CoreWeave vs Vast.ai

Ide

1. developerOpen Jupyter or connect my IDE (VS Code/Cursor) to the instance in one step

   weight 1 · round to Vast.ai
   CoreWeave—none0/10

   Evidence shows CoreWeave provisions bare-metal Kubernetes clusters, kubeconfig access, and Terraform/API management, but there is no mention of Jupyter notebooks, VS Code/Cursor remote-connect integration, or any one-step IDE/notebook connection workflow; developers would need to manually deploy and configure such tooling themselves via generic Kubernetes primitives.

   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”
   • [claimed-docs] “CoreWeave Kubernetes Service (CKS) offers a managed Kubernetes service that lets you run clusters on bare metal servers in CoreWeave Cloud.”
   • [claimed-docs] “CKS runs Kubernetes directly on bare metal Nodes, without a hypervisor. Customer clusters don't run Virtual Machines.”

   proof ↗⚑ flag

   Vast.ai~partialclaimed5/10

   Vast.ai documents secure SSH access to instances (key-based auth, run commands, transfer files) and one-click template launches that could include Jupyter environments, which together enable IDE remote-connection workflows (e.g., VS Code Remote-SSH). However, the evidence never explicitly names Jupyter notebook access or a documented one-step VS Code/Cursor connection flow. Missing for 10: explicit Jupyter launch/access docs, explicit VS Code/Cursor remote-connect guide or extension support, and any hands-on confirmation of a true 'one step' connect experience.

   • [claimed-docs] “Vast.ai instances are configured to accept keys only - Password authentication is disabled for improved security.”
   • [claimed-docs] “Log in securely, Run commands remotely, Transfer files without exposing your data”
   • [claimed-docs] “SSH (Secure Shell) is a protocol for safely connecting to remote servers. It encrypts your connection so you can: Log in securely, Run comma…”
   • [claimed-docs] “Launch prebuilt or custom templates with one click.”
   • [claimed-docs] “A template is how Vast helps you launch an instance, setting up your rented machine with whatever software and formatting you need.”
   • [claimed-docs] “Build a template - Package your environment so any GPU can run it with one click.”

   proof ↗⚑ flag

Networking

1. developerExpose ports to serve applications from my instance and connect instances over private networking

   weight 1 · round to Vast.ai
   CoreWeave—none0/10

   Evidence covers Kubernetes cluster management, node pools, storage, and Terraform/API access, but there is no mention of exposing ports/ingress for serving applications or of private networking/VPC connectivity between instances. Missing for 10: documentation on ingress/port exposure for serving apps, VPC or private networking setup between instances, and any hands-on confirmation of connectivity features.

   ⚑ flag

   Vast.ai✓fullclaimed7/10

   Docs explicitly describe exposing ports (search filter `direct_port_count>=1`, SSH access) and creating private overlay networks so instances on different machines can share a virtual LAN for multi-node workloads, directly matching both halves of the story. Missing for 10: independent/hands-on confirmation of port-forwarding for arbitrary web apps beyond SSH, and more detail on overlay network setup/limitations.

   • [claimed-docs] “Vast now supports creating *overlay* networks for instances, allowing client instances on different machines on the same physical LAN to sha…”
   • [claimed-docs] “Vast now supports creating overlay networks for instances, allowing client instances on different machines on the same physical LAN to share…”
   • [claimed-docs] “vastai search offers 'gpu_name=RTX_4090 num_gpus=1 verified=true direct_port_count>=1 rentable=true' -o 'dlperf_usd-'”
   • [claimed-docs] “Vast.ai instances are configured to accept keys only - Password authentication is disabled for improved security.”
   • [claimed-docs] “SSH (Secure Shell) is a protocol for safely connecting to remote servers. It encrypts your connection so you can: Log in securely, Run comma…”

   proof ↗⚑ flag

Ssh

1. developerSSH into my GPU instance with my own keys and get root-level control of the environment

   weight 3 · round to Vast.ai
   CoreWeave—none0/10

   CoreWeave's documented access model is Kubernetes-native (kubeconfig + API tokens via CKS) rather than traditional SSH-with-your-own-keys into a GPU instance; no docs mention SSH key injection, root shell access, or instance-level SSH at all — access is described purely in terms of kubectl/API authentication to clusters running on bare metal without VMs.

   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”
   • [claimed-docs] “CKS runs Kubernetes directly on bare metal Nodes, without a hypervisor. Customer clusters don't run Virtual Machines.”
   • [claimed-docs] “This page explains how to create, use, and manage API Access Tokens and the kubeconfig files generated alongside them, so you can authentica…”
   • [claimed-docs] “API Access Tokens authenticate users and grant access to resources such as CKS clusters and VPCs.”

   proof ↗⚑ flag

   Vast.ai~partialclaimed6/10

   Docs confirm SSH access is key-based only (password auth disabled) and lets you log in securely, run commands remotely, and transfer files, which implies developer control of the instance. However, the evidence pack never explicitly confirms root-level privileges or documents the workflow for adding a user's own SSH key. Missing for 10: explicit documentation of root/sudo access inside instances, and details on uploading/managing your own SSH keys.

   • [claimed-docs] “Vast.ai instances are configured to accept keys only - Password authentication is disabled for improved security.”
   • [claimed-docs] “Log in securely, Run commands remotely, Transfer files without exposing your data”
   • [claimed-docs] “SSH (Secure Shell) is a protocol for safely connecting to remote servers. It encrypts your connection so you can: Log in securely, Run comma…”

   proof ↗⚑ flag

Agent access

1. ai-native userPoint an agent at llms.txt or agent-oriented docs

   weight 2 · round drawn
   CoreWeave✓fullprobed9/10

   Direct probe confirms an llms.txt file is live at docs.coreweave.com/llms.txt with structured agent-oriented summary and links, and a companion MCP endpoint further confirms agent-reachable documentation. Missing for 10: independent third-party confirmation beyond the vendor's own probe/docs.

   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.coreweave.com/llms.txt # CoreWeave Docs > Learn how to deploy, manage, and observe your AI trainin…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the docs MCP endpoint https://docs.coreweave.com/mcp completed a FULL keyless JSON-RPC initialize hands…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed9/10

   Vast.ai publishes a live llms.txt at docs.vast.ai/llms.txt (confirmed via probe returning HTTP 200 with structured doc links) and additionally ships an explicit agent-oriented guide/skill ('vastai agent skill') documenting how AI coding assistants can drive the platform directly. This directly satisfies pointing an agent at llms.txt or agent-oriented docs. Missing for 10: no independent/community corroboration of agents actually consuming llms.txt in practice.

   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.vast.ai/llms.txt # Vast.ai Documentation: Affordable GPU Cloud Marketplace - [Welcome](https://doc…”
   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”

   proof ↗⚑ flag

2. ai-native userRun the product headlessly / in CI for automation

   weight 2 · round to Vast.ai
   CoreWeave✓fullprobed7/10

   CoreWeave provides a full REST API (Bearer-token gated, confirmed live via probe) for cluster/resource management, a Terraform provider for IaC, and kubeconfig/kubectl access, all of which support headless CI/automation workflows. missing for 10: no dedicated CI/CD pipeline examples (e.g., GitHub Actions integration) or first-party automation SDKs beyond generated API clients, and no independent hands-on report confirming a full CI pipeline running against CoreWeave.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”
   • [claimed-docs] “All API requests must include a CoreWeave API access token in the Authorization header as a Bearer token.”
   • [claimed-docs] “Use the CoreWeave Terraform provider in your OpenTofu configuration.”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed8/10

   Vast.ai offers a full CLI and REST API for scriptable, non-interactive control of the entire instance lifecycle (search, create, destroy), confirmed by both docs and hands-on runtime probes running keylessly. Scoped API keys are explicitly recommended for CI/CD and shared tooling, and the CLI was verified to run headlessly via uvx with no login. Missing for 10: no explicit CI pipeline example (e.g., GitHub Actions workflow) or third-party case study of running it in an automated CI pipeline.

   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “You can also create scoped keys with limited permissions using `vastai create api-key`, useful for CI/CD or shared tooling.”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”

   proof ↗⚑ flag

3. ai-native userConnect an agent via an official MCP server

   weight 3 · round to CoreWeave
   CoreWeave~partialprobed6/10

   A probe confirms CoreWeave operates a live, keyless MCP server at docs.coreweave.com/mcp that completes a JSON-RPC initialize handshake and exposes search/retrieval tools over its documentation, so an agent can officially connect via MCP. However, this MCP server only covers documentation search rather than actual platform management (clusters, node pools, inference), and there is no first-party doc page describing it or independent community verification beyond the probe. Missing for 10: a documented/announced MCP server (not just a discovered endpoint), MCP tools that let an agent actually operate CoreWeave resources (not just search docs), and independent community corroboration.

   • [probe] “PROBE runtime (recorded 2026-09-05): the docs MCP endpoint https://docs.coreweave.com/mcp completed a FULL keyless JSON-RPC initialize hands…”
   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.coreweave.com/llms.txt # CoreWeave Docs > Learn how to deploy, manage, and observe your AI trainin…”

   proof ↗⚑ flag

   Vast.ai—none0/10

   Vast.ai documents a CLI, REST API, Python SDK, and an 'agent skill' file for coding assistants, but no evidence anywhere describes an official MCP server or MCP integration. Since Vast.ai is a platform/service (not itself an agent), the MCP-server axis applies, and its absence from the evidence pack means the story is unmet.

   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”

   proof ↗⚑ flag

4. ai-native userUse an official CLI

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   CoreWeave's docs describe API access tokens, a Terraform/OpenTofu provider, generated gRPC/Connect clients, and kubectl via kubeconfig, but no evidence of a dedicated official CoreWeave CLI tool exists in the pack.

   ⚑ flag

   Vast.ai✓fullprobed9/10

   Vast.ai ships an official CLI (pypi `vastai`) with full documentation of the entire platform surface (search, instances, templates, volumes, serverless), and runtime probes confirm it installs and works keylessly for real marketplace queries. Missing for 10: independent third-party reviews specifically praising/critiquing the CLI's UX beyond vendor docs and probes.

   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “vastai search offers 'gpu_name=RTX_4090 num_gpus=1 verified=true direct_port_count>=1 rentable=true' -o 'dlperf_usd-'”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [probe] “official CLI documented at https://docs.vast.ai/cli/hello-world”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”

   proof ↗⚑ flag

5. ai-native userDrive the product through a documented public API

   weight 3 · round drawn
   CoreWeave✓fullprobed9/10

   CoreWeave documents a public REST/gRPC API for CKS clusters and Inference (create/list/update/delete), Bearer-token auth, generated clients in multiple languages, and a Terraform provider for IaC control — and a live runtime probe confirms the CKS API endpoint is reachable and Bearer-gated exactly as documented. missing for 10: independent third-party hands-on API usage reports/tutorials beyond CoreWeave's own docs.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “All API requests must include a CoreWeave API access token in the Authorization header as a Bearer token.”
   • [claimed-docs] “Clients are available for the Connect, gRPC, and Protobuf ecosystems across languages including Go, Python, TypeScript, Java, Kotlin, Rust, …”
   • [claimed-docs] “The CoreWeave Inference API provides programmatic control over inference gateways, model deployments, and capacity claims.”
   • [claimed-docs] “You can call the Inference API over several transport protocols, depending on your client tooling and performance needs”
   • [claimed-docs] “you can install a generated client instead of writing HTTP calls by hand.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed9/10

   Vast.ai publishes a documented REST API (with live OpenAPI spec) and CLI/SDK built on top of it, explicitly positioned for programmatic/agentic use, including scoped API keys for automation and a dedicated 'agent skill' for AI coding assistants. Runtime probes confirm the CLI and REST endpoints work keylessly for core operations like search/create/destroy instances, corroborating the documentation with hands-on evidence. Missing for 10: independent third-party developer reports specifically about API robustness/rate limits beyond docs and probes.

   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform.”
   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “create a scoped key with only the permissions that workload needs”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”

   proof ↗⚑ flag

6. ai-native userIssue scoped/least-privilege API credentials for an agent

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   CoreWeave documents API Access Tokens for authenticating to clusters/VPCs (coreweave-docs-8, coreweave-docs-17, coreweave-docs-26), but there is no evidence of scoped, least-privilege, or role-based permission configuration for these tokens — nothing describing granular scopes, IAM-style policies, or agent-specific credential issuance. missing for 10: documentation of configurable token scopes/permissions, role-based access control, or any mechanism to restrict a credential to a minimal set of actions for an autonomous agent.

   • [claimed-docs] “All API requests must include a CoreWeave API access token in the Authorization header as a Bearer token.”
   • [claimed-docs] “API Access Tokens authenticate users and grant access to resources such as CKS clusters and VPCs.”
   • [claimed-docs] “This page explains how to create, use, and manage API Access Tokens and the kubeconfig files generated alongside them, so you can authentica…”

   proof ↗⚑ flag

   Vast.ai✓fullclaimed8/10

   Docs explicitly describe creating scoped API keys with limited permissions for workloads/CI-CD/shared tooling via `vastai create api-key`, directly matching least-privilege agent credentialing, and this is reinforced by a dedicated api-keys reference page and the agent-skill docs describing agent-driven usage. Missing for 10: independent/hands-on verification of scope enforcement (e.g., testing that a scoped key actually blocks restricted actions) and detail on the granularity of available scopes.

   • [claimed-docs] “You can also create scoped keys with limited permissions using `vastai create api-key`, useful for CI/CD or shared tooling.”
   • [claimed-docs] “create a scoped key with only the permissions that workload needs”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”
   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”

   proof ↗⚑ flag

7. ai-native userBuild against official SDKs

   weight 2 · round to Vast.ai
   CoreWeave✓fullprobed7/10

   CoreWeave documents official generated clients for its Inference API across Connect/gRPC/Protobuf ecosystems in Go, Python, TypeScript, Java, Kotlin, Rust, and Swift, plus a first-party Terraform/OpenTofu provider and token-gated REST APIs for CKS and inference, all directly supporting AI-native programmatic/SDK access. Missing for 10: independent/hands-on developer corroboration of SDK quality, and explicit links to public SDK repos or version/release info.

   • [claimed-docs] “Clients are available for the Connect, gRPC, and Protobuf ecosystems across languages including Go, Python, TypeScript, Java, Kotlin, Rust, …”
   • [claimed-docs] “you can install a generated client instead of writing HTTP calls by hand.”
   • [claimed-docs] “You can call the Inference API over several transport protocols, depending on your client tooling and performance needs”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “Use the CoreWeave Terraform provider in your OpenTofu configuration.”
   • [claimed-docs] “The CoreWeave Inference API provides programmatic control over inference gateways, model deployments, and capacity claims.”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed9/10

   Vast.ai documents an official Python SDK (built alongside CLI) and REST API with OpenAPI spec, a documented migration path from the legacy vastai_sdk package, and runtime-probed confirmation that the CLI/SDK surface works live against the marketplace API. Documentation explicitly frames CLI/SDK as 'same operations, different syntax' with Python usage examples for serverless workflows, satisfying an AI-native builder's need for official SDK access. Missing for 10: independent (non-vendor) hands-on reports specifically validating the Python SDK's completeness/stability beyond CLI parity.

   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [claimed-docs] “Interacting with Vast Serverless is made easy through a powerful python SDK.”
   • [github] “Migrating from vastai-sdk? The old import still works: from vastai_sdk import VastAI”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”

   proof ↗⚑ flag

8. ai-native userSubscribe to events via webhooks

   weight 2 · round drawn
   CoreWeave—none0/10

   No evidence anywhere in the pack of a webhook subscription mechanism or event-driven notification system in CoreWeave's API or platform; the evidence covers cluster management, storage, inference API, and Terraform but nothing about webhooks or event subscriptions.

   ⚑ flag

   Vast.ai—none0/10

   No evidence of a webhook subscription mechanism anywhere in the docs, CLI, API reference, or probes; Vast.ai exposes REST API, CLI, and SDK for polling/imperative control but nothing describing event-driven push notifications or webhook callbacks. missing for 10: any documentation of webhook registration/endpoints, event types, or delivery/retry semantics.

   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”

   proof ↗⚑ flag

Agentic features

1. ai-native userSet up automations that run autonomously in the background

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   The axis applies to this product kind (peer products hold positive or none verdicts on this story), so lack of evidence for an applicable capability is "none", never "na". (na/none harmonized at arena bring-up — see pipeline/scripts/na-harmonize.ts.)

   ⚑ flag

   Vast.ai~partialprobed5/10

   Vast.ai exposes a CLI/SDK/REST API, scoped API keys for CI/CD, and an 'agent skill' letting AI coding assistants drive the platform (create instances, deploy endpoints, manage keys) — all usable to script automations, and Serverless endpoints let workloads run without manual GPU management. However, there is no evidence of a built-in scheduler, cron/trigger system, or persistent background job orchestration that runs autonomously without an external driver; automation depends on the user's own agent/script staying alive. Missing for 10: native scheduling/trigger mechanism, evidence of long-running autonomous background jobs, and independent confirmation of the agent-skill working unattended in production.

   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “You can also create scoped keys with limited permissions using `vastai create api-key`, useful for CI/CD or shared tooling.”
   • [claimed-docs] “create a scoped key with only the permissions that workload needs”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”
   • [claimed-docs] “Vast Serverless is an AI infrastructure platform that lets you run compute-intensive workloads without managing GPUs, paying for execution r…”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”

   proof ↗⚑ flag

2. ai-native userOperate the product with natural-language commands

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   CoreWeave's documented interfaces are structured (Cloud Console, Terraform/OpenTofu, REST/gRPC API, kubectl) with no evidence of a natural-language command interface for operating clusters, node pools, or inference deployments. The MCP endpoint only exposes documentation search/retrieval tools, not natural-language operation of the platform itself.

   • [claimed-docs] “You can create a cluster with the Cloud Console or with Terraform.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”
   • [probe] “PROBE runtime (recorded 2026-09-05): the docs MCP endpoint https://docs.coreweave.com/mcp completed a FULL keyless JSON-RPC initialize hands…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed7/10

   Vast.ai documents a dedicated 'vastai agent skill' explicitly designed so AI coding assistants can drive the platform on the user's behalf (create instances, deploy endpoints, manage keys, check balances) via natural-language-style interaction, backed by a full CLI/REST API surface. Missing for 10: independent or hands-on verification of the natural-language agent-skill flow itself (probes only confirm raw CLI/REST access, not NL interpretation), and no third-party account of an AI agent successfully using it end-to-end.

   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”

   proof ↗⚑ flag

Api quality

1. ai-native userExplore an interactive API reference with runnable examples

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   The evidence shows static API reference docs (CKS API, Inference API) and client libraries, but nothing describes an interactive, runnable API explorer (e.g., a 'try it' console or embedded runnable code samples). Absence of such evidence for an applicable capability (API reference documentation) yields none.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “The CoreWeave Inference API provides programmatic control over inference gateways, model deployments, and capacity claims.”
   • [claimed-docs] “You can call the Inference API over several transport protocols, depending on your client tooling and performance needs”
   • [claimed-docs] “you can install a generated client instead of writing HTTP calls by hand.”

   proof ↗⚑ flag

   Vast.ai~partialprobed4/10

   Vast.ai publishes a machine-readable OpenAPI spec and API-reference docs, and the CLI docs include copy-pasteable, runnable example commands that were verified to work live (search offers, create instance) without auth — showing some 'runnable example' quality. However there is no evidence of an interactive API console (e.g., embedded 'try it' Swagger-UI style playground) on the docs site itself. Missing for 10: an interactive in-browser API explorer/console, explicit 'try it now' runnable request execution within the API reference pages, and independent confirmation of such interactivity.

   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform.”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”

   proof ↗⚑ flag

2. ai-native userDownload a machine-readable API spec (OpenAPI or equivalent)

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   The evidence pack documents REST/gRPC/Protobuf APIs (CKS API, Inference API) and generated client libraries, but nowhere mentions an OpenAPI/Swagger spec or any downloadable machine-readable schema file that an AI agent could ingest directly.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “Clients are available for the Connect, gRPC, and Protobuf ecosystems across languages including Go, Python, TypeScript, Java, Kotlin, Rust, …”
   • [claimed-docs] “The CoreWeave Inference API provides programmatic control over inference gateways, model deployments, and capacity claims.”
   • [claimed-docs] “you can install a generated client instead of writing HTTP calls by hand.”
   • [claimed-docs] “You can call the Inference API over several transport protocols, depending on your client tooling and performance needs”

   proof ↗⚑ flag

   Vast.ai✓fullprobed9/10

   Vast.ai publishes a live OpenAPI spec at docs.vast.ai/api/openapi.json (verified HTTP 200 with an 'openapi' key), backed by a full REST API reference and CLI/SDK built on it. Missing for 10: no independent third-party confirmation that the spec is comprehensive/kept in sync beyond the probe check.

   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform.”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”

   proof ↗⚑ flag

3. ai-native userRely on versioned APIs with a documented deprecation policy

   weight 2 · round drawn
   CoreWeave—none0/10

   The evidence pack documents CoreWeave's various APIs (CKS API, Inference API, Terraform provider) but contains no mention of API versioning scheme or a documented deprecation policy anywhere in the docs or community sources.

   ⚑ flag

   Vast.ai—none0/10

   Evidence shows Vast.ai has a REST API, OpenAPI spec, and CLI/SDK, but there is no documentation of API versioning scheme or a deprecation policy anywhere in the pack; the only related signal is a backward-compatible import shim for the old SDK (vast-ai-gh-2), which is not a documented policy.

   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”
   • [github] “Migrating from vastai-sdk? The old import still works: from vastai_sdk import VastAI”

   proof ↗⚑ flag

1. ai-native userPerform bulk operations across many items at once

   weight 2 · round to Vast.ai
   CoreWeave~partialclaimed5/10

   CoreWeave's API and Terraform provider let you create/list/update/delete clusters, node pools, and other infra as code, which supports scripted bulk provisioning across many resources, and autoscaling lets node pools scale in bulk with demand. However, there is no explicit 'bulk operation' endpoint, batch API, or documented way to perform multi-item operations (e.g., bulk delete/update across many storage objects or inference deployments) in a single call — missing for 10: dedicated batch/bulk API endpoints, documented bulk operations for storage/inference resources, and independent evidence of large-scale bulk usage.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler...letting you scale CKS Node Pools in response to workload demand…”
   • [claimed-docs] “You can deploy multiple Node Pools within a single cluster, where each Node Pool contains any number of Nodes. This lets you run different t…”
   • [claimed-docs] “Use the CoreWeave Terraform provider in your OpenTofu configuration.”

   proof ↗⚑ flag

   Vast.ai~partialprobed6/10

   The CLI and REST API expose full programmatic control (search, create, destroy, manage keys) and are explicitly positioned for automation/scripting from shell or Python, which lets an AI-native user loop over many offers/instances (e.g., filtered searches returning many GPU offers at once). However, there is no documented native batch/bulk endpoint (e.g., one call to create or destroy N instances simultaneously) — only single-item CRUD operations that must be scripted in a loop. Missing for 10: explicit bulk/batch API endpoints or CLI flags for multi-item operations, and community/hands-on confirmation of bulk usage at scale.

   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [github] “vast.search_offers(query='gpu_name=RTX_4090 num_gpus>=4')”
   • [claimed-docs] “vastai search offers 'gpu_name=RTX_4090 num_gpus=1 verified=true direct_port_count>=1 rentable=true' -o 'dlperf_usd-'”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”

   proof ↗⚑ flag

2. ai-native userDefine rules that trigger actions automatically on events

   weight 3 · round to Vast.ai
   CoreWeave—none0/10

   CoreWeave's docs only describe fixed Kubernetes Cluster Autoscaler behavior reacting to resource demand, not a general-purpose rules/automation engine where users define custom event-trigger-action logic; no evidence of webhooks, alert-based actions, or configurable automation rules.

   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler...letting you scale CKS Node Pools in response to workload demand…”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler, letting you scale CKS Node Pools in response to workload demands…”
   • [claimed-docs] “letting you scale CKS Node Pools in response to workload demands for GPU, CPU, or memory resources”

   proof ↗⚑ flag

   Vast.ai~partialclaimed4/10

   Vast.ai documents 'autobilling', an explicit event-triggered rule (balance drops low → auto top-up from saved card), which is genuine automation-on-event. Beyond that, evidence only shows generic CLI/SDK/API scripting and agent-skill hooks for manual automation, not a broader rules/trigger engine (no webhooks, alerts, or conditional actions for instance events like preemption, cost thresholds, or health failures). Missing for 10: a general-purpose event/trigger system (webhooks, alerting, conditional rules) beyond the single billing use case, and independent confirmation of its reliability.

   • [claimed-docs] “Use the autobilling feature to have the system automatically top up your account using a saved credit card when it runs low.”
   • [claimed-docs] “Autobilling keeps your balance positive to protect your instances from interruption or deletion.”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”

   proof ↗⚑ flag

3. ai-native userSchedule recurring jobs or workflows

   weight 2 · round drawn
   CoreWeave—none0/10

   Evidence describes cluster provisioning, node pools, autoscaling, Slurm-on-Kubernetes (SUNK) for training/inference jobs, and Terraform/API management, but nothing documents recurring job scheduling, cron-style triggers, or workflow orchestration primitives for automating repeated runs.

   ⚑ flag

   Vast.ai—none0/10

   Vast.ai's evidence covers CLI/API/SDK for instance and serverless management but contains no mention of a built-in scheduler, cron-like recurring job feature, or workflow orchestration; users would have to bring their own external scheduler to hit the API/CLI repeatedly, which isn't documented as a first-party capability.

   ⚑ flag

Availability

1. ml engineerSee real-time GPU availability by type and region before I try to provision, instead of discovering stockouts by failure

   weight 2 · round to Vast.ai
   CoreWeave~partialclaimed5/10

   CoreWeave documents a 'Capacity Finder' tool that lets users compare placement availability across zones for a requested instance type and node count before creating a Spot Node Pool, plus quota visibility/error reporting in the Cloud Console — this directly supports pre-provisioning availability checks. However, evidence is limited to Spot Node Pools and doesn't clearly show real-time GPU availability by type/region across all provisioning paths (e.g., on-demand reserved instances), and there's no independent/hands-on confirmation of accuracy or granularity. missing for 10: broader coverage beyond Spot pools (on-demand/reserved capacity visibility), independent verification that Capacity Finder prevents stockouts in practice, API/programmatic access to availability data, and region-level (not just zone-level) granularity confirmation.

   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”
   • [claimed-docs] “use Capacity Finder in the Cloud Console to compare placement availability across Zones for your requested instance type and Node count”
   • [claimed-docs] “This page shows you how to view your quotas in the Cloud Console, how to read quota errors when a Node Pool exceeds its quota, and how to re…”
   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”

   proof ↗⚑ flag

   Vast.ai✓fullprobed8/10

   Vast.ai's search/offers API and CLI let engineers query live marketplace inventory (GPU type, count, verified status, price) before renting, and this was independently confirmed via a keyless live probe against both the CLI and the public REST bundles endpoint returning real-time offers with reliability data. This directly satisfies seeing availability before provisioning rather than hitting a stockout on create. missing for 10: explicit region/geolocation filter documentation and evidence of real-time refresh/alerting for capacity changes.

   • [claimed-docs] “vastai search offers 'gpu_name=RTX_4090 num_gpus=1 verified=true direct_port_count>=1 rentable=true' -o 'dlperf_usd-'”
   • [claimed-docs] “The search page is the main portal for finding good machines and creating instances on them.”
   • [github] “vast.search_offers(query='gpu_name=RTX_4090 num_gpus>=4')”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”

   proof ↗⚑ flag

Hardware

1. ml engineerChoose from current-generation datacenter GPUs (H100/H200/B200 class) as well as cheaper previous-generation options

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   The evidence pack covers CKS node pools, autoscaling, spot/on-demand pricing, and capacity finder across 'instance types' and 'Zones,' but never names specific GPU models or generations (H100/H200/B200 or older SKUs) nor confirms a menu of current- vs previous-generation GPU choices. Without any explicit mention of GPU SKU/generation selection, this applicable capacity-availability axis has no supporting evidence.

   ⚑ flag

   Vast.ai~partialprobed5/10

   Vast.ai's docs and CLI clearly support searching/filtering the marketplace by GPU model (e.g., `gpu_name=RTX_4090`) and offer tiered pricing (on-demand, reserved, interruptible) that would let an ML engineer pick pricier or cheaper GPU options, and 'Secure Cloud (Datacenter)' verification implies access to datacenter-grade hardware. However, none of the evidence explicitly confirms availability of current-generation H100/H200/B200-class GPUs alongside older cards — all concrete examples in the pack use RTX_4090 (a consumer card), not datacenter-tier flagship GPUs. Missing for 10: explicit documentation or listing showing H100/H200/B200 GPUs are actually rentable on the platform alongside cheaper previous-gen options.

   • [github] “vast.search_offers(query='gpu_name=RTX_4090 num_gpus>=4')”
   • [claimed-docs] “vastai search offers 'gpu_name=RTX_4090 num_gpus=1 verified=true direct_port_count>=1 rentable=true' -o 'dlperf_usd-'”
   • [claimed-docs] “Secure Cloud (Datacenter): These machines are verified and confirmed to be in a certified datacenter that meets our datacenter criteria.”
   • [claimed-docs] “Vast verifies that these machines are in a datacenter with a TIER 2/3 rating or ISO 27001. These offer cards have a blue label and are recom…”
   • [claimed-docs] “Reserved instances allow you to get significant discounts (up to 50%) by pre-paying for GPU time. You can convert any on-demand instance to …”
   • [claimed-docs] “Interruptible instances use a bidding system: **Lowest cost** (often 50%+ cheaper than on-demand)”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”

   proof ↗⚑ flag

Quotas

1. platform engineerSee documented quotas and instance limits and raise them through a defined process

   weight 2 · round to CoreWeave
   CoreWeave✓fullclaimed8/10

   CoreWeave docs explicitly cover viewing quotas in the Cloud Console, interpreting quota-exceeded errors, and requesting capacity increases, plus a Capacity Finder tool to check placement availability before requesting Spot Node Pools. Missing for 10: independent/hands-on corroboration of the request process turnaround or SLA, and no detail on approval workflow specifics.

   • [claimed-docs] “This page shows you how to view your quotas in the Cloud Console, how to read quota errors when a Node Pool exceeds its quota, and how to re…”
   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”
   • [claimed-docs] “use Capacity Finder in the Cloud Console to compare placement availability across Zones for your requested instance type and Node count”
   • [claimed-docs] “A Node Pool in CKS represents one or more instances that share a common configuration, such as the same labels, taints, and annotations.”

   proof ↗⚑ flag

   Vast.ai—none0/10

   The evidence pack covers billing, autobilling, reserved/interruptible instance pricing, API keys, and CLI/SDK usage, but contains no documentation of account quotas, per-user instance limits, or a defined process to request limit increases. This axis is applicable to any capacity-oriented GPU marketplace, so absence of evidence yields 'none'. Missing for 10: documented quota/limit tables, a support/ticket process for raising limits, any mention of default instance caps or how to request more capacity.

   ⚑ flag

Clusters

1. ml engineerProvision a multi-node GPU cluster with fast interconnect for distributed training without a sales cycle

   weight 3 · round to CoreWeave
   CoreWeave~partialprobed6/10

   CoreWeave documents self-service provisioning of multi-node GPU clusters via Console, Terraform, and API (docs-1,5,6,28), with Node Pools spanning many nodes, autoscaling, and Spot capacity available 'without long-term commitments or reservations' (docs-4,15,16,20,23) — all consistent with no-sales-cycle self-service. However, the evidence pack never explicitly mentions fast interconnect (e.g., InfiniBand/NVLink) specs, and quota pages imply some capacity increases require a request process (docs-12) which could reintroduce a sales-like step. missing for 10: explicit fast-interconnect/networking specs for multi-node training, and clearer confirmation that quota/capacity requests bypass sales entirely.

   • [claimed-docs] “You can create a cluster with the Cloud Console or with Terraform.”
   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”
   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “You can deploy multiple Node Pools within a single cluster, where each Node Pool contains any number of Nodes. This lets you run different t…”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler, letting you scale CKS Node Pools in response to workload demands…”
   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”
   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”
   • [claimed-docs] “This page shows you how to view your quotas in the Cloud Console, how to read quota errors when a Node Pool exceeds its quota, and how to re…”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”

   proof ↗⚑ flag

   Vast.ai!disputedcontradicted5/10

   Docs describe self-service, no-sales-cycle provisioning (CLI/API/search offers for multi-GPU nodes) and an overlay-network feature specifically for multi-node NCCL/PyTorch training, which addresses the 'no sales cycle' and 'multi-node' parts of the story. However, independent hands-on reports describe bandwidth as 'lousy' and 'all over the place compared to what's listed in the console,' directly undercutting the 'fast interconnect' claim needed for distributed training performance. Missing for 10: documented interconnect specs (e.g. InfiniBand/NVLink bandwidth guarantees), benchmarked multi-node throughput, and independent confirmation that overlay networking delivers low-latency performance at scale.

   • [claimed-docs] “Vast now supports creating *overlay* networks for instances, allowing client instances on different machines on the same physical LAN to sha…”
   • [claimed-docs] “Vast now supports creating overlay networks for instances, allowing client instances on different machines on the same physical LAN to share…”
   • [github] “vast.search_offers(query='gpu_name=RTX_4090 num_gpus>=4')”
   • [claimed-docs] “search for a GPU, rent it, wait for it to boot, connect to it, copy data, and clean up”
   • [community] “Prices for compute were competitive, but most instances had lousy bandwidth. You don't need bandwidth to run bitcoin miners, apparently. The…”
   • [community] “As a client (renting GPUs) I've found it to be pretty unreliable. I've tried renting GPUs a handful of times and hosts, bandwidth, and perfo…”

   proof ↗⚑ flag

Orchestration

1. platform engineerSchedule jobs on managed Slurm or Kubernetes instead of building my own scheduler on raw nodes

   weight 2 · round to CoreWeave
   CoreWeave✓fullclaimed8/10

   CoreWeave offers CKS as managed Kubernetes on bare metal with autoscaling node pools, and SUNK lets platform engineers run managed Slurm jobs inside that same Kubernetes cluster, directly delivering scheduling without building custom schedulers on raw nodes. missing for 10: independent/hands-on validation of Slurm-on-K8s (SUNK) at scale beyond CoreWeave's own docs, and details on job-queue features (priorities, preemption) comparable to a full HPC scheduler.

   • [claimed-docs] “Run training and inference workloads on the same cluster by running Slurm jobs inside Kubernetes.”
   • [claimed-docs] “Run training and inference workloads on the same cluster by running Slurm jobs inside Kubernetes. Share the same compute resources between w…”
   • [claimed-docs] “SUNK enables its users to do the following: Run training and inference workloads on the same cluster by running Slurm jobs inside Kubernetes…”
   • [claimed-docs] “CoreWeave Kubernetes Service (CKS) offers a managed Kubernetes service that lets you run clusters on bare metal servers in CoreWeave Cloud.”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler...letting you scale CKS Node Pools in response to workload demand…”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler, letting you scale CKS Node Pools in response to workload demands…”
   • [claimed-docs] “CKS runs Kubernetes directly on bare metal Nodes, without a hypervisor. Customer clusters don't run Virtual Machines.”

   proof ↗⚑ flag

   Vast.ai—none0/10

   Evidence shows Vast.ai provides raw GPU instance rental, templates, overlay networking for multi-node training, and CLI/API automation, but nothing about a managed Slurm or Kubernetes scheduling service — users would need to build their own scheduler on top of raw nodes. missing for 10: managed Slurm service, managed Kubernetes service, any orchestration layer beyond raw instance provisioning.

   • [claimed-docs] “Vast now supports creating *overlay* networks for instances, allowing client instances on different machines on the same physical LAN to sha…”
   • [claimed-docs] “Vast now supports creating overlay networks for instances, allowing client instances on different machines on the same physical LAN to share…”
   • [claimed-docs] “A template is how Vast helps you launch an instance, setting up your rented machine with whatever software and formatting you need.”
   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”

   proof ↗⚑ flag

1. ai-native userDo everything through the API that I can do in the UI

   weight 2 · round to Vast.ai
   CoreWeave~partialprobed6/10

   CoreWeave documents a broad API (CKS cluster CRUD, Terraform provider covering CKS/VPC/storage/inference, Inference API for gateways/deployments/capacity) confirmed live via a runtime probe (401 Bearer-gated), and notes some UI-only helpers like Capacity Finder that lack a documented API equivalent. Not all Cloud Console features (e.g., billing insights, quota views, Capacity Finder) are explicitly confirmed as API-accessible, so full UI/API parity isn't demonstrated. missing for 10: explicit API endpoints for billing insights/quota viewing, explicit API equivalent for Capacity Finder, independent third-party confirmation of full UI/API parity.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “The CoreWeave Inference API provides programmatic control over inference gateways, model deployments, and capacity claims.”
   • [claimed-docs] “This page shows you how to view your quotas in the Cloud Console, how to read quota errors when a Node Pool exceeds its quota, and how to re…”
   • [claimed-docs] “Billing insights: View billable resource usage and consumption breakdowns by resource type.”
   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”
   • [claimed-docs] “use Capacity Finder in the Cloud Console to compare placement availability across Zones for your requested instance type and Node count”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed9/10

   Vast.ai's REST API is explicitly documented as the foundation underlying both the CLI and SDK, giving full programmatic control over the entire platform (search, instance lifecycle, templates, volumes, serverless endpoints), and this is corroborated by a live OpenAPI spec and runtime probes showing the CLI/API perform the same search/rent/manage operations available in the UI, including a dedicated agent skill for AI assistants to drive the platform. Minor gap: no explicit UI-vs-API feature parity audit exists confirming literally every UI-only setting (e.g., some billing/account UI screens) has an API equivalent. Missing for 10: an explicit parity matrix or independent confirmation that 100% of UI actions (not just core lifecycle) are API-exposed.

   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform.”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”

   proof ↗⚑ flag

2. ai-native userExport all of my data in open formats and leave

   weight 3 · round drawn
   CoreWeave~partialclaimed3/10

   CoreWeave's object storage is S3-compatible (open standard) and infrastructure is managed via open Terraform/OpenTofu configs and standard kubeconfig/kubectl access, which support some data/infra portability, but there is no explicit documented 'export all your data' tool or account-exit workflow. missing for 10: a dedicated bulk data-export feature, documentation of exporting model weights/configs/billing history in open formats, and any statement about facilitating full account migration/leave.

   • [claimed-docs] “Store training data, checkpoints, and model weights with efficient S3-compatible access directly to your GPU compute resources.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “Use the CoreWeave Terraform provider in your OpenTofu configuration.”
   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”

   proof ↗⚑ flag

   Vast.ai~partialclaimed3/10

   Vast.ai instances run standard Linux with SSH access explicitly described as supporting file transfer ('Transfer files without exposing your data') and volumes are persistent, detachable storage, implying no proprietary lock-in on user data — but there is no explicit documentation of a bulk data-export feature, open-format guarantee, or account-closure data portability process. missing for 10: explicit data-export/portability documentation, statement on open formats, and evidence of full account data extraction on leaving.

   • [claimed-docs] “Log in securely, Run commands remotely, Transfer files without exposing your data”
   • [claimed-docs] “SSH (Secure Shell) is a protocol for safely connecting to remote servers. It encrypts your connection so you can: Log in securely, Run comma…”
   • [claimed-docs] “Volumes provide persistent storage that survives instance destruction and can be reattached to new instances.”

   proof ↗⚑ flag

Billing

1. platform engineerPull usage and billing breakdowns programmatically to attribute GPU spend by team or workload

   weight 2 · round drawn
   CoreWeave~partialclaimed4/10

   Docs mention a Billing Insights view for consumption breakdowns by resource type (coreweave-docs-13), but there is no evidence of a programmatic API/export for pulling usage or billing data, nor any mention of attributing spend by team, project, or workload tags. missing for 10: documented billing/usage API or export endpoint, evidence of team/workload-level cost attribution or tagging, and any programmatic (non-console) access to billing data.

   • [claimed-docs] “Billing insights: View billable resource usage and consumption breakdowns by resource type.”

   proof ↗⚑ flag

   Vast.ai~partialclaimed4/10

   Vast.ai's REST API and CLI provide general programmatic access to the platform, including account balance checks and billing pages (autobilling, credit top-up), and scoped API keys can be created per workload/team for CI/CD use, which could support some cost attribution. However there is no explicit evidence of a dedicated usage/billing breakdown endpoint, cost tagging, or per-team/workload spend reporting API. missing for 10: dedicated billing/usage export API, cost tagging or labels for spend attribution, documented reports/analytics endpoint for team-level breakdowns.

   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “Use the autobilling feature to have the system automatically top up your account using a saved credit card when it runs low.”
   • [claimed-docs] “create a scoped key with only the permissions that workload needs”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”
   • [claimed-docs] “Autobilling keeps your balance positive to protect your instances from interruption or deletion.”

   proof ↗⚑ flag

2. ml engineerI am billed at per-second or per-minute granularity and only while my instance is actually running

   weight 3 · round drawn
   CoreWeave—none0/10

   Docs mention pay-as-you-go, on-demand/spot capacity, and billing insights showing usage breakdowns, but no evidence specifies per-second or per-minute billing granularity or confirms billing only occurs while instances are actively running.

   • [claimed-docs] “Billing insights: View billable resource usage and consumption breakdowns by resource type.”
   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”
   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”

   proof ↗⚑ flag

   Vast.ai—none0/10

   Evidence shows Vast.ai quotes prices as hourly rates (dph/dph_total) and discusses reserved/interruptible pricing models and autobilling, but nowhere does the evidence pack confirm sub-hour (per-second or per-minute) billing granularity or explicitly state billing stops precisely when an instance is not running.

   ⚑ flag

Discovery

1. ai agentQuery the GPU catalog with live pricing and availability from a public or documented endpoint before committing any spend

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   Evidence shows a public pricing page and Capacity Finder UI for availability, but no documented/public API endpoint that returns live GPU catalog pricing and availability programmatically for an agent to query before committing spend.

   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”
   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”
   • [claimed-docs] “use Capacity Finder in the Cloud Console to compare placement availability across Zones for your requested instance type and Node count”
   • [claimed-docs] “This page shows you how to view your quotas in the Cloud Console, how to read quota errors when a Node Pool exceeds its quota, and how to re…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed9/10

   Vast.ai provides both a documented public REST endpoint (bundles/offers) and CLI (`vastai search offers`) that return live GPU pricing, specs, and availability without requiring an account or API key, verified via runtime probes. This directly matches the story of an agent querying the catalog before committing spend; missing for 10: independent third-party corroboration beyond vendor docs/probes and no explicit rate-limit/auth-tier documentation for heavy automated querying.

   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”
   • [claimed-docs] “vastai search offers 'gpu_name=RTX_4090 num_gpus=1 verified=true direct_port_count>=1 rentable=true' -o 'dlperf_usd-'”
   • [github] “vast.search_offers(query='gpu_name=RTX_4090 num_gpus>=4')”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”

   proof ↗⚑ flag

Pricing

1. platform engineerLock in reserved or committed-use discounts for sustained GPU capacity

   weight 1 · round to Vast.ai
   CoreWeave—none0/10

   Evidence mentions on-demand and spot capacity (no long-term commitment) and references a 'capacity-plans' section, but no doc excerpt describes reserved or committed-use discount pricing, contract terms, or commitment tiers for platform engineers to lock in.

   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”
   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”
   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”

   proof ↗⚑ flag

   Vast.ai✓fullclaimed7/10

   Vast.ai documents Reserved Instances that let users pre-pay for GPU time to lock in up to 50% discounts, convertible from any on-demand instance at any time, directly matching the committed-use discount story. Missing for 10: no independent/community corroboration of reserved-instance pricing behavior, no detail on commitment duration/terms flexibility, and no evidence of longer-term (e.g. multi-month/annual) contractual reserved capacity beyond simple pre-payment conversion.

   • [claimed-docs] “Reserved instances allow you to get significant discounts (up to 50%) by pre-paying for GPU time. You can convert any on-demand instance to …”
   • [claimed-docs] “You can convert any on-demand instance to a reserved instance at any time.”
   • [claimed-docs] “You can convert an on-demand instance into a reserved instance with a lower hourly rate by pre-paying.”
   • [claimed-docs] “Reserved instances allow you to get significant discounts (up to 50%) by pre-paying for GPU time.”

   proof ↗⚑ flag

2. ml engineerSee the published per-GPU-hour price for every GPU type on a public pricing page without talking to sales

   weight 3 · round to Vast.ai
   CoreWeave—none0/10

   The only pricing-related evidence is a marketing snippet from coreweave.com/pricing touting 'flexibility of great pricing' for on-demand/spot GPUs, but no evidence shows an actual published per-GPU-hour price table or rate card visible without contacting sales.

   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”

   proof ↗⚑ flag

   Vast.ai~partialprobed7/10

   The marketplace search page and public REST/CLI endpoints let anyone see live per-GPU-hour prices without an account or contacting sales (vast-ai-docs-41, vast-ai-probe-rt-2, vast-ai-probe-rt-3), which covers the core of the story. However, there's no dedicated static 'pricing page' listing every GPU type's rate at a glance—pricing is discovered via a dynamic search/filter interface, and a community user explicitly complained about the number of clicks needed to find even an approximate price (vast-ai-comm-3). Missing for 10: a single canonical pricing table/page enumerating all GPU types' rates, and stronger independent corroboration that pricing is easy to find without friction.

   • [claimed-docs] “The search page is the main portal for finding good machines and creating instances on them.”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”
   • [community] “As a potential user, I'm disliking how many clicks it takes to find even an approximate price. Prices should be easiest to find.”
   • [claimed-docs] “Vast typically delivers more competitive pricing than traditional cloud providers. The marketplace model creates natural price competition”

   proof ↗⚑ flag

Spot

1. ml engineerRent spot or interruptible GPU capacity at a deep discount with clearly documented preemption semantics

   weight 3 · round to Vast.ai
   CoreWeave~partialclaimed4/10

   CoreWeave docs confirm Spot Node Pools exist as pay-as-you-go, no-commitment bare-metal capacity with a Capacity Finder tool to check availability, but the evidence never documents actual preemption semantics (notice period, eviction behavior, discount percentage vs on-demand) that an ML engineer would need to plan around interruptions. Missing for 10: documented eviction/notice mechanics, explicit discount pricing tied to spot vs on-demand, and any hands-on or independent confirmation of how preemption actually behaves.

   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”
   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”
   • [claimed-docs] “use Capacity Finder in the Cloud Console to compare placement availability across Zones for your requested instance type and Node count”
   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”

   proof ↗⚑ flag

   Vast.ai~partialcommunity6/10

   Vast.ai clearly documents an interruptible/bidding instance type offering 50%+ discounts vs on-demand (vast-ai-docs-7/17/22/35/44), which satisfies the 'deep discount' part of the story. However, the evidence pack gives only the pricing angle and doesn't detail concrete preemption mechanics (notice period, how bidding/eviction actually triggers, restart behavior) — missing for 10: explicit preemption trigger/notice documentation, and independent confirmation that discount/preemption behavior matches docs (community notes reliability variance, e.g. vast-ai-comm-6, but this isn't a concrete dispute of interruptible semantics specifically).

   • [claimed-docs] “Interruptible instances use a bidding system: **Lowest cost** (often 50%+ cheaper than on-demand)”
   • [claimed-docs] “Lowest cost (often 50%+ cheaper than on-demand)”
   • [claimed-docs] “Interruptible instances use a bidding system: Lowest cost (often 50%+ cheaper than on-demand)”
   • [claimed-docs] “Interruptible instances use a bidding system... Lowest cost (often 50%+ cheaper than on-demand)”
   • [claimed-docs] “Vast.ai offers three instance types with different priority levels and pricing models to match your workload requirements and budget.”
   • [community] “As a client (renting GPUs) I've found it to be pretty unreliable. I've tried renting GPUs a handful of times and hosts, bandwidth, and perfo…”

   proof ↗⚑ flag

1. ai-native userChoose where my data is stored (region/residency)

   weight 2 · round drawn
   CoreWeave—none0/10

   The evidence describes CoreWeave's compute 'Zones' for GPU capacity placement (docs-23, docs-31) but never documents a customer-facing region/residency selection control for data storage (e.g., choosing a region for AI Object Storage or CKS clusters) or any data-residency compliance guarantees.

   ⚑ flag

   Vast.ai—none0/10

   The evidence pack shows Vast.ai lets users search/filter GPU offers by attributes like GPU model and reliability, and it distinguishes 'Secure Cloud' verified datacenters, but nowhere documents an explicit region/country selection or data-residency guarantee for where data is stored. A community report even flags a case where a host's advertised location (US) was misrepresented (actually China), undermining confidence that location can be reliably chosen or verified.

   • [claimed-docs] “Secure Cloud (Datacenter): These machines are verified and confirmed to be in a certified datacenter that meets our datacenter criteria.”
   • [claimed-docs] “Vast verifies that these machines are in a datacenter with a TIER 2/3 rating or ISO 27001. These offer cards have a blue label and are recom…”
   • [community] “Discussion flagged a Chinese host on vast.ai masquerading as a US host, raising concerns about where rented GPU compute is actually located.”

   proof ↗⚑ flag

2. ai-native userControl data retention and deletion

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   Evidence shows CoreWeave APIs can delete clusters/resources (coreweave-docs-5) but there's no documentation of data retention policies, data deletion guarantees for stored training data/checkpoints, or privacy controls governing customer data lifecycle — the core of the story is unaddressed.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “Store training data, checkpoints, and model weights with efficient S3-compatible access directly to your GPU compute resources.”

   proof ↗⚑ flag

   Vast.ai~partialprobed3/10

   Evidence shows users can destroy instances and volumes (which otherwise persist across instance destruction) via CLI/API, giving basic control over deletion of compute resources and data, but there is no explicit documentation of data retention policies, data-deletion guarantees, or privacy controls beyond simply terminating instances/volumes. missing for 10: explicit data retention/deletion policy docs, guarantees on backend data purge after deletion, account-level data deletion controls, independent verification of deletion behavior.

   • [claimed-docs] “Volumes provide persistent storage that survives instance destruction and can be reattached to new instances.”
   • [claimed-docs] “search for a GPU, rent it, wait for it to boot, connect to it, copy data, and clean up”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”

   proof ↗⚑ flag

3. ai-native userOpt out of telemetry and usage tracking

   weight 2 · round drawn
   CoreWeave—none0/10

   No evidence in the pack addresses telemetry opt-out or usage-tracking controls; the docs cover infrastructure, clusters, storage, and billing but nothing about user-level telemetry preferences. Missing for 10: any documentation of a telemetry/analytics opt-out setting, privacy controls dashboard, or usage-tracking disclosure.

   ⚑ flag

   Vast.ai—none0/10

   No evidence in the pack mentions telemetry, usage tracking, analytics opt-out, or privacy settings for Vast.ai's own platform/CLI; the docs focus on billing, instances, SSH, and API/CLI usage. missing for 10: any documentation of telemetry collection, an opt-out mechanism, or a privacy policy addressing usage tracking.

   ⚑ flag

Agent ops

1. ai agentProvision a GPU, monitor it, run a workload, and tear it down end to end through documented APIs, CLI, or MCP without a human in the console

   weight 3 · round to Vast.ai
   CoreWeave~partialprobed6/10

   CoreWeave documents a REST API (verified live and Bearer-token-gated in probe) and Terraform provider for creating/listing/updating/deleting CKS clusters (GPU node pools), plus Kubernetes-native monitoring, autoscaling, and quota/billing visibility, all of which an agent could drive without a human touching the Console. However, the MCP endpoint found is only for docs search/retrieval, not for provisioning or lifecycle actions, and there's no CLI or agent-specific tooling documented beyond kubectl/Terraform/API — full automated teardown and monitoring loop is implied but not shown end-to-end in a single agent-facing workflow. missing for 10: an agent-oriented CLI, an MCP server exposing actual provisioning/monitor/teardown actions (not just doc search), and a documented single end-to-end agent workflow example tying create→monitor→run→delete together.

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”
   • [claimed-docs] “All API requests must include a CoreWeave API access token in the Authorization header as a Bearer token.”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler...letting you scale CKS Node Pools in response to workload demand…”
   • [claimed-docs] “This page shows you how to view your quotas in the Cloud Console, how to read quota errors when a Node Pool exceeds its quota, and how to re…”
   • [claimed-docs] “Billing insights: View billable resource usage and consumption breakdowns by resource type.”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the docs MCP endpoint https://docs.coreweave.com/mcp completed a FULL keyless JSON-RPC initialize hands…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed8/10

   Vast.ai documents a full REST API, CLI, and Python SDK covering the entire instance lifecycle (search offers, create/rent, connect, destroy), confirmed by runtime probes showing the CLI and API working keylessly for search and an OpenAPI spec being live. It also ships a dedicated 'vastai agent skill' explicitly for AI coding assistants to create instances, deploy endpoints, and manage keys without human console interaction, plus scoped API keys for automation/CI. Missing for 10: no independent hands-on evidence of a full agent-driven create→monitor→run→teardown cycle end-to-end (only search was probed live), and no first-party MCP server confirmed beyond the 'agent skill' framing.

   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”
   • [claimed-docs] “search for a GPU, rent it, wait for it to boot, connect to it, copy data, and clean up”
   • [probe] “PROBE openapi: HTTP 200 at https://docs.vast.ai/api/openapi.json — contains "openapi" key”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”

   proof ↗⚑ flag

Manage

1. ml engineerSet auto-shutdown timers or spend limits so a forgotten instance can't silently run up a huge bill

   weight 2 · round drawn
   CoreWeave—none0/10

   Evidence shows quota management and billing usage dashboards (coreweave-docs-12, coreweave-docs-13), but nothing about auto-shutdown timers, idle-instance termination, or configurable spend limits/budget alerts that would stop a forgotten instance from running up costs.

   • [claimed-docs] “This page shows you how to view your quotas in the Cloud Console, how to read quota errors when a Node Pool exceeds its quota, and how to re…”
   • [claimed-docs] “Billing insights: View billable resource usage and consumption breakdowns by resource type.”

   proof ↗⚑ flag

   Vast.ai—none0/10

   The evidence pack documents autobilling (auto top-up to keep instances running), reserved/interruptible instance pricing, and scoped API keys, but none of these are auto-shutdown timers or spend caps — autobilling in fact works against this story by automatically refilling balance to prevent instance interruption rather than capping spend. No mention of idle-timeout, max-runtime, or budget-limit features exists anywhere in the pack.

   • [claimed-docs] “Use the autobilling feature to have the system automatically top up your account using a saved credit card when it runs low.”
   • [claimed-docs] “Autobilling keeps your balance positive to protect your instances from interruption or deletion.”
   • [claimed-docs] “Reserved instances allow you to get significant discounts (up to 50%) by pre-paying for GPU time. You can convert any on-demand instance to …”
   • [claimed-docs] “create a scoped key with only the permissions that workload needs”

   proof ↗⚑ flag

2. developerStart, stop, restart, and terminate instances programmatically and keep paying only for what is running

   weight 2 · round to Vast.ai
   CoreWeave~partialprobed6/10

   CoreWeave's CKS API supports programmatic create/list/update/delete of clusters and Node Pools, autoscaling to grow/shrink capacity, and Spot Node Pools are explicitly pay-as-you-go with no commitment, and billing dashboards show usage-based consumption — a live runtime probe even confirms the create/list/delete API is reachable and token-gated. However, the story asks specifically about start/stop/restart of individual instances, and evidence only documents cluster/node-pool-level create, delete, and autoscale operations rather than explicit stop/start/restart semantics for a single running instance. Missing for 10: explicit instance-level start/stop/restart API or docs (only pool-level scale/create/delete and Spot on-demand billing are evidenced).

   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler...letting you scale CKS Node Pools in response to workload demand…”
   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”
   • [claimed-docs] “Billing insights: View billable resource usage and consumption breakdowns by resource type.”
   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed8/10

   Vast.ai's CLI and REST API provide documented commands/endpoints for the full instance lifecycle (search, create, destroy) confirmed by runtime probes, plus billing docs (autobilling, per-hour pricing, reserved/interruptible discounts) demonstrate pay-only-for-what-runs. Community feedback corroborates cost savings but also notes reliability variance, which is a secondary concern rather than a lifecycle-control failure. Missing for 10: explicit documented 'stop'/'restart' CLI subcommands beyond create/destroy in the evidence, and independent hands-on confirmation of stop/restart specifically (only create/destroy verified at runtime).

   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”
   • [claimed-docs] “Use the autobilling feature to have the system automatically top up your account using a saved credit card when it runs low.”
   • [claimed-docs] “Autobilling keeps your balance positive to protect your instances from interruption or deletion.”
   • [claimed-docs] “Reserved instances allow you to get significant discounts (up to 50%) by pre-paying for GPU time. You can convert any on-demand instance to …”
   • [claimed-docs] “Interruptible instances use a bidding system: **Lowest cost** (often 50%+ cheaper than on-demand)”
   • [community] “Absolutely incredible how cheap the compute power is on this platform. Vast has saved me tens of thousands of dollars running ML experiments…”

   proof ↗⚑ flag

Provision

1. developerProvision an on-demand GPU instance from the console or API and be running code on it within minutes

   weight 3 · round to Vast.ai
   CoreWeave~partialprobed6/10

   CoreWeave documents Console/Terraform/API provisioning of Spot and On-Demand Node Pools with kubeconfig-based access to run kubectl/code, and the API is confirmed live and Bearer-token-gated by a runtime probe. However, provisioning is framed around Kubernetes clusters/Node Pools rather than a single quick 'instance' spin-up, and there is no first-party or independent evidence of actual end-to-end timing (minutes) or a simple single-VM/instance API akin to typical cloud on-demand GPU flows. missing for 10: evidence of a simple single-instance (non-cluster) on-demand GPU provisioning path, documented/observed time-to-running-code, and independent hands-on confirmation of the 'minutes' claim.

   • [claimed-docs] “You can create a cluster with the Cloud Console or with Terraform.”
   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”
   • [claimed-docs] “The API lets you create, list, update, and delete managed Kubernetes clusters on CoreWeave infrastructure.”
   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”
   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”
   • [claimed-docs] “Before you create a Spot Node Pool, use Capacity Finder in the Cloud Console to compare placement availability across Zones for your request…”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare GET to the documented CKS provisioning API https://api.coreweave.com/v1beta1/cks/clusters answer…”

   proof ↗⚑ flag

   Vast.ai✓fullprobed7/10

   Docs and runtime probes show the full quickstart flow — search offers via CLI/API keylessly, one-click templates, instance creation, SSH access with password auth disabled, and REST/CLI parity for automation — matching the 'search, rent, boot, connect' workflow described in vast-ai-docs-34. Community reviews largely corroborate ease and cost savings, though some flag friction/hoops versus AWS/GCP and inconsistent reliability, which tempers but doesn't contradict the core provisioning claim. Missing for 10: an independent hands-on timing test confirming actual boot-to-running-code within minutes, and resolution of community complaints about setup friction/reliability variance.

   • [claimed-docs] “This Quickstart will guide you through setting up your Vast.ai account and running your first instance in just a few steps.”
   • [claimed-docs] “search for a GPU, rent it, wait for it to boot, connect to it, copy data, and clean up”
   • [claimed-docs] “The search page is the main portal for finding good machines and creating instances on them.”
   • [claimed-docs] “Launch prebuilt or custom templates with one click.”
   • [claimed-docs] “The Vast.ai CLI gives you command-line access to the entire platform, authentication, GPU search, instance lifecycle, templates, volumes, se…”
   • [claimed-docs] “The Vast.ai REST API gives you programmatic control over the entire platform. It's the foundation that the CLI and SDK are built on.”
   • [claimed-docs] “SSH (Secure Shell) is a protocol for safely connecting to remote servers. It encrypts your connection so you can: Log in securely, Run comma…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official Vast.ai CLI (pypi `vastai`) installed via uvx and ran keylessly — `vastai --help` prints t…”
   • [probe] “PROBE runtime (recorded 2026-09-05): `vastai search offers 'gpu_name=RTX_4090 num_gpus=1' -o dph` searched the LIVE GPU marketplace with NO …”
   • [probe] “PROBE runtime (recorded 2026-09-05): a bare keyless GET to https://console.vast.ai/api/v0/bundles/ returned live marketplace offers as JSON …”
   • [community] “Absolutely incredible how cheap the compute power is on this platform. Vast has saved me tens of thousands of dollars running ML experiments…”
   • [community] “While they've done a pretty good job considering with their software stack it's still kind of a pain to use with significant limitations and…”

   proof ↗⚑ flag

Serverless

1. developerDeploy code to autoscaling serverless GPU workers that scale to zero, instead of managing always-on instances

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   CoreWeave's documented autoscaling is Kubernetes Cluster Autoscaler scaling Node Pools/bare-metal instances up/down with demand (coreweave-docs-3, coreweave-docs-16, coreweave-docs-30), and Spot/On-Demand node pools for burst capacity (coreweave-docs-4, coreweave-docs-20) — this is infrastructure-level cluster scaling, not a serverless 'deploy code and it scales to zero' abstraction. No evidence describes a serverless function/endpoint product, a scale-to-zero guarantee, or a developer simply pushing code without managing nodes/pools.

   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler...letting you scale CKS Node Pools in response to workload demand…”
   • [claimed-docs] “Spot Node Pools provide pay-as-you-go access to high-performance bare-metal compute resources without long-term commitments or reservations.”
   • [claimed-docs] “CKS supports scaling Node Pools by using the Kubernetes Cluster Autoscaler, letting you scale CKS Node Pools in response to workload demands…”
   • [claimed-docs] “Utilize on-demand GPU instances for additional workloads when you're not looking for long-term capacity commitments. Quickly spin-up burst O…”
   • [claimed-docs] “letting you scale CKS Node Pools in response to workload demands for GPU, CPU, or memory resources”

   proof ↗⚑ flag

   Vast.ai~partialclaimed6/10

   Vast.ai explicitly ships 'Vast Serverless', described as letting users run compute-intensive workloads 'without managing GPUs, paying for execution rather than GPU rental time,' with a Python SDK and CLI support for deploying endpoints — matching the pay-per-execution serverless model a developer would want instead of always-on instances. However, the evidence never explicitly confirms autoscaling behavior or scale-to-zero semantics, nor gives hands-on/independent corroboration of how workers scale under load. Missing for 10: explicit autoscaling/scale-to-zero documentation, independent or hands-on validation of serverless endpoint behavior, and details on cold-start/latency tradeoffs.

   • [claimed-docs] “Vast Serverless is an AI infrastructure platform that lets you run compute-intensive workloads without managing GPUs, paying for execution r…”
   • [claimed-docs] “Interacting with Vast Serverless is made easy through a powerful python SDK.”
   • [claimed-docs] “The vastai agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check b…”
   • [claimed-docs] “The `vastai` agent skill lets AI coding assistants drive Vast.ai on your behalf — create instances, deploy endpoints, manage SSH keys, check…”

   proof ↗⚑ flag

Data movement

1. developerMove data in and out efficiently — S3-compatible endpoints, cloud-storage sync, or documented transfer tooling

   weight 2 · round to CoreWeave
   CoreWeave~partialclaimed6/10

   CoreWeave documents S3-compatible Object Storage for moving training data, checkpoints, and model weights directly to GPU compute (coreweave-docs-10), plus a Terraform provider for managing storage buckets/policies as code (coreweave-docs-6). However, there is no documentation of cloud-storage sync tooling (e.g., rsync-like utilities, cross-cloud migration tools) or a dedicated CLI/SDK specifically for bulk data transfer beyond generic S3 API compatibility. missing for 10: dedicated data-migration/sync tooling, third-party or first-party CLI examples for bulk transfer, independent hands-on verification of transfer performance/throughput claims.

   • [claimed-docs] “Store training data, checkpoints, and model weights with efficient S3-compatible access directly to your GPU compute resources.”
   • [claimed-docs] “The CoreWeave Terraform provider lets you manage CoreWeave infrastructure as code, including CKS clusters, VPC networking, AI Object Storage…”
   • [claimed-docs] “All API requests must include a CoreWeave API access token in the Authorization header as a Bearer token.”

   proof ↗⚑ flag

   Vast.ai~partialcommunity5/10

   Vast.ai documents SSH-based secure file transfer to/from instances and a CLI workflow that includes 'copy data' as a step, plus persistent Volumes for data that survives instance destruction. However there is no evidence of S3-compatible endpoints or native cloud-storage (S3/GCS/Azure Blob) sync tooling, and a community comment notes persistent-disk workflows are still cumbersome over time. Missing for 10: S3-compatible storage endpoint, built-in cloud-storage sync/mirroring tooling, independent hands-on validation of transfer throughput/performance.

   • [claimed-docs] “Log in securely, Run commands remotely, Transfer files without exposing your data”
   • [claimed-docs] “SSH (Secure Shell) is a protocol for safely connecting to remote servers. It encrypts your connection so you can: Log in securely, Run comma…”
   • [claimed-docs] “search for a GPU, rent it, wait for it to boot, connect to it, copy data, and clean up”
   • [claimed-docs] “Volumes provide persistent storage that survives instance destruction and can be reattached to new instances.”
   • [community] “Love vast.ai. Only wish is for it to be easier to work with persistent disks over time. I only use it for one off expensive jobs now.”

   proof ↗⚑ flag

Storage

1. developerAttach persistent network storage that survives instance teardown, so datasets and checkpoints outlive any single GPU rental

   weight 3 · round to Vast.ai
   CoreWeave~partialclaimed5/10

   Docs confirm S3-compatible object storage for persisting training data, checkpoints, and model weights independent of GPU compute (coreweave-docs-10), which satisfies the 'outlives any single GPU rental' need, but there's no explicit evidence of attachable persistent block/network volumes (e.g., Kubernetes PersistentVolumes or NFS-style storage) for CKS nodes specifically surviving instance teardown. missing for 10: documentation of block/network-attached persistent volumes for CKS nodes, PVC/storage-class details, and independent confirmation of data survival across instance teardown.

   • [claimed-docs] “Store training data, checkpoints, and model weights with efficient S3-compatible access directly to your GPU compute resources.”

   proof ↗⚑ flag

   Vast.ai~partialcommunity6/10

   Vast.ai docs explicitly describe 'Volumes' as persistent storage that survives instance destruction and can be reattached to new instances, directly matching the story. However, a community user notes wishing it were 'easier to work with persistent disks over time' and now avoids it for anything but one-off jobs, indicating real friction in practice. Missing for 10: independent hands-on verification of volume reattachment reliability, docs detailing size/performance limits, and broader community corroboration beyond a single anecdotal complaint.

   • [claimed-docs] “Volumes provide persistent storage that survives instance destruction and can be reattached to new instances.”
   • [community] “Love vast.ai. Only wish is for it to be easier to work with persistent disks over time. I only use it for one off expensive jobs now.”

   proof ↗⚑ flag

Images

1. developerRun my own Docker image or custom machine template with my exact environment

   weight 2 · round to Vast.ai
   CoreWeave~partialclaimed3/10

   CKS is described as a managed Kubernetes service running directly on bare metal, which implies workloads (including custom Docker containers) can be deployed, but the evidence never explicitly documents how to submit a custom Docker image or create a custom machine/VM template with a specific environment. missing for 10: explicit docs on deploying custom container images to CKS, and any mention of custom machine templates/VM image support.

   • [claimed-docs] “CKS runs Kubernetes directly on bare metal Nodes, without a hypervisor. Customer clusters don't run Virtual Machines.”
   • [claimed-docs] “CoreWeave Kubernetes Service (CKS) offers a managed Kubernetes service that lets you run clusters on bare metal servers in CoreWeave Cloud.”
   • [claimed-docs] “You can deploy multiple Node Pools within a single cluster, where each Node Pool contains any number of Nodes. This lets you run different t…”

   proof ↗⚑ flag

   Vast.ai✓fullclaimed8/10

   Vast.ai's template system explicitly lets developers package their own Docker environment ('Package your environment so any GPU can run it with one click') and launch prebuilt or custom templates with one click, backed by CLI/API/SDK automation for creating and managing instances from these templates. Documentation confirms templates configure the rented machine with whatever software/formatting is needed, directly matching the story. missing for 10: no independent/hands-on confirmation of building and running a fully custom Docker image end-to-end (community evidence focuses on pricing/reliability, not template customization).

   • [claimed-docs] “Build a template - Package your environment so any GPU can run it with one click.”
   • [claimed-docs] “A template is how Vast helps you launch an instance, setting up your rented machine with whatever software and formatting you need.”
   • [claimed-docs] “Launch prebuilt or custom templates with one click.”
   • [claimed-docs] “PyTorch, vLLM, ComfyUI, agents, and more. Real deployments you can copy.”
   • [claimed-docs] “Automate with the CLI or SDK ... `vastai` from the shell or Python. Same operations, different syntax.”

   proof ↗⚑ flag

Templates

1. developerLaunch from pre-built ML templates (PyTorch, CUDA, vLLM, ComfyUI) instead of assembling an environment from scratch

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   The evidence pack covers CKS cluster/node-pool management, Terraform, storage, Slurm-on-K8s, and API access tokens, but contains no mention of pre-built ML environment templates or container images for PyTorch, CUDA, vLLM, or ComfyUI that a developer could launch directly. This is a fair axis for a GPU cloud platform, but no supporting evidence exists.

   ⚑ flag

   Vast.ai✓fullclaimed8/10

   Vast.ai's docs explicitly describe templates as a first-class launch mechanism ('Launch prebuilt or custom templates with one click', 'A template is how Vast helps you launch an instance') and directly name PyTorch, vLLM, and ComfyUI as available prebuilt templates ('PyTorch, vLLM, ComfyUI, agents, and more. Real deployments you can copy'). CUDA support is implied through GPU offer filtering and template packaging but not explicitly named as its own template. missing for 10: independent/hands-on confirmation of launching a specific named template (e.g., a community post describing a ComfyUI or vLLM template launch), and explicit CUDA-template documentation rather than inferred CUDA support.

   • [claimed-docs] “Launch prebuilt or custom templates with one click.”
   • [claimed-docs] “PyTorch, vLLM, ComfyUI, agents, and more. Real deployments you can copy.”
   • [claimed-docs] “A template is how Vast helps you launch an instance, setting up your rented machine with whatever software and formatting you need.”
   • [claimed-docs] “Build a template - Package your environment so any GPU can run it with one click.”

   proof ↗⚑ flag

Compliance

1. platform engineerVerify the provider's security and compliance posture (SOC 2, data handling, datacenter tiers) before putting proprietary models on it

   weight 2 · round to Vast.ai
   CoreWeave—none0/10

   The evidence pack contains no mention of SOC 2 certification, compliance attestations, data handling/privacy policies, or datacenter tier certifications; it only covers infrastructure/API features (CKS, Terraform, autoscaling, tokens) and unrelated financial/community discussion. Missing for 10: SOC 2 or ISO certifications, data handling/privacy documentation, physical datacenter tier/uptime certifications, any compliance trust page or audit report.

   ⚑ flag

   Vast.ai!disputedcontradicted4/10

   Vast.ai docs claim some infrastructure verification — 'Secure Cloud' machines are checked for TIER 2/3 datacenter rating or ISO 27001 certification (vast-ai-docs-14, vast-ai-docs-28) — but there is no mention of SOC 2 attestation, data-handling/privacy policies, or compliance documentation anywhere in the pack. Community evidence directly undercuts trust in these verifications: a host was found masquerading as US-based while actually located in China (vast-ai-comm-10), and another thread questions whether running consumer GPUs at scale even complies with Nvidia's own terms (vast-ai-comm-9) — concrete contradictions of the 'verified' security posture platform engineers would need to rely on. Missing for 10: SOC 2 or equivalent third-party compliance attestation, explicit data-handling/retention policy, and independent audit corroborating the datacenter tier claims.

   • [claimed-docs] “Secure Cloud (Datacenter): These machines are verified and confirmed to be in a certified datacenter that meets our datacenter criteria.”
   • [claimed-docs] “Vast verifies that these machines are in a datacenter with a TIER 2/3 rating or ISO 27001. These offer cards have a blue label and are recom…”
   • [community] “Running a fleet of RTX3090's on demand does not sound at all compliant with what little I know of NVidia's terms of use. There's surprisingl…”
   • [community] “Discussion flagged a Chinese host on vast.ai masquerading as a US host, raising concerns about where rented GPU compute is actually located.”

   proof ↗⚑ flag

Governance

1. platform engineerManage team members with roles and scoped API keys so credentials and spend stay controlled

   weight 1 · round drawn
   CoreWeave~partialclaimed4/10

   CoreWeave documents API Access Tokens that authenticate and scope access to specific resources (CKS clusters, VPCs) and billing insight views, showing some credential and spend visibility, but there is no evidence of team-member/user role management (RBAC for humans, org roles, or permission tiers) or spend controls/limits tied to specific keys. missing for 10: team member/role management (invite users, assign roles), scoped-key permission granularity beyond resource type, and spend-limiting or budget-control features tied to API keys.

   • [claimed-docs] “API Access Tokens authenticate users and grant access to resources such as CKS clusters and VPCs.”
   • [claimed-docs] “This page explains how to create, use, and manage API Access Tokens and the kubeconfig files generated alongside them, so you can authentica…”
   • [claimed-docs] “Create and download a kubeconfig for a specific cluster, so you can interact with the cluster using commands like kubectl.”
   • [claimed-docs] “Billing insights: View billable resource usage and consumption breakdowns by resource type.”

   proof ↗⚑ flag

   Vast.ai~partialclaimed4/10

   Vast.ai documents scoped API keys with limited permissions for CI/CD and shared tooling use cases (create api-key, per-workload scoping), which addresses the credential-control half of the story. However, there is no evidence of team/organization member management, role assignment, or multi-user account governance — the evidence pack only covers individual scoped keys, not team-based RBAC. Missing for 10: team member invitation/management, role definitions (admin/viewer/etc.), org-level spend controls tied to roles, and any audit trail for team activity.

   • [claimed-docs] “You can also create scoped keys with limited permissions using `vastai create api-key`, useful for CI/CD or shared tooling.”
   • [claimed-docs] “create a scoped key with only the permissions that workload needs”
   • [claimed-docs] “For shared tooling, CI/CD, or automation, create a scoped key with only the permissions that workload needs.”

   proof ↗⚑ flag