GLOBALATLAS AI · AI infrastructure
AI infrastructure
Training changes a model; inference executes it. Model serving, datasets and application APIs have separate resource requirements.
Select a stage of the AI lifecycle
Train and adapt
Serve a model
LLM hosting
A model service needs weights in memory, runtime buffers, a KV cache and an application endpoint.
Explore →AI inference
Inference executes a trained model. Input size, batching and context length determine working memory as well as model weights.
Explore →Infrastructure for DeepSeek
DeepSeek-R1 and its distilled models are different workloads. Distilled releases span 1.5B to 70B parameters; the full model needs a separate sizing exercise.
Explore →Infrastructure for Llama
Llama is a family of models. Llama 3.1 8B is an example of an 8-billion-parameter model, not a capacity promise for every Llama release.
Explore →Infrastructure for Mistral
Mistral 7B and other Mistral releases differ in size and architecture. Select a specific model card before sizing hardware.
Explore →Infrastructure for Qwen
Qwen includes text, code and multimodal model families. Start with an exact model release, supported serving runtime, precision and context budget; these choices determine the infrastructure requirements.
Explore →Infrastructure for Gemma
Gemma has multiple sizes and task-specific variants. Choose the exact model card, accept its applicable usage terms and verify runtime support before planning a self-hosted service.
Explore →Infrastructure for Stable Diffusion
Stable Diffusion serves image-generation workloads. SDXL Base, an optional refiner and additional workflow components have separate memory and storage requirements. Resolution and batch size influence the complete pipeline.
Explore →Infrastructure for FLUX
FLUX releases differ in architecture, inference code and model licensing. Select the exact release and supported workflow, then measure peak memory at the intended image resolution and precision.
Explore →Infrastructure for Whisper
Whisper is an open-source speech-recognition model family. Self-hosted transcription uses downloaded model weights and a local inference runtime; it is separate from a hosted speech API subscription.
Explore →Process content
Embedding infrastructure
Embedding pipelines encode input into vectors and usually combine model execution with a separate index or database.
Explore →Image generation infrastructure
Image generation memory depends on the model, resolution, batch size and inference pipeline.
Explore →Speech AI infrastructure
Transcription and speech synthesis have different latency, batching and audio-processing requirements.
Explore →