AWS GPU Pricing

GPU pricing on AWS is the one corner of the cloud where the old rule, that prices only ever fall, stopped holding. In June 2025 AWS cut on-demand P5 prices by about 44 percent. In January 2026 it raised Capacity Block prices for the same generation by about 15 percent, and in July it raised them again. The same H100 now costs anything from $3.10 to $6.88 an hour depending on how you buy it, and the hardware a generation up can cost $17 a GPU-hour or be unavailable at any price. For most teams the hardest part of AWS GPU pricing is not reading the rate card but understanding which of four rate cards applies.

This guide lays out AWS GPU pricing as it stands in October 2026 for US East in US dollars: on-demand rates for every current family, normalized to cost per GPU-hour so families can be compared, the four purchasing mechanisms and how they change the rate, worked examples for inference and training, the costs that sit outside the instance price, and the honest crossover between self-hosting on EC2 and buying inference per token through Bedrock. Rates are drawn from AWS On-Demand pricing, Capacity Blocks pricing, and Spot pricing, and they change often enough that you should confirm before budgeting.

Key Takeaways

• AWS GPU rates span 270x, from $0.53 an hour for a single T4 to $142 for eight Blackwell Ultra GPUs; normalize to cost per GPU-hour to compare families.
• The same H100 costs $6.88 per GPU-hour on-demand, $5.19 on a Capacity Block, about $3.83 on rare Spot, and toward $3.10 on a three-year commitment.
• GPU prices moved both ways in 2025 to 2026: P5 on-demand fell 44 percent, Capacity Blocks rose about 15 percent twice. Scarcity, not Moore's law, sets high-end GPU prices now.
• For inference, the g6.xlarge (L4) at $0.805 undercuts the older g5.xlarge (A10G) at $1.006 with a newer GPU; for large models, g6e (L40S) is the entry point.
• An idle GPU bills every hour; self-hosting beats Bedrock's per-token pricing only above roughly 40 to 50 percent utilization.

AWS GPU Instance Families and On-Demand Rates

Rates are US East (N. Virginia) Linux on-demand as of October 2026. Per-GPU-hour is the instance rate divided by GPU count, the number that lets families be compared.

Two things in that table are not obvious from AWS's own pages. First, the newer inference generation is cheaper than the older one: the L4-based g6 costs less per hour than the A10G-based g5 at every size while performing comparably or better, so teams still on G5 are overpaying. Second, high-end training capacity is increasingly sold only through Capacity Blocks; P5e has no public on-demand rate at all. Inference-optimized Inf2 instances using AWS's own Inferentia2 chips start near $0.76 an hour and can undercut G-series for supported model architectures, and Trainium instances serve the same role for training. For the wider EC2 catalog, see our EC2 instance types cost reference.

The Four Ways to Buy, and What Each Does to the Price

For a given GPU instance, how you purchase it changes the hourly rate by as much as the hardware choice does. The table shows the same p5.48xlarge under each mechanism.

The practical reading: Savings Plans and Reserved Instances lower the price but do not reserve capacity, which for P-series is the binding constraint. Capacity Blocks cost more per hour than a commitment but guarantee the GPUs will exist when your training job starts, which is why most large training customers use them. Spot is a real option for G-series inference and experimentation and a largely theoretical one for P5. The general commitment logic is in our guide to Reserved Instances, Savings Plans, and Spot; the GPU-specific twist is that availability, not discount, usually decides.

Why GPU Prices Moved Both Ways

Cloud customers learned to expect prices to fall. For high-end GPUs that expectation broke in 2025 and 2026. AWS cut P5 on-demand and Savings Plan rates by about 44 percent in June 2025, taking the H100 from roughly $12 to $6.88 per GPU-hour and undercutting its hyperscale rivals on-demand. Then in January 2026 it raised Capacity Block prices by about 15 percent for P5e (from $34.61 to $39.80 an hour) and P5en, and in July 2026 raised them again, taking P5e Capacity Blocks to $47.76. AWS attributes Capacity Block pricing to supply and demand, which is the point: the mechanism that guarantees capacity is the one where scarcity shows up in the price.

For budgeting, this means two things. On-demand GPU rates can be modeled from the rate card with the usual caveats. Capacity Block rates, which is what training workloads actually pay, should be budgeted with headroom and re-checked quarterly, because they have risen twice in a year and AWS's pricing page notes scheduled updates. Blackwell-generation P6 instances, where available, price at a premium over Hopper that reflects both performance and scarcity.

Worked Example One: Inference

A team serves a 13-billion-parameter open-weight model for an internal application, needing about one GPU's worth of capacity around the clock. On a g5.xlarge at $1.006 an hour, that is about $734 a month. On a g6.xlarge at $0.805, about $588, for equal or better throughput on the newer L4. Running the g6.xlarge on Spot at a typical 60 percent discount brings it to roughly $235 a month, viable because inference requests can be retried on interruption and a second instance in another pool covers the gap. Add a 100 GB gp3 volume ($8) and modest egress, and the realistic range is $250 to $600 a month depending on the purchasing choice, for the same workload.

Worked Example Two: Training

A team plans a fine-tuning run needing eight H100s for five days, about 120 hours. On-demand on a p5.48xlarge, if a launch succeeds, that is 120 times $55.04, about $6,600. Through a five-day Capacity Block at $41.53 an hour, about $4,980, paid up front, with the capacity guaranteed for the window. If the run instead needs the H200's memory, a p5e Capacity Block at $47.76 an hour is about $5,730, and it is the only way to get a P5e at all. Add roughly $300 for high-performance storage during the run and data transfer to pull the training set. For a one-off job the Capacity Block is both cheaper and safer; for a team training continuously, a three-year commitment at about $24.77 an hour cuts the compute to about $2,970 per five-day run, provided it can secure capacity when it needs it, which is the gamble commitments do not remove.

The Costs Outside the Instance Rate

• Idle time. A GPU instance bills every hour it exists, and GPUs sit idle far more than CPUs: between training runs, overnight in notebooks, waiting on data. At $55 an hour, a p5 left running over a weekend costs about $2,600 for nothing. Idle detection and scheduled shutdown matter more here than anywhere else in the estate.
• Storage. Training data and checkpoints on EBS or FSx for Lustre bill per provisioned gigabyte; FSx at the throughput training needs is not cheap. Model artifacts in S3 are inexpensive to store but not to repeatedly pull.
• Data transfer. Pulling datasets across regions or from the internet pays the standard egress and inter-region rates, and multi-node training that crosses Availability Zones pays cross-AZ transfer on every exchange. Keep training clusters in one zone.
• Operating system and software. Commercial Linux distributions add per-instance-hour license fees, and NVIDIA enterprise software may carry its own.
• Allocation. In a shared cluster, the GPU node is one bill and the teams using it are many. Our guide to

allocating GPU costs in shared Kubernetes clusters covers how to split it so that idle GPUs have an owner.

Self-Host on EC2 or Buy Tokens on Bedrock?

For inference, the alternative to renting GPUs is paying per token for a managed model through Amazon Bedrock. The economics turn on utilization. A g6e.xlarge at $1.86 an hour costs about $1,360 a month whether it serves one request or a million; Bedrock bills nothing when idle. At high, steady utilization a self-hosted open-weight model on EC2 costs less per token than any managed frontier model. At low or bursty utilization, the idle hours dominate and per-token pricing wins, often by a wide margin. The crossover for most teams sits somewhere around 40 to 50 percent sustained utilization, and below it self-hosting is usually a false economy unless compliance, customization, or model choice forces it. Our comparison of local LLMs by hardware covers which models fit which GPU memory.

How to Buy AWS GPUs Well

• Normalize everything to cost per GPU-hour and per unit of work (tokens served, training steps) before comparing families.
• Move inference off G5 onto G6 or G6e; the newer generation is cheaper and faster.
• Use Spot for G-series inference and experimentation with retry logic; do not plan P-series work around Spot.
• Book Capacity Blocks for training runs, budget them with headroom, and re-check rates quarterly.
• Commit only for GPU usage that is genuinely continuous, and remember commitments do not reserve capacity.
• Detect and shut down idle GPUs aggressively; the hourly rate makes idle time the largest controllable waste.
• Compare self-hosting against per-token pricing at your real utilization before renting a GPU for inference.

Seeing GPU spend by instance, by team, by utilization, and alongside the token spend it competes with is what Opslyft's cost visibility provides, and it is the input every decision above depends on. For how GPU cost fits the wider AI cost picture, see our guide to FinOps for AI, managing token and GPU costs.

The Bottom Line

AWS GPU pricing in 2026 is a rate card with four columns, not one, and the column you buy from matters as much as the GPU you choose. Inference is cheap and getting cheaper: the L4-based G6 family undercuts its predecessor, Spot is viable, and per-token alternatives beat self-hosting at low utilization. Training is expensive and getting scarcer: the H100 costs $6.88 an hour on-demand when you can get it, $5.19 on a Capacity Block when you must be sure, and Capacity Block prices have risen twice in a year. Normalize to cost per GPU-hour, buy by availability for training and by discount for inference, and treat idle GPU time as the first thing to eliminate. For how AWS compares with the specialist GPU clouds that have grown up around this scarcity, see our analysis of Nebius versus AWS GPU costs and our guide to Nvidia H100 and GPU pricing across providers.