H100 SXM5Available now
NVIDIA · Hopper- Memory
- 80 GB HBM3
- Bandwidth
- 3.35 TB/s
- Interconnect
- 900 GB/s NVLink 4
- Compute
- 3,958 TFLOPS FP8
- TDP
- 700 W
The volume workhorse. Deepest supply and the widest kernel coverage.
$2.14/GPU-hrAccelerators
Select a part, then configure the cluster. Reserved terms discount the on-demand rate.
H200 SXM5Available now
NVIDIA · Hopper- Memory
- 141 GB HBM3e
- Bandwidth
- 4.8 TB/s
- Interconnect
- 900 GB/s NVLink 4
- Compute
- 3,958 TFLOPS FP8
- TDP
- 700 W
Same compute as H100 with 76% more memory — the inference capacity upgrade.
$2.98/GPU-hrB200 SXMAvailable now
NVIDIA · Blackwell- Memory
- 192 GB HBM3e
- Bandwidth
- 8 TB/s
- Interconnect
- 1.8 TB/s NVLink 5
- Compute
- 9,000 TFLOPS FP8 · 18 PFLOPS FP4
- TDP
- 1000 W
Blackwell generation. Roughly 2.2× H100 training throughput at FP8.
$6.80/GPU-hrB300 SXMLimited capacity
NVIDIA · Blackwell Ultra- Memory
- 288 GB HBM3e
- Bandwidth
- 8 TB/s
- Interconnect
- 1.8 TB/s NVLink 5
- Compute
- 10,000 TFLOPS FP8 · 20 PFLOPS FP4
- TDP
- 1400 W
Blackwell Ultra. Memory headroom for long-context and MoE inference.
$8.90/GPU-hrGB200 NVL72Limited capacity
NVIDIA · Grace Blackwell- Memory
- 13.4 TB HBM3e / rack
- Bandwidth
- 8 TB/s per GPU
- Interconnect
- 130 TB/s all-to-all
- Compute
- 720 PFLOPS FP8 · 1.44 EFLOPS FP4 / rack
- TDP
- 1200 W
72 Blackwell GPUs and 36 Grace CPUs as one NVLink domain. Rack-scale only.
$11.24/GPU-hrGB300 NVL72Waitlist · Q4 2026
NVIDIA · Grace Blackwell Ultra- Memory
- 21 TB HBM3e / rack
- Bandwidth
- 8 TB/s per GPU
- Interconnect
- 130 TB/s all-to-all
- Compute
- 1.1 EFLOPS FP4 / rack
- TDP
- 1400 W
Blackwell Ultra at rack scale. Reserved allocation, waitlist by contract size.
$14.50/GPU-hrVera Rubin VR200 NVL144Announced · H2 2026
NVIDIA · Rubin- Memory
- HBM4 · 288 GB / GPU
- Bandwidth
- 13 TB/s
- Interconnect
- 260 TB/s NVLink 6
- Compute
- 3.6 EFLOPS FP4 / rack
- TDP
- 1800 W
Rubin generation paired with the Vera CPU. HBM4 and NVLink 6. Indicative pricing.
$21.00/GPU-hrRubin Ultra NVL576Announced · 2027
NVIDIA · Rubin Ultra- Memory
- HBM4e · 1 TB / package
- Bandwidth
- 32 TB/s
- Interconnect
- 1.5 PB/s domain
- Compute
- 15 EFLOPS FP4 / rack
- TDP
- 2300 W
576-GPU NVLink domain. Register interest — allocation opens on tape-out.
$38.00/GPU-hrInstinct MI355XLimited capacity
AMD · CDNA 4- Memory
- 288 GB HBM3e
- Bandwidth
- 8 TB/s
- Interconnect
- 1.07 TB/s Infinity Fabric
- Compute
- 5,000 TFLOPS FP8 · 10 PFLOPS FP4
- TDP
- 1400 W
ROCm 6.3+. Strong memory-per-dollar for inference and fine-tuning.
$5.40/GPU-hrTPU v6e (Trillium)Available now
Google · TPU v6- Memory
- 32 GB HBM
- Bandwidth
- 1.6 TB/s
- Interconnect
- 3,584 Gbps ICI
- Compute
- 918 TFLOPS BF16
- TDP
- 400 W
Pod-attached. JAX and PyTorch/XLA workloads only.
$1.85/GPU-hr