| GB300NVIDIA's Grace Blackwell Ultra superchip: a Grace CPU with Blackwell Ultra GPUs, 288 GB HBM3e each. | $13.44 | $6.72 | Verda (DataCrunch) | 2 |
| GB200NVIDIA's Grace Blackwell superchip, sold in NVL72 racks of 72 GPUs. | $10.50 | $6.22 | CoreWeave | 3 |
| B300NVIDIA's Blackwell Ultra GPU with 288 GB HBM3e, for training and inference at scale. | $6.94 | $0.99 | Runpod | 9 |
| B200NVIDIA's Blackwell GPU in HGX servers, 180 to 192 GB HBM3e. | $5.98 | $0.99 | Runpod | 12 |
| H200NVIDIA's Hopper GPU with 141 GB HBM3e: an H100 with more and faster memory. | $3.59 | $0.79 | Runpod | 11 |
| H100NVIDIA's Hopper GPU, the workhorse of 2023 to 2025 training, 80 GB HBM3 (SXM and PCIe). | $1.99 | $0.67 | Runpod | 15 |
| GH200NVIDIA's Grace Hopper superchip: a Grace CPU and a Hopper GPU, 96 GB. | $1.99 | – | Vultr | 3 |
| A100NVIDIA's Ampere data-center GPU, 40 or 80 GB, still common for training and inference. | $0.80 | $0.40 | Vast.ai | 11 |
| L40SNVIDIA's Ada data-center GPU for inference and graphics, 48 GB. | $0.67 | $0.40 | Vast.ai | 11 |
| RTX PRO 6000NVIDIA's Blackwell workstation-class server GPU, 96 GB. | $1.39 | $0.47 | Vast.ai | 6 |
| L4NVIDIA's low-power Ada GPU for inference and video, 24 GB. | $0.44 | $0.10 | Runpod | 4 |
| A10G / A10NVIDIA's Ampere inference GPU, 24 GB (the A10G is the AWS version). | $1.01 | $0.24 | AWS | 3 |
| RTX 6000 AdaNVIDIA's Ada workstation-class GPU, 48 GB. | $0.74 | $0.77 | Runpod | 3 |
| L40NVIDIA's Ada data-center GPU for graphics and inference, 48 GB. | $0.69 | $0.78 | Runpod | 2 |
| RTX A6000NVIDIA's Ampere workstation-class GPU, 48 GB. | $0.33 | $0.43 | Runpod | 3 |
| A40NVIDIA's Ampere data-center GPU for graphics and inference, 48 GB. | $0.35 | – | Runpod | 1 |
| RTX 4000 AdaNVIDIA's small Ada GPU, 20 GB. | $0.52 | – | Akamai (Linode) | 2 |
| V100NVIDIA's Volta GPU from 2017, 16 or 32 GB. | $0.24 | $0.12 | Verda (DataCrunch) | 2 |
| MI355XAMD's Instinct GPU on CDNA 4, 288 GB HBM3e. | $2.59 | $4.50 | Vultr | 3 |
| MI325XAMD's Instinct GPU on CDNA 3, 256 GB HBM3e. | $4.62 | – | Vultr | 2 |
| MI300XAMD's Instinct GPU on CDNA 3, 192 GB HBM3. | $2.59 | $1.11 | DigitalOcean | 5 |
| TPU IronwoodGoogle Cloud TPU seventh generation, per chip-hour. | $12.00 | – | Google Cloud | 1 |
| TPU v6e (Trillium)Google Cloud TPU, priced per chip-hour. | $2.70 | – | Google Cloud | 1 |
| TPU v5pGoogle Cloud TPU for training, per chip-hour. | $4.20 | – | Google Cloud | 1 |
| TPU v5eGoogle Cloud TPU for inference, per chip-hour. | $1.20 | – | Google Cloud | 1 |
| Trainium2AWS training chip (trn2), per chip-hour. | $6.45 | – | AWS | 1 |
| TrainiumAWS first-generation training chip (trn1). | $1.34 | $0.14 | AWS | 1 |
| Inferentia2AWS inference chip (inf2). | $0.76 | $0.08 | AWS | 1 |
Computeby Glyph