Intel backs SambaNova’s $350M bid to challenge GPUs in AI inference

Steven

Intel backs SambaNova’s $350M bid to challenge GPUs in AI inference

Summary

AI infrastructure company SambaNova has closed a $350 million funding round that includes investment from Intel Capital and other major funds. The partnership with Intel is a multi-year collaboration aimed at offering customers an alternative to GPU-based generative AI deployments, with hardware-software co-design and integration with Xeon CPUs.

SambaNova plans to ship its SN50 reconfigurable dataflow units (RDUs) later this year. The SN50 touts 2.5x higher FP16 performance and 5x higher FP8 performance over the SN40L, claiming 1.6 and 3.2 petaFLOPS respectively. Each RDU features a three-tier memory hierarchy (on-chip SRAM, HBM2E and DDR5), fast chip-to-chip fabric and the ability to scale a single inference worker across up to 256 accelerators.

Key Points

  • SambaNova raised $350M; Intel Capital is a lead investor and will collaborate on multi-year hardware-software co-design.
  • SN50 RDUs claim 2.5x FP16 and 5x FP8 gains vs the prior generation, with quoted 1.6 and 3.2 petaFLOPS figures.
  • Memory configuration: 432MB on-chip SRAM, 64GB HBM2E (1.8TB/s), and 256GB–2TB DDR5 per RDU to enable fast model swapping and large key-value caches.
  • SambaNova argues its dataflow architecture reduces data-movement overheads, letting fewer accelerators achieve higher real-world throughput than GPUs.
  • Company claims up to 5x higher per-user generation speed versus Nvidia’s B200 in some workloads, and strong inference performance already demonstrated by SN40L.
  • SN50 scales up to 256 accelerators per inference worker with 2.2TB/s bidirectional chip-to-chip bandwidth via a switched fabric.
  • Trade-offs: SN50’s peak dense FP8, HBM capacity and memory bandwidth are lower on paper than Nvidia Blackwell chips, but SambaNova emphasises achievable performance and utilisation economics.
  • SoftBank is signed up as an early customer; SambaNova intends to sell infrastructure rather than build a dedicated inference cloud.

Context and relevance

This is a notable vote of confidence from Intel into an alternative approach to GPU-dominated AI inference. Intel’s involvement brings scale, customers and deeper integration potential — which could accelerate adoption of dataflow architectures in production inference. For service providers and cloud operators struggling with per-rack economics as model customisation rises, SambaNova pitches better utilisation through fast model/context swapping and different scaling trade-offs.

The wider industry picture matters too: Nvidia remains dominant in training and increasingly in inference, but rivals and specialised startups are closing gaps on latency, cost and power for particular workloads. Intel backing could speed SambaNova’s market reach and put more pressure on GPU incumbents to optimise inference economics and adopt dataflow-like techniques.

Why should I read this?

Short version: Intel throwing cash and engineering weight at SambaNova could actually change the inference arms race. If you care about AI infrastructure costs, latency or how cloud providers will bill inference in 2026, this news is worth a skim (and a deeper read if you pick servers, buy racks or run inference fleets). We’ve done the heavy lifting — this piece tells you why the SN50 and Intel tie-up might matter to your budgets and SLAs.

Author style

Punchy: the article stresses the commercial stakes — SambaNova isn’t just benchmarking; it’s pushing a sellable alternative to GPUs, and Intel’s involvement gives credibility and distribution muscle that could be decisive.

Source

Source: https://www.theregister.com/2026/02/24/sambanova_intel_funding/