GPU compute is the raw accelerator capacity behind training and inference. Choosing the right class and tenancy model determines cost, start-up time, and how well a workload tolerates interruption — and for adult AI, it also determines where the work physically runs.
Accelerator classes
Two broad classes cover most workloads:
- Inference-optimized — tuned for throughput and cost per request on already-trained models. Lower memory per device, faster to schedule, and the default choice for serving.
- Flagship training — large-memory, high-interconnect devices for training and fine-tuning. Expensive and scarce; reserve them for jobs that actually need them.
Match the class to the phase. Serving a finished model on training hardware wastes money; training on inference hardware will not fit. The platform does not guarantee specific device models, so design against capability and memory requirements rather than a named part.
Virtual, bare-metal, and spot
| Tenancy | Isolation | Start-up | Interruption | Good for |
|---|---|---|---|---|
| Virtual | Shared host, isolated instance | Fast | None | Serving, small fine-tunes |
| Bare-metal | Whole host | Slower | None | Large training, consistent throughput |
| Spot | Shared, reclaimable | Fast | Can be reclaimed | Fault-tolerant batch and checkpointed training |
Spot is dramatically cheaper and fine for jobs that checkpoint frequently and can restart. Never put a latency-sensitive inference endpoint on spot without a reliable fallback, because reclamation is not something you control.
Autoscaling
Scale on the signal that reflects real load:
- Inference — scale on queue depth or concurrent requests, with a warm pool so new replicas are ready before the queue grows.
- Training — scale the job’s worker count, not a long-lived pool, and checkpoint so a reclaimed worker does not lose progress.
Set both a floor and a ceiling. A floor keeps latency-sensitive endpoints warm; a ceiling stops a traffic anomaly from consuming unbounded capacity. Scale-down should drain gracefully so in-flight requests finish.
Regional placement
Place accelerators where the data and the users are:
- Near users for interactive inference where round-trip time matters.
- Near data for training and batch jobs, so large datasets do not cross regions repeatedly.
- Where the law requires when residency or export rules apply to adult content or personal data.
Keep training data and checkpoints inside the region they belong to, and avoid moving them across boundaries for convenience.
Serverless inference versus dedicated GPUs
- Serverless inference abstracts the device entirely. You deploy a model and pay for what you use; the platform handles scheduling, scaling, and cold starts. Best for spiky or low-volume workloads and for teams that do not want to manage capacity.
- Dedicated GPUs give you a persistent device you control. Best for steady high-volume serving, custom runtimes, and training, where you need predictable performance and full control of the environment.
Many adult platforms run both: serverless for the long tail of models and burst traffic, dedicated capacity for the few models that serve continuously.
Where AdultInfra fits
AdultInfra places accelerator capacity for lawful adult workloads with inference-optimized and flagship classes, spot where it is safe, and regional placement that respects residency. To size a workload or move it off a general-purpose provider, start with adult AI inference and bring one job and its memory requirement.