Overview
Strengths & Best For
Novita has grown into a dual-product platform: a GPU cloud for training and fine-tuning (H100 SXM, A100, RTX 4090) and a high-volume LLM inference API serving 100+ open-source models at some of the lowest per-token rates in the market. In 2025–2026 Novita expanded its model catalog to include Llama 3.1/3.2, Qwen 2.5, DeepSeek V3/R1, and Mistral variants, and introduced batch inference pricing for high-volume workloads at up to 50% discount. The GPU cloud side offers fast provisioning (typically under 3 minutes), pre-built Docker environments for PyTorch and JAX, and persistent storage volumes. A practical choice for cost-conscious teams that want both raw GPU access and a cheap inference API from one provider.
- Among the lowest per-token LLM inference pricing in the market
- Batch inference API at up to 50% discount for high-volume workloads
- 100+ open-source models including Llama 3.2, Qwen 2.5, DeepSeek R1
- Fast GPU provisioning — typically under 3 minutes
- Pre-built PyTorch and JAX environments reduce setup time
- Smaller provider — cluster scale limited vs CoreWeave or Lambda for very large jobs
- Inference API latency trails Fireworks and Cerebras on speed benchmarks
- Less mature enterprise support and SLA options
Affiliate link — we may earn a commission.
No reviews yet — add yours ↗
Salad orchestrates idle consumer GPUs from homes and small businesses worldwide into a single API-driven platform. GPU catalog: RTX 5090, RTX 4090, RTX 5080, RTX 4080, RTX 3090. Four priority tiers: High (never preempted), Medium, Low, Lowest (batch, cheapest). vCPU and RAM included with every GPU. Products: Community Cloud (lowest-cost distributed pool), Salad Dedicated (reserved RTX PRO 6000 Blackwell bare metal on contract), AI Gateway (OpenAI-compatible LLM endpoint), Transcription API, Private AI Cloud (run on your own idle GPUs). SOC 2 Type I. Trusted by Civitai (10M images/day).
- RTX 4090 (Lowest priority) — cheapest RTX 4090 inference anywhere
- RTX 5090 (Lowest priority) — Blackwell at consumer prices
- Four priority tiers: High/Medium/Low/Lowest — balance cost vs reliability
- vCPU and RAM included with every GPU at no extra charge
- Per-second billing, no cold-start charges, no minimums
- Salad Dedicated: RTX PRO 6000 Blackwell bare metal on contract
- Consumer GPU network — not suitable for sub-10ms latency or NVLink multi-GPU training
- No data-residency guarantees on Community Cloud
- Interruptible nodes require fault-tolerant workload design
No reviews yet — add yours ↗
Live GPU Pricing
Region Coverage
Popular Comparisons
Provider pages
More Novita comparisons
Key differences
Novita
GPU cloud and LLM inference API — H100, A100, and 100+ open-source models
Salad
RTX 5090, RTX 4090 (Lowest priority) — distributed idle GPUs, per-second billing