SambaStack
The AI inference stack you don't have to build
SambaNova's licensed, validated software stack for SambaRack, with serving, monitoring, and user management ready to run behind an OpenAI-compatible API.
Dedicated AI infrastructure simplified
SambaStack™ is the industry’s leading software stack purpose-built for AI inference on RDUs.
With the flexibility to deploy on-premises or in the cloud, organizations are empowered to accelerate their AI innovation with dedicated SambaNova infrastructure.
As SambaNova's licensed, validated software stack for SambaRack, SambaStack handles serving, monitoring, and user management, all ready to run behind an OpenAI-compatible API.
Inference serving stack above SambaRack
What you get on day one.
SHIP WITHOUT REWRITING
Point your applications at it and go.
Most AI tooling is already written against the OpenAI API, and SambaStack speaks it. Your applications connect without a rewrite, your developers keep the frameworks they know, and teams already running Dynamo or llm-d keep those too.
- OpenAI-compatible API, so existing applications connect with almost no changes
- Validated serving on vLLM, with Dynamo and llm-d gateways supported
- Keep the agentic frameworks your teams already work in, including CrewAI and Cursor
OPERATE WITHOUT FRICTION
Installs in your cluster. Plugs into your monitoring tools.
You still operate the platform. What you skip is standing up a new one to do it. SambaStack is Kubernetes-native and installs in your cluster with a Helm chart, and it exposes inference metrics for whatever monitoring tools you already run.
- Inference metrics exposed for your monitoring tools, with Prometheus-format metrics available for Grafana dashboards
- Model deployment, cluster management, autoscaling, and load balancing included
- One vendor accountable for the hardware and the software together
CONTROL WHAT YOU SERVE
Set the tiers, the limits, and the queues.
Usage plans control who reaches which models and how much they can consume. Give production teams generous limits and keep trial users on tighter ones, and see where the usage actually goes.
- Service tiers, rate limits, and queue depths per usage plan
- Control access by model, not just by cluster
- Sell a premium tier or allocate capacity fairly, and see where the usage goes
Running in production today.
General Compute
- A high performance inference cloud provider focused on coding and voice agents, where response time is critical. General Compute chose SambaNova Platform with SambaStack for its air-cooled racks that can be deployed directly into existing data centers.
- 3X faster inference than GPUs. 4X faster time to first token. 5X faster response times MiniMax 2.7
Jason Goodison, CTO, General Compute
OVHcloud
- A European cloud provider partners with SambaNova, leveraging SambaStack to serve open models and power premium AI endpoints, hot-swapping large open source models at runtime with fast inference speeds in a small energy-efficient footprint.
- 4X better energy efficiency. About 10 kW per rack.
Octave Klaba, CEO, OVHcloud
Texas Advanced Computing Center
- An academic supercomputing center serving researchers across the sciences. TACC uses SambaStack on-premises to run open-source AI workloads, data analysis, and coding agents that support scientific research across thousands of academic disciplines.
- Air-cooled design, energy efficient with seamless integration with OpenAI-compatible APIs.
Niall Gaffney, Director of Data Intensive Computing, TACC
US National Laboratories
- U.S. national laboratories (including Argonne, Los Alamos, Oak Ridge, and Lawrence Livermore) deploy SambaNova systems to run open foundation models locally for daily scientific coding agents and research workflows in controlled, air-gapped data centers, keeping primary supercomputers free for large-scale training and simulations while maintaining full data control and security.
- At Argonne, SambaStack runs inside the Leadership Computing Facility, serving about 1,000 scientists a day and 200 billion tokens a month on 20 kW or less.
Rick Stevens, Associate Lab Director for Computing Environment and Life Sciences, Argonne National Laboratory
FAQs
What is SambaStack?
SambaNova's licensed software stack for SambaRack. It provides a management layer for operating the platform, a serving layer for running models, and a developer layer behind an OpenAI-compatible API.
Do I need SambaRack to run SambaStack?
Yes. SambaStack runs on the SambaRack foundation.
Can I use my own serving stack instead?
Yes. SambaRack runs leading open-source inference engines, so teams with their own platform can integrate directly. SambaStack is for teams who would rather license one.
Which serving frameworks are supported?
Leading open-source frameworks including vLLM.
Will my existing applications work?
Yes, in most cases with almost no changes. SambaStack speaks the OpenAI API, which most AI tooling is already written against.
What monitoring tools does it use?
Prometheus and Grafana, so inference metrics land in the dashboards your team already runs.
Can I control who accesses which models?
Yes. Usage plans set service tiers, rate limits, and queue depths per group, so production teams and trial users can be treated differently.