Renting an NVIDIA A100 in 2026 is easy: Ampere cards are well supplied and competitively priced. Millionminer puts mid-2026 on-demand listings from about $0.68 per GPU-hour. The median across 39 providers was $1.86 on 29 October 2026, and hyperscalers charge $3.43 (AWS) to $5.03 (Google Cloud) per hour for the 80GB card. The A100 rent you actually pay depends on the channel, the billing type and what gets billed around the GPU. Whether an A100 beats an H100 for you comes down to cost per request on your own workload, not the hourly sticker. This page covers the A100 rental decision itself; the wider question of renting vs owning cloud GPUs is worked through in our total cost analysis of cloud GPUs and on-premise accelerators.
What A100 GPU rent costs per hour, by provider tier
The A100 market splits into three tiers that behave almost like separate products: marketplaces at the floor, specialist clouds in the middle and hyperscalers at the top. Hyperscalers charge 3-5x more for the 80GB card, so the tier is the first thing to pin down when comparing quotes.
| Marketplace | Vast.ai, RunPod community tier | 40GB median $0.62/hr on Vast.ai across six verified US/CA hosts | A bare card on a third party’s hardware, reliability varying host by host, no uptime promise |
| Specialist cloud | Thunder Compute, Crusoe, CoreWeave, Lambda, RunPod secure | $1.09/hr (Thunder Compute, 80GB); $2.00/hr (Crusoe); $2.70/hr (CoreWeave, SXM) | The provider’s own fleet, more reliability and support without hyperscaler markup |
| Hyperscaler | AWS, Azure, Google Cloud | $3.43/hr (AWS, 80GB) to $5.03/hr (Google Cloud, 80GB); $3.40/hr per GPU on Azure’s 8x NDasr A100 v4 | An SXM card in a multi-GPU VM with bundled CPU, RAM and disk, an SLA and compliance coverage |
Across billing types, GetDeploying’s 29 October 2026 snapshot put the A100 median at $1.86 per GPU-hour on-demand, $1.53 reserved and $1.19 spot. At 720 hours a month, one A100 at that median on-demand price costs about $1,339.
Trackers disagree on the exact number. Ornn’s OCPI-A100 index, which records prices actually paid for on-demand A100 SXM4 rentals rather than list prices, stood at $1.10 per GPU-hour on 28 October 2026. In one Ornn comparison, published H100 rates at four GPU clouds ran 15% to 103% above the OCPI-H100 transaction value, a reminder that posted prices and paid prices diverge. GPUaaS (2026) makes the more useful point for a buyer: the spread between the cheapest and the most expensive A100 hour is wider than the A100-to-H100 gap at a single provider.
Is the A100 still easy to rent in 2026?
Yes, on most supply accounts. Thunder Compute’s October 2026 market analysis describes Hopper and Ampere cards as well supplied and competitively priced, while Blackwell hardware is largely reserved for hyperscalers and large enterprises with existing supply agreements. On the long view, GPUaaS reports that CoreWeave has committed to renting A100s through 2029, and that on 13 August 2026 Jensen Huang described the A100 fleet as mission-capable from 2020 through 2029.
The counter-evidence deserves a fair hearing. SemiAnalysis reported in early 2026 that on-demand GPU rental capacity was sold out across all GPU types, with holders of on-demand instances unwilling to release them despite price rises. Both pictures can be true at once: the A100 is listed widely, yet the cheapest listing may not be available at your scale or in your region, and when the cheapest tier is sold out where you need it, the real price is the next tier up. For a single card or a small cluster we would treat availability as a non-issue; for a sustained multi-node job we would confirm capacity before planning around a marketplace rate.
The kinds of provider are the three tiers in the table above. On AWS, the 80GB A100 comes through the p4de.24xlarge instance and the 40GB card through the older p4d.24xlarge.
A100 40GB or 80GB: which one to rent
Rent the 80GB card in almost every case, unless the 40GB is cheaper and your model fits it with room to spare. Thunder Compute’s October 2026 comparison found the 80GB A100 usually the better value, often costing less than a 40GB card, with its $1.09/hr 80GB listing undercutting 40GB cards on Paperspace ($3.09/hr) and Google Cloud ($3.67/hr).
On the hardware side, NVIDIA’s datasheet lists the A100 80GB with HBM2e at 1,935 GB/s (PCIe) and 2,039 GB/s (SXM) of memory bandwidth. NVIDIA claims up to 1.25x higher inference performance than the 40GB card, measured on MLPerf 0.7 RNN-T single-stream with one-seventh MIG slices, which is a narrow configuration and a vendor figure rather than an independent test.
The extra memory only pays when the workload needs it. GPUaaS (2026) puts it bluntly: an 80GB A100 running a 13B model sees no speedup over the 40GB variant, because memory bandwidth saturates long before capacity does. The 80GB card earns its place when the model weights or the KV cache need the headroom, which shows up as larger batches, longer context windows and no spilling to CPU offload. GetDeploying estimates that one A100 runs models up to roughly 31B parameters at 16-bit or 113B at 4-bit quantisation, assuming a 32K context.
Two other variant choices matter as much as memory size:
- PCIe or SXM. The SXM4 card mounts on NVIDIA’s NVLink baseboard at 400W, giving tighter GPU-to-GPU interconnect and better clocks and bandwidth than the 300W PCIe card. For multi-GPU work, a PCIe card and an NVLinked SXM card are effectively different machines.
- MIG partitioning. An A100 can be split into up to seven hardware-isolated GPU instances, each with its own memory, cache and compute. Thunder Compute notes this suits serving 7B to 13B models in multi-tenant setups where a full 80GB card would sit underused.
Spot, on-demand or reserved A100 rental
Spot gives the lowest A100 rate in exchange for interruption risk, reserved gives a lower rate in exchange for commitment, and on-demand gives flexibility at the headline price. Millionminer (mid-2026) puts reserved 1-to-12-month deals at 20% to 40% below on-demand and spot at 40% to 65% below. Thunder Compute reports spot at about half of on-demand, with reserved commitments saving about 25% on a one-year term and about 45% over three years. GPUaaS puts hyperscaler one-year commitments at 30% to 40% off on-demand.
Thunder Compute’s 2026 analysis reports A100 spot interrupting at 15% to 20%, against under 5% for H100 spot on AWS, with Google Cloud giving 30 seconds of preemption notice and AWS two minutes. An uncheckpointed training run that gets interrupted can turn a 60% discount into a net loss. Checkpointing every 15 to 30 minutes caps lost progress at half an hour of compute, and PyTorch Lightning and Ray Train both provide checkpoint-resume out of the box.
| Billing type | Fits | Avoid when |
|---|---|---|
| Spot | Checkpointed training, batch jobs, experiments that can restart | Production inference that cannot tolerate interruption |
| On-demand | Bursty or uncertain demand, development, short projects | The card runs flat out for many months, where a commitment discount pays |
| Reserved / committed | Steady inference or training at predictable utilisation for the whole term | Demand may move to a newer card or disappear before the term ends |
One more wrinkle changes the spot calculation: providers like Thunder Compute now offer on-demand A100 access below typical hyperscaler spot prices, with no interruption risk. The comparison of cheapest GPU cloud options works through that across GPU types.
The costs that sit outside the hourly A100 rate
Millionminer (2026) lists the traps in rough order of damage: comparing a hyperscaler’s whole-VM price (CPU, RAM and disk included) with a marketplace’s bare card, egress fees on moving weights and datasets out, idle billing on attached volumes and stopped instances, spot interruptions on uncheckpointed runs, headline rates that turn out to be commitments, and sold-out cheap tiers.
- Multi-GPU minimums. AWS’s p4d and Azure’s NDasr A100 v4 come as 8-GPU machines: GetDeploying lists AWS’s p4d.24xlarge (40GB cards) at $21.96 per hour on-demand, about $2.74 per GPU, and Azure’s 8x NDasr A100 v4 at $27.20 per hour. A team that needs one card pays for eight on those instances unless it picks a provider that sells single GPUs.
- Bundled CPU and RAM. The same 8-GPU AWS machine carries 96 vCPUs and 1.2TB of RAM in its price. That is good value if you use it and dead weight if you do not.
- Egress. AWS, Azure and Google Cloud charge $0.08 to $0.12 per GB for outbound data after the free tier, so moving a 100GB checkpoint out costs $8 to $12 on top of compute. Several specialist clouds charge nothing for egress.
- Low utilisation. GPUaaS, citing Cast AI’s 2026 State of Kubernetes Optimization report, puts average GPU utilisation across production clusters at about 5%, which makes the effective cost per useful compute hour twenty times the headline rate.
GPUaaS estimates that idle CPU, storage and networking charges inflate hyperscaler bills by 20% to 40% beyond the advertised rate, a figure from a GPU vendor that sells against hyperscalers and worth treating as an estimate.
Worked example: what idle time does to an A100’s effective price
Assume one A100 at the 29 October 2026 median on-demand rate of $1.86 per hour. If the card does useful work 25% of the time it is billed, the effective price per useful GPU-hour is $1.86 ÷ 0.25 = $7.44. That is above Google Cloud’s $5.03 list rate for an 80GB A100. The lesson is not that cheap channels are a trap; utilisation moves the real price further than the choice of provider does, so the first question is how busy the card will be.
Are A100 rental prices going down in 2026?
Not clearly, and the honest answer depends on which window you look at. Over the long run, A100 rates have fallen: an analysis of a16z’s Ornn-sourced chart estimates the A100 rental rate at about 58% of its level when observation began in early 2024, a decline of roughly 42%. AWS cut A100 on-demand prices by 33% on P4d and P4de from 1 June 2025. Millionminer reports analysts expecting A100 pricing to go below $1 broadly.
The 2026 data points the other way. GetDeploying’s A100 median was flat over the 90 days to 29 October 2026 and about 5% above a year earlier. A Bank of America report cited in August 2026 had A100 pricing flat at $1.65 an hour, up 17% year on year. The OCPI-A100 index rose 20.9% over thirty days to $1.10 per GPU-hour on 28 October 2026, from $0.91. SemiAnalysis concluded that GPU rental pricing was more likely to keep rising than falling, though its index tracks H100 contract pricing.
Two mechanisms explain the firming. Memory, not GPU die output, became the binding constraint in 2026, and as the newest chips’ memory bill rises, older cards like the A100 hold value instead of depreciating on schedule. GPUaaS adds that scarcity in current-generation supply pushes demand back into older parts. Silicon Data’s index work also shows A100 prices respond to inference demand, long-tail usage and secondary-market liquidity rather than to hyperscaler training waves, and that the A100 index fell more sharply than the H100’s in the mid-2025 correction.
Our reading: the multi-provider medians and the transaction-based index, which both show flat-to-rising A100 rates through late 2026, are stronger evidence for a near-term budget than the long-run decline. They carry caveats too: OCPI covers on-demand transactions only, and Ornn discloses a structural conflict of interest between the index and its affiliates’ commercial interests. We would budget at today’s rates, not wait for a drop, and keep commitments only as long as the reserved discount clearly pays.
Renting versus buying an A100
GPUaaS reported new A100 units at $5,000 to $15,000 in March 2026 depending on form factor, with used units 40% to 60% cheaper but carrying no warranty and limited driver support on a part NVIDIA stopped manufacturing two years earlier. Millionminer puts used A100s at $4,000 to $9,000.
Utilisation decides it. In Millionminer’s worked H100 case, break-even against renting lands near 14,000 GPU-hours, about 19 months at 100% utilisation and four years at 40%. At a 20% duty cycle, the same analysis finds a purchase never pays back before obsolescence. The arithmetic for buy vs rent breakeven utilisation is covered on its own page, and the parent cost analysis covers owning more broadly. For an A100 specifically, the support risk on used hardware is the extra line item a rental avoids.
When is a rented A100 cheaper per request than an H100?
A rented A100 beats an H100 when the H100’s speedup on your workload is smaller than its price premium. Thunder Compute frames the threshold at its own prices ($1.09/hr A100, $3.20/hr H100): the H100 lowers total job cost only when its speedup reaches roughly 3x or more. The A100 also lacks FP8 and the Transformer Engine, which is where much of the H100’s advantage on transformer workloads comes from.
Measured results show the gap varies. Hyperstack’s vLLM benchmark on Llama 3.1 70B found the H100 SXM5 generating 3,311 tokens per second against 1,148 on an A100 NVLink, a 2.8x throughput gain for 1.7x the price. NVIDIA’s own TensorRT-LLM benchmark reports H100 FP8 at up to 4.6x the A100’s maximum throughput, a vendor result at a stated peak configuration of 64 concurrent requests. At the other end, Thunder Compute reports the A100 winning on cost for inference below about 40% GPU utilisation, LoRA fine-tuning under 30B parameters, FP64 HPC and development work.
Worked example: cost per million tokens from the Hyperstack figures
Take Hyperstack’s H100 SXM5 at $2.40 per hour. The reported 1.7x price ratio implies an A100 at about $2.40 ÷ 1.7 ≈ $1.41 per hour.
- A100 output: 1,148 tokens/s × 3,600 s ≈ 4.13 million tokens per hour, so about $1.41 ÷ 4.13 ≈ $0.34 per million tokens.
- H100 output: 3,311 tokens/s × 3,600 s ≈ 11.92 million tokens per hour, so about $2.40 ÷ 11.92 ≈ $0.20 per million tokens.
- Now assume the same 2.8x throughput ratio but Thunder Compute’s prices of $1.09 (A100) and $3.20 (H100). The A100 comes to about $0.26 per million tokens and the H100 to about $0.27, essentially a tie.
The same throughput gap gives the H100 a clear win at Hyperstack’s price ratio and a draw at Thunder Compute’s. The calculation also assumes both cards run saturated: Thunder Compute finds the A100 cheaper for inference below about 40% GPU utilisation.
That is why we would not choose between them from published ratios. Cost per request is the target that lasts, and the context that decides it (your model, batch size, sequence length, precision and real utilisation) comes from profiling your own workload, not from a benchmark blog. The same reasoning applies when the comparison crosses vendors; see when the cost-per-inference calculus shifts between AMD and NVIDIA.
Where a model has to run close to sensors or on the device, the engineering problem moves to AI engineering for IoT and embedded devices, and the A100-versus-H100 question drops away entirely.
FAQ
Are GPU prices going down in 2026?
It depends on the card. SemiAnalysis reported H100 one-year contract prices rising almost 40%, from $1.70 to $2.35 per GPU-hour, between October 2025 and March 2026. Millionminer reports B200 and B300 listings roughly doubling in a year while previous-generation cards slid toward commodity pricing.
