80 GB GPUs you can deploy now

The 80 GB tier on DeployGPU means two cards, the A100 and the H100, with exactly the same memory size and very different compute underneath it. The table below covers both at once, with live prices.

Same memory, different compute

Specifications from NVIDIA's A100 and H100 datasheets. Because both cards carry exactly 80 GB, picking between them is not about what fits, it is about whether a workload benefits from the H100's extra bandwidth and FP8 support enough to justify its higher hourly rate. Long-context inference and FP8-native checkpoints tend to answer yes. Straightforward BF16 fine-tuning and training runs that are not bandwidth-bound often run well on the cheaper A100. The H100 also comes in a PCIe variant at the same 80 GB with lower bandwidth than SXM, and an NVL variant with 94 GB, both covered on the H100 page.

What fits in 80 GB

These are weights-only sizes of published checkpoints, calculated rather than measured, and the numbers are identical on either card since the memory size matches.

Where the two cards genuinely diverge is number formats. FP8 checkpoints run natively on the H100 and not on the A100. gpt-oss ships in MXFP4, which neither card runs natively (that takes Blackwell), so the serving engine handles it in software on both. The A100 and H100 pages cover each card's own precision support and variants in more depth.

Fine-tuning needs more than serving

Serving and fine-tuning follow different memory rules. gpt-oss-120b's own model card states that it can be fine-tuned on a single H100 node, while the smaller gpt-oss-20b fine-tunes on consumer hardware, a useful anchor for how much more headroom a fine-tuning run needs compared with the serving numbers above. Full-parameter fine-tuning in mixed precision generally costs around 18 bytes per parameter once gradients and optimizer state are included, which is why fine-tuning on a single 80 GB card usually means LoRA or a quantized method rather than updating every weight. The QLoRA paper reports fine-tuning models as large as 65B parameters on a single 48 GB GPU using 4-bit weights, which is often a bigger lever than moving to a bigger card.

Multi-GPU on the 80 GB tier

Both cards support NVLink at this tier, unlike most of the 48 GB tier below it. A100 SXM connects its GPUs at 600 GB/s, H100 SXM at 900 GB/s and bridged H100 NVL pairs at 600 GB/s, so a tensor-parallel job that spans 2 or more GPUs is a reasonable fit on either card, not just a fallback for when one GPU is not enough. The multi-GPU page covers how to confirm the interconnect on a specific offer before launching it.

Moving up or down a tier

For models that fit comfortably under 48 GB, the 48 GB tier costs less per hour. Once a single model needs more than 80 GB, or the job wants headroom for long context and many concurrent users, the 96 GB and larger tier covers the RTX PRO 6000, H200 and B200.

For A100 prices across the whole market, compare all providers on GpuPerHour. For a reference price over time, see the A100 rental index (GPUIndexes).