SambaNova Unveils SN50 AI Accelerator, Partners with Intel for Xeon-Based Inference Infrastructure

SambaNova has unveiled its new SN50 AI accelerator, positioning it as a purpose-built solution for inference workloads rather than training. The company claims the SN50 delivers five times more compute per accelerator and four times the networking bandwidth compared to its predecessor, while offering three times better efficiency than Nvidia's B200 GPU for inference tasks.

SN50 Architecture and Performance

The SN50 is built on SambaNova's Reconfigurable Data Unit (RDU) architecture and features a three-tier memory subsystem consisting of SRAM, HBM, and DDR5. This design enables multiple models to remain resident in memory for rapid hot-swapping, optimizing both memory utilization and power consumption.

The dual-chiplet processor is designed specifically for low-latency, real-time applications such as voice assistants, with a focus on memory and network bandwidth rather than pure compute performance. Each SN50 RDU processor delivers five times more compute than competing offerings, according to SambaNova, though the company hasn't disclosed which specific products it's comparing against.

SambaNova positions the SN50 primarily as part of its rack-scale SambaRack SN50 solution rather than as a standalone processor. Each 20 kW SambaRack SN50 contains 16 SN50 RDU processors, with up to 256 accelerators interconnectable across 16 racks using a multi-terabyte-per-second fabric. The company emphasizes that the 20 kW per rack power envelope operates within existing data center capabilities and relies on air cooling, eliminating the need for liquid cooling infrastructure.

For extremely large models, a cluster of 256 SN50 accelerators can handle configurations exceeding 10 trillion parameters with context windows of more than 10 million tokens. This capability is particularly relevant for reasoning-heavy and multi-model agentic AI workloads that demand both scale and responsiveness.

Performance Benchmarks

SambaNova cites SemiAnalysis's InferenceX benchmark results to support its performance claims. When running with FP8 precision, a Llama 3.3 70B model with 1K input and 1K output tokens reportedly achieves 895 tokens per second per user on the SN50, compared to 184 tokens per second per user on Nvidia's B200.

Across various configurations, throughput per-RDU is presented as significantly higher than per GPU, with an average advantage of approximately 3x when latency constraints are applied across Llama 70B, GPT-OSS 120B, and DeepSeek 671B models. This translates to improved cost-per-token economics for inference workloads.

Strategic Partnership with Intel

The week's announcements include a multi-year strategic collaboration between Intel and SambaNova aimed at building large-scale AI inference infrastructure around Intel Xeon platforms and SambaNova AI accelerators. Under this agreement, the companies will offer rack-scale solutions for AI workloads based on Intel Xeon processors and SambaNova AI accelerators for several years.

While SambaNova hasn't disclosed which specific CPU powers its current SambaRack SN50, the partnership indicates that future racks from the company will be Xeon-based. This collaboration targets select applications and customer types, including AI inference solutions for AI-native companies, model providers, as well as enterprises and government organizations worldwide.

The partnership makes strategic sense given that enterprises and governments represent Intel's traditional customer base. This coordinated go-to-market strategy through Intel's enterprise and cloud channels could provide SambaNova with access to production-ready inference systems for these sectors.

Intel emphasized that this agreement with SambaNova complements its existing GPU roadmap rather than replacing it, indicating the company will continue offering its own GPUs for inference and eventually for training workloads.

Controversial Investment Structure

One notable aspect of the announcement is Intel Capital's participation in SambaNova's Series E financing round. This investment structure raises questions given that SambaNova is chaired by Lip-Bu Tan, who also serves as Intel's chief executive.

While strategic investments between large companies and their ecosystem partners are common, having a company's top executive chair a business in which the company's investment arm takes a stake is unusual. This arrangement could potentially create conflicts of interest or governance challenges.

SoftBank Partnership

Separately, SambaNova has partnered with SoftBank to deploy the SN50 at the latter's data centers in Japan. SoftBank will be the first to roll out SN50 in its next-generation AI data centers, using the accelerator to drive low-latency inference for sovereign and enterprise customers across Asia-Pacific.

These data centers will run open-source and proprietary frontier models with strict performance requirements. The move expands SoftBank's existing partnership with SambaNova, which already operates SambaCloud in the region. The new clusters based on the SN50 will become standard for SoftBank's new data centers, effectively making SambaNova the core inference platform for SoftBank's sovereign AI programs and upcoming large-scale agentic deployments.

Hironobu Tamba, Vice President and Head of the Data Platform Strategy Division at SoftBank Corp., stated that standardizing on SN50 provides the ability to deliver world-class AI services with the performance of the best GPU clusters but with better economics and control.

Funding and Market Positioning

SambaNova has secured $350 million in strategic Series E funding to expand manufacturing and cloud capacity from investors including Vista Equity Partners, Cambium Capital, Intel Capital, and Battery Ventures. This funding positions SambaNova strongly in the AI inferencing market.

The company's focus on inference rather than training reflects a strategic bet that the inference market could be larger than the training market. As training workloads demand the best and fastest processors, inference introduces different requirements, with the ongoing Intel partnership set to benefit SambaNova, especially as agentic AI inferencing workloads demand fast CPUs.

SambaRack SN50 systems will be available in the second half of 2026, though pricing has not been disclosed. The company's emphasis on air cooling, existing data center compatibility, and competitive performance metrics positions it as a compelling alternative to GPU-based inference solutions for enterprises and governments seeking to deploy large-scale AI infrastructure without extensive data center modifications.