SambaNova
Fast per-token inference API (SambaCloud) for open models such as gpt-oss, Llama, GLM and MiniMax, running on SambaNova's own RDU chips
Overview
SambaNova builds its own AI chips, called RDUs, and sells inference on them. Developers reach it through SambaCloud, a per-token API for open models; enterprises can also buy SambaStack and SambaRack systems for their own data centers. The cloud model list in November 2026 includes gpt-oss-120b, Llama 3.3 70B, Gemma 4 31B, MiniMax M3.1 Flash and GLM-5.3, with prompt caching on some models and an Anthropic Messages API compatible endpoint added in July 2026. The company had a turbulent year: takeover talks with Intel, reported at about $1.6B, stalled around January 2026. SambaNova then raised a $350M+ Series E led by Vista Equity Partners and Cambium Capital with Intel Capital in February 2026, alongside its SN50 chip and an Intel partnership, and announced a $1B Series F first close at an $11B post-money valuation led by General Atlantic in July 2026. In October 2026 it said SN50 racks will be deployed in Oracle Cloud data centers. SambaNova claims SN50 is faster and cheaper to run than NVIDIA B200; that is vendor marketing we could not verify. Its closest rivals are other fast-inference chip makers, Cerebras and Groq, plus GPU clouds like Together AI and Fireworks AI.
The Verdict
Who Should Use SambaNova?
Best For
- Developers who need very fast output on popular open models
- Agent workloads that benefit from low latency and prompt caching
- Enterprises that want on-prem racks or Oracle Cloud capacity
Not Ideal For
- Teams that need closed frontier models or a huge model menu
- Teams that want to deploy their own custom fine-tunes cheaply
What's Great
- Very fast inference speeds on open models
- Easy to test models in the cloud playground
- Competitive per-token prices on open models
Watch Out For
- Short model list compared with GPU clouds
- Few public user reviews to judge reliability
- Leans toward enterprise deals; ownership news has shifted often
Pricing
View all features & details
Key Features
- SambaCloud per-token API for open models
- Models: gpt-oss-120b, Llama 3.3 70B, Gemma 4 31B, MiniMax M3.1 Flash, GLM-5.3
- Prompt caching on MiniMax models (Oct 2026)
- Anthropic Messages API compatibility (Jul 2026)
- RDU chips, newest SN50
- SambaStack and SambaRack for on-prem deployment
Company
- Intel takeover talks stalled (Jan 2026)
- Series E $350M+ led by Vista and Cambium (Feb 2026)
- Series F $1B first close, $11B valuation (Jul 2026)
- SN50 racks coming to Oracle Cloud (Oct 2026)
How It Compares
| Feature | SambaNova | Cerebras | Groq |
|---|---|---|---|
| Hardware | RDU chips (SN50) | Wafer-scale engine | LPU |
| Access | SambaCloud API, on-prem racks | Cloud API, systems | Cloud API |
| Focus | Fast open-model inference | Fast open-model inference | Fast open-model inference |
Sources: SambaNova · SambaCloud pricing (fetched Nov 2026) · SiliconANGLE (Feb 2026) · VKTR (Jul 2026) · Bloomberg Law (Jan 2026)
Is this your tool? Add a badge