Answers

40 short answers to what renters ask, each worked out from today's prices: the number first, then the reasoning and the offers behind it.

Prices and what they buy

• Are LLM input and output tokens priced equally?Of the 79 model API rates we track, 5 charge the same for input and output tokens.
• Can you rent a GPU by the minute?8 providers we track sell serverless GPUs that bill only for running time, across 16 GPU models.
• Do GPU rental prices include tax?$1.99 per GPU-hourFor example, the cheapest on-demand H100 today is $1.99 per GPU-hour before tax (Voltage Park).
• Does a lower cheapest GPU price mean the whole market fell?Compared with 6 October 2026, 49 of today's 443 offers are new listings and 52 listings from then are gone.
• GPU Price Per Month vs Per Hour: Which Should You Use?$1.99 an hourThe cheapest on-demand H100 at $1.99 an hour (Voltage Park) costs $1,453 for a full 730-hour month, or $350.24 at 8 hours a day on 22 working days.
• How do you compare monthly and annual GPU rental prices?49%Against the same provider's on-demand price for the same GPU, today's reserved rates run: 1-month terms 49% below (3 rates) and 12-month terms 65% below (3 rates).
• How Much Does an A100 Cost Per Hour?Compare cheapest A100 cloud rental prices per hour and the factors that move the number.
• How much does an eight-H100 node cost?The hourly price of a whole eight-GPU H100 node, from the eight-GPU configurations providers actually list, with the per-GPU rate beside it.
• How Much Does an H100 Cost Per Hour?Live hourly rental prices for the NVIDIA H100 and the factors that move the number.
• How Much Does an RTX 4090 Cost to Rent?Compare cheapest RTX 4090 cloud rental prices per hour and why the consumer card is such good value.
• How Much Does It Cost to Train an LLM?$19,900A run of 10,000 H100 GPU-hours costs $19,900 at the cheapest on-demand rate today (Voltage Park) and $33,300 at the median provider's rate, so the choice of provider alone moves the bill by $13,400.
• How should you compare GPU prices across regions?$3.37On-demand H100 offers with a published location cover 3 locations, from $3.37 (Latitude.sh, United States) to $12.29 (Microsoft Azure, East US).
• Is a GPU cloud deposit part of the rental price?$10Of the 18 providers whose billing terms we checked (7 July 2026), Jarvislabs ($10) and OVHcloud ($30 credit charge for first Public Cloud project) publish a minimum deposit; the other 16 publish none.
• Is Renting a GPU Cheaper Than Buying One?$35,316An H100 bought for $35,316 (the one listing we track) costs as much as 17,747 hours of its cheapest on-demand rental ($1.99 an hour), about 6.1 years at 8 hours a day.
• Is the A100 Still Worth Renting?$1.09 per GPU-hourAn 80 GB A100 rents from $1.09 per GPU-hour today (Thunder Compute), 55% of the cheapest on-demand H100 at $1.99 (Voltage Park).
• MI300X vs H100: Which Is Cheaper to Rent?Compare live rental prices for AMD's MI300X against the NVIDIA H100, plus how the two cards differ.
• Should checkpoint and restart time be in a GPU budget?$31.92 an hourOn the cheapest eight-GPU H100 node today ($31.92 an hour, Lambda), every 10 minutes spent writing checkpoints costs $5.32.
• What does 100 GB of GPU cloud egress cost?100 GBOf 18 providers whose terms we checked (7 July 2026), 6 publish free outgoing transfer, so 100 GB adds nothing: RunPod, Lambda, CoreWeave, Voltage Park, Crusoe and Massed Compute.
• What Is a GPU Hour?$0.10One on-demand GPU-hour costs from $0.10 (RTX A4000, TensorDock) to $16.11 (B200, Google Cloud), across 218 on-demand offers.
• What is a minimum GPU-hour commitment?Of the 37 reserved rates we track today, none is sold as a volume of GPU-hours; 23 are calendar terms of 1 to 60 months, most often 12 months (6), and 14 need a quote.
• What Is the Cheapest 140GB GPU to Rent?Compare the lowest listed rates on 140GB-class cloud GPUs like the H200, B200, and MI300X.
• What Is the Price Difference Between the H100 and H200?How H200 rental rates compare to the H100, and when the premium is worth it.
• Why Are GPU Prices Different Between Providers?$1.99Each provider's cheapest on-demand rate, today: H100 from $1.99 (Voltage Park) to $9.80 (Google Cloud) across 24 providers, 4.9 times apart.
• Why is a GPU price listed when it is out of stock?Of today's 443 offers, 11 are reported out of stock and 19 on a waitlist, each still with its published price; 42 are reported in stock, and the rest do not report stock at all.

Memory and models

• Can you add VRAM across rented GPUs?$2.70 an hourTwo 80 GB cards rented one at a time cost from $2.70 an hour today (2 × A100, Massed Compute); one MI300X with 192 GB on a single card costs from $2.59 (DigitalOcean).
• Does four-bit quantization use a quarter of 16-bit memory?25%Across the 33 models we estimate at 4,096 tokens of context, the 4-bit total is 25% to 43% of the 16-bit total.
• How does context length affect GPU memory estimates?0.5 GBFor DeepSeek R1 Distill Llama 8B, the KV cache takes 0.5 GB at 4,096 tokens and 17.2 GB at 131,072 tokens, 4.3 times its 4-bit weights (4.0 GB).
• How Much VRAM Do I Need to Run DeepSeek V3?336 GBDeepSeek V3, with 671 billion parameters, needs about 336 GB at 4-bit and 1,342 GB at 16-bit (weights alone, before cache and headroom).
• Is 24GB of VRAM Enough for Running an LLM?24 GBAt 4-bit with 4,096 tokens of context, 28 of the 33 models we estimate fit in 24 GB; the largest is Qwen2.5 32B at 20.9 GB.
• What GPU Do I Need for a 70B Model?43.9 GBA 70B model such as DeepSeek R1 Distill Llama 70B needs about 43.9 GB at 4-bit and 171 GB at 16-bit (our inference estimate at 4,096 tokens with 20% headroom).
• What GPU Do I Need to Run Llama 405B?203 GBLlama 3.1 405B needs about 203 GB at 4-bit and 812 GB at 16-bit (weights alone, before cache and headroom).
• Why does an LLM need more memory than its weights?35.3 GBFor DeepSeek R1 Distill Llama 70B at 4-bit, the weights take 35.3 GB, the cache for 4,096 tokens 1.3 GB, and with 20% runtime headroom the estimate is 43.9 GB.

Billing, fees and terms

• Do storage charges stop when a rented GPU stops?Of 18 providers whose terms we checked (7 July 2026), 8 keep billing storage after the GPU stops (RunPod, Lambda, Hyperstack, Nebius, Crusoe, Jarvislabs, Microsoft Azure and AWS), Vultr keeps billing the instance, Modal bills nothing while stopped, and 8 do not publish it.
• Does GPU rental pricing include CPU and RAM?48 of today's 443 offers publish the vCPUs that come with the GPU, and 34 publish the RAM.
• Serverless GPU vs Instance Pricing: What Is the Difference?$2.50The cheapest serverless H100 costs $2.50 for each hour a worker runs, against $1.99 for an always-on instance, so serverless is cheaper while its workers run less than 80% of the time.
• What Is the Difference Between Secure Cloud and Community Cloud?19 community-cloud rates can be matched with the same provider's secure or on-demand rate for the same GPU.
• When does a reserved GPU cost less than on-demand?35%Reserved rates cost 35% to 85% of the same provider's on-demand rate (median 70%, 16 rates from Hyperstack and Latitude.sh).

More questions

• Can you rent two GPUs instead of an eight-GPU node?Of today's 20 on-demand H100 offers with a published node size, 9 rent a single GPU, 3 rent 2 GPUs at once, 2 rent 4 GPUs at once and 6 rent 8 GPUs at once.
• Multi-GPU Node Pricing ExplainedHow providers price 2, 4, and 8 GPU nodes, and what the per-hour rate includes.
• What Is the Difference Between Spot and On-Demand GPU Pricing?How spot and on-demand GPU rates differ, and when each one is the right choice.

Figures from our price checks of 13 October 2026; each answer shows its sources and the offers behind it.