SambaNova has raised $1 billion in its Series F funding round, valuing the company at $11 billion, a substantial increase from the $1.6 billion Intel considered for acquisition in December 2025. The round includes a $3.5 billion order from VC2 and sees JPMorganChase as an on-prem inference customer, marking the largest public commitment to SambaNova's Reconfigurable Dataflow Unit (RDU) architecture to date. This indicates a growing trend among regulated enterprises to purchase non-GPU silicon for sovereign, air-cooled racks.

JPMorganChase will deploy SambaNova's SN40 and SN50 systems in a multi-year agreement, running inference within existing data centers on standard Ethernet and Red Hat Enterprise Linux, without the need for liquid cooling. The vertically integrated stack comprises RDU silicon, SambaRack enclosures, and SambaStack orchestration software, which offers standard API endpoints and integrates with Kubernetes and existing inference platforms. At Computex, SambaNova demonstrated disaggregated inference with Intel Xeon 6 CPUs, Nvidia Blackwell GPUs for prefill, and SN40 chips for decode, claiming 2×–3× the per-chip performance of a GPU-only baseline while reusing GPUs' onboard HBM. A co-development agreement with Intel in February aims to further integrate CPU-to-RDU.

SambaNova's $11B valuation positions it ahead of Groq but below Cerebras in the AI-chip funding race.
FIG. 02 SambaNova's $11B valuation positions it ahead of Groq but below Cerebras in the AI-chip funding race. — Cerebras, Groq, and Intel public records

SambaNova asserts that the SN50 can handle agentic workloads at about one-third the cost of GPU-based systems, supports clusters up to 256 chips, and manages models at multi-trillion-parameter scale. The $3.5 billion VC2 order, aimed at ultra-low-latency metro token delivery, will be deployed over three years using the Nvidia-plus-SambaNova disaggregated architecture. With 40 to 50 named customers, including Saudi Aramco and Intel, and a $350 million-plus Series E in 2026, SambaNova anticipates profitability in 2027.

SambaNova SN50 claims ~3× cost advantage on agentic AI workloads versus GPU-based inference.
FIG. 03 SambaNova SN50 claims ~3× cost advantage on agentic AI workloads versus GPU-based inference. — SambaNova technical claims, source [6]

Production evidence for the SN50 is pending, with shipments expected to start in the second half of 2026, making JPMorgan's and VC2's commitments future-oriented. All performance and cost claims are vendor-supplied and lack independent MLPerf or public benchmark verification. Architects must consider whether SambaStack's abstraction layer introduces latency or compatibility issues with standard vLLM, TGI, or TensorRT-LLM serving stacks when inference spans different silicon vendors.

CEO Rodrigo Liang bets that enterprises will not quantize multi-trillion parameter models and will need memory capacity beyond what standard GPU configurations can economically provide. Grand View Research estimates the AI inference market at $97.24 billion in 2024, projecting growth to $253.75 billion by 2030. SambaNova's capital efficiency hinges on shipping racks rather than building data centers, but the company allocated Series F capital to secure the supply chain for the next 12 months, acknowledging that silicon and system availability, not just software maturity, is the immediate constraint.

Architects should consider the disaggregated prefill-decode pattern, inserting a specialized decode accelerator into existing GPU-populated data centers to reuse stranded HBM, avoiding the need to replace the entire fleet, which is worth prototyping if the p99 latency budget can accommodate cross-vendor orchestration overhead.

Written and edited by AI agents · Methodology