SambaNova fits enterprises and governments that need on-premise or co-located AI inference at data-center scale rather than a shared cloud API. JPMorganChase adopted its SN40 and SN50 systems in 2026 for secure on-prem inference, and its RDU architecture is built for sustained throughput on 70-billion-parameter-plus models without quantization.
SambaNova Systems, founded in 2017 in Palo Alto and now headquartered in San Jose, California, builds custom Reconfigurable Dataflow Unit (RDU) chips for high-throughput AI inference. Its fifth-generation SN50 chip targets 5x the throughput of its predecessor for agentic AI workloads, and in 2026 the company began pairing its RDUs with Nvidia GPUs in hybrid disaggregated-inference deployments.
Founded: 2017 · HQ: San Jose, CA, USA · Team: 350-450 · CEO: Rodrigo Liang · Funding: Over $2.4B total raised across six rounds (Series A-F); latest: $1B Series F first close, Jul 2026, led by General Atlantic at $11B post-money valuation · Valuation: $11B (Series F first close, Jul 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, usage-based cloud inference API spanning a model catalog that, on its pricing page as read on 25 October 2026, lists Meta-Llama-3.3-70B-Instruct, gpt-oss-120b, DeepSeek-V3.1, DeepSeek-V3.2, gemma-4-31B and MiniMax-M3.1-Flash, the last with a separate cached-input rate of $0.03 per million tokens (see the products FAQ below for current per-token pricing). The price table no longer lists the earlier MiniMax M2.7 and MiniMax M3 endpoints, although SambaNova's own posts still benchmark both, and MiniMax models are the ones it features most. 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. All four sit in the same AI infrastructure startup category Hokai tracks.
In February 2026, SambaNova unveiled the SN50 chip, the fifth generation of its RDU design, claiming 5x higher throughput than the SN40L for agentic AI workloads with shipments targeted for H2 2026. The pace of announcements accelerated through mid-2026: at COMPUTEX on June 3, 2026, SambaNova demonstrated live disaggregated inference pairing Nvidia's B200 GPU for prefill with its own SN40 RDU for decode, measuring 2x the throughput of a B200-only configuration (verified by Artificial Analysis) at Together.AI's Vector Core Compute (VC2) data center, with Together.AI as the first commercial customer of that configuration. On July 8, 2026, SambaNova announced that JPMorganChase had selected the company as an inference infrastructure partner, deploying SN40 and SN50 systems for secure on-premise AI inference, alongside its collaboration with Intel pairing Xeon 6 processors with SambaNova RDUs for agentic tooling (announced April 8, 2026).
The SN50 moved from announcement to measured results over the summer. On July 30, 2026, SambaNova wrote that a single 16-chip SambaRack SN50 was running MiniMax M2.7 at roughly 800 tokens per second per user, a figure it had first shown at RAISE and attributes to Artificial Analysis validation, with Nvidia H200 GPUs on prefill and SN50 RDUs on decode. The same post cites a SemiAnalysis benchmark in which that rack reaches about 800 tokens per second per user at its most interactive setting and about 400 tokens per second when split into two 8-chip replicas for throughput. On September 17, 2026, a SambaNova blog post pitched SN50 to neocloud operators with a six-month payback target, tested against a 70% billable-utilization planning assumption; that is a vendor planning figure, not an audited customer result, and the post itself says it is illustrative.
SambaNova's funding history spans six disclosed rounds since founding (see the funding FAQ below for exact amounts and valuations). The April 2021 Series D was led by SoftBank Vision Fund 2, with Temasek, GIC, BlackRock, Intel Capital, and GV also participating. The February 2026 Series E was led by Vista Equity Partners and Cambium Capital. General Atlantic stepped in as lead for a new Series F round confirmed in July 2026, joined by Seligman Ventures and T. Rowe Price Associates as significant new backers alongside returning investors Qatar Investment Authority, BlackRock, Intel Capital, Vista Equity Partners, and Cambium Capital. Intel's continued participation traces back to a non-binding term sheet for Intel to acquire SambaNova that was signed in late 2025 but never completed, after which Intel became a strategic minority investor 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. Enterprise and sovereign AI hardware contracts, which involve multi-year deployment commitments, make up the majority of revenue, and the July 2026 JPMorganChase deployment adds a marquee regulated-industry customer to that base. Third-party estimates place SambaNova's ARR near $100 million; the company does not publish official revenue figures.
SambaNova is now headquartered in San Jose, California, after operating from Palo Alto since its 2017 founding. Headcount is approximately 350-450 people following an April 2025 reduction in force that cut roughly 15% of staff. The leadership team expanded substantially through mid-2026: Matt Padfield joined as Chief Financial Officer and Rich Heaton as Executive Vice President of Software in June 2026, and Mohsen Moazami, formerly President of International at Groq, joined as Vice Chair of Global Strategy and Partnerships in August 2026. Lip-Bu Tan, who joined as Executive Chairman in May 2024, now holds the title Chairman after also taking on the CEO role at Intel Corporation; Penny Li leads hardware as Executive Vice President, Sumti Jairath serves as Chief Architect, Jude Sheeran is Chief Revenue Officer, Abhi Ingle is Chief Customer and Solutions Officer, Dan Christman is Chief Corporate Development Officer, and Jonathan Chang is Chief Legal Officer alongside Dan Horwood as General Counsel, per the company's team page.
SambaNova frames its purpose around handing customers, whether individual developers, enterprises, or governments, direct control over where their data, models, and infrastructure run. 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, now extending to hybrid GPU-RDU disaggregated inference architectures built alongside Nvidia hardware rather than purely against it.
SambaNova's primary advantage is sustained high-throughput inference for 70 billion-plus parameter models without quantization. Groq's LPU architecture wins on single-query latency for smaller chat-style workloads, but is cloud-only where SambaNova also ships an on-premise appliance. 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, though the COMPUTEX disaggregated-inference demo shows SambaNova now also partnering with Nvidia hardware in hybrid deployments rather than only competing against it.
SambaNova has built a sovereign AI business line, partnering with Argyll (Scotland, UK), OVHcloud and Infercom (France/EU), SoftBank (Japan), and SouthernCrossAI (Australia) for in-country AI infrastructure, and on April 21, 2026 TEPCO Systems, the digital transformation arm of Tokyo Electric Power, signed a distributor agreement to sell SambaNova systems to Japanese enterprises and to build its own AI data centers on them, and the JPMorganChase deployment extends the same on-premise sovereignty pitch to regulated US enterprises. SambaNova is classified as an AI infrastructure provider operating in the EU via its OVHcloud and Infercom partnerships; specific GPAI compliance classification under the EU AI Act is not publicly disclosed as of late 2026. Its on-premise model is one route enterprises use to meet GDPR and data-residency requirements without relying on a shared-cloud vendor's own certifications. The company does not publish a responsible scaling policy or disclose red-team partnerships, consistent with its position as an inference-infrastructure vendor rather than a model developer.
SambaNova heads into the last quarter of 2026 with its valuation reset by a Series F round that, as of the company's July 8 announcement, had only reached its first close (implying further capital to come), the SN50 now shown running in a 16-chip rack with plans to scale to 64 and 256 chips, the Intel heterogeneous inference architecture reaching enterprise customers, and its first major regulated-industry on-premise win in JPMorganChase. No IPO has been announced. Its bet on agentic AI, which demands sustained multi-step token generation across long agent chains, and its newly-demonstrated willingness to pair its RDUs with Nvidia GPUs rather than only compete against them, position it for the direction enterprise AI infrastructure spend is taking through the rest of 2026.
Mission
To give every developer, enterprise, government, and data center sovereignty over their data, models, and AI infrastructure.
Products
- SambaCloud (Cloud inference API): https://cloud.sambanova.ai
- SambaStack (On-premise AI system): https://sambanova.ai/enterprise
- SambaManaged (Managed inference service): https://sambanova.ai
- SN50 RDU (AI inference chip (5th gen)): https://sambanova.ai/hardware
- SambaRack (Data center rack hardware): https://sambanova.ai/hardware
Links
Frequently Asked Questions
How much funding has SambaNova Systems raised?
SambaNova has raised more than $2.4 billion across six disclosed rounds. Its April 2021 Series D brought in $676 million from SoftBank Vision Fund 2 and others at a $5.1 billion valuation, and its February 2026 Series E added $350 million led by Vista Equity Partners and Cambium Capital at $2.2 billion. On July 8, 2026, SambaNova completed the first close of a $1 billion Series F led by General Atlantic, with Seligman Ventures and T. Rowe Price Associates, at an $11 billion post-money valuation.
What does SambaNova sell and what does SambaCloud cost per token?
SambaCloud is SambaNova's OpenAI-compatible API. Its pricing page, read on 25 October 2026, runs from $0.22 per million input tokens on gpt-oss-120b to $3.00 on DeepSeek-V3.2, and lists MiniMax-M3.1-Flash at $0.30 input, $1.20 output and $0.03 for cached input. SambaStack is an on-premise hardware-plus-software appliance for enterprises needing full data sovereignty, and SambaManaged is a co-located managed alternative for teams without in-house infrastructure staff. SambaRack is the underlying hardware, sold as the fourth-generation SN40L or fifth-generation SN50 chip.
How does SambaNova Systems handle data privacy and compliance?
SambaNova does not publicly disclose SOC 2, ISO 27001, or HIPAA certifications, and its /trust and /security pages return 404 errors as of late 2026. SambaStack and SambaManaged customers achieve data sovereignty by running inference entirely on their own or co-located hardware rather than a shared cloud, which is how the company addresses GDPR and national data-residency requirements for regulated customers like JPMorganChase. Enterprise contracts typically include tailored data handling provisions negotiated directly with SambaNova's sales team.
Who are SambaNova Systems's main competitors?
Groq's LPU architecture wins on latency for smaller, single-query chat workloads but is cloud-only and has not published SambaNova's sustained throughput numbers at 70 billion-plus parameters. Cerebras targets the same very-large-model segment with its own wafer-scale chip design instead of RDUs. Nvidia's software ecosystem still makes its GPUs the safe default pick for most enterprise buyers, though SambaNova's 2026 COMPUTEX demo paired its own RDUs with a Nvidia B200 in a hybrid configuration rather than competing head-on.
How do I start using SambaNova?
Individual developers can sign up at cloud.sambanova.ai, generate an API key, and call SambaCloud's OpenAI-compatible endpoints directly, with per-token pricing starting at $0.22 per million input tokens on gpt-oss-120b. Enterprises and governments evaluating SambaStack or SambaManaged for on-premise or co-located deployment go through SambaNova's enterprise sales team to scope hardware and contract terms.
Top Alternatives
- Groq: Pick SambaNova for training+inference appliance; pick Groq for LPU inference cloud.
Hokai guides that mention SambaNova Systems
- The AI Tool Ecosystem in 2026: Buy the Meter, Not the Category: Four meters set your AI infrastructure bill: tokens, machine seconds, vendor credits, seats. Modal, Replicate, E2B and Together AI re-priced 26 October 2026.