Cloud inference

Where to rent GPUs — or skip them for a managed API

Three product shapes solve “I need inference”: rent a GPU and run your stack, use serverless GPU functions, or call a managed model API. Pick the shape first; then compare providers inside it.

GPU marketplace

You choose the card and image. Best when control and $/hour matter more than an API SLA.

Serverless GPU

You ship code; the platform scales hardware. Best for bursty jobs without babysitting VMs.

Managed inference API

You pay for tokens or predictions on open-model serving APIs in this hub. Frontier closed APIs (OpenAI, Anthropic, DeepSeek, Gemini) are compared on /ai-models — this page is not a complete managed-API catalog.

Run the provider finder Local GPUs instead Frontier model APIs Open-weight models Buy vs rent guideEditorial scores last reviewed September 3, 2026

Start here

Pick the path before you pick the logo

Most bad shortlists mix a marketplace pod, a serverless function, and a token API as if they were the same product.

Rent a GPU (marketplace / cluster)

When: You need a specific SKU, custom serving stack, or sustained utilization that beats token APIs.

Next: Start with RunPod or Vast.ai for flexible rental; size VRAM on /gpus first. Remember stop ≠ terminate — tear down storage when the experiment ends.

Serverless GPU functions

When: Work arrives in bursts, your code is container-shaped, and you refuse to babysit VMs overnight.

Next: Compare Modal and Fal.ai against cold-start tolerance for your workload; treat DGX Cloud Lepton as NVIDIA marketplace capacity, not a Modal clone.

Managed inference API

When: You want tokens or predictions tomorrow, not CUDA drivers — and the model is already on a catalog.

Next: For open-model serving, shortlist Together, Fireworks, Groq, Replicate, Hugging Face Endpoints, or Baseten. For OpenAI / Anthropic / DeepSeek / Gemini token APIs, use /ai-models — those vendors are not rows on this hub.

Buy local hardware instead

When: Privacy, offline use, or always-on personal inference makes cloud rental look like a subscription you cannot pause.

Next: Use the /gpus capacity tables and GPU finder before you rent indefinitely. Match VRAM to model size, then pick rental vs buy.

Before you rent

Preflight checklist

Questions that prevent the usual bill shocks and cold-start surprises.

1Name the product shape first: marketplace GPU, serverless function, managed API, or local card — do not mix them in one shortlist.
2Write down cold-start tolerance in user-visible seconds. If the answer is “near zero,” scale-to-zero is a product risk.
3Forecast spend from utilization: idle pods, token volume, or per-prediction size — not from a homepage hero number.
4Confirm the exact model ID or container image before architecture lock-in.
5Plan a failover provider for anything user-facing.
6Size VRAM on /gpus before you rent — a 70B Q4 fit check beats discovering OOM after the pod is warm.
7On marketplaces: stop ≠ terminate. Tear down volumes/disks when the experiment ends or storage keeps billing.
8Prefer interruptible/community only for jobs that can checkpoint; use secure/on-demand for anything user-facing.

Compare chart

Fit by decision axis

Bars show how each provider scores on the axis that usually drives the choice — not a single overall winner.

GPU choice

Pick exact GPUs

Time-to-serving (editorial)

Warm-path fit

Price clarity

Easy to forecast

Production ops

Less DIY ops

Open-model breadth

Catalog depth

Scores are editorial planning ratings (1–10) for product shape — not published $/hour quotes or vendor SLAs. Verify current pricing on each provider site.

At a glance

Provider matrix

Billing shape and GPU control differ more than marketing copy. Scan the matrix, then open a head-to-head.

Scroll the matrix sideways to compare billing, GPU choice and fit.

GPU inference providers compared by category and billing
Provider	Category	Billing	GPU choice	Best for
RunPod	GPU marketplace	Pods: per-second compute and local disk billing; network volumes: hourly storage billing. Serverless Flex/Active workers: per-second lifecycle time.	Broad consumer and datacenter SKUs	Teams that want to pick a specific GPU SKU and control the container stack
Vast.ai	GPU marketplace	Per-second marketplace billing — interruptible, on-demand, or reserved	Wide, host-dependent inventory	Cost-sensitive batch jobs and experiments that can tolerate host variability
Modal	Serverless GPU	Functions: per-second CPU, memory and GPU resources; Shared Endpoints: per-token. Storage and network egress have separate allowances and rates.	Choose supported GPU types through the function GPU setting, including T4, L4, A10, L40S, A100, RTX PRO 6000, H100, H200, B200 and B300; allocation and multi-GPU limits vary.	Developers who want GPU code as functions with autoscaling
Together AI	Managed inference API	Serverless per-token; dedicated inference per GPU-minute for ready replicas; GPU clusters and provisioned throughput have separate contracts.	Managed serverless; dedicated deployment profiles with selected hardware; separate GPU clusters for infrastructure control.	Serving open-weight models without operating your own GPU fleet
Fireworks AI	Managed inference API	Serverless: model/tier-specific tokens. Reserved Throughput: prepaid dollar-per-minute reservation plus overage. Dedicated: per-GPU-second.	Managed serving stack plus on-demand GPU-hour SKUs	Production chat and agent backends that need snappy open-model inference
Replicate	Managed inference API	Official models: model-specific input/output units; community/public models: active hardware time; private models: setup, idle and active hardware time, with labeled fast-boot exceptions.	Official/public model hardware is managed or model-declared; custom deployments choose supported hardware and instance bounds.	Product teams shipping model-backed features without owning GPU ops
Lambda	Cloud GPU cluster	On-demand instances priced hourly, billed in one-minute increments; 1-Click Clusters reserved and billed weekly; filesystems billed separately.	Datacenter NVIDIA focus	Training and sustained inference on reserved or on-demand GPU instances
Hugging Face Endpoints	Managed inference API	Dedicated hardware rates displayed per hour, calculated per minute while initializing/running; paused or scaled-to-zero endpoints are not charged.	Choose supported CPU, GPU or accelerator instances by cloud, size and quota; available catalog options vary.	Teams already living in the Hub who want a managed endpoint per model
Groq	Managed inference API	GroqCloud model-specific token, audio-hour or character pricing; eligible asynchronous Batch requests receive a 50% discount. Dedicated platform terms differ.	GroqMetal dedicated LPU infrastructure; GroqCore adds an inference stack and GroqAssured enterprise controls. These are separate from the shared GroqCloud API.	Latency-sensitive chat, agents, and batch inference on Groq’s supported model set
DeepInfra	Managed inference API	Shared: model-specific tokens, output units or execution time. Private deployments: GPU-hours with minute-granularity billing, invoiced weekly.	Shared APIs manage hardware. Private deployments select A100, H100, H200, B200 or B300 GPUs; dedicated clusters are a separate contact-sales option.	Cost-first hosting of popular open models (Llama, Qwen, DeepSeek, embedding models)
Cerebras Inference	Managed inference API	Shared Free Trial/Pay as You Go per-model tokens; enterprise Dedicated Inference reserves capacity under organization-specific pricing.	Cerebras wafer-scale serving infrastructure; hardware generation depends on the service	Supported shared models or enterprise dedicated deployments where measured latency/throughput justify the serving path
Nebius Token Factory	Managed inference API	Public serverless per-token; dedicated deployments per GPU-hour with per-minute granularity.	Public shared capacity is platform-managed; dedicated deployments select supported GPU types and GPUs per replica.	Shared model APIs or configurable dedicated open-weight serving with region and capacity requirements
Fal.ai	Serverless GPU	Model APIs: per-model output unit/request; Serverless custom endpoints: per-second runner/GPU time, including billable idle/keep_alive; Compute: per-instance GPU-hour while running.	Model APIs manage hardware; custom Serverless chooses supported GPU machine types. Compute offers 1x or 8x H100 SXM instances with full SSH.	Product teams shipping image, video, or audio features without operating GPU fleets
Baseten	Managed inference API	Model APIs: model-specific token/tool pricing. Dedicated deployments, image builders and training: instance minutes while the workload is running.	Select hardware per deployment	ML teams shipping bespoke models who want managed serving without building the ops stack
CoreWeave	Cloud GPU cluster	Serverless per-token; Dedicated Inference per GPU-hour; CKS/cloud capacity uses on-demand, spot, reserved or Flex terms.	Dedicated Inference explicitly chooses GPU class and zone; serverless manages hardware; CKS exposes the full cluster serving stack.	Teams choosing between managed custom-weight inference, serverless model APIs and full GPU-cluster control
NVIDIA DGX Cloud Lepton	GPU marketplace	Partner GPU marketplace / DGX Cloud usage	NVIDIA and partner GPU classes exposed through DGX Cloud Lepton	Teams wanting to deploy/customize models to managed endpoints over NVIDIA Cloud Partner/GPU marketplace capacity.

Some outbound links use a first-party redirect hop for click counting. Commission is only claimed when a partner programme is active for that specific link — most vendor hops here are not paid placements. Affiliate disclosure.

Providers

16 places teams actually rent or buy inference

Grouped by product shape — marketplace rental, serverless GPU, managed API, or cluster. Each card now shows GPU choice, cold-start reality, and the main watch-out.

GPU marketplace

Pods: per-second compute and local disk billing; network volumes: hourly storage billing. Serverless Flex/Active workers: per-second lifecycle time.

RunPod

7.4/10 editorial avg

Rent selectable GPUs through Pods, or run autoscaling inference workers through the separate Serverless product.

GPU choice: Broad consumer and datacenter SKUs
Cold start: Depends on pod spin-up and image cache

Best for: Teams that want to pick a specific GPU SKU and control the container stack

Watch out: Pod compute stops when a Pod stops, but persistent volume disks ($0.20/GB-month stopped) and network volumes ($0.07/GB-month under 1TB; $0.05/GB-month above 1TB) continue billing until deleted. Temporary container disk is erased on stop. Serverless worker billing and always-on worker settings differ from Pod billing.

Product facts checked October 2, 2026

Per-second marketplace billing — interruptible, on-demand, or reserved

Vast.ai

6.4/10 editorial avg

A marketplace of host-offered GPUs with competitive spot-style pricing for interruptible or carefully filtered capacity.

GPU choice: Wide, host-dependent inventory
Cold start: Instance provisioning + your stack

Best for: Cost-sensitive batch jobs and experiments that can tolerate host variability

Watch out: Host inventory/reliability and offer duration vary. Interruptible instances may be paused when outbid or when on-demand capacity is requested. On-demand and reserved are high-priority and non-interruptible during the rental period, but this is an instance priority/resource guarantee, not a host-level uptime SLA. Reserved capacity is prepaid and may discount up to 50% based on commitment. Stopped-instance storage keeps billing until the instance is deleted.

Product facts checked October 2, 2026

Partner GPU marketplace / DGX Cloud usage

NVIDIA DGX Cloud Lepton

6.0/10 editorial avg

NVIDIA managed AI platform that combines partner/cloud GPU discovery with model deployment and inference endpoints; it is broader than a GPU-only marketplace.

GPU choice: NVIDIA and partner GPU classes exposed through DGX Cloud Lepton
Cold start: Zero replicas cannot serve requests or automatically wake on traffic under the current documentation. Scale the deployment up again, or configure a nonzero minimum; provisioning/model-load delay depends on capacity.

Best for: Teams wanting to deploy/customize models to managed endpoints over NVIDIA Cloud Partner/GPU marketplace capacity.

Watch out: Endpoints default to one replica and scale to zero after one hour inactive. The documentation says automatic scale-up from zero is unsupported: an operator must scale the deployment up before it can serve requests again. Partner inventory, rates and SLAs vary.

Product facts checked October 2, 2026

Serverless GPU

Functions: per-second CPU, memory and GPU resources; Shared Endpoints: per-token. Storage and network egress have separate allowances and rates.

Modal

7.0/10 editorial avg

Serverless containers and GPUs billed for what you run — strong fit for Python inference jobs and bursty workloads without managing VMs.

GPU choice: Choose supported GPU types through the function GPU setting, including T4, L4, A10, L40S, A100, RTX PRO 6000, H100, H200, B200 and B300; allocation and multi-GPU limits vary.
Cold start: Can be noticeable on cold containers; benchmark the real image and model.

Best for: Developers who want GPU code as functions with autoscaling

Watch out: You can choose a specific supported GPU type, but containers and scaling remain managed. Warm idle containers consume billable resources. Starter compute credits do not cover Shared Endpoints tokens.

Product facts checked October 2, 2026

Model APIs: per-model output unit/request; Serverless custom endpoints: per-second runner/GPU time, including billable idle/keep_alive; Compute: per-instance GPU-hour while running.

Fal.ai

6.8/10 editorial avg

Generative media Model APIs, autoscaling custom Serverless endpoints and continuously running Compute GPU instances with SSH access, training and batch support.

GPU choice: Model APIs manage hardware; custom Serverless chooses supported GPU machine types. Compute offers 1x or 8x H100 SXM instances with full SSH.
Cold start: Cold-start and queue latency vary by model/runner; fal reports caching reduces cold starts, not eliminates all cold starts.

Best for: Product teams shipping image, video, or audio features without operating GPU fleets

Watch out: Choose the product first: Model API output units, Serverless runner lifetime and Compute instance hours bill differently. Compute requires enabled account access; its numeric rates are supplied in the dashboard or by sales.

Product facts checked October 2, 2026

Managed inference API

Serverless per-token; dedicated inference per GPU-minute for ready replicas; GPU clusters and provisioned throughput have separate contracts.

Together AI

7.2/10 editorial avg

Managed inference and fine-tuning oriented around open models — serverless token APIs plus dedicated / GPU-cluster options (together.ai/pricing), not tokens-only.

GPU choice: Managed serverless; dedicated deployment profiles with selected hardware; separate GPU clusters for infrastructure control.
Cold start: Dedicated replicas need provisioning/model loading after restart or zero replicas; ready warm replicas continue billing while idle. No universal startup latency is established.

Best for: Serving open-weight models without operating your own GPU fleet

Watch out: Serverless token APIs, dedicated GPU hours and provisioned throughput have different cost models. Each PTU represents reserved capacity; tokens per minute depend on the model and token type. Together announced $800M in Series C funding on July 1, 2026; this does not guarantee endpoint availability or service levels.

Product facts checked October 2, 2026

Serverless: model/tier-specific tokens. Reserved Throughput: prepaid dollar-per-minute reservation plus overage. Dedicated: per-GPU-second.

Fireworks AI

7.0/10 editorial avg

Low-latency inference API focused on fast serving of popular open and partner models.

GPU choice: Managed serving stack plus on-demand GPU-hour SKUs
Cold start: Shared Serverless is documented as warm. A dedicated deployment at zero replicas initiates scale-up on a request but immediately returns 503 DEPLOYMENT_SCALING_UP; requests are not queued. Retry with backoff or keep a minimum replica running.

Best for: Production chat and agent backends that need snappy open-model inference

Watch out: Standard shared Serverless has no uptime/latency SLA. Reserved Throughput provides a throughput SLA within the purchased dollar-per-minute reservation; unused reservation does not roll over and overage has no SLA. On-demand dedicated rates checked October 2: H100/H200 $8, B200 $13, B300 $15 and GB300 $20 per GPU-hour, billed per second; regional restriction premium 1.5x.

Product facts checked October 2, 2026

Official models: model-specific input/output units; community/public models: active hardware time; private models: setup, idle and active hardware time, with labeled fast-boot exceptions.

Replicate

6.6/10 editorial avg

Run public and private models via simple API predictions — popular for image, video, and community model demos.

GPU choice: Official/public model hardware is managed or model-declared; custom deployments choose supported hardware and instance bounds.
Cold start: Official models are documented as always warm; community models/zero-minimum custom deployments may incur model-download and startup delay. Set a minimum instance for an always-on custom deployment and pay for that warm capacity.

Best for: Product teams shipping model-backed features without owning GPU ops

Watch out: Per-prediction economics and cold models can surprise at scale

Product facts checked October 2, 2026

Dedicated hardware rates displayed per hour, calculated per minute while initializing/running; paused or scaled-to-zero endpoints are not charged.

Hugging Face Endpoints

7.0/10 editorial avg

Deploy Hub models to dedicated inference endpoints with Hugging Face’s hosting and scaling controls. NVIDIA has announced an agreement to acquire Hugging Face; this review has not established transaction completion.

GPU choice: Choose supported CPU, GPU or accelerator instances by cloud, size and quota; available catalog options vary.
Cold start: Scale-to-zero defaults to one hour of inactivity and needs min replicas 0. The next request cold-starts; the proxy can return 503 while initializing. X-Scale-Up-Timeout can hold a request until readiness or its specified timeout.

Best for: Teams already living in the Hub who want a managed endpoint per model

Watch out: Endpoint sizing, autoscaling, idle replicas and scale-to-zero behavior drive cost. Confirm current product terms and transaction status before relying on future roadmap commitments. Scope Hub tokens and private-model access to the deployment.

Product facts checked October 2, 2026

GroqCloud model-specific token, audio-hour or character pricing; eligible asynchronous Batch requests receive a 50% discount. Dedicated platform terms differ.

Groq

6.4/10 editorial avg

GroqCloud serves a curated model API on Groq LPU hardware. Groq also offers dedicated infrastructure, an inference-stack layer and enterprise controls; published speed claims depend on model and workload.

GPU choice: GroqMetal dedicated LPU infrastructure; GroqCore adds an inference stack and GroqAssured enterprise controls. These are separate from the shared GroqCloud API.
Cold start: GroqCloud is a shared API; no universal first-request latency is established by the captured documentation. Dedicated capacity has its own deployment terms.

Best for: Latency-sensitive chat, agents, and batch inference on Groq’s supported model set

Watch out: Confirm the exact served model, endpoint and current quota before committing architecture. Groq’s first-party announcement describes a non-exclusive NVIDIA technology licensing agreement. Its August 17, 2026 release confirms a $350M Series A; neither announcement establishes a fixed throughput or capacity guarantee for your workload.

Product facts checked October 2, 2026

Shared: model-specific tokens, output units or execution time. Private deployments: GPU-hours with minute-granularity billing, invoiced weekly.

DeepInfra

3.8/10 editorial avg

Shared model APIs with model-specific token, output-unit or execution-time pricing, plus custom LLM/LoRA deployments on selectable dedicated GPUs.

GPU choice: Shared APIs manage hardware. Private deployments select A100, H100, H200, B200 or B300 GPUs; dedicated clusters are a separate contact-sales option.
Cold start: Not verified for shared endpoints; measure cold-start behavior on the selected model and deployment.

Best for: Cost-first hosting of popular open models (Llama, Qwen, DeepSeek, embedding models)

Watch out: Shared API defaults to 200 concurrent requests per model; burst throughput/rate limits and per-model billing differ. Private deployments use dedicated GPU capacity and GPU-hour billing; verify GPU availability and total instance uptime costs.

Product facts checked October 2, 2026

Shared Free Trial/Pay as You Go per-model tokens; enterprise Dedicated Inference reserves capacity under organization-specific pricing.

Cerebras Inference

3.0/10 editorial avg

Managed inference built on Cerebras wafer-scale hardware. Published throughput figures depend on the model, context, precision and serving configuration.

GPU choice: Cerebras wafer-scale serving infrastructure; hardware generation depends on the service
Cold start: Shared API and reserved dedicated deployments have different capacity behavior; no universal first-request latency is verified here.

Best for: Supported shared models or enterprise dedicated deployments where measured latency/throughput justify the serving path

Watch out: Confirm the served model and access route. Cerebras’s August 18, 2026 CS-4 announcement reports three WSE-3T wafers, up to 2x CS-3 speed and up to 10x throughput per watt; these are vendor-reported, workload-dependent figures. Its August 13 OpenAI Ultrafast announcement reports up to 750 output tokens/sec for GPT-5.6 Sol, initially in limited preview through the OpenAI API. The September 1 Finland announcement describes phased construction toward 165 MW.

Product facts checked October 2, 2026

Public serverless per-token; dedicated deployments per GPU-hour with per-minute granularity.

Nebius Token Factory

3.6/10 editorial avg

Token Factory serves open-weight models through shared OpenAI-compatible APIs and isolated dedicated endpoints. Dedicated deployments can select supported GPU configurations, regions and eligible custom weights.

GPU choice: Public shared capacity is platform-managed; dedicated deployments select supported GPU types and GPUs per replica.
Cold start: Dedicated capacity is governed by configured replicas and lifecycle; no universal shared or dedicated startup latency is established in the captured docs.

Best for: Shared model APIs or configurable dedicated open-weight serving with region and capacity requirements

Watch out: Confirm exact checkpoints, regions and quotas in Token Factory. Nebius’s LPX announcement cites Artificial Analysis measuring 3,400 output tokens/sec for a single user on Gemma 4 31B; this specific model/hardware result does not guarantee throughput for other models, concurrent traffic or end-to-end latency.

Product facts checked October 2, 2026

Model APIs: model-specific token/tool pricing. Dedicated deployments, image builders and training: instance minutes while the workload is running.

Baseten

6.8/10 editorial avg

Hosted Model APIs for supported models plus managed dedicated deployments for custom/open/fine-tuned weights, hardware selection and autoscaling.

GPU choice: Select hardware per deployment
Cold start: Dedicated endpoints can scale to zero; requests may wait for a replica to load. Keep capacity warm or accept cold-start latency; verify request parking/timeout behavior for the selected endpoint.

Best for: ML teams shipping bespoke models who want managed serving without building the ops stack

Watch out: Dedicated billing includes image-builder workloads, cold starts/model loading and idle warm replicas. Monthly budgets notify by default; optional enforcement stops Model API requests only, while dedicated and training workloads continue accruing charges.

Product facts checked October 2, 2026

Cloud GPU cluster

On-demand instances priced hourly, billed in one-minute increments; 1-Click Clusters reserved and billed weekly; filesystems billed separately.

Lambda

7.0/10 editorial avg

Cloud GPUs and clusters aimed at ML teams that want familiar instance-style access with strong NVIDIA inventory.

GPU choice: Datacenter NVIDIA focus
Cold start: Instance boot + your serving stack

Best for: Training and sustained inference on reserved or on-demand GPU instances

Watch out: Capacity and lead times depend on the selected cluster and contract. Lambda announced the closing of a $926M term loan B on August 27, 2026 for a committed customer deployment. That financing does not establish that a particular GPU SKU is currently available to other customers.

Product facts checked October 2, 2026

Serverless per-token; Dedicated Inference per GPU-hour; CKS/cloud capacity uses on-demand, spot, reserved or Flex terms.

CoreWeave

6.6/10 editorial avg

GPU cloud with three inference paths: curated serverless token APIs, managed dedicated custom-weight inference and self-managed serving on CoreWeave Kubernetes Service. Cluster capacity also has on-demand, spot and reserved options.

GPU choice: Dedicated Inference explicitly chooses GPU class and zone; serverless manages hardware; CKS exposes the full cluster serving stack.
Cold start: Dedicated model deployments load weights and need ready replicas; self-managed clusters require provisioning and your serving stack. No universal startup delay is established.

Best for: Teams choosing between managed custom-weight inference, serverless model APIs and full GPU-cluster control

Watch out: Check the live SKU, contract and billing unit. CoreWeave reports a 10x tokens/sec-per-megawatt result for Vera Rubin NVL72 versus GB200 NVL72 on DeepSeek R1 at matched per-user interactivity; the result is workload-specific. Its October 2 price table lists 8-GPU nodes at $49.24/hr (H100), $50.44 (H200) and $68.80 (B200); inference-single rates are separately billed per GPU.

Product facts checked October 2, 2026

Head-to-head

Comparisons that clarify the trade-off

Marketplace vs serverless, API vs API, and cluster vs rental — each pair answers a real shortlist question.

FAQ

Questions teams ask before the first invoice

Qualitative answers only — verify live pricing on each provider.

Is renting a GPU cheaper than a managed API?

It depends on the exact model, API tariff, GPU rate, utilization and operating effort. Warm idle instances add costs; APIs and scale-to-zero functions can suit intermittent traffic. Compare total completed-work cost using the actual billing units.

What is the difference between RunPod and Together AI?

RunPod offers self-managed Pods and autoscaling Serverless workers. Together offers shared model APIs, managed dedicated inference, provisioned throughput and GPU clusters. Compare the specific product path: neither provider has only one billing or operating model.

When do cold starts matter?

Whenever a user is waiting. Container startup and model loading can take seconds or minutes, depending on image, weights and capacity. Check whether requests wait, fail with a retryable error, or require an operator to restore replicas; Lepton currently cannot automatically wake from zero.

Should I rent cloud GPUs or buy a local card?

Rent for burst capacity, team sharing, and SKUs you will not keep busy for years. Buy local when privacy, offline use, or always-on personal inference makes cloud hours look like rent. Many teams do both.

Does stopping a rented GPU stop the bill?

Not always. Stopping a pod can leave disks or network volumes attached. Terminate the instance and confirm storage teardown — especially on RunPod and Vast.ai — or you keep paying after the GPU is idle.

How do I rent a GPU for a coding agent or local model without getting burned?

Size the model on /gpus first, pick marketplace vs managed API, start with a short prepaid or capped budget, pull a known image, run one smoke prompt, then shut down fully (stop + delete storage). Keep a failover provider for anything user-facing.

Are the fit scores prices?

No. They are editorial planning ratings for product shape. Always verify current pricing, regions, and model availability on the provider site before you commit spend.

Decision rule

If you need a specific GPU SKU, rent. If you need a model endpoint tomorrow, buy an API.

Mixing those goals in one shortlist is how teams overpay. Use the fit chart for shape, the matrix for ops reality, and a head-to-head before you commit spend.

Example marketplace: RunPod Compare local GPUs GPU finder