H100
Today
H100
$1.35/ GPU·h
Read Independent · read dailyEvery GPU price.
Verified
.
Prices from 173 clouds, read daily, converted to the true cost per GPU-hour and ranked on price, value and trust.
Prices read daily from providers’ own pages and APIs
Today's ranking · 75 GPUs ranked · on-demand · readings of the last 45 daysThree rankings for every GPU
Pick a GPU. Three independent rankings follow, ten places each: the lowest true cost, the most throughput per dollar, the most trusted provider. A GPU is ranked once at least 3 providers hold an eligible on-demand offer read in the last 45 days; every row carries its read date.
Prices readHow we rank →
Datacenter 14 of 28
Workstation 7 of 15
Consumer 3 of 32
number = providers with an eligible on-demand offer in the last 45 days
H10080 GB192 eligible on-demand offers from 75 providers · median $3.09 · not ranked: 67 unavailable, 27 resold, 29 older readingsH100 page →All H100 offers →
Ranking 01
Cheapest
Lowest true cost per GPU-hour, in stock or stock unknown, newest reading per provider
Ranking 02
Best value
Most throughput per dollar for LLM inference · Trust ≥ 60 first, then provisional · estimated
Ranking 03
Most trusted
Highest Trust Score among providers with an eligible offer · unscored providers follow, cheapest first
GPUBank1× · France · SingaporeTrust 93$1.35
Northflank1× · Oct 31Trust 92$2.74- 4
Nebius1× · Oct 31Trust 89$4.50 - 5
Baseten1× · Oct 31Trust 87$3.75
Listed, not ranked (fewer than 3 providers with an eligible on-demand offer in the last 45 days): Titan Xp, GTX 1070, GTX 1070 Ti, RTX 6000, GTX 1080, Titan RTX, RTX 4090 D, RTX 5880 Ada, Gaudi 2, GTX 1660.
True cost = list price plus the mandatory add-ons the provider publishes · one floor per marketplace · one provider per ranking · the newest reading per provider, GPU and mode wins · rankings are computed, never sponsoredAll 164 GPUs →
Offers explorer · 17,489 current offers · 164 providers · 4 rental modes · last 45 daysEvery current offer, filtered your way
Each row is one configuration read on the provider's own page, in the dated study or in the market report within the last 45 days: the newest reading per provider, GPU and rental mode. Filter by GPU, mode, provider type, region, stock, Trust Score, price ceiling, GPU count, memory and reading age; sort; open the receipt behind any number.
ReferenceFull comparison tool →
359 on-demand offers for the H100 · 93 providers · median $3.47readings of the last 45 days, newest per provider, GPU and mode · reference · prices per GPU-hour · true cost = list price + mandatory add-ons
| Provider | GPU · configuration | Region · stock | True cost | Value | |
|---|---|---|---|---|---|
 io.netMarketplace · Trust 42 | 1× H100 PCIE80 GB per GPU | —Unavailable | $1.00no add-ons | — | Receiptread 23:48 UTC |
| io.netMarketplace · Trust 42 | 1× H100 PCIE80 GB per GPU | —Unavailable | $1.00no add-ons | — | Receiptread 23:48 UTC |
| GPUBankOperator · Trust 93 | 1× H100 PCIE16 vCPU · 188 GB RAM | France · SingaporeIn stock | $1.35no add-ons | Value 84 | Receiptread Oct 25 |
| io.netMarketplace · Trust 42 | 1× H100 PCIE80 GB per GPU | —Unavailable | $1.59no add-ons | — | Receiptread 23:48 UTC |
 GcoreOperator · Trust 80 | 1× H100 SXM80 GB per GPU | —Stock unknown | $1.71no add-ons | Value 80 | Receiptread 23:56 UTC |
| GcoreOperator · Trust 80 | 2× H100 SXM80 GB per GPU | —Stock unknown | $1.71no add-ons | Value 80 | Receiptread 23:56 UTC |
| GcoreOperator · Trust 80 | 4× H100 SXM80 GB per GPU | —Stock unknown | $1.71no add-ons | Value 80 | Receiptread 23:56 UTC |
| GcoreOperator · Trust 80 | 8× H100 SXM80 GB per GPU | —Stock unknown | $1.71no add-ons | Value 80 | Receiptread 23:56 UTC |
 GPUNETReseller · not scored · resold | 1× H100 PCIE16 vCPU · 128 GB RAM | USUnavailable | $1.75no add-ons | — | Receiptread 23:55 UTC |
 LiumMarketplace · Trust 40 · floor | 2× H100 SXM384 vCPU · 503.4 GB RAM | RUIn stock | $1.79no add-ons | Value 79 | Receiptread 23:48 UTC |
 HyperAIOperator · not scored | 1× H100 SXM80 GB per GPU | —Stock unknown | $1.80no add-ons | Value 79 | Receiptread 23:50 UTC |
 CUDO ComputeOperator · Trust 80 | 1× H100 SXM1 vCPU · 4 GB RAM | NorwayUnavailable | $1.82no add-ons | — | Receiptread 23:47 UTC |
 BeamOperator · Trust 76 | 1× H100 SXM80 GB per GPU | —Stock unknown | $1.83no add-ons | Value 78 | Receiptread 23:56 UTC |
| GPUBankOperator · Trust 93 | 1× H100 SXM20 vCPU · 125 GB RAM | France · SingaporeIn stock | $1.85no add-ons | Value 78 | Receiptread Oct 25 |
 Airon AIOperator · not scored | 8× H100 SXM2,048 GB RAM | SwedenStock unknown | $1.88no add-ons | Value 77 | Receiptread 23:55 UTC |
| GPUNETReseller · not scored · resold | 8× H100 PCIE128 vCPU · 1,536 GB RAM | USUnavailable | $1.88no add-ons | — | Receiptread 23:55 UTC |
 Voltage ParkOperator · Trust 89 | 8× H100 SXM208 vCPU · 928 GB RAM | USUnavailable | $1.89no add-ons | — | Receiptread 23:48 UTC |
| Voltage ParkOperator · Trust 89 | 1× H100 SXM26 vCPU · 116 GB RAM | USUnavailable | $1.89no add-ons | — | Receiptread 23:48 UTC |
| Voltage ParkOperator · Trust 89 | 4× H100 SXM104 vCPU · 464 GB RAM | USUnavailable | $1.89no add-ons | — | Receiptread 23:48 UTC |
| Voltage ParkOperator · Trust 89 | 2× H100 SXM52 vCPU · 232 GB RAM | USUnavailable | $1.89no add-ons | — | Receiptread 23:48 UTC |
 Vast.aiMarketplace · Trust 78 · floor | 1× H100 SXM24 vCPU · 283.4 GB RAM | IndiaIn stock | $1.93no add-ons | Value 76 | Receiptread 23:41 UTC |
 GPU.aiOperator · not scored | 1× H100 PCIE28 vCPU · 180 GB RAM | CanadaIn stock | $1.98no add-ons | Value 76 | Receiptread 23:49 UTC |
| GPU.aiOperator · not scored | 2× H100 PCIE60 vCPU · 360 GB RAM | CanadaIn stock | $1.98no add-ons | Value 76 | Receiptread 23:49 UTC |
| GPU.aiOperator · not scored | 4× H100 PCIE124 vCPU · 720 GB RAM | CanadaUnavailable | $1.98no add-ons | — | Receiptread 23:49 UTC |
| GPU.aiOperator · not scored | 8× H100 PCIE252 vCPU · 1,440 GB RAM | CanadaUnavailable | $1.98no add-ons | — | Receiptread 23:49 UTC |
- io.net$1.001× H100 PCIE · —UnavailableTrust 42Receipt →
- io.net$1.001× H100 PCIE · —UnavailableTrust 42Receipt →
- GPUBank$1.351× H100 PCIE · France · SingaporeIn stockTrust 93Value 84Receipt →
- io.net$1.591× H100 PCIE · —UnavailableTrust 42Receipt →
- Gcore$1.711× H100 SXM · —UnknownTrust 80Value 80Receipt →
- Gcore$1.712× H100 SXM · —UnknownTrust 80Value 80Receipt →
- Gcore$1.714× H100 SXM · —UnknownTrust 80Value 80Receipt →
- Gcore$1.718× H100 SXM · —UnknownTrust 80Value 80Receipt →
- GPUNET$1.751× H100 PCIE · USUnavailable
- Lium$1.792× H100 SXM · RUIn stockTrust 40Value 79Receipt →
- HyperAI$1.801× H100 SXM · —UnknownValue 79Receipt →
- CUDO Compute$1.821× H100 SXM · NorwayUnavailableTrust 80Receipt →
- Beam$1.831× H100 SXM · —UnknownTrust 76Value 78Receipt →
- GPUBank$1.851× H100 SXM · France · SingaporeIn stockTrust 93Value 78Receipt →
- Airon AI$1.888× H100 SXM · SwedenUnknownValue 77Receipt →
- GPUNET$1.888× H100 PCIE · USUnavailable
- Voltage Park$1.898× H100 SXM · USUnavailableTrust 89Receipt →
- Voltage Park$1.891× H100 SXM · USUnavailableTrust 89Receipt →
- Voltage Park$1.894× H100 SXM · USUnavailableTrust 89Receipt →
- Voltage Park$1.892× H100 SXM · USUnavailableTrust 89Receipt →
- Vast.ai$1.931× H100 SXM · IndiaIn stockTrust 78Value 76Receipt →
- GPU.ai$1.981× H100 PCIE · CanadaIn stockValue 76Receipt →
- GPU.ai$1.982× H100 PCIE · CanadaIn stockValue 76Receipt →
- GPU.ai$1.984× H100 PCIE · CanadaUnavailable
- GPU.ai$1.988× H100 PCIE · CanadaUnavailable
Showing 25 of 359 · Open in the comparison tool →
Marketplaces count once at their floor price in medians and rankings; every host offer stays listed · a reading older than 45 days, or superseded by a newer one for the same provider, GPU and mode, is kept on the GPU pages and never rankedBrowse by GPU →
Providers · 15 scored clouds with the most current offers · 173 with price readings · 262 listed in allPopular GPU providers
The scored clouds with the most configurations read in the last 45 days. Entry price, H100 price, catalogue breadth, footprint, billing and compliance as published, and the Trust Score built from their own documents.
Vast.aiMarketplace · Los AngelesVery wide catalogue: 81 GPU cards on the pricing page (site claims 68+ GPU types), from GTX 1060 to B300, with 'from' and median prices visible without an account
- From
- $0.03 / GPU·h
- H100
- $1.93
- Catalogue
- 92 GPU types
- Footprint
- 52 countries
Since 2016SpotSOC 2Per-second billing
Trust 787,995 current
GPUBankOperator · TallinnGPU rental, whole machine, traffic included, billed per second.
- From
- $0.23 / GPU·h
- H100
- $1.35
- Catalogue
- 9 GPU types
- Footprint
- 4 countries
Since 2026SpotReservedISO 27001Per-second billing
Trust 9363 current
AceCloudOperator · NoidaTwo India regions (Noida, Mumbai) plus Atlanta, with India-hosted data and INR billing with GST invoices
- From
- $0.97 / GPU·h
- H100
- —
- Catalogue
- 12 GPU types
- Footprint
- 2 countries
Since 2012SpotReservedISO 27001
Trust 731,624 current
Google CloudHyperscaler · Mountain ViewWidest GPU range of the set read: GB300 (A4X Max), GB200 (A4X), B200, H200, H100, A100 80GB/40GB, RTX PRO 6000, L4, T4, P4, V100, P100
- From
- $0.35 / GPU·h
- H100
- $11.06
- Catalogue
- 11 GPU types
- Footprint
- 22 countries
SpotReservedSOC 2ISO 27001Per-second billing
Trust 78357 current
VerdaOperator · HelsinkiSelf-service Blackwell access: GB300 at 9.72 USD/GPU-h, B300 at 8.45, B200 at 6.85 on demand in 1x-8x sizes
- From
- $0.24 / GPU·h
- H100
- $4.79
- Catalogue
- 15 GPU types
- Footprint
- 1 country
Since 2020SpotReservedServerlessSOC 2
Trust 81333 current
Widest GPU range in the study set: GB200, H200, H100 SXM, H100 NVL, MI300X, A100 (three variants), RTX PRO 6000 Blackwell, A10, T4, V100 plus AMD V710/V620/MI25 and NVIDIA M60 graphics GPUs
- From
- $0.34 / GPU·h
- H100
- $11.06
- Catalogue
- 14 GPU types
- Footprint
- 32 countries
SpotReservedSOC 2Per-second billing
Trust 81318 current
Wide self-serve catalogue: 17 GPU models bookable in the public console on 2026-09-29, from RTX A4000 to A100 SXM 80GB, in 1x to 10x VM configurations.
- From
- $0.44 / GPU·h
- H100
- $2.75
- Catalogue
- 27 GPU types
- Footprint
- 7 countries
Since 2018Per-minute billing
Trust 65318 current
Public, unauthenticated offers API (/api/gpu-offers) listing 331 live offers across 24 GPU types and 7 upstream providers at capture
- From
- $0.44 / GPU·h
- H100
- $2.64
- Catalogue
- 23 GPU types
- Footprint
- 12 countries
Spot
Trust 53272 current
RunpodOperator · MoorestownBroad self-serve catalogue: 21 Pod GPU configurations priced publicly, 43 GPU models in the GPU types reference including B300, B200, H200 and MI300X
- From
- $0.23 / GPU·h
- H100
- $2.69
- Catalogue
- 41 GPU types
- Footprint
- 11 countries
Since 2022ServerlessSOC 2ISO 27001Per-second billing
Trust 95268 current
AWSHyperscaler · SeattleLargest GPU catalogue among the providers reviewed: B300, B200, GB200, H200, H100, A100, L40S, L4, A10G, T4, RTX PRO 6000, RTX PRO 4500, V100 and fractional L4
- From
- $0.53 / GPU·h
- H100
- $6.88
- Catalogue
- 18 GPU types
- Footprint
- 6 countries
SpotReservedISO 27001Per-second billing
Trust 79233 current
Low published on-demand rates on older and consumer GPUs: V100 16GB $0.25, RTX 4090 $0.39, A100 SXM4 80GB $1.05 per GPU-hour
- From
- $0.25 / GPU·h
- H100
- —
- Catalogue
- 10 GPU types
- Footprint
- 8 countries
ReservedPer-second billing
Trust 61233 current
LeaderGPUOperator · AmsterdamBare-metal single-tenant servers with full root access, no hypervisor, no preemption (homepage FAQ).
- From
- $0.23 / GPU·h
- H100
- —
- Catalogue
- 21 GPU types
- Footprint
- 1 country
SpotReservedISO 27001
Trust 59194 current
HyperstackOperator · LondonPer-GPU list prices published for 13 GPU SKUs from RTX A4000 (0.15 USD/h) to B300 (7.40 USD/h), with separate reservation and spot tables
- From
- $0.50 / GPU·h
- H100
- $2.50
- Catalogue
- 10 GPU types
- Footprint
- 3 countries
Since 2020SpotReservedSOC 2Per-minute billing
Trust 81173 current
SeewebOperator · FrosinoneLow published EU prices: H100 SXM EUR 1.89/h (USD 2.166), spot EUR 0.76/h (USD 0.871), 12-month reservation EUR 1.61/h (USD 1.845).
- From
- $0.32 / GPU·h
- H100
- $2.11
- Catalogue
- 11 GPU types
- Footprint
- 1 country
Since 1998SpotReservedISO 27001
Trust 69138 current
UpCloudOperator · HelsinkiPrices for all GPU plans visible without an account in EUR and USD, per server-hour and as a monthly estimate
- From
- $0.64 / GPU·h
- H100
- $2.00
- Catalogue
- 6 GPU types
- Footprint
- 1 country
SpotISO 27001
Trust 75124 current
Ordered by configurations read in the last 45 days · Trust Score v0.1 reads six dimensions from each provider's own pages; “not disclosed” earns zero pointsMost trusted providers →
Models · 15 flagships across 12 categories · catalogue of 479 open models · news checked 2026-10-27Popular open models
The open-weight models teams deploy now: the newest flagship per family, then the image, video, speech and embedding models that run on rented GPUs. What each needs, and the cheapest real configuration that runs it today, per month at 730 hours.
DeepSeek V4.1 FlashDeepSeek · Chat and reasoningUpdated763.2B · 16B active · 1M ctx · mit · Oct 17, 2026
Runs from$5,676 / month
8× RTX PRO 6000 at UpCloudINT4checkpoint 510 GB · INT4-packed + FP8/BF16
Kimi K3Moonshot AI · Chat and reasoning2.8T · 104B active · 1M ctx · Kimi K3 License · Jul 27, 2026
Runs from$13,140 / month
8× MI325X at TensorWaveINT4checkpoint 1,560 GB · INT4-packed
Muse Spark Open 816B-A42BMeta · Chat and reasoningNew816B · 42B active · 1M ctx · Muse Model License · Oct 15, 2026
Runs from$8,972 / month
4× MI325X at Cyfuture AIFP8checkpoint 1,644 GB · BF16
MiniMax-M3.1-FlashMiniMax · Chat and reasoningNew
428B · 23B active · 1M ctx · MiniMax Community License · Oct 23, 2026
Runs from$1,754 / month
4× A100 80GB at Vast.aiINT4checkpoint 250 GB · INT4-packed
Beam 1.0 501B-A23BReflection AI · Chat and reasoningNew
501B · 23B active · 1M ctx · Apache-2.0 · Oct 16, 2026
Runs from$5,676 / month
8× RTX PRO 6000 at UpCloudFP8checkpoint 1,003 GB · BF16
Step 5 PreviewStepFun · Chat and reasoningNew
600B · 27B active · 1M ctx · StepFun Model License 1.0 · Oct 15, 2026
Runs from$8,972 / month
4× MI325X at Cyfuture AIFP8checkpoint 650 GB · FP8
Qwen3.8 27BQwen · Chat and reasoning27.8B · 256k ctx · apache-2.0 · Aug 2026
Runs from$78 / month
4× RTX 3060 at Vast.aiFP8checkpoint 56 GB · BF16
GLM-5.3 FlashZ.ai · Chat and reasoning321B · 18B active · 1M ctx · mit · Aug 25, 2026
Runs from$4,536 / month
2× MI325X at Cyfuture AIFP8checkpoint 328 GB · FP8
Gemma 4 31BGoogle · Chat and reasoning31.3B · 256k ctx · apache-2.0 · Apr 2, 2026
Runs from$78 / month
4× RTX 3060 at Vast.aiFP8checkpoint 63 GB · BF16
gpt-oss-120bOpenAI · Chat and reasoning117B · 5.1B active · 128k ctx · apache-2.0 · Aug 5, 2025
Runs from$196 / month
4× Tesla V100 32GB at Vast.aiINT4checkpoint 65 GB · INT4-packed
Mistral Large 3 675B Instruct 2512Mistral AI · Chat and reasoning675B · 41B active · 288k ctx · apache-2.0
Runs from$8,972 / month
4× MI325X at Cyfuture AIFP8checkpoint 682 GB · FP8
Qwen3 Coder NextQwen · Code79.7B · 3B active · 256k ctx · apache-2.0
Runs from$196 / month
4× Tesla V100 32GB at Vast.aiFP8checkpoint 159 GB · BF16
FLUX 3 Image KleinBlack Forest Labs · Image generationNew
5.6B · FLUX Kommunity License v1.0 · Oct 15, 2026
Runs from$26 / month
1× RTX 3060 at Vast.aiFP8checkpoint 43 GB · BF16
Wan2.2 T2V A14BWan · Video
27B · 14B active · apache-2.0 · Jul 28, 2025
Runs from$78 / month
4× RTX 3060 at Vast.aiFP8checkpoint 114 GB · BF16
whisper-large-v3-turboOpenAI · Speech to text0.809B · mit · Oct 1, 2024
Runs from$15 / month
1× GTX 1050 Ti at Clore.aiFP8checkpoint 2 GB · FP16
Selection rule: one flagship open-weight release per family, newest first, then the most deployed image, video, speech and embedding models · “New” within 45 days of a documented release or update · sizing = weights at the shown precision + 20% runtime headroomBrowse by category →
Workload finder · 477 open models · 477 cross-checked on Hugging Face Oct 28, 2026 · 312 with KV cache from config.json · 110 GPUs pricedWhich GPU fits your model?
Search any model of the catalogue, from the newest frontier MoE to image, video, speech and embedding models. We size the memory it needs at the chosen precision and context, then show the smallest real configuration on every GPU, priced from the cheapest eligible on-demand offer for that GPU count across every provider on file, with the alternatives, the read date and the receipt. Every step of the arithmetic is shown under the table.
Prices readAll models →
Model 1 in the catalogue
Precision published
Context
Qwen3 32BQwen · 32.8B · Chat and reasoning
44GB needed
weights 36.0 GB headroom 7.2 GB
Checked on Hugging Face Oct 28, 2026 · 32.8B params in the safetensors headers · 65.5 GB on disk · config read
73 configurations fit 44 GB at FP8 · 110 GPUs checked · prices from the cheapest eligible on-demand offer per GPU count, 45-day window, reference Nov 1, 2026
- RTX 3060Cheapest fit$0.114× 12 GB · 90% used, tight$78 / monthVast.aiTrust 78Receipt →
- Tesla V100 32GB$0.142× 32 GB · 68% used$99 / monthVast.aiTrust 78Receipt →
- Titan Xp$0.164× 12 GB · 90% used, tight$119 / monthVast.aiTrust 78Receipt →
- Tesla P100$0.194× 16 GB · 68% used$138 / monthVast.aiTrust 78Receipt →
- RTX 3090$0.222× 24 GB · 90% used, tight$157 / monthVast.aiTrust 78Receipt →
- Quadro RTX 8000$0.231× 48 GB · 90% used, tight$167 / monthVast.aiTrust 78Receipt →
- RTX 6000 Ada$0.251× 48 GB · 90% used, tight$183 / monthClore.aiTrust 46Receipt →
- Titan RTX$0.272× 24 GB · 90% used, tight$196 / monthVast.aiTrust 78Receipt →
| GPU | Smallest config | VRAM used | $ / hour · month | Cheapest at · alternatives | tok/s per user |
|---|---|---|---|---|---|
| RTX 3060Cheapest fit | 4× 12 GB48 GB total | 90% tight | $0.11$78 / month$0.03 per GPU | Vast.aiTrust 78read Oct 31 · receipt · also Clore.ai $0.17, Edgevana $0.40 · +1 more | — |
| Tesla V100 32GB | 2× 32 GB64 GB total | 68% | $0.14$99 / month$0.07 per GPU | Vast.aiTrust 78read Oct 31 · receipt · also Cyfuture AI $1.13, io.net $2.52 · +8 more | — |
| Titan Xp | 4× 12 GB48 GB total | 90% tight | $0.16$119 / month$0.04 per GPU | Vast.aiTrust 78read Oct 31 · receipt · no other provider at this count | — |
| Tesla P100 | 4× 16 GB64 GB total | 68% | $0.19$138 / month$0.05 per GPU | Vast.aiTrust 78read Oct 31 · receipt · also Edgevana $0.43 | — |
| RTX 3090 | 2× 24 GB48 GB total | 90% tight | $0.22$157 / month$0.11 per GPU | Vast.aiTrust 78read Oct 31 · receipt · also Clore.ai $0.29, Edgevana $0.40 · +2 more | — |
| Quadro RTX 8000 | 1× 48 GB48 GB total | 90% tight | $0.23$167 / month$0.23 per GPU | Vast.aiTrust 78read Oct 31 · receipt · also TensorDock $0.32 | — |
| RTX 6000 Ada | 1× 48 GB48 GB total | 90% tight | $0.25$183 / month$0.25 per GPU | Clore.aiTrust 46read Oct 31 · receipt · also gcube $0.40, Vast.ai $0.47 · +15 more | — |
| Titan RTX | 2× 24 GB48 GB total | 90% tight | $0.27$196 / month$0.13 per GPU | Vast.aiTrust 78read Oct 31 · receipt · also Edgevana $0.52 | — |
How these numbers are computed for Qwen3 32B
- Weights. The published checkpoint (BF16) weighs 36.0 GB on disk, for 32.8B parameters (safetensors files summed on Hugging Face).
- Runtime headroom. +20% = 7.2 GB for the CUDA context, activations and allocator fragmentation. Serving stacks reserve memory the same way (vLLM takes 90% of a card by default), so a fit above 90% is flagged tight.
- KV cache. 2 (K and V) × 8 KV heads × 128 dims × 2 bytes = 4,096 B per token per layer, over 64 full-attention layers. Pick a context above to add it; weights-only is the floor for short prompts. config.json ↗
- Fit and price. The smallest count among 1, 2, 4 and 8 GPUs of one type whose total memory covers the need (tensor parallelism splits weights and cache across the cards). The price is the cheapest eligible on-demand offer for exactly that GPU count among all 110 GPUs and every provider on file (in stock or stock unknown, newest reading per provider, no aggregator or outlier rows), times the count; monthly = 730 hours. The alternatives column shows the next providers at the same count.
- Not included. Batching beyond one sequence, speculative-decoding drafts, LoRA adapters, multi-node machines above 8 GPUs, and provider-side reservations. Throughput figures, where shown, are the internal study's single-stream estimates, dated and marked as such.
Qwen/Qwen3-32B ↗cross-checked Oct 28, 2026 · status: matchModel page with every precision and context →
Cheapest fit: RTX 3060 × 4 at $0.11 per hour at Vast.ai (read Oct 31).Sizing method →
Weights: published checkpoint size on Hugging Face at its own precision, other precisions estimated at 1 and 0.6 bytes per parameter · +20% runtime headroom · KV cache per token from config.json (GQA, MLA and sliding-window layers handled), one sequence, FP16 cache · prices: cheapest eligible on-demand offer for exactly that GPU count, 45-day window · throughput: internal study estimates, datedHow sizing works →
Market pulse · 92 days of daily medians · 33 GPUs tracked · 8 dated eventsWhere prices are heading
Three baskets, one index each, rebased to 100 on the first day of the history; add any tracked GPU in dollars per GPU-hour. Zoom, pan, pull the range handles, read every point, toggle the dated events and a trend projection.
Index · Aug 1, 2026 = 100 · 92 days to Nov 1, 2026 · drag to zoom, hold the hand tool to pan, pull the handles below
Datacenter6 GPUs
$3.92/ GPU·h median
+4.4%30 d+11.1%90 d
H100 · H200 · B200 · B300 · L40S · L4
Workstation4 GPUs
$0.52/ GPU·h median
—30 d—90 d
RTX PRO 6000 · RTX 6000 Ada · RTX A5000 · RTX A4000
Consumer5 GPUs
$0.30/ GPU·h median
+15.8%30 d—90 d
RTX 5090 · RTX 4090 · RTX 3090 · RTX 5080 · RTX 5070 Ti
What moved this month105 GPUs tracked · 12 published above 4 providers · 8 rising · 2 falling
- 1Tesla V100 32GBlow confidence$1.41−45.2%4
- 2L40Shigh confidence$1.31+22.1%10
- 3RTX 6000 Adamedium confidence$1.02+32.5%8
- 4B200medium confidence$7.67+30.4%9
- 5H200high confidence$4.63+15.7%13
- 6RTX PRO 6000high confidence$2.14+28.5%16
- 7RTX 3090low confidence$0.24+9.1%5
- 8H100high confidence$3.55+9.2%19
Median fresh on-demand price now against 30 and 90 days ago · n = providers behind the median · bars are proportional to the 30-day moveAll 12 movers →
Latest moves, sourcedeach entry links to the document it was read from
Timeline →- methodgpurank's own daily readings begin
From 2026-10-30, each day's median is computed from gpurank's own readings (the providers' pages and APIs read by its collector); earlier points come from a dated snapshot of 2026-10-28. The two use different provider sets, so a step at this date is a change of source, not of price.
gpurank methodology · checked Nov 1, 2026 → - priceNebius moves preemptible capacity to dynamic spot pricing
Spot H200 from $0.79, repriced every 15 minutes up to one cent below on-demand.
docs.nebius.com · checked Oct 16, 2026 ↗ - priceNebius raises on-demand prices by 17–21%
H100 $3.85 → $4.50, H200 $4.50 → $5.40, B200 $7.15 → $8.50, B300 $7.85 → $9.50 per GPU-hour.
docs.nebius.com · checked Oct 16, 2026 ↗
30 more dated entries in the internal chronology (financing, launches, capacity), on the market page
Index published only when a basket has at least 4 providers with current offers · movers need n ≥ 4 · 30- and 90-day changes as of Oct 28, 2026, from a dated snapshot series; gpurank's own daily readings (since Oct 30, 2026) take over once they cover 30 days · moves beyond ±50% are not published · the raw series are downloadable as CSVHow the index is built →
How we rank · same rules for every providerThree questions, one rule each
Rankings are computed, never negotiated. Here is what each podium measures, with today's answer for the H100, the six rules that apply to every row, and what a receipt looks like.
- Question 01
Cheapest
The lowest true cost per GPU-hour among offers read in the last 45 days and not marked unavailable, newest reading per provider. True cost is the list price plus the mandatory add-ons a provider publishes.
Rulelist price + storage while stopped + egress + public IPv4Today for the H100GPUBank$1.35/ GPU·h - Question 02
Best value
Throughput per dollar. A spec-based performance estimate (memory bandwidth for LLM inference, FP16 for training and image) divided by true cost, expressed as a percentile across all current offers.
Ruleproviders under Trust 60, or not scored, rank after the trusted ones, marked provisionalToday for the H100GPUBankValue 84/100 - Question 03
Most trusted
Trust Score v0.1: six dimensions read from the provider's own documents, with the evidence and its date. Company, pricing transparency, reliability, security, compliance, support.
Rule“not disclosed” earns zero pointsToday for the H100RunpodTrust 95/100
Six rules, applied to every rowApplied today to 25,339 current offers from 173 providers · versioned, changes logged
Full methodology →- 0145 days, dated rowsOnly readings from the last 45 days can be ranked, the newest per provider, GPU and mode; every row shows its read date. Older offers stay listed, in grey.
- 02One floor per marketplaceOn Vast.ai, Lium or Akash the floor price counts once in medians and podiums, never every host.
- 03One provider per podiumA provider appears once per podium; its other configurations are still in the table.
- 04Outliers suspendedPrices outside 0.3× to 3× the median are held back until a second reading confirms them.
- 05Resold capacity attributedWhen a provider resells another cloud's machines, the offer is listed under the operator and marked resold.
- 06Unavailable means unrankedAn offer the provider marks unavailable stays listed, in grey, and enters no ranking, median or podium until it is back in stock.
Receipt · offer #9000010As attached to every number
- Offer
- GPUBank · 1× H100 PCIE
- List price
- $1.35 / GPU·h
- Storage while stopped
- not published
- Egress
- not published
- Public IPv4
- not published
- True cost
- $1.35 / GPU·h
- Read
- Oct 25, 2026 · 12:00 UTC
- Source
- provider pricing page →
whole machine: vCPU, RAM, NVMe, IPv4 and traffic included
Independent, no sponsored rows, no pay-to-rank · some outbound links carry a commission, disclosed on each pageWhy every number has a receipt →
The month in numbers · November 2026 · computed from 25,339 current readingsWhat moved this month
Six signals computed from the dataset, no commentary: the two indices, the H100 median, the biggest riser and faller among published movers, and where Blackwell Ultra can be rented today.
Datacenter index+4.4%6 GPUs · median of daily medians · 30 daysConsumer index+15.8%5 GPUs · RTX 5090 to 3080 · 30 daysH100 median$3.09/ GPU·h−8.5%75 providers with a fresh offer · 30-day changeBiggest riserL40S+34.6%$0.97 → $1.31 · n = 10 · high confidenceBiggest fallerTesla V100 32GB−42.7%$2.46 → $1.41 · n = 4 · low confidenceB300 on demand$5.57/ GPU·h from29 providers rent Blackwell Ultra today · cheapest Seeweb
Questions people ask
Methodology →Where do the prices come from?
From the providers' own pricing pages and APIs, read on a schedule. Each number keeps the URL it was read from and the time of the reading. Prices relayed by third-party aggregators are shown for context but never ranked.
What is “true cost”?
The list price per GPU-hour plus the mandatory add-ons the provider publishes: storage billed while the instance is stopped, egress, a public IPv4, a minimum charge. The receipt attached to every number shows each line.
How fresh are the numbers?
Rankings only use readings from the last 45 days, the newest per provider, GPU and rental mode; every row shows its read date. Older offers stay visible in grey. The reference time of the current dataset is stamped on the rankings, the explorer and the bottom of every page.
Why is a cheaper offer not ranked first?
Because it is marked unavailable, is an older reading, is resold capacity, sits outside 0.3–3× the median, or is one host among many on a marketplace. The line under each podium counts what was left out and why.
Do providers pay to be listed or ranked?
No. Listing is free and rankings are computed from the rules above. Some outbound links carry a commission; this is disclosed on each page and has no effect on any ranking.
What does the Trust Score measure?
Six dimensions read from the provider's documents, each with its evidence and date: company, pricing transparency, reliability, security, compliance and support. It is not a review score, and anything not disclosed earns zero.
Indices: median of the members' daily medians, rebased in the market chart · movers published above 4 providers only · reference reading Nov 1, 2026Dated events and sources →