GPUレンタル価格

NVIDIA GB200 (per GPU, NVL72)

New floor: GMI Cloud now lists GB200 on-demand at $8.00/GPU-hr, marked 'AVAILABLE NOW' on GMI's own pricing page (confirmed today) — the first sub-$10 on-demand GB200 rate seen, undercutting CoreWeave's long-standing $10.50/GPU-hr listing (still live, $42/hr for a 4-GPU instance). Oracle Cloud bare-metal remains the best-corroborated mid-point at ~$16.00/GPU-hr, which pulls the reasoned median down to ~$13.25/GPU-hr (midpoint of the two confirmed on-demand configs) from last cycle's Oracle-anchored $16.00 read. Azure's ND GB200 v6 bare-metal is unchanged at $108.16/hr for a 4-GPU ND128isr node (re-confirmed today) = $27.04/GPU-hr, still the ceiling. AWS's 1-year reserved rate of $10.58/GPU-hr is unchanged, still the only published contract figure for this SKU.

出典: GetDeploying — GB200 price comparison, CoreWeave pricing, GMI Cloud pricing, Spare Cores — Azure ND128isr NDR GB200 v6 pricing

NVIDIA B200 (180GB)

Floor holds: DeepInfra's $3.69/GPU-hr on-demand listing (reconfirmed today, 'no wait times, no complex configs') remains the cheapest confirmed on-demand price, edging out Packet.ai's $3.75/GPU-hr Dynamic-tier rate. Verda's headline $3.12/GPU-hr is still a spot rate, not on-demand — Verda's own on-demand price is $6.23/GPU-hr, so it stays excluded from the floor. Typical price ticks up slightly to ~$6.25/GPU-hr (from $6.01): AIMultiple's tracked index (~$6.52) and Vast.ai's marketplace median (~$6.82) pull the cross-provider read above CoreWeave ($4.26-$8.60), Nebius ($3.95-$7.15) and Civo ($3.79-$4.49), while Silicon Data's index reads lower at ~$5.45 — within normal cross-tracker spread, not a directional signal. Ceiling unchanged: Google Cloud's a4-highgpu-8g still prices at $16.11/GPU-hr (re-confirmed today), above Oracle's bare-metal $14.00/GPU-hr. Still no clean published 1-year rate. Nov 2 refresh: GetDeploying shows a $7.20 provider-start median and $7.99 all-config median, with Packet.ai $3.75 lowest and AWS p6-b200 $14.24; our blended median moves to $6.90.

出典: GetDeploying — B200 price comparison, ComputePrices — B200 GPU price, DeepInfra — GPU instances, Verda pricing

AMD MI355X (288GB)

New floor: Daytona now lists MI355X on-demand at $5.99/GPU-hr (confirmed on Daytona's own pricing page today, distinct from its $3.44/GPU-hr preemptible tier) — a ~16% drop from the prior $7.10 floor and the first sub-$7 confirmed on-demand rate for this SKU. DigitalOcean also added MI355X this cycle, but only as a spot SKU at $4.50/GPU-hr, so it doesn't count as on-demand. Typical pricing eases to ~$7.00/GPU-hr as the new floor pulls the cluster down, with OVHcloud/Vultr on-demand ~$7.10 and CoreWeave ~$7.20 still anchoring the middle. Oracle bare-metal remains the ceiling at $8.60/GPU-hr. CoreWeave's $4.25/GPU-hr 1-year reservation remains the only published contract rate for this SKU.

出典: Daytona pricing, ComputePrices — MI355X price comparison, Spheron — AMD MI300X & MI355X pricing 2026, CoreWeave pricing

NVIDIA H200 (141GB)

Floor moves to Beam's serverless H200 rate of $2.09/GPU-hr (confirmed live on Beam's own pricing page today) — FluidStack's prior $2.30 floor could not be re-verified this cycle: FluidStack no longer has a discoverable public on-demand price page and appears to be sales-quote only, so it drops out of the floor. Typical price holds at ~$4.40/GPU-hr (GetDeploying's tracked median, re-confirmed today), with cheap capacity (GMI Cloud $2.60, Jarvislabs $3.99 on-demand/$1.99 spot) offset by mid-tier providers (Hyperbolic, Crusoe, DigitalOcean, Nebius) clustering $4.0-4.6. Hyperscaler ceiling unchanged: Azure's ND96isr H200 v5 still prices at $110.24/hr for 8 GPUs = $13.78/GPU-hr — no reserved SKU exists for that instance. The prior unsourced ~$3.72 reserved-rate estimate is replaced by a concrete figure: DigitalOcean's published 12-month reserved rate of $3.40/GPU-hr (confirmed live today). Nov 2 refresh: GetDeploying median $4.66/GPU-hr across 38 providers (lowest $2.09 unverified stock, Azure $13.78 ceiling). Nov 6 refresh: GetDeploying shows a $5.00 median across 38 providers (lowest $2.09 Beam unverified, $3.59 Runpod in stock; Azure $13.78 ceiling); our blended median moves to $4.80.

出典: GetDeploying — H200 price comparison, Vantage — Azure ND96isr H200 v5 instance pricing, Beam — pricing, DigitalOcean — GPU Droplets pricing

NVIDIA H100 (80GB)

Floor unchanged at Lium.io's $1.30/GPU-hr pool (not re-verified this cycle); ceiling unchanged at Azure ND96isr H100 v5 ($98.32/hr ÷ 8 = $12.29/GPU-hr), with AWS ~$6.88 and Google Cloud ~$11 below it. Typical price edged up to ~$3.49/GPU-hr (from $3.29): first-party lists show increases — Runpod Secure Cloud H100 SXM $3.99/hr (was $3.29; Community $2.69), Lambda H100 SXM $3.99-4.29, and Nebius HGX H100 raised from $3.85 to $4.50/GPU-hr effective Oct 1, 2026 — reversing part of the 2025 decline. DigitalOcean's 12-month reserved rate of $3.26/GPU-hr (last confirmed 28 Sep) is retained as the contract figure. Nov 2 refresh: GetDeploying median across 46 providers is $3.83/GPU-hr (lowest listed $1.75 PCIe, Google Cloud a3-highgpu-8g $11.07), so the median moves up from $3.49. Nov 6 refresh: GetDeploying lists a $3.99 provider median (lowest $1.80 HyperAI in stock, Google Cloud 8x SXM $11.07); our blended median moves to $3.90.

出典: GetDeploying — H100 price comparison, Vantage — Azure ND96isr H100 v5 instance pricing, Lium.io pricing, DigitalOcean — GPU Droplets pricing, Runpod pricing, Lambda pricing

AMD MI300X (192GB)

Floor and ceiling hold: TensorWave's sales-quote rate remains the effective floor at ~$1.71/GPU-hr (no self-serve checkout, re-confirmed today via GetDeploying and Thunder Compute), and Azure's ND96isr MI300X v5 list price is unchanged at $48-96/hr across regions ($6.00-$12.00/GPU-hr per cloudprice.net), still the ceiling. Typical price ticked up to ~$2.99/GPU-hr — GetDeploying's tracked median today, up from last cycle's $2.79 (+7%), within normal week-to-week noise rather than a directional signal — with the same low-cost cluster (Vultr ~$1.85, Hot Aisle ~$1.99, Runpod/GPU Outlet ~$2.39, DigitalOcean ~$2.59) still anchoring the middle below the hyperscaler premium. DigitalOcean's 12-month reserved rate holds at $1.91/GPU-hr (re-confirmed today).

出典: GetDeploying — MI300X price comparison, Thunder Compute — MI300X pricing, DigitalOcean — GPU Droplets pricing, cloudprice.net — Azure ND96isr MI300X v5

NVIDIA A100 (80GB)

Legacy workhorse holding roughly flat: floor stays at Vast.ai's marketplace ~$0.53/GPU-hr (range $0.39-0.67 depending on host reliability, unchanged), well below Thunder Compute's own $1.09/GPU-hr anchor and CoreWeave's newly-tracked $2.21/GPU-hr PCIe rate (confirmed today via CoreWeave's own pricing page, a new data point in the mid-cluster). GetDeploying's tracked median ticked up slightly to ~$1.75/GPU-hr (from last cycle's $1.69, +3.6%, within normal noise) with GCP a2-ultragpu still listing ~$5.07/GPU-hr at the top, ceiling unchanged. Civo's tiered 80GB rate card is unaffected — $1.79 on-demand, $1.59 at 12 months, down to $1.39 at 36 months; most other contract demand has moved to H100/H200.

出典: GetDeploying — A100 price comparison, CoreWeave pricing, Vast.ai live GPU pricing, Civo — NVIDIA A100 GPU cloud pricing

NVIDIA L40S (48GB)

Floor holds at $0.38/GPU-hr (Lium.io, re-confirmed today via ComputePrices' live tracker) — note Lium's own pricing page currently marks its lower $0.35/GPU-hr rate as 'reference price, unavailable' with no pods in stock right now, so $0.38 remains the operative floor rather than $0.35. Typical pricing holds at ~$1.53/GPU-hr — GetDeploying's tracked median reads flat (~$1.50) on a fresh check, no change; Runpod (~$0.79), gpu.ai/Vast.ai (~$0.80) and CoreWeave (~$0.99) still cluster below it. Ceiling holds near $3.5-3.8/GPU-hr (AWS us-east-1 8x config at $3.77; Oracle bare-metal at $3.50). Correction: the contract figure was previously sourced to Civo's 36-month tier ($0.89); Civo's full rate card, re-confirmed live today, is $1.29 on-demand → $1.19 at 6 months → $1.09 at 12 months → $0.99 at 24 months → $0.89 at 36 months, so the 1-year field now correctly uses the 12-month rate of $1.09.

出典: AIMultiple — GPU cloud price index, Lium.io pricing, ComputePrices — L40S price comparison, Civo — NVIDIA L40S GPU cloud pricing

NVIDIA RTX 5090 (32GB)

Consumer flagship on marketplaces only (no hyperscaler offering): Salad Cloud still floors the market at ~$0.25/GPU-hr, unchanged since launch and re-confirmed today via ComputePrices. Median edges up further to ~$0.67/GPU-hr (ComputePrices' live tracker today, corroborating GetDeploying's own read) from last cycle's $0.65 — a modest continuation of the demand-tightening trend flagged last cycle, not a new spike. Ceiling unchanged: Beam's serverless rate ($1.09/GPU-hr, re-confirmed on Beam's own pricing page today) remains the priciest widely-published on-demand option — Beam's own on-demand (non-serverless) tier still lists from ~$0.72/GPU-hr, well below its serverless ceiling. Nov 2 refresh: GetDeploying median $0.72/GPU-hr across 14 providers (lowest $0.35 HyperAI).

出典: GetDeploying — RTX 5090 price comparison, Vast.ai — RTX 5090 pricing, Beam — pricing, ComputePrices — RTX 5090 price comparison

NVIDIA RTX 4090 (24GB)

Ceiling corrected: the prior $1.61/GPU-hr top traced to a Beam guide page that has since gone dead (404, checked today); Beam's live pricing page now shows RTX 4090 at $0.69/GPU-hr serverless and $0.44/GPU-hr on-demand, both well below the old figure. The highest confirmed on-demand rate today is RunPod's Secure Cloud tier at $0.74/GPU-hr (RunPod Community Cloud runs $0.34/GPU-hr), so the ceiling moves there — a correction, not a market move. Vast.ai's marketplace floor is holding near ~$0.13-0.14/GPU-hr — marketplace pricing is volatile and hour-to-hour reads vary. Typical pricing holds near ~$0.42/GPU-hr, continuing a slow downward drift as Theta EdgeCloud, GPU.ai and Novita add cheap capacity. Novita's 12-month reserved rate of $0.27/GPU-hr remains the only contract figure tracked for this SKU.

出典: GetDeploying — RTX 4090 price comparison, RunPod — RTX 4090 pricing, Vast.ai live GPU pricing, Beam — pricing