GPU Pricing & Availability
Comparison of 38 AI accelerators — specs, cloud pricing, and availability across providers. Data-dense, sorted by manufacturer and release date.
Showing 38 of 38| GPU | Manufacturer | Architecture | Memory | Bandwidth | FP8 TFLOPS | Power | Status | Cloud $/hr |
|---|---|---|---|---|---|---|---|---|
| NVIDIA H100 SXM5 | NVIDIA | Hopper | 80GB HBM3 | 3.35 TB/s | 3958 | 700W | Shipping (allocation constrained) | from $2.23 |
| NVIDIA H100 PCIe | NVIDIA | Hopper | 80GB HBM3 | 2.0 TB/s | 3026 | 350W | Shipping | from $1.98 |
| NVIDIA H200 SXM5 | NVIDIA | Hopper (refresh) | 141GB HBM3e | 4.8 TB/s | 3958 | 700W | Shipping | from $2.49 |
| NVIDIA H200 PCIe | NVIDIA | Hopper (refresh) | 141GB HBM3e | 4.0 TB/s | 3026 | 350W | Shipping | from $2.1 |
| NVIDIA B200 | NVIDIA | Blackwell | 192GB HBM3e | 8.0 TB/s | 4500 | 1000W | Shipping (limited initial allocation) | from $3.5 |
| NVIDIA B100 | NVIDIA | Blackwell | 192GB HBM3e | 8.0 TB/s | 3600 | 700W | Shipping | from $2.8 |
| NVIDIA GB200 NVL72 | NVIDIA | Blackwell (Grace+Blackwell) | 13.8TB HBM3e (72 GPU system) | 8.0 TB/s per GPU | 324000 | 27000 (system)W | Shipping to select customers | TBD |
| NVIDIA GB200 NVL36 | NVIDIA | Blackwell (Grace+Blackwell) | 6.9TB HBM3e (36 GPU system) | 8.0 TB/s per GPU | 162000 | 13500 (system)W | Shipping | TBD |
| NVIDIA L40S | NVIDIA | Ada Lovelace | 48GB GDDR6 | 0.864 TB/s | 733 | 350W | Shipping (widely available) | from $0.79 |
| NVIDIA L40 | NVIDIA | Ada Lovelace | 48GB GDDR6 | 0.864 TB/s | 362 | 300W | Shipping | from $0.65 |
| NVIDIA L4 | NVIDIA | Ada Lovelace | 24GB GDDR6 | 0.300 TB/s | 242 | 72W | Shipping (widely available) | from $0.49 |
| NVIDIA A100 SXM4 80GB | NVIDIA | Ampere | 80GB HBM2e | 2.0 TB/s | N/A | 400W | Shipping (mature, widely available) | from $0.75 |
| NVIDIA A100 PCIe 80GB | NVIDIA | Ampere | 80GB HBM2e | 1.94 TB/s | N/A | 300W | Shipping (widely available) | from $0.55 |
| NVIDIA A30 | NVIDIA | Ampere | 24GB HBM2 | 0.933 TB/s | N/A | 165W | Shipping | TBD |
| NVIDIA A10 | NVIDIA | Ampere | 24GB GDDR6 | 0.600 TB/s | N/A | 150W | Shipping | from $0.4 |
| NVIDIA T4 | NVIDIA | Turing | 16GB GDDR6 | 0.320 TB/s | N/A | 70W | Shipping (legacy) | from $0.35 |
| NVIDIA V100 SXM2 | NVIDIA | Volta | 32GB HBM2 | 0.900 TB/s | N/A | 300W | End of life (still deployed) | from $0.3 |
| NVIDIA B300 | NVIDIA | Blackwell Ultra | 288GB HBM3e | 8.0 TB/s | 5000 | 1200W | Shipping (HGX B300 available now) | TBD |
| NVIDIA Vera Rubin | NVIDIA | Vera Rubin | TBD (expected 288GB+ HBM4) | TBD (expected 13+ TB/s) | TBD | TBD (expected 1400W+)W | Announced — shipping H2 2026 to hyperscalers only | TBD |
| NVIDIA Rubin Ultra | NVIDIA | Rubin Ultra | TBD (expected 576GB+ HBM4e) | TBD (expected 13+ TB/s) | TBD | TBD (expected 1800W+)W | Announced (not yet shipping) | TBD |
| AMD Instinct MI300X | AMD | CDNA 3 | 192GB HBM3 | 5.3 TB/s | 2614 | 750W | Shipping | from $1.99 |
| AMD Instinct MI300A (APU) | AMD | CDNA 3 | 128GB HBM3 | 5.3 TB/s | 2614 | 760W | Shipping | TBD |
| AMD Instinct MI350 | AMD | CDNA 4 (3nm) | 288GB HBM3E | 8.0 TB/s | 10000 | 1000W | Shipping (limited initial allocation) | TBD |
| AMD Instinct MI325X | AMD | CDNA 3 (refresh) | 288GB HBM3E | 6.0 TB/s | 2614 | 750W | Shipping | from $2.49 |
| AMD Instinct MI400 (Announced) | AMD | CDNA 5 (next-gen) | TBD (expected 384GB+ HBM4) | TBD (expected 10+ TB/s) | TBD | TBD (expected 1200W+)W | Announced (not yet shipping) | TBD |
| Intel Gaudi 3 | Intel | Habana | 128GB HBM2e | 3.2 TB/s | 1835 | 600W | Shipping (limited) | TBD |
| Intel Gaudi 2 | Intel | Habana | 96GB HBM2e | 2.45 TB/s | 1458 | 600W | Shipping (mature) | TBD |
| Google TPU Trillium (v6e) | Trillium | 32GB HBM3 | 1.6 TB/s | 1836 | TBDW | Operational (GCP only) | from $2.7 | |
| Google TPU v5p | TPU v5 (Pod) | 95GB HBM3 | 2.77 TB/s | 918 | TBDW | Operational (GCP only) | from $4.2 | |
| Google TPU v5e | TPU v5 (Lite) | 16GB HBM3 | 0.81 TB/s | 393 | TBDW | Operational (GCP only) | from $1.2 | |
| Google TPU v4 | TPU v4 | 32GB HBM3 | 1.2 TB/s | N/A | TBDW | Operational (GCP only) | from $3.22 | |
| Google Ironwood (TPU v7) | Ironwood | TBD (expected 64GB+ HBM3e) | TBD (expected 3+ TB/s) | TBD | TBDW | Announced (not yet shipping) | TBD | |
| Cerebras CS-3 (Wafer-Scale) | Cerebras Systems | WSE-3 (wafer-scale) | 44GB SRAM (on-die) + external | 21 PB/s (on-die) | 250 | TBD (system)W | Shipping (limited) | TBD |
| Groq LPU (Inference Engine) | Groq | LPU (Tensor Streaming) | TBD (SRAM-based) | 80 TB/s (on-die SRAM) | 1500 | 300W | Shipping (limited) | TBD |
| SambaNova SN40L | SambaNova Systems | RDU (Reconfigurable Dataflow) | 64GB HBM3 + 1.5TB DDR5 | 1.7 TB/s (HBM) | 1275 | TBDW | Shipping (limited) | TBD |
| Tenstorrent Grayskull | Tenstorrent | Wormhole (RISC-V) | 32GB LPDDR4X | 0.137 TB/s | 466 | 200W | Shipping (limited) | TBD |
| Huawei Ascend 910B | Huawei | Da Vinci (7nm) | 64GB HBM2e | 1.6 TB/s | 640 | 400W | Shipping (China domestic only) | TBD |
| Huawei Ascend 910C | Huawei | Da Vinci (enhanced) | 80GB HBM3 | 2.0 TB/s | 800 | 450W | Shipping (China domestic only) | TBD |