AI & GPU Compute

GPU-accelerated instances and an AI platform for large-model training and inference, image processing, and scientific computing — usable compute without the wait.

Core capabilities

Training and inference want different hardware

Training is bound by VRAM and inter-GPU bandwidth: parameters, gradients, and optimizer state all live in memory, and once you go multi-GPU, how fast the cards talk to each other often matters more than raw per-card throughput. Confirm the model fits before comparing compute.

Inference rewards cost-per-card and throughput. Once a model is live, quantization and request batching usually raise per-card throughput more cheaply than adding cards.

GPU cost has more levers than CPU cost

First, use it or release it — GPU hours are expensive, so free the instance when a training run finishes instead of letting it idle. Second, run fault-tolerant batch work on interruptible instances: materially cheaper, at the cost of possible reclamation, which means checkpointing properly.

Third, split the workloads: keep steady inference on reserved capacity for stability and send bursty batch jobs to temporary instances. Separating the two is cheaper than buying one fleet sized for the peak.

The supporting resources people overlook

CPU, memory, and disk consume more than expected around GPUs: preprocessing is CPU-bound, dataset loading is memory- and IO-bound, and checkpointing writes to storage. Under-provision those and GPU utilization drops — expensive silicon waiting on cheap resources.

Pairs well with

Frequently asked questions