltage Park's documented fleet is NVIDIA HGX H100 SXM5 on Dell PowerEdge XE9680 servers. The marketplace's GPU machines run from T4 and L4 through A100, H100, H200 and B200. TPUs appear on the pricing page and in a Studio screenshot, with no product documentation beyond that." } ], "gap": "Lightning AI sells GPU compute two ways. It owns a fleet: the Voltage Park merger closed January 21, 2026, and Voltage Park's help center documents nodes of eight NVIDIA H100 SXM5 GPUs on Dell PowerEdge XE9680 servers, rented on demand as virtual machines (one to eight GPUs on 100 Gbps Ethernet) or as bare metal in multiples of eight, self-serve up to 1,016 GPUs on 3.2 Tbps InfiniBand, with 12-month Dedicated Reserve contracts beyond that. It also brokers: the lightning.ai GPU marketplace runs Studios, Jobs and Deployments on Lightning's own cloud or on AWS, Google Cloud, Lambda and Nebius from one console and one bill, on demand, interruptible or reserved. The SDK's Machine API reference documents H200 in one, two, four and eight GPUs and B200 in one or eight, the same names on every cloud with availability by cloud and account, and the pricing page lists B200 on demand. The buyer gets current NVIDIA capacity without choosing a cloud first.\n\nThe architect's concern is what the strong grade rests on. Voltage Park's Blackwell page advertises B200, GB200 NVL72, B300 and GB300 NVL72 for reservation, but the help center documents H100 only, so the owned fleet scored here is Hopper. Blackwell enters the cell through the marketplace's B200 machine type, which may run on Lightning's cloud or another provider's. Bring-your-own-cloud compute sits in the enterprise's own AWS or Google Cloud account and is scored on those rows, not this one.\n\nCalibration: CoreWeave and Nebius read strong on the same rented-NVIDIA-fleet shape, with B300 documented; Hugging Face reads gap as a service vendor. Lightning reads in the infrastructure-as-a-service (IaaS) shape, because the enterprise rents and sizes the machines.", "borrowedJudgment": "Rented, so the runtime call is the operator's. Voltage Park decides which nodes a bare-metal or VM rental lands on and when they're maintained; on the marketplace, the underlying cloud places the instance. The enterprise picks the machine type, count and, through the SDK's cloud_provider argument, the provider. That pin overrides Lightning's brokering choice, not the host placement the grade rests on: vendor decides, visible, not overridable, Ceded. Portability: the H100 nodes and marketplace instances are standard NVIDIA machines a CUDA or PyTorch workload leaves without a rebuild, Retained.", "notes": "Named, not scored (checked November 13, 2026): Voltage Park's Blackwell fleet (HGX B200, GB200 NVL72, HGX B300, GB300 NVL72) on its product page with 'Reserve Now' and no help-center documentation; TPUs on the lightning.ai pricing page and a getting-started screenshot; Lightning's managed cluster documentation (GPU Cloud > Clusters, with Kubernetes, SLURM, Networking and Storage pages, archived October 25, 2026 and since withdrawn; every page now redirects to a Cluster overview that reads 'Documentation for Cluster is on its way.'). The GPU marketplace doc says 'over 7 compute providers'; the SDK's CloudProvider enum names five (AWS, GCP, Lambda Labs, Nebius, Lightning). Bring your own cloud (Enterprise tier) runs Lightning's platform in the enterprise's AWS or Google Cloud account; that compute is scored on those rows. The B200 machine type rests on the SDK's Machine API reference (lightning.ai/docs/sdk/stable/api/machine), which governs over the pricing page; which provider serves a B200 request isn't documented. Voltage Park's own Blackwell fleet, Lightning's January 21, 2026 merger post ('H100s, B200s, and GB300s'), and the pricing page's B200 cluster reservations are vendor statements without documentation.", "components": [ { "id": "lightning-ai.0a.voltage-park-gpu-compute", "component": "Voltage Park GPU Compute (8x H100 SXM5 on Dell PowerEdge XE9680; VMs of 1 to 8 GPUs on Ethernet; Bare Metal in Multiples of 8, Self-Serve to 1,016 GPUs on InfiniBand; Dedicated Reserve, 12+ Months)", "detail": "Instances the enterprise rents on Lightning's own fleet through the Voltage Park console and API. Standard NVIDIA nodes: a CUDA or PyTorch workload moves to another cloud's NVIDIA instances or an OEM server without a rebuild. Retained, the CoreWeave HGX reading.", "dapm": "Retained", "ga": { "basis": "offered", "url": "https://support.voltagepark.com/article/deploy-gp-us", "quote": "Access GPU deployment solutions on-demand for AI, ML, HPC, and enterprise-grade applications.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.0a.lightning-gpu-marketplace-machines", "component": "Lightning GPU Marketplace Machines (Lightning Cloud, AWS, Google Cloud, Lambda, Nebius; T4 Through H200 and B200; On-Demand, Interruptible, Reserved)", "detail": "Machines the enterprise rents through Lightning's console, SDK and bill, placed on Lightning's cloud or a named provider. Standard NVIDIA instances, Retained on the same reading; the Studio, Job and Deployment runtimes that sit on them are read at Layer 2.", "dapm": "Retained", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/gpu-cloud/gpu-marketplace", "quote": "Lightning AI's GPU marketplace lets you run any type of workload (Studio, Job, Pipeline, Deployment, etc...) on any cloud.", "checked": "November 13, 2026" } } ] }, { "id": "layer0b", "label": "Layer 0B", "shortName": "Fabric", "title": "Network Fabric", "purpose": "Interconnects, switching and routing, network virtualization", "status": "strong", "statusLabel": "3.2 Tbps InfiniBand on Voltage Park Bare Metal, Self-Serve to 1,016 GPUs", "authority": { "decides": "vendor", "visible": true, "overridable": false, "boundary": "vendor", "direction": "Ceded" }, "nvidia": [ { "component": "NVIDIA Quantum-2 InfiniBand", "detail": "InfiniBand is an open standard of the InfiniBand Trade Association; Voltage Park's implementation is NVIDIA Quantum-2 with Mellanox adapters, and its setup guide installs MLNX_OFED and runs NCCL all-reduce tests across the mlx5 devices. The H100 nodes' NVLink domain is the eight GPUs inside each HGX board." } ], "gap": "Voltage Park sells a GPU fabric. Its help center documents bare-metal H100 nodes on 3.2 Tbps InfiniBand or 100 Gbps Ethernet, with virtual machines on Ethernet only, and its pricing page lists self-serve InfiniBand rentals of 8 to 1,016 GPUs up within 15 minutes and reserved configurations of 32 to 8,000-plus. An InfiniBand setup guide walks through OFED drivers, RDMA tooling and NCCL tests. On lightning.ai, an on-demand multi-node job is placed on collocated nodes with high-performance networking, and Lightning says it looks for clouds with InfiniBand and benchmarks them. The buyer gets a training fabric without building one.\n\nThe architect's concern is how little of it is documented as theirs to shape. Voltage Park's product pages call the fabric non-blocking NVIDIA Quantum-2 and describe VPC-level isolation, but the help center documents no partitioning, topology or isolation controls; firewall rules are something its network engineers implement on request. Off Voltage Park, the marketplace inherits whatever the chosen cloud provides, and Lightning's own docs say some clouds don't offer high-performance networking.\n\nCalibration: CoreWeave and Nebius read strong on the same InfiniBand shape, with partition-key isolation and topology labels documented on Nebius. Lightning's documented controls are thinner, but strong is pegged to the best shipping fabric the enterprise can deploy, and a GA 3.2 Tbps InfiniBand fabric rented at thousand-GPU scale stands with theirs.", "borrowedJudgment": "Ceded where the grade rests. Voltage Park's InfiniBand fabric makes the routing, congestion and isolation calls; the enterprise picks InfiniBand or Ethernet at deploy time and tunes NCCL on its own nodes, and no documented control overrides the fabric: vendor decides, visible, not overridable, Ceded. Portability: InfiniBand is a standard, but the enterprise can't take a fabric with it, the CoreWeave and Nebius network fabric reading.", "notes": "Named, not scored (checked November 13, 2026): Quantum-X800 InfiniBand, Spectrum-X Ethernet and ConnectX-8 on Voltage Park's Blackwell page, with the Blackwell systems themselves undocumented; GB200 and GB300 NVL72 rack fabrics on the same page. Marketing pages describe a non-blocking fabric and VPC-level isolation; the help center doesn't document either control. Bring your own cloud uses the enterprise's own cloud networking and an optional VPN, scored on those clouds' rows. Withdrawn (checked November 13, 2026): Lightning's GPU Cloud > Clusters pages, archived October 25, 2026, described InfiniBand HDR, NDR and NDR400 and 200G and 400G RoCE v2 support, topology-aware placement, per-fabric NCCL tuning and dedicated network partitions across '8+ partner clouds'; every one now redirects to a Cluster overview that reads 'Documentation for Cluster is on its way.' Withdrawn docs don't score.", "components": [ { "id": "lightning-ai.0b.voltage-park-infiniband-fabric", "component": "Voltage Park InfiniBand Fabric (3.2 Tbps; Bare Metal in Multiples of 8 GPUs, Self-Serve to 1,016; Reserved to 8,000+)", "detail": "NVIDIA InfiniBand on Lightning's own fleet, chosen at deploy time for bare-metal rentals. Ceded, the CoreWeave and Nebius network fabric reading.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://support.voltagepark.com/article/bare-metal", "quote": "Select your networking option (Ethernet or InfiniBand) > Bare Metal", "checked": "November 13, 2026" } }, { "id": "lightning-ai.0b.voltage-park-ethernet-and-public-networking", "component": "Voltage Park Ethernet and Public Networking (100 Gbps Ethernet for VMs and Bare Metal; NAT and Port Forwarding on VMs; Provider-Side Firewall Rules on Request)", "detail": "Lightning's network primitives in its own data centers, operated by Voltage Park. Ceded, the CoreWeave VPC and Nebius Virtual Networks reading.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://support.voltagepark.com/article/deploy-gp-us", "quote": "Our Virtual Machines (VM) instances are only available in 100 Gbps Ethernet.", "checked": "November 13, 2026" } } ] }, { "id": "layer0c", "label": "Layer 0C", "shortName": "Storage", "title": "Storage Substrate", "purpose": "Arrays, storage servers and software, front-end protocols: where the data sits", "status": "moderate", "statusLabel": "VAST-Backed NFS Volumes and a Cloud-Backed Drive, With No AI-Throughput Path Documented", "authority": { "decides": "vendor", "visible": true, "overridable": false, "boundary": "vendor", "direction": "Ceded" }, "nvidia": [], "gap": "Lightning sells storage on both halves of the company. Voltage Park's On-Demand Storage rents shared volumes backed by VAST Data's Universal Storage, which the help center calls all-flash and purpose-built for GPU workloads; a volume attaches to many bare-metal instances, mounts over NFS, and appears as a persistent volume claim on Managed Kubernetes clusters. Each bare-metal node also carries five unmounted 3.2 TB NVMe drives the enterprise can pool with LVM. On lightning.ai, the teamspace Drive gives every Studio, Job and Deployment one shared filesystem, backed by the cloud provider underneath; high-performance volumes on Lightning Cloud ($0.30 per GiB per month) mount read/write as fixed-size network volumes; and cloud-agnostic folders on Lightning Storage reach across clouds without egress fees, with LitData to stream object storage fast. The buyer gets storage that follows the workload across clouds.\n\nThe architect's concern is the throughput claim. Voltage Park's mount instructions are NFSv3 over TCP with nconnect and VAST's multipath option, not NFS over RDMA or GPUDirect Storage, and neither Voltage Park nor Lightning publishes a throughput figure for this service. VAST's own GPUDirect results measure configurations Voltage Park doesn't document. Lightning's own fast tiers are priced against EFS and Filestore, and its data overview offers Lustre and 'more performant' cluster filesystems only on request.\n\nCalibration: 0C strong asks for GA storage built for accelerator throughput with a documented GPU-direct path or published AI-scale performance. CoreWeave reads strong on a public CAIOS benchmark, Nebius on documented shared-filesystem bandwidth, and Supermicro and Cisco on the VAST AI OS sold as a licensed product with VAST's own credentials. A storage service on VAST, documented only as NFS volumes, reads moderate on the general-NAS line: real, deployable, and short of the documentation the strongs carry.", "borrowedJudgment": "Ceded where the grade rests. VAST's engine places and protects the bytes in a Voltage Park volume, and Lightning's backends do the same for the Drive and its volumes; the enterprise picks a size, resizes and deletes, which is configuration, and no per-file placement handle is documented: vendor decides, visible, not overridable, Ceded. Portability: Voltage Park's volumes sit behind NFS, a multi-vendor door a workload leaves without a rebuild, Delegated; Lightning's Drive and Lightning Storage are a proprietary filesystem reached through Lightning's runtimes, CLI and SDK, Ceded; node-local NVMe goes with the node, Retained.", "notes": "Named, not scored (checked November 13, 2026): Lustre and 'more performant, enterprise-grade data solutions' for clusters, contact-sales only in the data overview; the GPU Cloud > Clusters > Storage page, which redirects to the Cluster overview ('Documentation for Cluster is on its way.') and has no archived snapshot, so any throughput figure it once carried can't be verified. Connected S3, GCS and R2 buckets, and EFS and Filestore folders in the enterprise's own accounts, are the enterprise's storage on those clouds' rows. Studio disks are environment storage inside the Studio runtime, read at Layer 2. VAST's GPUDirect benchmarks (vastdata.com) are VAST's; they'd attach here if Voltage Park documented NFS over RDMA or GPUDirect on its service.", "components": [ { "id": "lightning-ai.0c.voltage-park-on-demand-storage", "component": "Voltage Park On-Demand Storage (VAST-Backed Shared Volumes; NFSv3 With nconnect and VAST Multipath; Attach to Many Bare-Metal Instances; Persistent Volume Claims on Managed Kubernetes)", "detail": "A managed shared-storage service on Lightning's own fleet, reached through NFS. Delegated, the CoreWeave Dedicated VAST and Distributed File Storage reading: a proprietary engine behind a multi-vendor interface the workload can leave.", "dapm": "Delegated", "ga": { "basis": "offered", "url": "https://support.voltagepark.com/article/deploy-storage", "quote": "Voltage Park's On-Demand Storage solution is a high-performance, scalable, and cost-efficient storage layer purpose-built for GPU-accelerated workloads.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.0c.lightning-teamspace-drive-and-lightning-storage", "component": "Lightning Teamspace Drive and Lightning Storage (Proprietary Distributed Filesystem Shared by Every Studio, Job and Deployment; Cloud-Agnostic Folders; High-Performance Volumes on Lightning Cloud as Network Mounts)", "detail": "Lightning's own storage, which its FAQ calls 'a proprietary distributed filesystem.' Workloads see it as files inside Lightning's runtimes; outside them it's reached through lit:// URLs, the lightning CLI and the SDK, and the high-performance volumes are documented only as 'network mounts,' with no protocol named. Copying files out is data portability, not an exit: the teamspace layout, access scoping and lit:// references don't run anywhere else. Ceded, with nothing multi-vendor documented at the door.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/data/drive", "quote": "All Studios in a teamspace share the same drive.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.0c.node-local-nvme-on-voltage-park-bare-metal", "component": "Node-Local NVMe on Voltage Park Bare Metal (Five 3.2 TB Data Drives per Node, Pooled by the Enterprise With LVM)", "detail": "Drives inside the rented node, configured by the enterprise. Retained, the CoreWeave local-storage reading.", "dapm": "Retained", "ga": { "basis": "offered", "url": "https://support.voltagepark.com/article/local-storage-configuring-an-lvm-volume-on-voltage-park-on-demand-bare-metal-servers", "quote": "By default, Voltage Park On-Demand bare metal servers have one NVMe SSD mounted at the root partition.", "checked": "November 13, 2026" } } ] }, { "id": "layer1a", "label": "Layer 1A", "shortName": "Data", "title": "Data Foundation & Governance", "purpose": "The data itself, organized and governed: the Governance Catalog that Layer 2C queries", "status": "moderate", "statusLabel": "Teamspace-Scoped Data Access and a Model Registry; No Data Catalog", "authority": { "decides": "code", "visible": true, "overridable": true, "boundary": "vendor", "direction": "Retained" }, "nvidia": [], "gap": "Lightning's 1A function is access governance, organized around the teamspace. Each teamspace has one isolated Drive, and connected S3, Google Cloud Storage and R2 buckets are scoped to the teamspace that added them, with write access a per-connection switch on S3. Organizations carry Admin, Member and Guest roles, with custom roles on the Enterprise tier, and SAML 2.0 or OIDC single sign-on. The model registry stores models and versions under ORG/TEAMSPACE/model:version, shared by model or by version, on Lightning's cloud or the enterprise's. The buyer gets a clean boundary at the teamspace. Inside it, access is broad: every role, Guest included, can add data and use what the teamspace shares.\n\nThe architect's concern is that nothing answers what the data is. A dataset on Lightning is a Studio renamed to hold one, or a folder in the Drive; versioning is copying the folder, with the web-app version marked coming soon. No catalog carries classification, sensitivity or lineage at the object or field level, and while the Enterprise tier lists 'Audit logs,' the documentation doesn't say whether the enterprise can query them or what they cover. The stores themselves hold the bytes and read at 0C.\n\nCalibration: governance over data without a catalog that says what the data is reads moderate. Nebius reads moderate on access permits and bucket policies, and Hugging Face moderate on an artifact registry with resource groups and gating; CoreWeave reads strong only on the VAST Catalog in its Dedicated tier.", "borrowedJudgment": "Lightning enforces the teamspace membership, roles and connection permissions the enterprise set, as written: code decides, visible, overridable, Retained. Portability: the teamspace and role model and the registry's naming and sharing are Lightning's, Ceded; the LitModels client is Apache 2.0, but it opens onto one registry; SAML and OIDC federation is a multi-vendor standard, Delegated.", "notes": "Not scored (checked November 13, 2026): dataset versioning in the web app ('coming soon'); transfer of a removed member's work ('coming soon'). 'Audit logs' appear in the Enterprise column of lightning.ai/pricing; the security audit-logs page describes logging of user activity and configuration changes under SOC 2, with no documentation of customer access or queryability. Connected buckets are the enterprise's data on those clouds' rows; Lightning scores the teamspace gate on them, not the buckets. Voltage Park's console adds organization accounts and SSH keys and no data governance.", "components": [ { "id": "lightning-ai.1a.teamspaces-roles-and-data-access", "component": "Teamspaces, Roles and Data Access (One Isolated Drive per Teamspace; Teamspace-Scoped S3, GCS and R2 Connections, Per-Connection Write Access on S3; Admin, Member and Guest Roles; Custom Roles on Enterprise)", "detail": "Lightning's account and access model over the data on its paper. Ceded, the Nebius identity-and-access reading: the teamspace and role opinions don't run anywhere else.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/data/drive", "quote": "Each teamspace has a single isolated drive. Use this to control access to data.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.1a.single-sign-on", "component": "Single Sign-On (SAML 2.0 and OIDC Federation With the Enterprise's Identity Provider)", "detail": "Federation the enterprise configures at its own identity provider and at Lightning. Delegated, the Hugging Face SSO reading: a multi-vendor standard the same identity provider serves to any other platform.", "dapm": "Delegated", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/security/security-features/configure-sso", "quote": "Lightning AI supports SSO integrations via SAML v2 and OIDC.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.1a.model-registry", "component": "Model Registry (LitModels; ORG/TEAMSPACE/model:version; Share a Model or One Version; Host on Lightning's Cloud or the Enterprise's)", "detail": "Versioned model storage and sharing through the browser, CLI and the Apache 2.0 LitModels client. The weights download as files, but the registry's naming, versions and sharing live on one service: a standard-shaped client with nothing substitutable behind it, Ceded.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/inference/custom-models/model-registry", "quote": "Upload models from your browser, command line (CLI), or via code with Python scripts.", "checked": "November 13, 2026" } } ] }, { "id": "layer1b", "label": "Layer 1B", "shortName": "Retrieval", "title": "Context Management & Retrieval", "purpose": "Low-latency retrieval for RAG — vector/hybrid search, context windows", "status": "moderate", "statusLabel": "GraphN Knowledge Bases: Managed Chunking, Embedding, Vector Search and Reranking Over Text, Images and Video; Dense Search Only, Upload-Fed, Workspace-Scoped", "authority": { "decides": "vendor", "visible": true, "overridable": true, "boundary": "vendor", "direction": "Delegated" }, "nvidia": [], "gap": "Lightning's retrieval lives in GraphN, its agent-workflow platform at graphn.ai. A GraphN knowledge base is a managed document store an agent, a function or the API can search: GraphN chunks the content with configurable size and overlap, embeds it with BGE-M3 for text or Qwen3-VL for text, images and video segments, stores the vectors in Qdrant, and answers a search with ranked chunks, reranked by a cross-encoder or a vision-language reranker unless the request turns it off. Documents arrive by upload or API, with optional metadata, through durable asynchronous batch ingestion with idempotent retries; OCR and markdown conversion feed PDFs and images in. Embedding, reranking, ingestion and vector storage are each metered on the published price list, and a separate embed endpoint returns BGE-M3 vectors for the enterprise's own use. The buyer gets retrieval for an agent without standing up a vector database.\n\nEach search sets its result count, a minimum score threshold and a metadata filter, and can switch reranking off or pick the reranker, per the generated OpenAPI specification GraphN publishes as its source of truth.\n\nThe architect's concern is how far the service reaches. Search is dense-vector only, with no hybrid or keyword retrieval documented; access is scoped to the workspace, not to the document's source permissions; and content arrives by upload, with no connectors that sync a source system. An enterprise whose second requirement is hybrid search over a governed source brings another tool. On the Lightning platform itself, retrieval remains a pattern the enterprise builds; LitServe's embeddings spec stays a sub-threshold library.\n\nCalibration: Nebius reads moderate on hosted embedding and rerank endpoints beside managed Qdrant and pgvector. GraphN goes further by running the whole ingest-embed-search-rerank path as one managed service, with multimodal retrieval Nebius doesn't document. The strongs (Elastic's native hybrid, permission-aware retrieval; the hyperscalers' managed retrieval portfolios; IBM Cloud's assembled kit) carry hybrid search, filtering and source permissions GraphN doesn't.", "borrowedJudgment": "Delegated on the request. GraphN's engine decides what the knowledge base returns, through the embedding model chosen when the knowledge base was created, Qdrant's similarity ranking and the reranker; each search can filter by metadata, set a score threshold and a result count, and turn reranking off or choose the reranker, which changes the outcome of that ranking per request: vendor decides, visible, overridable, Delegated, the Nebius 1B reading. Portability: the knowledge-base API is GraphN's own, and the index and chunking choices stay with the service, Ceded; the embed endpoint serves open BGE-M3 weights the enterprise can run itself, so its vectors stay in a space it can reproduce, Delegated.", "notes": "GraphN (graphn.ai, 'GraphN by Lightning AI') runs under Lightning AI's terms with published pricing; the only maturity badge in its docs is on LoRA adapters. Ruled November 13, 2026 (Keith): LitServe's OpenAI-compatible embeddings spec alone is 'functionally a gap,' a thin slice with no clear support statement, and an open-source library scores only where the vendor manages it or provides an SLA; GraphN's knowledge bases are a managed, metered service on Lightning's paper. Source hierarchy: the task-oriented API reference page lists query, top_k and rerank; the generated OpenAPI specification (github.com/voltagepark/graphn-openapi, version 0.39.0, 'Source of truth') adds reranker_model, score_threshold and metadata_filter, and governs as the more specific surface. The specification's repository sits in the voltagepark GitHub organization: GraphN is built on the Voltage Park side of the merged company. Lightning's Model APIs catalog still documents no embeddings endpoint; the Agent object's 'knowledge' parameter is undocumented.", "components": [ { "id": "lightning-ai.1b.graphn-knowledge-bases", "component": "GraphN Knowledge Bases (Managed Chunking, Embedding and Qdrant Vector Search; BGE-M3 Text and Qwen3-VL Multimodal Embeddings Over Text, Images and Video Segments; Cross-Encoder and Vision-Language Reranking; OCR and Markdown Conversion; Asynchronous Batch Ingestion; Workspace-Scoped API, SDK, CLI and MCP)", "detail": "Lightning's managed retrieval service inside GraphN, metered per embedding, reranking, ingestion and stored vector. Ceded: the knowledge-base API is GraphN's own, and the index, chunking choices and stored vectors don't move; rebuilding elsewhere means re-ingesting the documents.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://graphn.ai/faq", "quote": "A GraphN knowledge base is a searchable document collection that agents and functions can query for workspace-specific context.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.1b.graphn-embed-endpoint", "component": "GraphN Embed Endpoint (POST /v1/{workspace}/embed; BAAI/bge-m3, MIT License, by Default, or a Knowledge Base's Embedding Model; Normalized Vectors)", "detail": "Embeddings from open weights served on GraphN's own path. Delegated, the Nebius open-weights embeddings reading: the enterprise can run the same BGE-M3 checkpoint itself or at another provider and keep its vectors in the same space; the single-vendor path costs a client change, not a re-embedding.", "dapm": "Delegated", "ga": { "basis": "offered", "url": "https://graphn.ai/docs/api-reference", "quote": "Returns vectors and dimensions.", "checked": "November 13, 2026" } } ] }, { "id": "layer1c", "label": "Layer 1C", "shortName": "Pipelines", "title": "Data Movement & Pipelines", "purpose": "Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering", "status": "moderate", "statusLabel": "Pipelines and Batch Jobs Run the Enterprise's Data Jobs as a Managed Service on a Schedule; Bucket Connectors Only, No CDC, No Lineage", "authority": { "decides": "code", "visible": true, "overridable": true, "boundary": "vendor", "direction": "Retained" }, "nvidia": [], "gap": "Lightning moves and transforms data by running the enterprise's code. Pipelines, defined in the Lightning SDK, chain Job, multi-machine Job, Deployment and Deployment Release steps, each on its own container and machine type, sharing one filesystem; they run on a cron schedule, take a new version each time the defining SDK script is re-executed, and are created only from the SDK. Lightning runs the enterprise's data prep across many machines as Jobs, scaled by plan tier, and the jobs commonly use LitData, Lightning's Apache 2.0 library for parallel map, chunked optimize, and streaming from S3, Google Cloud Storage or local disk. The library itself isn't scored: Lightning's terms govern open-source software by its own license, outside the agreement's warranty and support, so its value on this row is what Lightning runs and supports, not the bits. Movement between places is a copy between folders Lightning mounts into the Studio (the Move data guide is cp -R between Studio, S3, EFS, GCS and Filestore folders), plus a one-time Studio transfer between clouds and cloud-agnostic folders without ingress or egress fees except when read from AWS or Google Cloud, where transfer fees apply. The buyer gets a managed place to run any data job, at any scale, on a schedule, with the Enterprise tier's 99.95% uptime SLA behind it.\n\nThe architect's concern is everything a pipeline product usually carries. The documented connectors are mounted S3, Google Cloud Storage and R2 buckets; Snowflake and Databricks are named on the Data connections page without pages of their own, and custom connectors 'for any data service are available to enterprise customers' on request. There's no change data capture (CDC), no lineage, and no KV-cache tiering. Pipelines run what the enterprise wrote and nothing more; the second requirement, such as pulling from a database or tracing a column back to its source, brings another tool.\n\nCalibration: Hugging Face reads moderate on the same shape, Jobs as a hosted runner with schedules, with no connectors, CDC or lineage; Nebius moderate on object transfers; CoreWeave gap; Dataiku strong on a Flow with connectors and column-level lineage.", "borrowedJudgment": "The enterprise's code decides. A Pipeline runs the steps, order and schedule the enterprise defined, and the job code inside each step is the enterprise's; the engine chooses nothing about what to move or when beyond that: code decides, visible, overridable, Retained, the Hugging Face Jobs reading. Portability: Pipeline definitions are written against Lightning's SDK, steps and Studios, Ceded, though the scripts inside the steps, LitData calls included, are the enterprise's and move with a change of runner.", "notes": "Not scored (checked November 13, 2026): artifacts from jobs submitted through Docker ('currently supported in early access'). Ruled November 13, 2026 (Keith): for open-source software a maturity label doesn't indicate GA status (CNCF projects graduate years after whole industries run on them); an open-source capability scores on a vendor's row only where the vendor adds value beyond the bits, by managing it or standing behind it with an SLA. LitData (Apache 2.0; PyPI classifier '4 - Beta') is therefore named, not scored: Lightning's terms place open-source software outside the agreement's warranty and support sections and under its own license. The Move data guide's commands are plain cp between mounted folders. Named, not scored: Snowflake and Databricks connections (listed on the Data connections page; the Snowflake link resolves to the S3 page) and enterprise custom connectors ('Custom connectors for any data service are available to enterprise customers'; 'If you need a connector, please get in touch'), availability statements without product documentation.", "components": [ { "id": "lightning-ai.1c.lightning-pipelines-and-batch-jobs", "component": "Lightning Pipelines and Batch Jobs (SDK-Defined Steps: JobStep, MMTStep, DeploymentStep, DeploymentReleaseStep; Per-Step Container and Machine; Shared Filesystem; Cron Schedules; Automatic Versioning)", "detail": "Lightning's managed runner for the enterprise's data and training jobs, a Service under Lightning's agreement with its support terms and the Enterprise uptime SLA. Ceded, the Hugging Face Jobs reading: the pipeline definitions bind to Lightning's SDK, steps and Studios, while the code inside each step stays the enterprise's.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/inference/pipelines", "quote": "Pipelines let you connect different workload types into a single pipeline.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.1c.cross-cloud-data-placement", "component": "Cross-Cloud Data Placement (Studio Transfer Between Clouds; Cloud-Agnostic Lightning Storage Folders, No Ingress or Egress Fees Except From AWS or GCP)", "detail": "Lightning's movement of a Studio's data from one provider to another, and its storage that reaches across clouds. Ceded: the transfer is Lightning's action on Lightning's proprietary storage.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/data/switch-cloud-providers", "quote": "The most popular way to move a Studio from one cloud provider to another is to transfer it.", "checked": "November 13, 2026" } } ] }, { "id": "layer2a", "label": "Layer 2A", "shortName": "Orchestration", "title": "Infrastructure Orchestration", "purpose": "GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization", "status": "moderate", "statusLabel": "Lightning Places Studios, Jobs, Multi-Node Runs and Deployments Across Clouds, and Voltage Park Sells Managed Kubernetes; Budgets, Not Fair-Share", "authority": { "decides": "vendor", "visible": true, "overridable": false, "boundary": "vendor", "direction": "Ceded" }, "nvidia": [ { "component": "NVIDIA GPU and MPI Operators on Voltage Park Kubernetes", "detail": "Voltage Park's Managed Kubernetes installs the NVIDIA GPU Operator for drivers and runtime and includes the MPI Operator for distributed training. Lightning's own scheduler carries no NVIDIA dependency beyond the machines it places." } ], "gap": "Lightning's orchestration is a placement engine for its own runtimes. Studios, Jobs, multi-machine training (MMT) and Deployments each take a machine type and, optionally, a cloud, and Lightning finds the capacity: it collocates nodes for multi-node jobs, scans every region on a cloud account for the fastest allocation, starts work on interruptible machines and saves it on the two-minute warning, keeps pre-provisioned pools warm on the enterprise's own cloud accounts, and autoscales Deployments on CPU, GPU, concurrent requests, requests per minute or latency, down to zero after a cooldown the enterprise sets. Bring your own cloud (Enterprise tier) runs the same control plane in the enterprise's AWS or Google Cloud account through a stack that grants Lightning a role. Organizations toggle which providers teams may use, grant cloud accounts per teamspace, and set budgets by transferring credits to a teamspace; plans cap concurrent GPUs. Voltage Park sells Managed Kubernetes as an add-on to bare metal: upstream Kubernetes with the NVIDIA GPU and MPI Operators, Prometheus and Grafana GPU dashboards, the VAST CSI driver, OIDC authentication with Kubernetes RBAC, managed patching and encrypted etcd snapshots. The buyer gets GPUs found and started without a scheduler to run.\n\nThe architect's concern is what isn't there. Nothing documents queues, priorities, preemption between teams or fair-share; a teamspace budget is a spending cap, not a share of the fleet. Voltage Park's Kubernetes documents no cluster autoscaler or topology labels, and Lightning's managed Kubernetes and SLURM clusters were documented until late October 2026 and withdrawn. An enterprise with two teams contending for one reservation brings Kueue or Volcano onto the Kubernetes add-on.\n\nCalibration: CoreWeave and Nebius read strong on managed Kubernetes and Slurm together, with topology-aware scheduling and autoscaling documented. Hugging Face reads moderate on per-object lifecycle controls over its hosted products with no scheduler over a fleet. Lightning sits between: a real cross-cloud placement engine and a thin managed Kubernetes, without the scheduler layer the strongs carry.", "borrowedJudgment": "Ceded where the grade rests. By default Lightning places a Deployment on the cloud and region with the cheapest instance of the type chosen, and its engine decides which nodes a Studio, Job, MMT or Deployment gets, how multi-node runs are collocated, and when an autoscaled Deployment grows or shrinks. The enterprise can pin a cloud or region, reassign it after creation, or fix the replica count, the same choice of where and how many an enterprise makes when it picks an EC2 region; the placement and scheduling beneath that choice stay Lightning's: vendor decides, visible, not overridable, Ceded, the AWS, GCP and OCI 2A reading. Portability, read separately: Lightning's scheduler, budgets and bring-your-own-cloud control plane are its own, Ceded; Voltage Park's Kubernetes is the upstream API the enterprise's manifests already speak, Delegated, the CoreWeave CKS and Nebius reading.", "notes": "Withdrawn (checked November 13, 2026): Lightning's GPU Cloud > Clusters pages (Kubernetes, SLURM, Networking, Storage), archived October 25, 2026 and now redirecting to 'Documentation for Cluster is on its way.' The SDK's K8sCluster class documents monitoring and billing usage for clusters on Lightning; it reports, it doesn't schedule. Not scored: Voltage Park's Observability page (marketing only). Quota requests at the provider (AWS service quotas) are the cloud's, linked from Lightning's UI. Ruled November 13, 2026 (Keith): choosing a cloud, region or fixed replica count is the EC2-region choice, not an override of the placement the grade rests on; 2A authority holds Ceded, matching AWS, GCP and OCI.", "components": [ { "id": "lightning-ai.2a.lightning-workload-placement-and-scaling", "component": "Lightning Workload Placement and Scaling (Studios, Jobs, Multi-Machine Training and Deployments Across Lightning's Cloud and Marketplace Providers; Collocated Multi-Node Placement; Region Scanning; Interruptible Machines With Save on Interruption; Pre-Provisioned Pools; Deployment Autoscaling on Five Metrics With Scale to Zero)", "detail": "Lightning's scheduler for its own runtimes. Ceded: the placement and scaling opinions live in Lightning's control plane and its job, MMT and Deployment definitions, with no second implementation.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/train/multi-node-training", "quote": "When you submit an on-demand multi-node job, Lightning AI takes care of finding collocated nodes", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2a.bring-your-own-cloud-control-plane", "component": "Bring Your Own Cloud Control Plane (Enterprise Tier; Lightning-Managed Stack and Role in the Enterprise's AWS or Google Cloud Account; Per-Teamspace Access; Multi-Region)", "detail": "Lightning's orchestration running against compute the enterprise owns in its own cloud account. Ceded: the compute is the enterprise's, but the control plane and its opinions are Lightning's.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/gpu-cloud/overview", "quote": "Lightning AI lets organizations connect their cloud VPCs.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2a.organization-compute-controls", "component": "Organization Compute Controls (Provider Toggles; Cloud Accounts Granted per Teamspace; Teamspace Credit Budgets; Plan Concurrency Caps; Studio Auto-Sleep and Interruptible Defaults)", "detail": "Who may use which cloud and how much they may spend, set in Lightning's organization settings. Ceded, the Hugging Face resource-group reading.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/team-management/organizations/manage-organization-clusters", "quote": "To limit the cloud that your organization uses, toggle on/off the cloud providers you prefer.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2a.voltage-park-managed-kubernetes", "component": "Voltage Park Managed Kubernetes (Add-On to Bare Metal; Upstream Kubernetes; NVIDIA GPU and MPI Operators; Prometheus and Grafana; VAST CSI; OIDC With Kubernetes RBAC; Managed Patching and Encrypted etcd Snapshots)", "detail": "A managed control plane behind the Kubernetes API. Delegated, the CoreWeave CKS and Nebius Managed Kubernetes reading: the enterprise's manifests, Helm charts and operators lift to any conformant cluster.", "dapm": "Delegated", "ga": { "basis": "offered", "url": "https://support.voltagepark.com/article/add-on-managed-kubernetes", "quote": "Our control-plane exposes the standard Kubernetes OIDC authentication flow", "checked": "November 13, 2026" } } ] }, { "id": "layer2b", "label": "Layer 2B", "shortName": "Runtime", "title": "Application Runtime & Execution", "purpose": "Model serving, agent execution, inference APIs, distributed inference", "status": "strong", "statusLabel": "Managed Deployments Serve Any Container, Official vLLM Included; GraphN Adds OpenAI-Compatible Open-Model Inference With Tool Calling and a Workflow Engine for Agents, Functions and MCP Tools; Sandboxes for Agent Code", "authority": { "decides": "model", "visible": true, "overridable": true, "boundary": "model", "direction": "Delegated" }, "nvidia": [ { "component": "Open Engines on NVIDIA GPUs, No NIM in the Path", "detail": "The documented serving path is the enterprise's own container or the official vllm/vllm-openai image on NVIDIA machines, with tensor parallelism across eight GPUs in the example; no NIM, TensorRT-LLM or Triton surface appears in Lightning's documentation. The dependency is the GPU, not NVIDIA's serving software." } ], "gap": "Lightning's runtime is a managed home for the enterprise's own serving stack. Deployments turn any container into an autoscaling API: the documented self-serve path runs the official vllm/vllm-openai image with the enterprise's own vllm serve command (the example sets --tool-call-parser, --enable-auto-tool-choice and --reasoning-parser), on a chosen machine and cloud, with replicas from zero to N, autoscaling on five metrics (only max concurrent requests scales straight from zero to N; the others step through one replica first), multi-node inference with the distributed environment set for it, token or basic authentication, custom domains, websockets and release strategies, and weights pulled from the model registry. LitServe servers and custom engines deploy the same way, and turn-key models from the AI Hub deploy onto a chosen cloud account in one click. Model APIs sit beside them: pay-per-token access to a catalog that includes OpenAI's and Anthropic's models and open models such as DeepSeek V4.1 Flash and GLM 5.3, through an OpenAI-compatible endpoint and an Anthropic-compatible base URL that Claude Code and OpenCode point at directly. GraphN, Lightning AI's agent-workflow platform at graphn.ai, adds a second managed runtime: an OpenAI-compatible inference endpoint (model.graphn.ai/v1) over built-in open models (DeepSeek V4.1 Flash, gpt-oss 120B, Kimi K2.5, Qwen, Mistral Large 3, GLM 5.3 and others) where tool calling works with every built-in chat model and structured outputs and reasoning controls are documented per model, plus custom weights run on GraphN-managed GPUs and billed per GPU-second, and imported endpoints the enterprise operates; and a workflow engine that runs the enterprise's YAML workflows of agent, function, MCP-tool, connector, conditional, loop, parallel, handoff-router and judge-loop steps, publishes each version as an API and an MCP endpoint, with synchronous, asynchronous event, streaming (SSE and WebSocket) and batch modes available according to the configured factory, and returns traces with spans, timing and token usage. GraphN functions are the enterprise's Python behind a one-line decorator, run in isolated Firecracker microVMs. Sandboxes give agent-generated code an isolated Linux environment with Node or Python runtimes, snapshots and persistence, from the Python or JavaScript SDK or the CLI. LitLogger, the experiment tracker built into the platform, records metrics, files, media and models per run and logs inference from LitServe endpoints. The buyer gets the enterprise's own agent loop served at production scale on whichever open model it picks.\n\nThe architect's concern is which door carries the loop. Deployments and GraphN's inference endpoint both do: on a Deployment, tool calling, structured output and reasoning controls are whatever the enterprise's vLLM configuration supports and the latency is the machine it chose; on GraphN they're documented per model and the latency is a property of Lightning's managed service, to be measured against the workload. Lightning's own Model APIs are thinner: the docs document no tool calling, structured outputs or reasoning controls, rate limits stop at 300 requests a minute on Enterprise, and the catalog itself renders only in the console. The capture is decoupled, the CoreWeave Forge shape: the container, the endpoint and the weights leave, while the experiment history in LitLogger and the deployment settings stay. GraphN's workflows, agents and published versions live in its DSL and control plane, a second captive layer beside LitLogger's history. The SDK's Agent object, an assistant endpoint backed by a model, a prompt template and 'knowledge,' is documented only in the API reference and is a component, not the price of the grade.\n\nCalibration: CoreWeave reads strong on open serving (vLLM on CKS) plus an OpenAI-compatible catalog with tool calling; Nebius and Cerebras strong on OpenAI-compatible open-model serving with tool use and structured outputs documented. Hugging Face reads moderate on maturity (an engine in maintenance, experimental agent libraries, tool support that varies by provider). Lightning's Deployments carry the loop the way CoreWeave's CKS does, and GraphN's endpoint the way Nebius's and Cerebras's do.", "borrowedJudgment": "Delegated at the model boundary. The model decides whether and how to call a tool, but the loop runs in code the enterprise controls: its own client against an OpenAI-compatible endpoint (a vLLM Deployment, GraphN's inference API or Model APIs), or its own GraphN workflow, where conditional guards, functions and judge loops it wrote can stand before any effect: model decides, visible, overridable, Delegated, the CoreWeave, Nebius and Cerebras reading. Portability: a Deployment's container and GraphN's inference endpoint sit behind the interface other platforms implement, Delegated, as do Model APIs; the deployment settings, GraphN's workflow DSL and control plane, LitLogger's experiment history, Sandboxes and the Agent object are Lightning's own surfaces, Ceded; the enterprise's tools, functions and MCP servers are its own, Retained.", "notes": "Not scored (checked November 13, 2026): lightning code ('Closed beta: lightning code is in closed beta'), coding agents on code.lightning.ai; Docker job artifacts in early access; 'Let us deploy for you' managed model deployment for enterprise customers (contact sales, no documentation of the service). LitServe as a library is named, not scored, under the November 13 open-source ruling; it scores here only as one of the servers Deployments run. Model APIs' catalog at platform.lightning.ai/lightning-ai/model-apis/models renders client-side; the models named here are the ones the docs name. The AI Hub is a catalogue of deployable models and apps; its one-click deploy is a Deployment, and a catalogue earns no grade on its own. Clouds-and-regions mentions NVIDIA DGX Cloud as a provider Lightning is 'working with'; not scored. Named November 13, 2026: Lightning's 'Own Your AI' positioning (lightning.ai/own-your-ai, client-rendered, unreadable; the September 2026 recap frames it as 'smaller, specialized models trained on your data, running on infrastructure you own'); no public documentation of a hosted fine-tuning or post-training service. Fine-tuning runs as the enterprise's code on Studios, Jobs and MMT. GraphN (graphn.ai) runs under Lightning AI's terms with published pricing; the only maturity badge in its docs is on LoRA adapters. Named, not scored: GraphN 'inference guardrails' that 'can inspect chat-completion traffic' (one line on the platform page, no configuration documentation), read at 2C. Not scored: GraphN LoRA adapters, up to eight per base model ('LoRA adapters are in public beta', graphn.ai/docs/lora-adapters; no GA date).", "components": [ { "id": "lightning-ai.2b.lightning-deployments-the-served-workload", "component": "Lightning Deployments: the Served Workload (Any Container; Official vLLM Image With the Enterprise's vllm serve Command; LitServe and Custom Engines; Multi-Node Inference; the OpenAI-Compatible Endpoint the Container Exposes)", "detail": "The enterprise's serving stack run as a managed service under Lightning's agreement and the Enterprise uptime SLA. Delegated, the CoreWeave CKS open-serving and Hugging Face Inference Endpoints reading: the container and the endpoint it exposes run on any other platform. The deployment lifecycle around it reads Ceded, on its own chip here and in the placement-and-scaling chip at 2A.", "dapm": "Delegated", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/inference/overview", "quote": "Lightning AI lets you bring your own container and configure every detail of auto-scaling", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.lightning-deployment-control-surface", "component": "Lightning Deployment Control Surface (Autoscaling Configuration on Five Metrics; Token and Basic Auth; Custom Domains; Release Strategies; Registry-Sourced Weights; Health Checks)", "detail": "The settings that wrap a served container on Lightning, configured through its console and SDK. Ceded: no other provider consumes them, so moving the workload means rebuilding them, the Hugging Face Inference Providers control-surface and Azure Foundry platform-surface reading.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/inference/lightning-inference/autoscaling", "quote": "By default, a deployment adds a new replica when CPU usage exceeds 95% for over 60 seconds.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.model-apis", "component": "Model APIs (Pay-per-Token Catalog Including OpenAI, Anthropic, DeepSeek and GLM Models; OpenAI-Compatible Endpoint and Anthropic-Compatible Base URL; Plan Rate Limits to 300 Requests per Minute)", "detail": "A hosted model service consumed through the OpenAI and Anthropic interface standards. Delegated, the inference-interface reading: the enterprise's client points at another provider without a rebuild.", "dapm": "Delegated", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/inference/model-apis", "quote": "Add AI to your apps instantly with model APIs.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.graphn-managed-inference", "component": "GraphN Managed Inference (OpenAI-Compatible model.graphn.ai/v1; Built-In Open Models With Tool Calling, Structured Outputs and Reasoning Controls; Custom Weights on GraphN-Managed GPUs per GPU-Second; Imported Endpoints)", "detail": "Open-weight models served by Lightning behind the OpenAI interface standard. Delegated, the inference-interface reading: the client points at another provider without a rebuild.", "dapm": "Delegated", "ga": { "basis": "offered", "url": "https://graphn.ai/docs/inference-quickstart", "quote": "tool calling work with every built-in chat model", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.graphn-workflow-engine", "component": "GraphN Workflow Engine (YAML DSL: Agent, Function, MCP-Tool, Connector, Conditional, Loop, Parallel, Handoff-Router and Judge-Loop Steps; Versioned Publishing With Resource Snapshots; Sync, Async, Streaming, Batch and MCP Endpoints; Execution Traces)", "detail": "Lightning's runtime for the enterprise's agent workflows. Ceded: the DSL, the agent and resource definitions, and the published versions have no second implementation, the Azure Foundry platform-surface reading.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://graphn.ai/faq", "quote": "A GraphN workflow is an executable graph with defined inputs, ordered or branching steps, and an output.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.customer-created-tools-functions-and-mcp-servers", "component": "Customer-Created Tools, Functions and MCP Servers (Tool Calls Executed in the Enterprise's Code Against Its Deployment; Python Functions Behind GraphN's @function Decorator in Firecracker microVMs; the Enterprise's Own MCP Servers)", "detail": "Logic the enterprise writes. Retained, the customer-tools and thin-wrapper readings: a function registered by a one-line decorator moves with its imports, and MCP is a multi-vendor protocol. Functions that call GraphN's chat, kb or storage modules bind to GraphN and read with the workflow engine.", "dapm": "Retained", "ga": { "basis": "offered", "url": "https://graphn.ai/faq", "quote": "GraphN agents can call tools exposed by Model Context Protocol servers.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.lightning-sandbox", "component": "Lightning Sandbox (Isolated Linux Environments for Agent-Generated Code; node24, node22, python313 or Custom OCI Images; Snapshots and Persistence; Python SDK, JavaScript SDK and CLI; Teamspace-Scoped Keys)", "detail": "Lightning's execution environment for code an agent writes. Ceded, the CoreWeave Sandboxes reading: the sandbox API, snapshots and lifecycle are Lightning's own.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/build/sandbox", "quote": "Lightning AI Sandbox gives every workload its own isolated Linux environment", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.lightning-agents", "component": "Lightning Agents (SDK Agent Object: an Assistant Endpoint Backed by a Model, Base URL, Prompt Template, Prompt Suggestions and Knowledge; Publish Status)", "detail": "A prompt-configured assistant endpoint documented in the SDK API reference. Ceded: the configuration lives in Lightning's agent object with no second implementation.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/sdk/stable/api/agent", "quote": "Agents expose an AI assistant endpoint backed by a custom model and prompt configuration.", "checked": "November 13, 2026" } }, { "id": "lightning-ai.2b.litlogger-experiment-tracking", "component": "LitLogger Experiment Tracking (the Experiment Tracker Built Into Lightning AI; Metrics, Files, Media and Models Logged per Experiment; Resume and Retrieval by Name; PyTorch Lightning and Fabric Integration; Inference Logging From LitServe Endpoints)", "detail": "Lightning's hosted record of the enterprise's training and inference runs; the client is Apache 2.0, but the history it writes lives in Lightning's service. Ceded, the CoreWeave W&B Models reading: the weights leave, and the experiment history that compounds doesn't. The model registry the history references is scored as a facet at 1A and inside the Deployment control surface here.", "dapm": "Ceded", "ga": { "basis": "offered", "url": "https://lightning.ai/docs/platform/train/experiment-management", "quote": "LitLogger is the experiment tracker built into Lightning AI.", "checked": "November 13, 2026" } } ] }, { "id": "layer2c", "label": "Layer 2C", "shortName": "Reasoning", "title": "Agentic Infrastructure — The Reasoning Plane", "purpose": "Policy-driven placement and resource coordination — the Autonomy Layer", "status": "gap", "statusLabel": "Agent Governance Lives Inside GraphN Workflows (Approved Tools, Handoff Routers, Judge Loops, Workspace Keys, Traces); a Sandbox Egress Firewall Fences One Environment; No Plane Over the Enterprise's Agents", "authority": { "decides": "absent", "visible": null, "overridable": null, "boundary": "vendor", "direction": "Absent" }, "nvidia": [], "gap": "Lightning governs agents inside GraphN, its agent-workflow platform, and not beyond it. Within a GraphN workflow, each agent calls only the functions and MCP tools configured for it, a handoff router hands work between specialist agents, a judge loop has an evaluator agent score a generator's output against a pass condition, conditional guards the enterprise writes decide which steps run, workspace-scoped keys and secrets bound what the workflow can reach, and every execution returns a trace of spans, timing, token usage and node results. On the Lightning platform, teamspaces, roles, single sign-on and scoped keys decide who may start a Studio, Job, Deployment or sandbox, and the Sandbox egress firewall, set at creation as allow-all, deny-all or a CIDR allowlist, drops and logs any connection its policy denies. Outside workflows, GraphN's inference endpoint serves the OpenAI and Anthropic APIs that Claude Code, Codex, Cursor and OpenCode already speak, and every request lands in the workspace's logs with its model, key, tokens, cache hits and cost. The buyer gets concrete controls on the agents GraphN runs and a metered record of the model calls that pass through it.\n\nThe architect's concern is that there's still no plane. The workflow controls are scoped to the application that runs the agent. The inference endpoint reaches agents GraphN doesn't run, but what it documents is a workspace key and a usage log, not policy: no role-based access beyond the key, no documented blocking rules, and an audit of calls rather than of agent actions. Nothing documents an agent identity distinct from a workspace key, a gateway that applies the enterprise's policy to its agents' model and tool calls, a registry of agents across runtimes, or orchestration across platforms. A GraphN line says configured inference guardrails can inspect chat-completion traffic, with no documentation of what they check or how they're configured. There's no documented per-request inference placement: GraphN's home page advertises an 'auto' model ('each request goes to a model that can handle it'), but the models reference and inference quickstart don't document it, and Deployments' cheapest-cloud default within editable restrictions is 2A placement. Routing isn't reasoning, a judge loop's evaluator is a model scoring a model, and a firewall answers whether a connection is legal, never whether the outcome was right; the deterministic outcome validator is absent here as it is almost everywhere.\n\nCalibration: legs scoped inside one application are the runtime's own governance and read at 2B (the September 15, 2026 addendum): Cloudera's Agent Studio, Confluent's Streaming Agents and WEKA's AIDP all hold gap on that line, and GraphN is the same shape. The moderate floor is a GA gateway every agent and model call passes through with role-based access and full audit (the Nutanix line); CoreWeave, Nebius, Hugging Face, Cerebras and Groq hold gap with account controls and no plane.", "borrowedJudgment": "Nothing offered as a plane. GraphN's tool approvals, handoff routing, judge loops and traces are scored at 2B, where they're the workflow engine's own governance and strengthen the override reading; the enterprise that wants agent identity, a request gateway, a registry or orchestration over agents beyond GraphN brings its own: Absent, the Cloudera, Confluent and WEKA reading.", "notes": "Sub-threshold signals (checked November 13, 2026): GraphN 'inference guardrails' ('Configured guardrails can inspect chat-completion traffic before model requests proceed', graphn.ai/ai-workflow-platform; no configuration documentation); GraphN workspace scoping of workflows, agents, tools, models, data, secrets and keys; the Sandbox NetworkPolicy egress firewall (create-time only, IP-based, not carried by snapshot restore, so a restored sandbox defaults to allow-all unless the policy is passed again); teamspace-scoped sandbox API keys; Model APIs plan rate limits. Named, not scored: 'Audit logs' appears in the Enterprise column of lightning.ai/pricing, and the security audit-logs page describes logging of user activity and configuration changes under SOC 2; neither documents customer access, queryability, or coverage of agent actions. lightning code is in closed beta. Named, not scored: GraphN 'auto' model routing ('Let the prompt pick the model. Send auto and each request goes to a model that can handle it', graphn.ai home page; absent from the models reference and quickstart; routing, not reasoning, even if documented). GraphN inference request logs (model, key, tokens, cache hits, cost per request, graphn.ai/docs/coding-agents) are usage records, scored with the inference endpoint at 2B.", "components": [] }, { "id": "layer3", "label": "Layer 3 (+1)", "shortName": "Applications", "title": "AI Application Layer — The Value Plane", "purpose": "AI-powered business capabilities — business logic, workflow automation", "status": "gap", "statusLabel": "No Packaged Business Application: GraphN's 72 Blueprints Are Lightning-Authored Starting Points to Customize, the AI Hub Is a Catalogue of Deployable Models and APIs, and the Coding Agent Is a Closed-Beta Developer Tool", "authority": { "decides": "absent", "visible": null, "overridable": null, "boundary": "vendor", "direction": "Absent" }, "nvidia": [], "gap": "Lightning's closest thing to a value plane is GraphN's blueprint catalog. Seventy-two Lightning-authored workflow packages span thirteen categories, from a Contract Risk Reviewer, an Invoice Data Extractor & Matcher, a Resume Screener and a Lead Scorer to Production Incident Response and a Clinical Note Summarizer; each bundles the workflow graph with its agents, functions, MCP servers, sample inputs and output contracts, and deploys into the enterprise's GraphN workspace in a click. GraphN frames them as starting points: 'Pick one, customize it, and deploy,' and 'deploy a working baseline and then customize it.' On the Lightning platform, the AI Hub presents itself as a place to 'deploy models or containers as autoscaling APIs' and to 'build and share AI APIs,' Studios publish as templates to an organization or a public gallery, and the coding agent at code.lightning.ai is a closed-beta developer tool. The buyer gets a fast start on dozens of business tasks.\n\nThe architect's concern is who finishes them. The flagship Support Router runs on 'mock data bundled with the blueprint'; no blueprint documents a connector to a system of record (an ERP, an HRIS, a CRM, a ticketing system), so each becomes a business capability only after the enterprise wires it to its own data and tools and owns the result. What then runs is the enterprise's adapted workflow on GraphN's engine, scored at 2B. The value plane credits a business application the vendor delivers and keeps delivering: an agent that does the work counts, but a builder's starting point the enterprise copies, wires and finishes doesn't, however capable the agents inside it. The blueprints sit with Dataiku's Solutions ('templated projects with sample data'), NVIDIA's Blueprints and Kamiwaza's 'deployment patterns and templates, not packaged applications,' all named and not scored.\n\nCalibration: IBM reads strong and IBM Cloud moderate on watsonx Orchestrate's prebuilt domain agents, which do sales, procurement and HR work out of the box against SAP SuccessFactors, Workday and Salesforce, the line a packaged application crosses. CoreWeave reads moderate on Monolith's first-party engineering-test modules. Nebius, Hugging Face, Cerebras and Groq read gap on catalogues, playgrounds and templates.", "borrowedJudgment": "Nothing offered at the value plane. A deployed blueprint becomes the enterprise's workflow definition, adapted and owned by the enterprise and run on GraphN's engine, where its authority and portability are read at 2B; apps and APIs deployed from the AI Hub are their authors', with nothing documenting procurement and support through Lightning's agreement: Absent.", "notes": "Ruled November 13, 2026 (Keith): gap. The question was whether GraphN's 72 Lightning-authored blueprints are first-party business applications (the case for credit: authored, deployable, metered and run on Lightning's paper, broader than CoreWeave's Monolith modules) or agent-builder starting points (the Dataiku Solutions and NVIDIA Blueprints reading; the case: 'Pick one, customize it, and deploy,' mock data in the flagship, no system-of-record connectors). Starting points: 'Use this blueprint' opens a template editor and saving creates the enterprise's own copy, the flagship runs on mock data, and the capability is GraphN's engine and general models, credited at 2B. CoreWeave's Monolith modules differ in kind (a maintained first-party product doing the business function on real data with trained domain models), not in count. Not scored: lightning code ('Closed beta'); 'Let us deploy for you' (contact sales). The AI Hub renders client-side; its page metadata describes deploying models and containers as APIs, and the platform overview's older 'prebuilt enterprise apps' language is set aside for the live surface.", "components": [] } ] }