# SAMBANOVA IR snapshot — retained record

> Retained ir record from ir:SAMBANOVA.

Canonical: https://highsignal.app/data/records/9cd22e4e548e901d

[← sources](https://highsignal.app/data)


source: [ir](https://highsignal.app/data/ir) (ir:SAMBANOVA)  published 2026-06-05 09:29 ingested 2026-06-05 09:35

[SambaNova Systems](https://highsignal.app/entities/SAMBANOVA)[original source ↗](https://sambanova.ai/)

## extract

Pushing the AI frontier with premium inference
Run the largest models by maximizing dataflow efficiency with
high speed and sustained throughput.
[Connect with Experts](https://highsignal.app/contact)
The first disaggregated inference demo for AI agents is live
At [COMPUTEX](https://newsroom.intel.com/artificial-intelligence/intel-announces-new-ai-innovations-at-computex), SambaNova demonstrated what the next era of AI inference looks like: Premium inference for AI agents powered by GPUs and RDUs, running live in the newly-announced VC2 data center for the first time.
Introducing the SN50 RDU - our fifth-generation AI chip!
Only chip to deliver the speed and throughput required for agentic AI. Purpose-built for agentic inference delivering the best tokens per watt.
Inference stack by design
Inference at scale
The groundbreaking dataflow technology and memory architecture delivers the performance and speed required for ever-growing AI models.
[Learn more →](https://sambanova.ai/products/rdu-ai-chips)
[Inference at scale](https://sambanova.ai/products/rdu-ai-chips)
Energy efficiency
Generating the maximum number of tokens per watt with the highest power efficiency naturally enables fast inference and scalability.
[Learn more →](https://sambanova.ai/products/sambarack)
[Energy efficiency](https://sambanova.ai/products/sambarack)
Infrastructure flexibility
SambaStack switches between multiple frontier-scale models, enabling complex agentic AI workflows to execute end-to-end on one node.
[Learn more →](https://sambanova.ai/products/sambastack)
[Infrastructure flexibility](https://sambanova.ai/products/sambastack)
Why Modern Al Infrastructure Demands Model Bundling
Not One-Model-Per-Node Thinking
[Learn more](https://highsignal.app/blog/modern-ai-infrastructure-demands-model-bundling)
The Goldilocks Zone for agents
The SN50 delivers 3X the savings compared to competitive chips for agentic inference. Co-Founder and Chief Technologist Kunle Olukotun explains how SN50 tiered memory allows agents to have access to a cache for models and prompts, further improving efficiency.
Build with the best open-source models
MiniMax M2.7
The latest frontier model from MiniMax, MiniMax M2.7 is the first in the M-series that deeply participates in its own evolution.
Built for complex agent harnesses, multi-agent collaboration, and high-fidelity productivity tasks, M2.7 hits 56.22% on SWE-Pro, 55.6% on VIBE-Pro for end-to-end project delivery, and 57.0% on Terminal Bench 2.
Best of all, SambaNova leads on speed at 435 output tokens/s, >3x faster than any other provider.
DeepSeek
We support the groundbreaking DeepSeek models, including the 671-billion-parameter DeepSeek-V3.1, which excels in coding, reasoning, and mathematics at a fraction of the cost of other models.
On our SambaNova RDU, DeepSeek-V3.1 achieves remarkable speeds of up to 200 tokens / second, as measured independently by Artificial Analysis.
Llama
As a launch partner for Meta's Llama 4 series, we've been at the forefront of open-source AI innovation. SambaCloud was the first platform to support all three variants of Llama 3.1 (8B, 70B, and 405B) with fast inference.
We are excited to work with Meta to deliver fast inference on both Scout and Maverick models.
OpenAI gpt-oss-120b
OpenAI recently released gpt-oss-120b, a model that delivers high accuracy in just 120-billion parameter with a Mixture of Experts (MoE) architecture.
As a small but efficient model, it runs extremely fast on SambaNova RDUs at over 600 tokens per second, making it a great choice for near real-time agentic AI.
The only chips-to-model computing built for AI
Inference | Bring Your Own Checkpoints
SambaNova provides simple-to-integrate APIs for Al inference, making it easy to onboard applications. Our APIs are OpenAI compatible allowing you to port your application to
SambaNova in minutes.
Auto Scaling | Load Balancing | Monitoring | Model Management | Cloud Create | Server Management
SambaOrchestrator simplifies managing AI workloads across data centers. Easily monitor and manage model deployments and scale automatically to meet user demand.
SambaRack™ is a state-of-the-art system that can be set up easily in data centers to run Al inference workloads. SambaRack SN40-16 is our fourth generation system optimized for low power inference (average of 10 kWh) and running many models simultaneously.
SambaRack SN50 is our fifth-generation system optimized for fast agentic inference at a fraction of the cost running the largest models, like gpt-oss-120b and DeepSeek.
SN40 | SN50 RDU
At the heart of SambaNova's innovation lies the RDU. With a unique three-tier memory architecture and dataflow processing, RDU chips are able to achieve much faster inference using a lot less power than other architectures.
-
Complete AI platform that provides a fully integrated end-to-end agentic AI stack – spanning across agents, models, knowledge, and data.
-
Composable AI platform that is open, unifies structured and unstructured data, queries in any environment, and deploys on any AI model. Build or use pre-built AI agents — all with business-aware intelligence.
-
Sovereign AI platform that keeps data secure and governed while business teams query in any environment. IT stays in control, while business teams self-serve AI — and both can focus on what matters.

## retained text



```
AI Inference.
Blazing Fast. Superior Economics.
We power the world’s largest AI models with the fastest, most efficient inference engine, delivering unmatched speed and industry-best economics.
Inference is going to be the center of the AI economy
SambaNova CEO Rodrigo Liang joins CNBC's Squawk on the Street to explain why the economics of inference, and not training benchmarks, now decide what actually gets deployed. His case: 10x lower cost of inference, a six-month payback, and no new data center buildout required.
SambaNova Expands TACC Partnership
SAN JOSE, Calif. — Nov. 17, 2026— [SambaNova](https://sambanova.ai/), an AI chip company delivering the highest performance inference, today announced the expansion of its partnership with the [Texas Advanced Computing Center](https://tacc.utexas.edu/) (TACC) at The University of Texas at Austin.
Size your AI data center deployment
Our virtual walkthrough helps you visualize energy requirements, floor space, and scale out to multi-megawatt clusters with the SambaRack SN50 hardware.
The Air-Cooled Token Factory: Intel GPU D-Series + SambaRack SN50
Over the last several months, many customers have been evaluating and deploying disaggregated inference workloads to help optimize the speed and economics of AI.
Inference stack by design
[Inference at scale
The groundbreaking dataflow technology and memory architecture delivers the performance and speed required for large AI models.
Learn more →](/products/rdu-ai-chips)
[Energy efficiency
Generating the maximum number of tokens per watt with the highest power efficiency naturally enables fast inference and scalability.
Learn more →](/products/sambarack)
[Infrastructure flexibility
SambaStack switches between multiple frontier-scale models, enabling complex agentic AI workflows to execute end-to-end on one node.
Learn more →](/products/sambastack)
Dataflow Architecture Explained
RMSNorm
QKV
Attention
Projection
FFN
The Goldilocks Zone for agents
The SN50 delivers 3X the savings compared to competitive chip hardware for agentic inference. Co-Founder and Chief Technologist Kunle Olukotun explains how SN50 tiered memory allows agents to have access to a cache for models and prompts, further improving efficiency.
Serve the models your nation 
actually needs
The most capable models are now trillions of parameters, and they are the ones citizens and enterprises want to use.
Frontier open-weight models are optimized to run on RDUs, so you can offer what developers are asking for without a per-model engineering project each time.
[More on SambaRack](/products/sambarack)
The only chips-to-model computing built for AI
Inference | Bring Your Own Checkpoints
SambaNova provides simple-to-integrate APIs for Al inference, making it easy to onboard applications. Our APIs are OpenAI compatible allowing you to port your application to SambaNova in minutes.
Auto Scaling | Load Balancing | Monitoring | Model Management | Cloud Create | Server Management
SambaOrchestrator simplifies managing AI workloads across data centers. Easily monitor and manage model deployments and scale automatically to meet user demand.
SambaRack™ is a state-of-the-art hardware platform that can be set up easily in data centers to run Al inference workloads. SambaRack SN40-16 is our fourth-generation platform optimized for low power inference (average of 20 kWh) and running many models simultaneously.
SambaRack SN50 is our fifth-generation system optimized for fast agentic inference at a fraction of the cost running the largest models.
SN40 | SN50 RDU CHIP
At the heart of SambaNova's innovation lies the RDU chip. With a unique three-tier memory architecture and dataflow processing, RDUs are able to achieve much faster inference using a lot less power than other architectures.
- Complete AI platform that provides a fully integrated end-to-end agentic AI stack – spanning across agents, models, knowledge, and data.
- Composable AI platform that is open, unifies structured and unstructured data, queries in any environment, and deploys on any AI model. Build or use pre-built AI agents — all with business-aware intelligence.
- Sovereign AI platform that keeps data secure and governed while business teams query in any environment. IT stays in control, while business teams self-serve AI, and both can focus on what matters.
```



document fetched 2026-11-22 02:56
