SambaNova Systems delivers AI chips that outpace Nvidia by 4x in energy efficiency, as validated by Stanford researchers. The startup’s Reconfigurable Dataflow Unit (RDU) runs massive models like DeepSeek-R1 at 200 tokens per second.
- SambaNova Systems delivers AI chips that outpace Nvidia by 4x in energy efficiency, as validated by Stanford researchers.
- The startup’s Reconfigurable Dataflow Unit (RDU) runs massive models like DeepSeek-R1 at 200 tokens per second.
- Backers included Intel Capital, GV, BlackRock, Temasek, and GIC.
- SambaRack systems average 10 kW for inference on gpt-oss-120B models.
Keep reading for the full breakdown on AI chip 4x energy efficient — everything you need to know is covered below.
This edge positions SambaNova to erode Nvidia’s grip on AI hardware markets.
SambaNova origins challenge Nvidia’s dominance
SambaNova Systems launched in 2017 in Palo Alto, California, and was founded by Stanford professors Kunle Olukotun, Rodrigo Liang, and Christopher Ré. One founder holds a MacArthur Genius Award for AI contributions. The team built the RDU from scratch for AI workloads, skipping graphics roots that define Nvidia GPUs.
Investors poured $676 million into a Series D round on April 13, 2021, led by SoftBank Vision Fund 2, pushing the valuation past $5 billion. Backers included Intel Capital, GV, BlackRock, Temasek, and GIC. Early funding, traced to a $56 million Series A in 2018, was led by Walden International.
Rodrigo Liang, founder and CEO of SambaNova Systems, remarked: “Today’s announcement signifies an extension of our existing collaboration with SoftBank Corp., highlighting SambaNova’s superior performance for swift inference.” Partnerships expanded to SoftBank Corp. on March 5, 2025, deploying SambaNova Cloud in APAC.
RDU technology targets Nvidia’s dominance
SambaNova’s RDU features a three-tier memory architecture optimised for large AI models. Independent tests show 198-255 output tokens per second on DeepSeek 671B with 16 RDUs, matching Nvidia’s eight GPUs but handling full 16-bit precision. Stanford’s Hazy Research measured RDU at 4X better Intelligence per Joule than Nvidia’s Blackwell B200.
Unlike Nvidia’s CUDA-locked GPUs, SambaNova offers a full-stack platform with OpenAI-compatible APIs, deployable on-premises or in the cloud in minutes. SambaRack systems average 10 kW for inference on gpt-oss-120B models.
Su Le, SambaNova’s Chief Growth Officer, stated: “Our goal is to deliver the fastest, most advanced vertically-integrated AI inference and training solutions to customers.”
The platform supports Llama 3.1 variants (8B, 70B, 405B) and DeepSeek-R1 at complete precision, enabling AI agents to run in seconds. Canvass AI partnered on October 10, 2024, to verticalize deployments in South Asia and EMEA.
Canvass CEO Humera Malik noted: “We’re delivering cutting-edge generative AI solutions and services alongside SambaNova, driving greater AI adoption across a diverse customer base.”
Intel acquisition talks Intensify Nvidia fight
Intel entered advanced talks to buy SambaNova for $1.6 billion, including debt, announced on December 13, 2025, with a potential closing in January 2026. A non-binding term sheet signed on December 9, 2025, advanced negotiations amid regulatory reviews. Intel CEO Lip-Bu Tan chairs SambaNova and backed its Series A via Walden.
This price marks a 68% drop from SambaNova’s 2021 peak of $5 billion, reflecting cooled AI startup hype. Intel seeks RDU tech and talent to counter Nvidia in inference markets. SambaNova’s full-stack beats Nvidia’s partial offerings, which demand heavy integration effort.
Hume AI uses SambaNova for real-time voice agents, while Hugging Face CTO Julien Chaumond said, “We are excited to partner with SambaNova and bring faster inference on Open Source models directly to our developer community.”
Blackbox.ai CEO Robert Rizk added, “The speed at which the SambaNova team responded to and supported us during the testing and the production phase is outstanding and was a real differentiator.”
SambaOrchestrator is used for monitoring and auto-scaling of DeepSeek, Llama, and Whisper on SambaNova Cloud. SambaStack is used on-site by government organisations to safeguard data. Data centres use SambaManaged to launch inference at low power in 90 days.
Benchmarks validate RDU’s lead in tokens per kWh on the larger models. Intel’s move highlights SambaNova’s threat to Nvidia’s dominance. SambaNova targets sovereign AI demands, enabling states to establish secure data centres.