01EGPU Rental Prices· as of 2026-10-24
GPU rental prices vs. depreciation
Do rented GPUs lose value on schedule? Each provider's own list price for one offer, followed through 277 archived quarterly captures from 2021Q4 to 2026Q3, set against the straight line an asset written off over 6 years would follow. Archived list prices from 9 providers' own pricing pages and price lists, not transactions; kept apart from the site's current quotes.
What the archive shows
- Prices rose in 2026. 20 of 23 specialist-cloud offers list a higher price at their latest 2026 capture (2026Q3) than at their first 2026 capture, a median change of +33%; 3 held and none fell. The rises span B300, B200, H200, H100, A100 80GB, MI300X, L40S and V100.
- Most sit above a 6-year write-down. 19 of 20 specialist offers tracked for at least 2 quarters list above the 6-year straight line from their first archived price (median +60%), and 20 of 20 above a 5-year line.
- Before 2026, several were below it. 5 of 20 specialist offers were priced under the 6-year line at most of their captures after the first (H100 at Hyperstack, H100 at Nebius, H100 at Verda, L40S at Verda, V100 at Verda). Their latest captures are above the line again only after the 2026 rises.
- Big-cloud list prices barely move. 11 of 11 big-cloud offers kept the same list price through 2026. Across the whole archive 3 of 10 ever changed, each by a cut. A list price that stands still sits above the line by construction, and it leaves out negotiated discounts, so their distance from the line says little about the market.
Same-offer list price change in 2026
Each specialist-cloud offer from its first 2026 capture to its latest. Same provider, same configuration, one comparable per-GPU-hour basis, so a provider joining or leaving cannot register as a price move. Big-cloud list prices are left out: 11 of 11 did not change.
List price against a straight-line write-down
Each line is one offer’s list price as a share of its own first archived price (100%), by years since that capture. The dashed lines are the straight-line paths from 100% to zero over 6 and 5 years, the same for every offer in these units: a line above them has held its price better than a write-down on that schedule. Points are quarterly captures; the line between two captures does not show when a price changed.
Every offer, exact values
Prices are USD per GPU-hour, linked to the archived copy that shows them. The path columns need at least 2 quarters between the first and latest capture; “Below” counts the captures after the first priced under the 6-year path. Every quarter of each offer is on the archived prices page.
| Offer | First archived | Latest | 6-yr path at latest | vs 6-yr path | vs 5-yr path | Below 6-yr path | Change in 2026 |
|---|---|---|---|---|---|---|---|
| B300 · VerdaSpecialist cloud · 1x B300 SXM6 268GB | $4.95 2026Q1 | $7.50 2026Q30.5 yr later | $4.54 | +65% | +68% | 0 of 2 | +52%since 2026Q1 |
| B300 · AWSBig cloud · p6-b300.48xlarge | $17.80 2026Q2 | $17.80 2026Q30.3 yr later | — | — | — | — | 0%since 2026Q2 |
| B200 · LambdaSpecialist cloud · NVIDIA B200 SXM6 (8x tab) | $4.99 2025Q4 | $6.69 2026Q30.7 yr later | $4.38 | +53% | +57% | 0 of 3 | +34%since 2026Q1 |
| B200 · NebiusSpecialist cloud · NVIDIA HGX B200 | $5.50 2025Q3 | $7.15 2026Q31.0 yr later | $4.60 | +55% | +62% | 0 of 4 | +30%since 2026Q1 |
| B200 · Together AISpecialist cloud · NVIDIA HGX B200 | $5.50 2025Q4 | $8.19 2026Q30.8 yr later | $4.80 | +71% | +76% | 0 of 3 | +49%since 2026Q1 |
| B200 · VerdaSpecialist cloud · 1x B200 SXM6 180GB | $3.64 2025Q3 | $6.11 2026Q31.0 yr later | $3.05 | +101% | +109% | 0 of 4 | +53%since 2026Q1 |
| B200 · AWSBig cloud · p6-b200.48xlarge | $14.24 2025Q3 | $14.24 2026Q31.0 yr later | $11.87 | +20% | +25% | 0 of 4 | 0%since 2026Q1 |
| H200 · HyperstackSpecialist cloud · NVIDIA H200 SXM | $3.50 2025Q3 | $3.99 2026Q31.0 yr later | $2.93 | +36% | +42% | 0 of 4 | +14%since 2026Q1 |
| H200 · JarvislabsSpecialist cloud · NVIDIA H200 SXM | $3.80 2026Q2 | $3.99 2026Q30.3 yr later | — | — | — | — | +5%since 2026Q2 |
| H200 · NebiusSpecialist cloud · NVIDIA HGX H200 | $3.50 2024Q4 | $4.50 2026Q31.8 yr later | $2.47 | +82% | +98% | 0 of 7 | +29%since 2026Q1 |
| H200 · Together AISpecialist cloud · NVIDIA HGX H200 | $3.79 2025Q4 | $5.99 2026Q30.8 yr later | $3.31 | +81% | +87% | 0 of 3 | +58%since 2026Q1 |
| H200 · VerdaSpecialist cloud · 1x H200 SXM5 141GB | $3.59 2024Q4 | $4.00 2026Q31.7 yr later | $2.55 | +57% | +70% | 3 of 7 | +34%since 2026Q1 |
| H200 · AWSBig cloud · p5en.48xlarge | $10.60 2025Q1 | $7.91 2026Q31.4 yr later | $8.09 | −2% | +4% | 5 of 6 | 0%since 2026Q1 |
| H200 · Google CloudBig cloud · a3-ultragpu-8g | $10.60 2025Q2 | $10.60 2026Q31.4 yr later | $8.14 | +30% | +39% | 0 of 5 | 0%since 2026Q1 |
| H100 · HyperstackSpecialist cloud · NVIDIA H100 SXM | $3.75 2024Q3 | $3.20 2026Q31.9 yr later | $2.53 | +26% | +40% | 7 of 8 | +33%since 2026Q1 |
| H100 · JarvislabsSpecialist cloud · NVIDIA H100 SXM | $2.69 2026Q2 | $2.69 2026Q30.3 yr later | — | — | — | — | 0%since 2026Q2 |
| H100 · LambdaSpecialist cloud · NVIDIA H100 SXM (8x tab) | $2.59 2023Q3 | $3.99 2026Q33.0 yr later | $1.31 | +206% | +281% | 0 of 12 | +33%since 2026Q1 |
| H100 · NebiusSpecialist cloud · NVIDIA HGX H100 | $4.85 2023Q4 | $3.85 2026Q32.8 yr later | $2.59 | +49% | +80% | 6 of 11 | +31%since 2026Q1 |
| H100 · Together AISpecialist cloud · NVIDIA HGX H100 | $2.99 2025Q4 | $3.99 2026Q30.8 yr later | $2.61 | +53% | +58% | 0 of 3 | +33%since 2026Q1 |
| H100 · VerdaSpecialist cloud · 1x H100 SXM5 80GB | $4.57 2024Q1 | $3.25 2026Q32.4 yr later | $2.71 | +20% | +39% | 9 of 10 | +42%since 2026Q1 |
| H100 · AWSBig cloud · p5.48xlarge | $12.29 2023Q3 | $6.88 2026Q33.0 yr later | $6.14 | +12% | +40% | 3 of 12 | 0%since 2026Q1 |
| H100 · Google CloudBig cloud · a3-highgpu-8g | $11.06 2024Q2 | $11.06 2026Q32.3 yr later | $6.75 | +64% | +88% | 0 of 9 | 0%since 2026Q1 |
| H100 · Google CloudBig cloud · a3-megagpu-8g | $11.68 2025Q2 | $11.68 2026Q31.4 yr later | $8.97 | +30% | +39% | 0 of 5 | 0%since 2026Q1 |
| A100 80GB · HyperstackSpecialist cloud · NVIDIA A100 SXM | $1.60 2025Q3 | $1.60 2026Q31.0 yr later | $1.34 | +20% | +24% | 0 of 4 | 0%since 2026Q1 |
| A100 80GB · JarvislabsSpecialist cloud · NVIDIA A100 80GB | $1.49 2026Q2 | $1.49 2026Q30.3 yr later | — | — | — | — | 0%since 2026Q2 |
| A100 80GB · LambdaSpecialist cloud · NVIDIA A100 SXM (8x tab) | $1.50 2023Q1 | $2.79 2026Q33.5 yr later | $0.63 | +343% | +512% | 0 of 14 | +56%since 2026Q1 |
| A100 80GB · VerdaSpecialist cloud · 1x A100 SXM4 80GB | $2.20 2021Q4 | $1.79 2026Q34.6 yr later | $0.50 | +260% | +1040% | 0 of 14 | +39%since 2026Q1 |
| A100 80GB · AWSBig cloud · p4de.24xlarge | $5.12 2022Q2 | $3.43 2026Q34.3 yr later | $1.49 | +130% | +348% | 0 of 17 | 0%since 2026Q1 |
| A100 80GB · Google CloudBig cloud · a2-ultragpu-8g | $5.07 2023Q1 | $5.07 2026Q33.6 yr later | $2.01 | +152% | +261% | 0 of 14 | 0%since 2026Q1 |
| MI300X · Hot AisleSpecialist cloud · VM Small 1x MI300x | $1.99 2025Q3 | $2.99 2026Q31.0 yr later | $1.67 | +79% | +86% | 0 of 4 | +50%since 2026Q1 |
| L40S · VerdaSpecialist cloud · 1x L40S 48GB | $1.36 2024Q2 | $1.37 2026Q32.3 yr later | $0.84 | +63% | +86% | 6 of 9 | +51%since 2026Q1 |
| L40S · AWSBig cloud · g6e.xlarge | $1.86 2024Q3 | $1.86 2026Q32.0 yr later | $1.24 | +50% | +67% | 0 of 8 | 0%since 2026Q1 |
| V100 · VerdaSpecialist cloud · 1x Tesla V100 16GB | $0.85 2021Q4 | $0.17 2026Q34.6 yr later | $0.19 | −12% | +180% | 9 of 14 | +21%since 2026Q1 |
| V100 · AWSBig cloud · p3.16xlarge | $3.06 2021Q4 | $3.06 2026Q34.7 yr later | $0.64 | +379% | +1880% | 0 of 19 | 0%since 2026Q1 |
How this is measured, and what it does not show
Same offer. A series is one provider’s on-demand list price for one configuration, read from archived copies of its own pricing page (the Wayback Machine capture nearest mid-quarter) or its price list (AWS’s dated Price List versions). Prices on another basis, such as per-second serverless rates or week-minimum chassis, are left out, and so are the providers whose terms bar reuse of their prices. Where a page renamed an offer, the archive’s own record of the instance shape decides whether the two labels are one series (18 series are rejoined; details below).
The path. The benchmark starts at the offer’s first archived price and falls in a straight line to zero over a useful life: 6 years, the latest life Microsoft, Alphabet, CoreWeave and Oracle state for their servers, and 5 years, the latest for Amazon and Nebius (Amazon’s for part of its fleet). The clock starts at the first archived capture, not at the chip’s launch, which keeps the path as high as the data allows: a price above it is above it on the benchmark’s most favourable reading.
A benchmark, not a forecast. Straight-line depreciation is how an owner writes an asset’s cost down in its accounts. A rental price also covers power, space, staff, financing and margin, so it need not fall to zero even on a written-off chip, and a price above the line does not by itself mean an owner earns back the chip’s cost. The comparison asks one thing: has the price fallen as fast as the book value of a chip on that schedule would?
List prices, sampled quarterly. These are advertised on-demand rates, not what buyers paid: discounts, reserved terms and negotiated contracts are not visible, and transaction-based indexes can differ. One capture a quarter misses changes in between. A posted price never shows available capacity. The current daily quotes are on the GPU rental prices page; they are a separate dataset and are not joined to these.
Server useful lives in company filings
The lives large cloud operators state for the servers that hold these chips, each from the filing that states the change (checked Oct 27, 2026). Oldest change first.
| Company | Life | Effective | Filing |
|---|---|---|---|
| MicrosoftServer and network equipment | 4 → 6 yr | Jul 1, 2022 | Form 10-K, fiscal year ended June 30, 2022“In July 2022, we completed an assessment of the useful lives of our server and network equipment.” |
| AlphabetServers (certain network equipment from five years) | 4 → 6 yr | Jan 1, 2023 | Form 10-K, fiscal year 2023“adjusted the estimated useful life of our servers from four years to six years” |
| CoreWeaveComputing equipment utilized in data centers | 5 → 6 yr | Jan 1, 2023 | Form 10-K, fiscal year 2025“Effective January 1, 2023, the Company changed its estimate of the useful life for its computing equipment … utilized in data centers from five to six years.” |
| AmazonServers | 5 → 6 yr | Jan 1, 2024 | Form 10-K, fiscal year 2024“We had previously increased the useful life of our servers from five years to six years effective January 1, 2024.” |
| OracleServers and networking equipment | 5 → 6 yr | Jun 1, 2024 | Form 10-K, fiscal year ended May 31, 2025“increased the estimate of the useful lives from five years to six years, … effective at the beginning of fiscal 2025.” |
| AmazonA subset of servers and networking equipment | 6 → 5 yr | Jan 1, 2025 | Form 10-K, fiscal year 2024“effective January 1, 2025, from six years to five years” |
| MetaMost servers and network assets | 5 → 5.5 yr | Jan 1, 2025 | Form 10-Q, first quarter 2025“an increase in the estimated useful lives of most servers and network assets to 5.5 years, effective January 1, 2025.” |
| NebiusServer and network equipment | 4 → 5 yr | Jan 1, 2026 | Form 6-K, first quarter 2026 results“Starting the first quarter of 2026, we revised the useful life for our server and network equipment … from four years to five years” |
Renamed offers joined into one series
- A100 80GB · Lambda
- “8x NVIDIA A100” → “8x NVIDIA A100 SXM” → “On-demand 8x NVIDIA A100 SXM” → “NVIDIA A100 SXM (8x tab)”Lambda's on-demand 8-GPU A100 SXM instance, per GPU-hour: priced per instance (÷ 8) through 2024Q2, per GPU from 2024Q3; the label changed with each page redesign.
- H100 · Lambda
- “8x NVIDIA H100 SXM New” → “8x NVIDIA H100 SXM New Price” → “On-demand 8x NVIDIA H100 SXM” → “NVIDIA H100 SXM (8x tab)”Lambda's on-demand 8-GPU H100 SXM instance, per GPU-hour: priced per instance (÷ 8) through 2024Q2, per GPU from 2024Q3; the label changed with each page redesign.
- B200 · Lambda
- “On-demand 8x NVIDIA B200 SXM6” → “NVIDIA B200 SXM6 (8x tab)”Lambda's on-demand 8-GPU B200 SXM6 instance, per GPU-hour, before and after the move to lambda.ai/pricing.
- A100 80GB · Verda
- “1A100.22V” → “1x A100 SXM4 80GB”Verda's (formerly DataCrunch) 1-GPU A100 80GB instance with 22 vCPUs, on-demand: the instance code on datacrunch.io, the product name on verda.com.
- H100 · Verda
- “1H100.80S.30V” → “1x H100 SXM5 80GB”Verda's 1-GPU H100 SXM5 instance with 30 vCPUs, on-demand (the 45-vCPU 2023 instance is a separate series).
- H200 · Verda
- “1H200.141S.44V” → “1x H200 SXM5 141GB”Verda's 1-GPU H200 SXM5 instance with 44 vCPUs, on-demand.
- B200 · Verda
- “1B200.30V” → “1x B200 SXM6 180GB”Verda's 1-GPU B200 SXM6 instance with 30 vCPUs, on-demand (the 31-vCPU 2025Q2 instance is a separate series).
- B300 · Verda
- “1B300.30V” → “1x B300 SXM6 268GB”Verda's 1-GPU B300 SXM6 instance with 30 vCPUs, on-demand (pay-as-you-go column).
- L40S · Verda
- “1L40S.20V” → “1x L40S 48GB”Verda's 1-GPU L40S instance with 20 vCPUs, on-demand.
- V100 · Verda
- “1V100.6V” → “1x Tesla V100 16GB”Verda's 1-GPU V100 instance with 6 vCPUs, on-demand.
- H100 · Hyperstack
- “NVIDIA H100 SXM 80GB” → “NVIDIA H100 SXM”Hyperstack's on-demand H100 SXM, per GPU-hour; the page dropped the memory size from the label in 2025Q3.
- H100 · Nebius
- “H100 SXM5” → “8 × H100 SXM” → “NVIDIA H100 GPU” → “NVIDIA HGX H100”Nebius's on-demand H100 SXM (HGX) GPU, per GPU-hour: listed per GPU on nebius.ai from 2023Q4 to 2024Q3, per 8-GPU instance (÷ 8) in 2024Q4, per GPU since.
- H200 · Nebius
- “8 × H200 SXM” → “NVIDIA H200 GPU” → “NVIDIA HGX H200”Nebius's on-demand H200 SXM (HGX) GPU, per GPU-hour: per 8-GPU instance (÷ 8) in 2024Q4, per GPU since.
- H100 · Together AI
- “NVIDIA HGX H100 SXM” → “NVIDIA HGX H100”Together AI's on-demand HGX H100, per GPU-hour; the page dropped “SXM” from the label in 2026Q2.
- A100 80GB · Jarvislabs
- “A100-80GB (On-Demand VM)” → “NVIDIA A100 80GB”Jarvislabs' on-demand 1-GPU A100 80GB VM; the 2026Q3 page reads it with the on-demand, 1× option selected.
- H100 · Jarvislabs
- “H100 (On-Demand VM)” → “NVIDIA H100 SXM”Jarvislabs' on-demand 1-GPU H100 VM; the 2026Q3 page reads it with the on-demand, 1× option selected (its earlier bare-metal VM is a separate series).
- H200 · Jarvislabs
- “H200 (On-Demand VM)” → “NVIDIA H200 SXM”Jarvislabs' on-demand 1-GPU H200 VM; the 2026Q3 page reads it with the on-demand, 1× option selected (its earlier bare-metal VM is a separate series).
- MI300X · Hot Aisle
- “1xVM (1x MI300x Virtual Machine)” → “Small 1x 192GB MI300x Virtual Machine (on-demand)” → “VM Small 1x MI300x”Hot Aisle's 1-GPU MI300X virtual machine with no contract (its one-week 8-GPU chassis and Shadeform container are separate series).