Cost per embedding at scale is a throughput problem. Cold-start latency breaks production retrieval. Both are solvable with the right GPU configuration.
API-based embedding at scale is expensive. At 10M documents × 512 tokens, your embedding bill rivals your inference bill. You need your own GPU.
CPU-based embedding is too slow for large corpus ingestion. Even a single A100 can embed millions of documents per hour. CPU cannot.
Serverless embedders cold-start on every request when idle. For interactive RAG, that's 2–5 seconds on the first query. Dedicated endpoints don't cold-start.
Corpus ingestion is a batch workload. You don't need low latency — you need to process 50M chunks as cheaply and quickly as possible.
Submit an async job, configure batch size, point it at your storage bucket. Spot pricing applies. Output lands back in your bucket when done.
POST your job spec and dataset location. No blocking call. Webhooks or polling for completion.
The platform packs input sequences into optimal batches for your GPU. BGE-large-en processes thousands of sequences per second on an A100.
Batch embedding is checkpointable by design. Spot instances cut cost significantly. Interrupted jobs resume from last completed shard.
Embeddings write directly to your S3-compatible private bucket. No shared staging, no egress through our storage.
A production RAG pipeline has three GPU-adjacent components. Aircloud handles two of them; you bring your vector DB of choice.
Batch-embed your corpus using a dedicated GPU endpoint. BGE, E5, Nomic, or your custom fine-tuned embedder. Private endpoint means consistent latency on live queries.
Write embeddings to Pinecone, Weaviate, Qdrant, pgvector, or any vector store. Aircloud handles compute; you keep control of your vector index.
Retrieved chunks go into a private LLM endpoint — GLM 5.1, Gemma 4, Kimi 2.5, or your fine-tuned model. Single-tenant, consistent TTFT, no shared-GPU variability.
Model
Strong English retrieval. Best on BEIR. L40S handles 4K+ sequences/sec.
Model
Instruction-tuned. Top multilingual performance. Needs A100 for high-throughput batch.
Model
Open source, Matryoshka dims. Fast inference. Fits comfortably on L40S.
Model
Bring your own model, trained on your domain. Deploy as a private endpoint the same way.
Batch jobs on spot. Dedicated endpoints for live queries. Both on isolated GPU, without the API pricing model.