Novita AI drops RTX 6000 Ada 48GB from self-serve GPUs, adds a base RTX 5090 tier
Novita AI removed the RTX 6000 Ada 48GB instance from its self-serve GPU lineup and added a new base RTX 5090 32GB on-demand tier at $0.73/hr, while its model catalog grew from 172 to 177 listings.
Self-serve `/en/gpus` lineup: RTX 4090 24GB, RTX 4090 24GB (HF), RTX 5090 32GB (HF) at 0.72/hr on-demand (0.36/hr spot), and RTX 6000 Ada 48GB at 0.77/hr on-demand (0.39/hr spot); 172 models listed on the catalog.
Self-serve `/en/gpus` lineup: RTX 4090 24GB, RTX 4090 24GB (HF), RTX 5090 32GB (HF) at 0.72/hr, plus a new base RTX 5090 32GB tier at 0.73/hr on-demand (0.37/hr spot); RTX 6000 Ada 48GB no longer listed; 177 models on the catalog.
Novita AI’s self-serve GPU-instance page (/en/gpus) swapped one SKU for another between the 2026-07-29 and 2026-08-04 captures. RTX 6000 Ada 48GB — the highest-VRAM card in the self-serve lineup at $0.77/hr on-demand — was removed entirely, continuing a pattern of narrowing the self-serve page to consumer-class RTX 4090/5090 silicon (L40S and H100 SXM were pulled from the same page back in June). In its place, a second RTX 5090 32GB listing appeared: a base on-demand tier at $0.73/hr ($0.37/hr spot), sitting alongside the existing “High frequency” RTX 5090 variant at $0.72/hr ($0.36/hr spot) — so the base SKU is now priced a cent above the high-frequency one, an inversion of the RTX 4090 pair, where the high-frequency variant costs roughly double the base rate.
Dedicated endpoints, bare-metal nodes, and the Agent Sandbox rate card were all unchanged in this capture. The model catalog grew modestly from 172 to 177 listings, adding SKUs including Deepseek V3.2, Deepseek V4 Flash, DeepSeek-OCR 2, and PaddleOCR-VL.
The /en/gpus self-serve lineup lost RTX 6000 Ada 48GB (was $0.77/hr on-demand, $0.39 spot) and gained a new base RTX 5090 32GB on-demand tier at $0.73/hr ($0.37 spot) — priced a cent above the existing 'High frequency' RTX 5090 variant at $0.72/hr, an inversion of the RTX 4090 pair where the HF variant costs roughly double the base rate. Dedicated endpoints, bare-metal, and Agent Sandbox rate cards held steady; the model catalog grew from 172 to 177 with additions including Deepseek V3.2, Deepseek V4 Flash, DeepSeek-OCR 2, and PaddleOCR-VL.