AI Inference Acceleration
SambaNova SambaRack SN50: how do I request a quote and what facilities info (power/cooling) do you need?
10 min read
Most infrastructure teams look at SambaRack SN50 once they’re ready to run the largest models for agentic inference at scale, but the practical questions come fast: how do you request a quote, and what power and cooling information do you need ready for facilities and data center planning?
Quick Answer: To request a SambaRack SN50 quote, you contact SambaNova’s sales team through the “Get Started” flow and share your workload, deployment timeline, and data center requirements. Facilities teams typically need rack footprint, power feed requirements, cooling assumptions, and any constraints on power density and redundancy so SambaNova can size the configuration correctly.
The Quick Overview
- What It Is: SambaRack SN50 is SambaNova’s fifth-generation, rack-level AI inference system, built around RDU chips with a three-tier memory architecture for fast, efficient agentic inference on the largest models.
- Who It Is For: Platform, infra, and data center teams that need to run multi-model, agentic LLM workloads (like DeepSeek and gpt-oss-120b) at high throughput and predictable power envelopes.
- Core Problem Solved: It removes the “one-model-per-node” bottleneck by letting you bundle and switch between multiple frontier-scale models on a single node, while keeping power, cooling, and operational overhead within data center limits.
How It Works
SambaRack SN50 is delivered as a data-center-ready rack system integrating SambaNova’s RDUs, networking, and management into a single deployment unit. The stack is designed for inference-first workloads: SambaStack runs on RDUs with custom dataflow processing and a three-tier memory architecture, while SambaOrchestrator provides autoscaling, load balancing, monitoring, and model management across your racks or data centers.
You typically go from quote request to deployment planning in three phases: scoping, facilities review, and configuration & pricing.
- Scoping the Workload & Deployment Model:
- Define target models (e.g., DeepSeek-R1, gpt-oss-120b, Llama family).
- Describe expected traffic patterns (tokens/sec, concurrent sessions, regions).
- Clarify deployment model: on-premises, co-lo, or sovereign/partner data center.
- Facilities & Data Center Review:
- SambaNova’s team works with your infra and facilities leads to understand rack space, power feeds, cooling capacity, and redundancy requirements.
- This is where you align on power density, cooling modality (air vs. liquid in your environment), and controls/monitoring expectations.
- Configuration & Quote:
- SambaNova proposes an SN50 configuration (rack count, RDU capacity, networking, orchestration) matched to your constraints.
- You receive commercial terms, deployment timelines, and a technical overview that facilities can baseline against their power and cooling budgets.
Features & Benefits Breakdown
| Core Feature | What It Does | Primary Benefit |
|---|---|---|
| Fifth-generation SambaRack SN50 system | Integrates RDUs, networking, and management into a rack-level inference system optimized for fast agentic inference on the largest models. | High-throughput agentic workloads at a fraction of the cost and power versus traditional GPU stacks. |
| RDU with three-tier memory architecture | Uses custom dataflow processing and tiered memory to keep models and prompts “hot,” minimizing data movement. | More tokens per watt and better real-world agent loop performance (fewer stalls on memory and I/O). |
| SambaStack + SambaOrchestrator | Full-stack inference with orchestration for Auto Scaling | Load Balancing |
| OpenAI-compatible APIs via SambaCloud | Lets you access SN50-backed inference using familiar APIs and port existing applications in minutes. | Low switching cost from existing OpenAI-based apps to SambaNova infrastructure. |
| Model bundling & multi-model support | Switch between frontier-scale models on one node and run many models simultaneously. | Avoids one-model-per-node patterns, reducing hardware sprawl and data center overhead. |
Ideal Use Cases
-
Best for production agentic inference on frontier-scale models:
Because SambaRack SN50 is “optimized for fast agentic inference at a fraction of the cost running the largest models, like gpt-oss-120b and DeepSeek,” it’s ideal when you’re orchestrating multi-step agents, tools, and routing across models and need consistent throughput. -
Best for data center or sovereign AI deployments with tight power/cooling envelopes:
Because the RDU’s dataflow architecture and tiered memory maximize tokens per watt, facilities teams can support high-density inference without blowing past power or cooling budgets.
Limitations & Considerations
-
Facilities constraints drive configuration options:
If your data center has strict limits on rack power density, total power draw, or cooling capacity, SambaNova will tailor SN50 configurations accordingly. Early sharing of these constraints speeds up quoting and avoids designs that can’t be deployed. -
Lead times and deployment coordination:
SambaRack SN50 is a rack-level system, not a single card or instance. Plan for procurement, delivery, racking, and integration with your existing monitoring and control planes. Engaging facilities and network teams during the quote phase is essential.
Pricing & Plans
SambaRack SN50 is priced as an integrated rack-level system rather than a per-instance cloud SKU. Pricing depends on:
- Number of racks and RDUs required to hit your target tokens/sec and concurrency.
- Redundancy requirements (N+1, multi-rack, multi-site).
- Deployment model (on-premises, co-location, or partner/sovereign data center).
- Support, orchestration, and any managed-services overlay.
You start the process by engaging SambaNova’s team through the “Get Started” flow. From there, configurations are customized to your workload and facilities profile.
- Inference-Optimized Rack Configurations: Best for enterprises and service providers needing sustained production inference on large models with predictable power usage.
- Sovereign / Co-lo-Aligned Configurations: Best for organizations with strict data residency, security, or regional requirements that still need SN50’s agentic inference performance.
How to Request a SambaRack SN50 Quote
From a practical operations perspective, requesting a quote is straightforward, but the quality of the quote depends on how clearly you describe your workload and facilities constraints.
Step 1: Initiate Contact
- Go to the SambaNova site and use the Get Started / contact form:
Get Started - Indicate you’re interested in SambaRack SN50 and specify that you are looking for on-prem / data center deployment.
Step 2: Provide Workload & Architecture Context
Be prepared to share:
- Workload type:
- Agentic inference, RAG, chat, tools/agents, batch generation, etc.
- Target models:
- Example: DeepSeek-R1, gpt-oss-120b, Llama family, or your own checkpoints (bring-your-own-checkpoints supported).
- Performance targets:
- Tokens/sec required for steady-state, peak traffic, and tail latency expectations (P95/P99 if you have them).
- Multi-model behavior:
- Number of models, switching behavior, and whether they need to be resident at once (key for model bundling on SN50).
- Integration approach:
- Direct on-prem deployment or via SambaCloud/OpenAI-compatible APIs with local or hybrid integration.
The more concrete your throughput and concurrency requirements, the more precise the SN50 sizing and quote.
Step 3: Share Facilities & Data Center Details
This is where power and cooling come into focus. Data center and facilities teams should be ready with:
- Location & type of deployment:
- Internal data center vs. co-location vs. sovereign/partner facility.
- Available rack space:
- How many racks you can allocate today and any constraints on rack height or footprint.
- Power availability per rack:
- Maximum kW per rack (continuous and peak).
- Number and type of power feeds (e.g., dual-feed, voltage, and phase).
- Existing PDU configurations and redundancy requirements (e.g., N+1, 2N).
- Total power budget:
- How much additional power the site can support for AI infrastructure overall.
- Cooling capacity and design:
- Cooling per rack (kW) you can support.
- Whether your environment is purely air-cooled or supports higher-density or liquid-assisted solutions.
- Hot aisle / cold aisle layout details and any known thermal hot spots or constraints.
- Environmental and compliance requirements:
- Specific standards, regional regulations, or internal policies that affect equipment choice, cabling, or layout.
Even though SambaRack SN50 is not explicitly stated as using the 10 kWh profile (that’s called out for SN40L-16), it is still an inference-optimized system designed for efficient, large-model workloads. Power and cooling info allows SambaNova to right-size SN50 so you stay within site limits.
Step 4: Network & Integration Considerations
While not strictly facilities, including network details in the quote process helps avoid later redesigns:
- Uplink bandwidth per rack and desired topology.
- Latency constraints within and across data centers.
- Integration with your existing observability stack for monitoring and logging.
- Any segmentation or network security policies that influence rack placement.
Step 5: Review Proposed Configuration & Quote
Once SambaNova has:
- Your workload profile (models, tokens/sec, concurrency).
- Your facilities constraints (rack count, power density, cooling capacity).
- Your network and integration context.
They’ll propose a SambaRack SN50 configuration and provide:
- Hardware configuration (rack count, RDU density, networking).
- Expected performance (tokens/sec, throughput envelopes for key models).
- Power/cooling profile aligned with your facility.
- Pricing, support options, and deployment timelines.
You can then iterate with your infra, facilities, and procurement teams to align on final terms.
Facilities Power & Cooling Information: Checklist
To streamline the quote and avoid back-and-forth, have this ready:
- Rack & Space
- Number of racks available
- Rack dimensions and any non-standard constraints
- Power
- Max kW per rack (continuous)
- Power feed specs (voltage, phase, plug type)
- Redundancy requirements (N, N+1, 2N)
- Site-wide additional power budget for AI
- Cooling
- Cooling capacity per rack (kW)
- Cooling design (air only, hot/cold aisle, liquid availability)
- Known limits on inlet temperatures / delta-T
- Operations
- Data center access and change windows
- Monitoring and alerting integration requirements
- Any special compliance or safety constraints
Providing this early lets SambaNova design an SN50 deployment that your facilities team can sign off on without surprises.
Frequently Asked Questions
How do I specifically request a SambaRack SN50 quote rather than a generic AI solution?
Short Answer: Use the SambaNova contact form, specify SambaRack SN50, and include both workload and facilities details in your initial message.
Details:
Go to https://sambanova.ai/contact and indicate that you are evaluating SambaRack SN50 for on-prem or data-center deployment. In the description, outline:
- Your target models (e.g., gpt-oss-120b, DeepSeek-R1).
- Expected tokens/sec or user concurrency.
- Where you want to deploy (data center/region).
- Any known power/cooling constraints or rack limits.
This routes you to the right technical and commercial teams, so your first follow-up includes realistic SN50-based options rather than generic cloud or smaller-system proposals.
What power and cooling information do you absolutely need before SambaNova can size an SN50 deployment?
Short Answer: SambaNova needs your per-rack power limit, cooling capacity, and how many racks your data center can allocate, plus redundancy and compliance constraints.
Details:
While SambaNova can start conversations with rough estimates, meaningful SN50 sizing requires:
- Per-rack power limit: kW available under your normal operating profile and any absolute maximum.
- Cooling per rack: How many kW of heat your cooling system can reliably remove per rack.
- Rack count and layout: How many racks you can dedicate to SN50, and whether they must be grouped or spread.
- Redundancy and standards: Dual feeds, PDU setups, and any facility standards (e.g., power quality, environmental thresholds).
With this, SambaNova can propose SN50 configurations that meet your throughput goals while staying within your mechanical and electrical envelopes.
Summary
SambaRack SN50 is purpose-built for fast, large-model agentic inference in real data centers, where power, cooling, and operational complexity are non-negotiable constraints. To request a quote, you start a conversation via SambaNova’s Get Started flow and come prepared with two things: a clear picture of your LLM workloads (models, tokens/sec, concurrency) and concrete facilities information (rack space, power limits, cooling capacity, redundancy).
With that information, SambaNova can design a SambaRack SN50 configuration that hits your performance targets, respects your power and cooling budgets, and slots into your existing operations with OpenAI-compatible APIs and full-stack orchestration.