GPU rent vs buy: H100, A100 and L4

Cloud list-price snapshot 2026-10-23 05:37:50 UTC · prices are USD per GPU-hour

Rent or buy? For H100, A100 80 GB and L4, the cheapest official on-demand and spot list prices on AWS, Azure and Google Cloud are compared with an explicitly sourced card-price and maximum-TDP assumption. The card purchase figure is not the cost of a functioning server: a GPU can require a compatible chassis, CPUs, RAM, cooling and facility power. The result is a card-only lower-bound illustration, not a buying recommendation.

The three electricity scenarios come from the U.S. Energy Information Administration, Electric Power Monthly, table 5.6.A: U.S. July 2026 national averages, released 24 September 2026 (preliminary; industrial USD 0.0977, commercial USD 0.1453, residential USD 0.1831 per kWh). They are not a forecast or your local power tariff.

H100: rent and buy

Cheapest on-demand: USD 6.8800 per GPU-hour (AWS us-east-1, p5.48xlarge). Cheapest spot: USD 1.2953 per GPU-hour (Google Cloud australia-southeast2, a3-highgpu-8g). These are different capacity types, not guarantees of availability; A100 here means the 80 GB variant, while H100 cloud rows here are the 80 GB variant only (the distinct 94 GB H100 NVL is excluded). The cited H100 purchase price does not specify card form factor, and rated SXM power is not necessarily the power of a rented machine.

List prices per chip across providers

Purchase assumption: USD 31,000 per GPU; CloudZero, H100 cost guide (updated 9 October 2026): about USD 31,000 for a new 80 GB card; card form factor unspecified. This is an illustrative hardware price, not an official MSRP or a complete host/server purchase.

Power assumption: 700 W per card at rated maximum; NVIDIA, H100 product specifications (undated; accessed 18 October 2026): H100 SXM maximum TDP up to 700 W; PCIe/NVL variants differ. TDP is a ceiling/design figure, not measured energy consumption of a particular workload.

Electricity for one running GPU-hour

First break-even month by utilisation and electricity rate

On-demand uses the lowest on-demand row across the providers above; spot uses the lowest spot row. The two prices may be from different regions and sizes. At zero resale value and PUE 1, the first month in which cumulative rental list price reaches card purchase plus electricity is shown. “Not within horizon / no list price” means beyond the API's 120-month limit, no positive hourly saving, or no listed cloud price; it is not a promise that owning will pay off. Spot interruption, time to re-run work and availability are not priced here.

A100 80GB: rent and buy

Cheapest on-demand: USD 3.4309 per GPU-hour (AWS us-east-1, p4de.24xlarge). Cheapest spot: USD 0.6788 per GPU-hour (Azure eastus, NC24ads_A100_v4). These are different capacity types, not guarantees of availability; A100 here means the 80 GB variant, while H100 cloud rows here are the 80 GB variant only (the distinct 94 GB H100 NVL is excluded). The cited H100 purchase price does not specify card form factor, and rated SXM power is not necessarily the power of a rented machine.

List prices per chip across providers

Purchase assumption: USD 10,000 per GPU; Jarvislabs, A100 price guide (updated 19 April 2026): the lower end of its new 80 GB SXM USD 10,000–15,000 range. This is an illustrative hardware price, not an official MSRP or a complete host/server purchase.

Power assumption: 400 W per card at rated maximum; NVIDIA, A100 product specifications (undated; accessed 18 October 2026): standard A100 80 GB SXM maximum TDP 400 W. TDP is a ceiling/design figure, not measured energy consumption of a particular workload.

Electricity for one running GPU-hour

First break-even month by utilisation and electricity rate

On-demand uses the lowest on-demand row across the providers above; spot uses the lowest spot row. The two prices may be from different regions and sizes. At zero resale value and PUE 1, the first month in which cumulative rental list price reaches card purchase plus electricity is shown. “Not within horizon / no list price” means beyond the API's 120-month limit, no positive hourly saving, or no listed cloud price; it is not a promise that owning will pay off. Spot interruption, time to re-run work and availability are not priced here.

L4: rent and buy

Cheapest on-demand: USD 0.7045 per GPU-hour (Google Cloud us-east4, g2-standard-4). Cheapest spot: USD 0.3068 per GPU-hour (Google Cloud asia-northeast3, g2-standard-4). These are different capacity types, not guarantees of availability; A100 here means the 80 GB variant, while H100 cloud rows here are the 80 GB variant only (the distinct 94 GB H100 NVL is excluded). The cited H100 purchase price does not specify card form factor, and rated SXM power is not necessarily the power of a rented machine.

List prices per chip across providers

Purchase assumption: USD 2,000 per GPU; Jarvislabs, L4 price guide (updated 19 April 2026): the lower end of its single-card USD 2,000–3,000 range. This is an illustrative hardware price, not an official MSRP or a complete host/server purchase.

Power assumption: 72 W per card at rated maximum; NVIDIA, L4 product specifications (undated; accessed 18 October 2026): maximum TDP 72 W. TDP is a ceiling/design figure, not measured energy consumption of a particular workload.

Electricity for one running GPU-hour

First break-even month by utilisation and electricity rate

On-demand uses the lowest on-demand row across the providers above; spot uses the lowest spot row. The two prices may be from different regions and sizes. At zero resale value and PUE 1, the first month in which cumulative rental list price reaches card purchase plus electricity is shown. “Not within horizon / no list price” means beyond the API's 120-month limit, no positive hourly saving, or no listed cloud price; it is not a promise that owning will pay off. Spot interruption, time to re-run work and availability are not priced here.

Method and exclusions

One modelling month is 730 hours (365 days × 24 hours ÷ 12 months), and utilisation is the share of those hours actively billed/used. For each chip the monthly saving is (cloud list USD/GPU-hour − watts/1000 × USD/kWh) × 730 × utilisation; break-even is the first whole month with accumulated savings at least the assumed purchase price, searched through 120 months. This follows the GPU prices API with resale_pct=0 and pue=1; for H100 use min_gpu_mem_gb=80 and inspect gpu_mem_gb in the returned rows to exclude 94 GB NVL (the API has no exact-memory filter). It omits idle electricity, CPU/host purchase and power, cooling above PUE 1, hosting, financing, maintenance, taxes, networking, egress, discounts and interruption costs. Cloud rates include machine CPU/RAM share as described in the source snapshot, but the ownership estimate does not. This asymmetry makes ownership look cheaper than a like-for-like build.

Run this comparison for your inputs

The API answers it for your purchase price, watts, tariff, GPU count, utilisation and horizon, on the newest snapshot. Use pricing="on_demand" or pricing="spot" to keep capacity types separate; pricing="any" takes the lower priced kind. The response contains monthly_cloud_usd, monthly_power_usd, monthly_ownership_usd and break_even_month.

POST https://tanod.dev/v1/compute/gpu-prices
Content-Type: application/json

{"chip": "H100", "pricing": "on_demand", "rent_vs_buy": {"purchase_usd": 31000.0, "watts_per_gpu": 700.0, "usd_per_kwh": 0.1453, "hours_per_month": 365.0, "months": 120, "resale_pct": 0, "pue": 1}}

See H100, A100 and L4 price guides; all guides. Estimates, not advice.