SambaNova: Fastest AI Inference Chips, $2.2B (2026)

SambaNova (founded 2017, Palo Alto) builds RDU chips for AI inference. Raised $1.49B total; valued at $2.2B in Feb 2026. Runs Llama 4 at 461 tokens per second.

SambaNova Systems (founded 2017, Palo Alto, CA) makes Reconfigurable Dataflow Unit chips and runs the SambaCloud inference platform. The company raised $350 million in a February 2026 Series E at a $2.2 billion valuation. SambaCloud serves open-source models (Llama, DeepSeek, MiniMax) at record speeds for developers, enterprises, and sovereign AI programs in Japan, the UK, and Australia.

SambaNova Systems, founded in 2017 in Palo Alto, California, builds custom Reconfigurable Dataflow Unit (RDU) chips for AI inference and runs the SambaCloud inference platform. The company has raised $1.49 billion total, with its latest $350 million Series E in February 2026 valuing it at $2.2 billion. SambaCloud runs Llama 4 at 461 tokens per second and DeepSeek-R1 671B at 198 tokens per second, without quantization, on its SN40L chips.

Founded: 2017 · HQ: Palo Alto, CA, USA · Team: 350-450 · CEO: Rodrigo Liang · Funding: $1.49B total raised (Series A-E); latest: $350M Series E, Feb 2026, led by Vista Equity Partners and Cambium Capital · Valuation: $2.2B (Series E, Feb 2026)

About SambaNova Systems

SambaNova Systems was co-founded in November 2017 in Palo Alto, California, by Rodrigo Liang (CEO), Kunle Olukotun (Chief Technologist), and Christopher Re. Liang and Olukotun had worked together at Sun Microsystems and Oracle on high-performance microprocessor design, including the world's first multi-core processors, while Re is a Stanford professor and MacArthur Fellow known for machine learning systems research. The founding thesis was that GPU architecture, designed for graphics and general parallel computing, is structurally inefficient for the memory-access patterns of large AI model inference. SambaNova's answer was the Reconfigurable Dataflow Unit (RDU), a purpose-built inference chip that routes computation through a dataflow graph rather than a conventional instruction pipeline, enabling higher throughput per watt for large language models. The company's product line has four components sharing a common RDU foundation. SambaCloud is an OpenAI-compatible cloud inference API running models including Llama 4, DeepSeek-R1 671B at 198 tokens per second, and Llama 3.1 70B at 461 tokens per second at full precision without quantization. SambaStack is the on-premise alternative: an integrated hardware-plus-software appliance that enterprises deploy inside their own data centers, giving them complete data sovereignty over AI inference. SambaManaged is a co-located managed service for organizations that want dedicated hardware without an in-house infrastructure team, and SambaRack is the physical rack hardware available in fourth-generation SN40L and fifth-generation SN50 configurations. In February 2026, SambaNova unveiled the SN50 chip, the fifth generation of its RDU design. The SN50 delivers 5x higher throughput than the SN40L for agentic AI workloads, supports up to 256 accelerators linked via a multi-terabyte-per-second interconnect, and offers a 3x lower total cost of ownership than GPU-based alternatives according to company benchmarks. Shipments are scheduled for H2 2026, with SoftBank Corporation confirmed as the first customer for deployment in sovereign AI data centers in Japan. SambaNova also launched a multi-year collaboration with Intel in February 2026, pairing Intel Xeon 6 processors with SambaNova RDUs in a heterogeneous inference architecture targeted at enterprise data centers, expected to be available to customers in H2 2026. SambaNova has raised approximately $1.49 billion across five funding rounds. The April 2021 Series D raised $676 million from SoftBank Vision Fund 2, Temasek, GIC, BlackRock, Intel Capital, and GV, valuing the company at $5.1 billion and briefly making it the world's best-funded AI startup at that time. The February 2026 Series E raised $350 million from Vista Equity Partners, Cambium Capital, and Intel Capital at an implied valuation of $2.2 billion, a 57% decline from the 2021 peak. Intel's participation in the Series E came after a non-binding term sheet for Intel to acquire SambaNova for approximately $1.6 billion was signed in late 2025 but never completed, resulting in Intel investing as a strategic minority partner instead. SambaNova earns revenue from three channels: API token pricing and subscriptions on SambaCloud, multi-year hardware and software contracts for SambaStack, and managed service fees for SambaManaged. SambaCloud grew approximately 40% year-over-year and represented roughly 25% of ARR as of early 2026. Enterprise and sovereign AI hardware contracts, which involve multi-year deployment commitments, make up the majority of revenue. Third-party estimates place SambaNova's ARR near $100 million; the company does not publish official revenue figures. SambaNova employed approximately 500 people before conducting a 15% reduction in force in April 2025, laying off 77 employees to refocus resources on inference infrastructure. As of mid-2026, headcount is approximately 397 to 410. Lip-Bu Tan, former CEO of Cadence Design Systems, joined as Executive Chairman in May 2024, adding semiconductor industry depth to the executive team. Engineering is led by SVP Penny Li and concentrated in Palo Alto, with support teams active in the regions where sovereign AI partnerships operate. SambaNova's stated mission is to give every developer, enterprise, government, and data center sovereignty over their data, models, and AI infrastructure. The company does not train its own foundation models and does not publish AI safety research, operating as inference infrastructure for models trained by Meta, DeepSeek, and MiniMax. Its primary technical output is applied hardware engineering: five generations of RDU chip design, throughput and efficiency benchmarks, and systems-level work on dataflow computing. SambaNova's primary advantage is sustained high-throughput inference for 70 billion-plus parameter models without quantization: the SN40L processes Llama 3.1 70B at 461 tokens per second at full precision, where GPU-based alternatives typically require quantization to reach comparable speeds. Groq's LPU architecture offers lower latency for smaller single-query workloads but does not match SambaNova's sustained throughput at 400B-plus scale. Together AI provides a broader model catalog on GPU infrastructure. Cerebras competes on wafer-scale throughput for the very largest models. Nvidia H100 and H200 GPUs remain the default enterprise AI accelerator choice due to their software depth, which represents SambaNova's biggest structural challenge in the market. SambaNova has built a sovereign AI business line, partnering with Argyll (Scotland, UK), OVHcloud and Infercom (France/EU), SoftBank (Japan), and SouthernCrossAI (Australia) to provide national AI infrastructure that operates outside US hyperscaler control. This positioning aligns with EU AI Act implementation and tightening national data residency mandates worldwide. The company has not published a public trust center or disclosed SOC 2 or ISO 27001 certifications as of mid-2026; customers seeking compliance documentation must request it directly from SambaNova's enterprise sales team. SambaNova enters H2 2026 with three concrete catalysts: SN50 chip shipments beginning, the Intel heterogeneous inference architecture reaching enterprise customers, and SoftBank's Japan sovereign AI deployment going live. The company must grow revenue to support an exit above the $2.2 billion February 2026 valuation. No IPO has been announced and no active acquisition discussions are known as of mid-2026. Its focus on agentic AI, which demands sustained multi-step token generation across long agent chains, positions it well for the direction enterprise AI investment is taking through 2026 and beyond.

Mission

To give every developer, enterprise, government, and data center sovereignty over their data, models, and AI infrastructure.

Products

Links

Website · GitHub · Twitter · LinkedIn · Blog · Docs

Frequently Asked Questions

What is SambaNova Systems and what do they build?

SambaNova Systems is an AI chip and cloud inference company founded in November 2017 in Palo Alto, California, by Rodrigo Liang, Kunle Olukotun, and Christopher Re, all with backgrounds in microprocessor engineering and machine learning systems research at Stanford University and Sun/Oracle. The company builds Reconfigurable Dataflow Unit (RDU) chips that deliver higher token throughput than GPU-based alternatives for large language models by routing computation through a dataflow graph rather than a conventional instruction pipeline. Its main products are SambaCloud (an OpenAI-compatible cloud inference API), SambaStack (an on-premise integrated hardware-plus-software AI appliance), SambaManaged (a co-located managed inference service), and SambaRack (the physical rack hardware in SN40L and SN50 form factors). SambaCloud runs DeepSeek-R1 671B at 198 tokens per second and Llama 3.1 70B at 461 tokens per second at full precision without quantization, which are among the fastest published speeds for those model sizes. The company has raised $1.49 billion in total funding and was last valued at $2.2 billion in its February 2026 Series E. Developers can access SambaCloud by creating an API key at cloud.sambanova.ai and calling the OpenAI-compatible endpoints directly.

Who founded SambaNova Systems and who is the CEO?

SambaNova Systems was co-founded in November 2017 by three people: Rodrigo Liang (CEO), Kunle Olukotun (Chief Technologist), and Christopher Re. Rodrigo Liang holds bachelor's and master's degrees in Electrical Engineering from Stanford University and spent years at Sun Microsystems and Oracle developing high-performance microprocessors, including work on the world's first multi-core processors. Kunle Olukotun is a Stanford professor whose decades of research on chip multiprocessors and parallel computing architectures directly informed the RDU chip design at the core of SambaNova's products. Christopher Re is a Stanford MacArthur Fellow whose research on weak supervision and machine learning systems brought ML expertise to the founding team; he remains primarily in an academic role. Lip-Bu Tan, former CEO of Cadence Design Systems, joined as Executive Chairman in May 2024, adding deep semiconductor industry and customer network experience to the leadership team. No major CEO change has occurred since founding; Rodrigo Liang has led the company since day one. The engineering organization is led by SVP Penny Li, and CFO Geoff Ribar oversees finance.

How much funding has SambaNova Systems raised?

SambaNova has raised approximately $1.49 billion in total across five disclosed funding rounds. The largest single round was the April 2021 Series D: $676 million led by SoftBank Vision Fund 2, with Temasek, GIC, BlackRock, Intel Capital, GV (Google Ventures), Walden International, and WRVI participating, valuing SambaNova at $5.1 billion. The most recent round, a $350 million Series E closed in February 2026, was led by Vista Equity Partners and Cambium Capital, with Intel Capital investing, at an implied valuation of approximately $2.2 billion, a 57% decline from the 2021 peak. Intel's investment came after it had signed a non-binding term sheet to acquire SambaNova for approximately $1.6 billion in late 2025, a deal that ultimately failed to close, resulting in Intel becoming a strategic minority investor rather than an acquirer. The company does not publish official revenue figures; third-party estimates place ARR near $100 million as of early 2026. No IPO has been announced and SambaNova remains private.

What products does SambaNova Systems make?

SambaNova offers five main products built on its Reconfigurable Dataflow Unit chip architecture. SambaCloud is the company's cloud inference API, providing OpenAI-compatible endpoints for running models including Llama 4, DeepSeek-R1 671B, MiniMax, and GPT-OSS-120B; developers sign up at cloud.sambanova.ai, generate an API key, and call the endpoints like any other LLM API. SambaStack is the on-premise option: a bundled hardware-plus-software appliance enterprises install in their own data centers, providing full data sovereignty for regulated industries and government customers. SambaManaged is a co-located managed service where SambaNova operates dedicated hardware on behalf of customers who do not want to manage infrastructure directly. SambaRack is the physical data center hardware, currently available as the fourth-generation SN40L and the fifth-generation SN50 (announced February 2026, expected to ship in H2 2026 with 5x throughput gains over SN40L). SambaCloud is accessible to individual developers; SambaStack and SambaManaged are sold via enterprise sales with multi-year contracts.

Where is SambaNova Systems headquartered and how big is the team?

SambaNova Systems is headquartered in Palo Alto, California, in the Stanford Research Park area of Silicon Valley. The company had approximately 500 employees before conducting a 15% reduction in force in April 2025, in which 77 people were laid off as SambaNova refocused its engineering resources on inference infrastructure rather than broader AI platform development. As of mid-2026, headcount has stabilized at approximately 397 to 410 employees, based on LinkedIn data and third-party company profile sources. Engineering and chip design teams are concentrated in Palo Alto, and the company has international teams supporting sovereign AI partnerships in Japan (SoftBank), the United Kingdom (Argyll in Scotland), France and the EU (OVHcloud, Infercom), and Australia (SouthernCrossAI). No additional large-scale layoffs or significant office expansions have been reported as of mid-2026.

What is SambaNova's mission or research focus?

SambaNova's stated mission is to give every developer, enterprise, government, and data center sovereignty over their data, models, and AI infrastructure, and to be the most efficient and adaptable AI platform on the planet. In practical terms, the company's research is applied hardware and systems engineering: designing RDU chips (five generations from SN10 through SN50) that achieve higher token throughput per watt and per dollar than GPU-based alternatives for large language model inference. SambaNova does not train its own foundation models and does not publish AI safety research, distinguishing it from research-focused companies such as Anthropic, Google DeepMind, or OpenAI. The sovereignty framing refers specifically to data control: SambaStack and SambaManaged allow enterprise and government customers to run AI inference entirely on their own or co-located hardware, never sending data to a shared cloud. SambaNova has not published a responsible scaling policy or disclosed red-team evaluation partnerships.

Is SambaNova Systems compliant with SOC 2, GDPR, or HIPAA?

SambaNova does not publicly disclose SOC 2, ISO 27001, or HIPAA certifications, and the company does not maintain a public trust center as of mid-2026; its /trust and /security pages return 404 errors. For cloud API customers using SambaCloud, the data handling and retention policy is not publicly detailed, so enterprise teams with compliance requirements should request specific documentation from SambaNova's sales team before committing. SambaStack and SambaManaged customers achieve their own data sovereignty by running inference entirely on their own or co-located hardware, which structurally addresses GDPR and national data residency requirements without relying on SaaS compliance certifications. The company's legal-agreements page is available at sambanova.ai/legal-agreements and covers terms of service. Enterprise contracts for SambaStack typically include tailored data handling provisions, which is a common approach for on-premise AI infrastructure vendors targeting government and regulated sectors.

How does SambaNova compare to Groq in 2026?

SambaNova and Groq both offer specialized AI inference chips with much faster throughput than standard GPU clouds, but they optimize for different workloads. Groq uses its Language Processing Unit (LPU) architecture, which excels at time-to-first-token latency for single queries on models up to 70B parameters, making it a better fit for chat applications and interactive workloads where initial response speed matters most. SambaNova's RDU architecture is optimized for sustained high-throughput generation at 70B-plus scale, running Llama 3.1 70B at 461 tokens per second and DeepSeek-R1 671B at 198 tokens per second at full precision, which makes it better suited for batch inference, agentic AI pipelines, and multi-step token generation workloads. SambaNova also offers on-premise deployment via SambaStack, while Groq is cloud-only as of mid-2026. Groq has a simpler developer onboarding and tends to have lower pricing for short single-query usage. Pick SambaNova if you need maximum sustained throughput at 70B-plus scale or an on-premise sovereign deployment; pick Groq if you need the lowest latency for chat-style single-query workloads.