SambaNova scales deployment through Intel and Supermicro

I see SambaNova moving from research into massive infrastructure deployments. JPMorgan Chase selected the company as an inference infrastructure partner to deploy SN40 and SN50 systems. Supermicro and SambaNova also plan to build disaggregated AI inference systems using 10,000 SN50 RDU chips. These systems combine NVIDIA GPUs for the prompt processing phase with SambaNova RDUs to handle the response generation. The hardware works in existing air-cooled data centers. The Intel collaboration provides a blueprint for agentic AI. The design uses Intel Xeon 6 processors as the host CPU for task coordination, workload distribution, and tool and API execution, while SambaNova RDUs function as the dedicated inference fabric for high-throughput, low-latency decode. This combination of hardware works because it separates the tasks. Intel Xeon 6 provides the memory bandwidth, PCIe lane density, and on-die accelerators. The Xeon 6 also acts as the action CPU to compile code and validate results. The SN50 RDU provides a software-optimized, reconfigurable dataflow processor that maps the AI model graph to the chip to optimize inference. The SN50 delivers 5x maximum speed and 3x throughput advantage over NVIDIA Blackwell B200 GPUs on agentic inference patterns for models such as Meta’s Llama 3.3 70B. The hardware targets NeoClouds and sovereign clouds before it reaches wider enterprise data centers. I would note that the reliance on NVIDIA GPUs for prefill still leaves a dependency on a primary competitor.

Samba-1 meets the trillion parameter requirement

SambaNova launched Samba-1, a one trillion parameter generative AI model for the enterprise. This model uses a Composition of Experts architecture to combine over 50 open source models. The system achieves 10X more efficiency than other models of similar size. The SN40L chip supports this model with a 1/10th hardware footprint compared to traditional setups. The SN40L chip manages both HBM and DRAM from a single unit to improve efficiency. The chip supports models up to 5 trillion parameters. The SambaNova Suite provides access to Llama2 variants, DeepSeek Coder, and the BLOOM 176B model. This includes Llama2 7B, 13B, and 70B, as well as DeepSeek Coder 1.3B, 6.7B, and 33B. Accenture and NetApp use these capabilities to solve business problems. The SN40L provides 256k sequence length on a single system node. The SN40L includes both dense and sparse compute and contains both large and fast memory. This system allows the model to function as a single Composition of Experts to solve business problems.

Infrastructure teams should watch these deployment patterns. The company delivers the model in private clouds or on-premises environments. This prevents customers from disclosing private data to public domains. The company faces strong competition from Cerebras and Groq in the inference market. I’ve noticed that the companies must prove they can win repeatable deployments rather than just showing interest in new options. You should observe how these hardware-agnostic deployment options evolve. Do the companies really need another way to run massive models in private?

Sovereign AI and huge funding rounds

SambaNova raised $1 billion in a Series F round led by General Atlantic. This funding brings the company valuation to $11 billion. The round included investments from Seligman Ventures, T. Rowe Price Associates, and Capital Group. New and existing investors also included A&E Investment, Assam Ventures, funds and accounts managed by BlackRock, Intel Capital, and the Qatar Investment Authority. The company will use this cash to expand capacity and scale deployments. The company has raised more than $1.14 billion in total funding. This amount comes five months after the company raised $350 million in a Series E round.

The company holds significant wins in the sovereign AI sector. SoftBank Corp. is the first customer for the SN50 chip in its Japanese data centers. Other customers include Argyll in the UK, Infercom with OVHcloud in the EU, and SouthernCrossAI in Australia. These deployments address the need for data residency and technology independence. These sovereign AI wins rely on localized support rather than just raw performance. The $350 million Series E round in February came from Vista Equity Partners and Cambium Capital. This earlier round included a $35 million investment from Intel. Intel previously considered an acquisition of SambaNova for $1.6 billion. SambaNova also generates revenue through SambaCloud and SambaManaged. The company employs 437 people to drive this international traction.