← Back to research
•·5 min read·product

SambaNova

SambaNova builds custom RDU chips and enterprise AI platforms for on-premise and cloud inference. After 2025 layoffs and a pivot to inference, it raised a $350M Series E with Vista and Intel in February 2026.

https://sambanova.ai ↗
ai-inferencecustom-siliconhardwareenterpriseon-premise

Key takeaways

  • Custom Reconfigurable Dataflow Unit (RDU) chip designed specifically for AI workloads; fifth-generation SN50 launched for agentic inference
  • Pivoted from training to inference in April 2025, cutting ~15% of staff (77 employees)
  • Raised a $350M Series E in February 2026 led by Vista Equity Partners with Intel as strategic investor (~$2.2B valuation, down from $5.1B in 2021) after Intel's ~$1.6B acquisition offer collapsed
  • SambaCloud provides an OpenAI-compatible inference API; sovereign AI data center partners live in Australia, Europe, and the UK

FAQ

What is SambaNova?

A company building custom RDU (Reconfigurable Dataflow Unit) chips and enterprise AI platforms for on-premise and cloud deployments.

What is an RDU?

A Reconfigurable Dataflow Unit — custom silicon that can be reconfigured for different AI workload patterns, unlike fixed GPU architectures.

Does SambaNova offer cloud inference?

Yes. SambaCloud provides an OpenAI-compatible inference API for models like DeepSeek-V3.2, Llama 3.3, and gpt-oss-120b, in addition to on-premise SambaRack hardware deployments.

Is SambaNova still independent?

Yes. Intel's reported ~$1.6B acquisition offer collapsed in late 2025; instead, SambaNova raised a $350M Series E in February 2026 led by Vista Equity Partners, with Intel taking roughly 9% ownership as a strategic investor alongside a multiyear collaboration agreement.

Company Overview

SambaNova builds custom AI chips (RDU — Reconfigurable Dataflow Unit) and enterprise AI platforms.[1] Founded by researchers from Stanford, the company targets enterprise and government customers who need AI infrastructure they can deploy on their own premises.

The company has been through a turbulent stretch. In April 2025 it laid off 77 employees (~15% of its roughly 500-person workforce) and pivoted away from training workloads to focus on inference.[2] Reported acquisition talks with Intel at roughly $1.6 billion collapsed in late 2025; instead, SambaNova closed a $350 million Series E in February 2026 led by Vista Equity Partners, with Intel investing for roughly 9% ownership plus a multiyear collaboration agreement — at an implied valuation around $2.2 billion, well below its $5.1 billion 2021 peak.[3] As of June 2026 the refocused company has shipped its fifth-generation SN50 chip and expanded sovereign-AI deployments with four data center partners live across Australia, Europe, and the UK (SouthernCrossAI, Infercom, OVHcloud, Argyll).[1]

What It Does

  • SambaCloud — OpenAI-compatible inference API serving open models including DeepSeek-V3.1/V3.2, Llama 3.3-70B, MiniMax M2.7, and gpt-oss-120b[4]
  • SambaRack systems — On-premise hardware with RDU chips: the fourth-generation SN40-16 (low-power inference) and fifth-generation SN50 (agentic inference)[1]
  • SambaStack — Integrated chips-to-model computing platform for model deployment and management
  • SambaManaged — Managed deployment option for enterprises and sovereign data center partners

How It Works

The RDU (Reconfigurable Dataflow Unit) differs from both GPUs and other custom silicon:

  • Reconfigurable — Hardware dataflow can be configured for different model architectures
  • Dataflow architecture — Data moves through compute units rather than being fetched from memory
  • Terabytes of memory — Large memory capacity for serving multiple models
  • Software-defined — SambaFlow compiler optimizes models for RDU automatically

For cloud users, SambaCloud provides standard OpenAI-compatible inference APIs.[4] For enterprise, SambaRack systems deploy in customer data centers. The new SN50 RDU is pitched specifically at agentic inference — chained LLM calls where latency compounds — with SambaNova claiming up to 435 output tokens/s on DeepSeek-V3.1.[5]

Pricing

  • SambaCloud — Free developer tier with API key signup; per-token pricing for production (published at cloud.sambanova.ai/plans/pricing)[4]
  • Enterprise hardware — Custom pricing for SambaRack systems
  • Managed deployments — Enterprise contracts with support via SambaManaged

Strengths

  • On-premise option — Critical for regulated industries (government, healthcare, finance)
  • Reconfigurable architecture — Adapts to different model types without hardware changes
  • Large memory — Can serve multiple large models simultaneously
  • Enterprise relationships — Established in government and regulated sectors
  • Full stack — Hardware + software + support as integrated platform
  • Stanford research pedigree — Strong technical foundation

Weaknesses / Risks

  • Company turbulence — 15% layoffs in April 2025, a collapsed Intel acquisition, and a valuation down ~57% from its 2021 peak raise viability questions[2][3]
  • Abandoned training market — The pivot to inference-only narrows the addressable market and concedes training to NVIDIA
  • Limited public cloud presence — SambaCloud is newer and less proven than competitors
  • Enterprise-only pricing — On-prem systems require significant investment
  • Smaller developer community — Less mindshare than Groq or GPU platforms
  • Ecosystem maturity — Fewer supported models and integrations than GPU alternatives
  • Competition from NVIDIA — NVIDIA's own enterprise offerings (DGX) compete directly
  • Custom silicon risk — Long hardware development cycles vs fast-moving GPU roadmap

What Developers Say

Developer discussion of SambaNova is thinner than for Groq or Cerebras; the most substantive thread compares its RDU serving to Groq's LPU approach.[6]

"SambaNova run the models at full precision, and on a single node, whilst Groq runs a single model on hundreds of chips." — snhbsub, Hacker News[6]

"'nearly as fast' as the next best competitor is not going to cut it" — anon291, Hacker News, on competing against NVIDIA's dominance[6]

"This offers no insight into the architecture of either system or the chips they developed." — cry-oscillator, Hacker News, on SambaNova's benchmark marketing[6]

Competitive Landscape

vs. Cerebras: Both custom silicon with enterprise focus. Cerebras has the larger chip; SambaNova offers reconfigurability.

vs. Groq: Groq targets cloud API developers with speed. SambaNova targets enterprise on-prem — and after its 2025 pivot, all three custom-silicon players (with Cerebras) now sell LLM tokens-as-a-service from their own hardware.[2]

vs. NVIDIA DGX: NVIDIA has the broader ecosystem. SambaNova claims architectural advantages for AI-specific workloads.

vs. Baseten/Modal: GPU cloud platforms are more flexible and accessible. SambaNova wins for on-prem requirements.

Ideal User

  • Government and defense organizations requiring on-premise AI
  • Regulated industries (healthcare, finance) with data sovereignty requirements
  • Enterprise teams wanting integrated hardware+software AI platform
  • Organizations running multiple large models needing high memory capacity

Bottom Line

SambaNova is the enterprise on-prem play in custom AI silicon, now refocused entirely on inference after the April 2025 layoffs and pivot.[2] The Vista-led $350M Series E with Intel as strategic partner stabilizes the balance sheet, but at a ~$2.2B valuation — less than half its 2021 peak — and a collapsed acquisition behind it, the company is in prove-it mode.[3] The SN50 and sovereign-AI data center expansion show real momentum in its niche.[1] Recommended for organizations where on-premise or sovereign deployment is a hard requirement. Not recommended as a general-purpose inference API over Groq, Cerebras, or GPU clouds given thinner ecosystem and company risk. Outlook: the Intel partnership is the wildcard — it could supply manufacturing scale and enterprise channels, or SambaNova remains a niche survivor in a consolidating custom-silicon field.

Sources

Share on XShare on LinkedIn