SambaNova

  Executive Briefing

SambaNova Systems is an American AI infrastructure company headquartered in Palo Alto, California, that designs and manufactures custom silicon processors — called Reconfigurable Dataflow Units (RDUs) — along with a vertically integrated software platform and cloud inference service built entirely on top of that proprietary hardware. Founded in 2017 by three Stanford University-affiliated technologists — Rodrigo Liang (Co-Founder and CEO), Kunle Olukotun (Co-Founder and Chief Technologist), and Christopher Ré (Co-Founder) — the company was built on a single architectural conviction: that GPUs and CPUs are structurally inefficient for AI workloads because they were designed for general-purpose computation, and that a processor purpose-built around the dataflow patterns of neural networks could eliminate the memory bottleneck and deliver far greater speed and efficiency. Lip-Bu Tan, a prominent semiconductor industry figure, joined as Executive Chairman in May 2024.

The core insight behind the RDU architecture is the replacement of conventional fetch-decode-execute instruction cycles with a reconfigurable dataflow pipeline that maps an entire AI model's computation graph directly onto hardware. Where a GPU must repeatedly shuttle activations and weights between compute units and off-chip memory, the RDU fuses operators and streams data through on-chip pathways in a single pass. The most visible expression of this approach is the fourth-generation SN40L chip, which combines 520 MB of on-chip SRAM, 64 GB of HBM3, and 1.5 TB of DDR5 DRAM in a three-tier memory hierarchy — enabling models as large as 5 trillion parameters to run from a single socket and allowing multiple large models to remain simultaneously resident with microsecond hot-swap. SambaNova's independently benchmarked inference speeds for open-weight models — including a reported world record of 132 tokens per second for Llama 3.1 405B at launch and 198–255 tokens per second for DeepSeek-R1 671B — have made speed the company's primary market differentiator.¹²

SambaNova sells its technology through three channels: on-premises hardware systems (DataScale and SambaRack) deployed at national laboratories, sovereign cloud providers, and large enterprises; a managed enterprise service (SambaManaged); and a public pay-as-you-go cloud inference API (SambaNova Cloud) that gives developers access to the fastest open-model inference available at prices that reportedly undercut major API providers by several times.³ The government and sovereign-AI segment is a core focus, with systems deployed at Los Alamos, Lawrence Livermore, and Argonne National Laboratories, and national cloud partnerships in Europe, the UK, Germany, and Australia.

In February 2026, SambaNova simultaneously announced its fifth-generation SN50 RDU — claiming 5x the compute and 4x the network bandwidth of the SN40L, with support for 10-trillion-parameter models and 10-million-token contexts — along with a $350 million Series E funding round co-led by Vista Equity Partners and Cambium Capital, a multi-year collaboration agreement with Intel, and SoftBank as the first SN50 customer.⁴ The round valued the company at roughly $2.24 billion (estimated from disclosed share prices), a recovery from a trough reached during Intel acquisition talks in late 2025 but still well below the $5.1 billion peak reached in 2021. With a roster of sovereign and national-lab deployments, a competitive cloud API, and a next-generation chip approaching production, SambaNova is one of the few companies attempting to build the full stack — from silicon to service — outside of NVIDIA's ecosystem.

  At a Glance

  Origins & Founding

SambaNova was incorporated in November 2017 in Palo Alto, incubated out of research and executive networks centered on Stanford University. The three co-founders brought together rare and complementary expertise: Rodrigo Liang had spent more than two decades at Sun Microsystems and Oracle leading SPARC processor and ASIC development teams, giving SambaNova hands-on semiconductor engineering leadership from day one. Kunle Olukotun is the Cadence Design Professor of Electrical Engineering and Computer Science at Stanford, a pioneer of chip-multiprocessor (CMP) architecture, and the founder of Afara Websystems, a chip startup acquired by Sun Microsystems in 2002; he directs the Stanford Pervasive Parallelism Lab and is one of the world's leading authorities on parallel computing architecture. Christopher Ré is a Stanford ML professor best known for data mining and weak supervision, and as co-creator of the Snorkel data-labeling framework, though his role has been less operationally visible than Liang and Olukotun's since the company's early years.⁵⁶

The founding thesis was that the dominant GPU and CPU architectures had hit fundamental limits for AI workloads — specifically the data-movement bottleneck: GPUs were designed for graphics and later repurposed for deep learning, meaning their memory hierarchy and instruction execution model were poorly matched to the streaming, operator-fusion patterns of large neural networks. The founders argued that a processor designed from first principles around the dataflow structure of AI models could eliminate wasted cycles, dramatically reduce off-chip memory access, and achieve far higher utilization of raw compute capacity. This architectural bet — which required building both the hardware and all the software to program it — shaped every subsequent technical and business decision the company has made.

SambaNova emerged from stealth in April 2018 alongside a $56 million Series A led by Walden International and GV (Google Ventures), with the DataScale on-premises system as its first commercial product.⁷

  History & Timeline

    2017–2020: Stealth, seed, and series A

SambaNova was seeded with $2 million in November 2017 and spent its first two years building the first-generation RDU (SN10) and the SambaFlow compiler stack. In April 2018 the company emerged from stealth with its $56 million Series A and a pitch centered on enterprise and government AI at a time when few organizations had the software maturity or hardware budget to tackle trillion-parameter models. A $250 million Series C led by SoftBank Vision Fund closed in April 2020, funding SN10-based DataScale production and the company's early national-laboratory engagements.⁸

    2020–2021: DataScale launch and peak funding

The DataScale system reached commercial availability in December 2020, and in April 2021 SambaNova closed its $676 million Series D — led again by SoftBank Vision Fund, with Temasek joining as a new investor — at a post-money valuation of $5.1 billion.⁹ This was the high-water mark of SambaNova's funding trajectory, coming during the peak of the AI infrastructure investment cycle. During this period the company announced its first major government partnerships, with the Department of Energy and National Nuclear Security Administration (NNSA) deploying DataScale systems at Los Alamos (for quantum chemistry workloads in the "Darwin" system), Lawrence Livermore, and Argonne National Laboratory's ALCF AI Testbed.¹⁰¹¹¹²

    2022–2023: SN40L and the inference pivot

The third-generation SN30 was deployed in training configurations at Argonne. In September 2023 SambaNova announced the SN40L — the fourth-generation RDU — fabricated on TSMC's 5nm node in a 2.5D CoWoS-S package, featuring 1,040 compute cores, 102 billion transistors, and the three-tier memory architecture that would define SambaNova's inference differentiation story.¹³ As AI workloads shifted rapidly toward inference, SambaNova increasingly positioned itself around serving the largest open-source models at unprecedented speeds, rather than training. Valuation began to compress relative to the 2021 peak as competition intensified and the funding environment cooled.

    2024–2025: SambaNova Cloud, DeepSeek, and Intel talks

May 2024 brought the appointment of Lip-Bu Tan as Executive Chairman, lending semiconductor industry credibility to the board as SambaNova navigated a more challenging capital environment. In September 2024 SambaNova launched the SambaNova Cloud public inference API, immediately posting what were independently reported as world-record output speeds for Llama 3.1 405B at 132 tokens per second.¹ The real acceleration came in February 2025 when SambaNova deployed DeepSeek-R1 671B in full, native precision — serving it from a single SambaRack housing 16 RDU chips, compared to an estimated 40 racks of GPUs required for equivalent throughput — and Artificial Analysis independently measured 198–255 tokens per second, the fastest among all providers benchmarked at the time.²

By mid-2025, with valuation reportedly declining toward ~$1.6 billion, Intel entered acquisition negotiations. A term sheet was signed in late 2025; talks subsequently collapsed without a deal closing.¹⁴

    2026: SN50, Series E, and the Intel partnership

On 24 February 2026 SambaNova announced three things simultaneously: the SN50 fifth-generation RDU chip, a $350 million+ Series E funding round, and a multi-year collaboration agreement with Intel under which Intel Capital became a strategic investor (reported at approximately 9% ownership).⁴ The Series E was co-led by Vista Equity Partners and Cambium Capital, with new participants including T. Rowe Price Associates, Battery Ventures, Gulf Development Public Company, Saudi First Data, and the Qatar Investment Authority (reported), alongside existing investors including BlackRock, GV, Atlantic Bridge, and Redline Capital. SoftBank was named as the first SN50 customer, with initial deployments planned for Japanese data centers. SN50 shipments are expected to begin in H2 2026.

  What They Offer — Products & Platform

SambaNova's offering spans custom silicon, a proprietary software stack, on-premises systems, and a public cloud inference API — all built and operated as a vertically integrated unit, which distinguishes it from inference providers that run third-party GPU hardware.

Hardware systems come in two form factors. The SambaRack houses 16 RDU chips in a standard air-cooled rack, consuming approximately 10 kWh on average and integrating into conventional data-center facilities without the liquid-cooling infrastructure required by high-end GPU clusters. DataScale is the full on-premises deployable system combining SambaRack hardware, the SambaFlow software stack, and high-speed interconnects, sold to national laboratories, defense agencies, and large enterprises that require data residency and air-gapped operation.

SambaFlow is the compiler and software stack that automates the partitioning and mapping of AI model computation graphs onto the RDU's dataflow architecture. Rather than requiring manual kernel tuning per model (as is common with GPU inference frameworks), SambaFlow performs operator fusion, memory allocation across the three-tier hierarchy, and pipeline scheduling automatically, enabling SambaNova to add new models to its catalog rapidly.

SambaStudio is a visual development environment that allows organizations to move from model experimentation to production, covering fine-tuning, deployment management, and monitoring. SambaManaged packages hardware, software, and professional services into a managed enterprise deployment offering for customers prioritizing speed-to-production.

SambaNova Cloud is the public inference API offering pay-as-you-go access to the largest open-source models at the speeds enabled by RDU hardware. Free-tier accounts receive $5 in initial credits; developer accounts pay per-token; enterprise accounts operate on subscription terms. The API is compatible with OpenAI-style client libraries, reducing integration friction.

Sovereign AI is a packaged deployment path for national cloud providers and government agencies requiring full data sovereignty and on-premises or dedicated infrastructure. SambaNova claims a 90-day deployment path to production, and the sovereign offering currently powers inference for OVHcloud (Europe), Argyll Data Development (Scotland/UK), Infercom (Germany), and SouthernCrossAI (Australia).¹⁵

  Technology & Infrastructure

    RDU Architecture

The Reconfigurable Dataflow Unit is SambaNova's answer to what its founders identified as the fundamental mismatch between GPU architecture and AI model computation. A GPU executes Single Instruction Multiple Thread (SIMT) programs that repeatedly dispatch kernels to compute units, requiring frequent round-trips between DRAM and compute — each round-trip consuming energy and time proportional to data volume. The RDU instead treats an AI model's compute graph as a dataflow program: the entire computation pipeline is compiled once by SambaFlow and instantiated as a static, reconfigurable circuit on the chip, through which activations and weights flow without kernel-launch overhead or redundant memory access. Operator fusion across the full model graph — not just within a single kernel — eliminates the intermediate memory writes that dominate GPU inference latency.

    SN40L (Fourth Generation, 2023)

The current production chip was fabricated at TSMC on the 5nm node in a 2.5D CoWoS-S multi-chip package, and was presented at IEEE ISSCC 2024.¹³ Key specifications:

The three-tier memory hierarchy is the SN40L's most distinctive feature: the 520 MB of on-chip SRAM handles the hottest activations and attention weights; 64 GB of HBM3 buffers model layers and KV-cache; and 1.5 TB of DDR5 provides enough capacity for the full weight matrices of models up to 5 trillion parameters without paging. Critically, multiple models can reside in this hierarchy simultaneously, with switching between them in microseconds — enabling the multi-model serving that SambaNova's cloud API relies on. A single 16-chip SambaRack served DeepSeek-R1 671B at production inference speeds in February 2025, replacing what would have required an estimated 40 racks of GPU hardware.²

    SN50 (Fifth Generation, announced February 2026)

SambaNova announced the SN50 on 24 February 2026, with shipments expected to begin in H2 2026. The process node has not been publicly disclosed. SambaNova claims:

SN50 performance claims against competing hardware are SambaNova's own benchmarks pending independent third-party verification on shipping systems.⁴

    SambaNova Cloud Infrastructure

SambaNova Cloud currently runs on SN40L-based SambaRack systems. The February 2026 Intel collaboration agreement adds Intel Xeon-based compute capacity and joint reference architectures for enterprise buyers, expanding the platform beyond pure RDU inference to complementary workloads.

  Model Catalog & Performance

SambaNova Cloud serves open-source models selected for enterprise and developer demand. All models run on RDU hardware; the catalog is not static and has expanded rapidly since the September 2024 launch. Key models currently or recently served include:

Performance figures below are independently benchmarked by Artificial Analysis unless otherwise noted.¹²

The gpt-oss-120b (low) figure of approximately 724 tokens per second represents the fastest model in the current SambaNova Cloud catalog by output speed as of early 2026.

  Pricing & Performance Position

SambaNova Cloud pricing as of early 2026 (approximate, per million tokens input/output):

SambaNova's pricing is reported to undercut major API providers by 5–7x on average for comparable models as of March 2026, though this is a company-cited figure and the competitive landscape shifts frequently.³ The value proposition is speed and price together: SambaNova occupies the "fastest and cheapest" positioning for open-model API inference, rather than competing on proprietary model capability.

The three SambaNova Cloud service tiers are: a Free tier with $5 in initial credits; a Developer tier with pay-as-you-go per-token billing; and an Enterprise tier on subscription terms with SLA guarantees and dedicated capacity. Enterprise pricing is negotiated separately and is not publicly listed.

  People & Leadership

Rodrigo Liang (Co-Founder and CEO) is the operational center of SambaNova. He spent more than 20 years at Sun Microsystems and Oracle leading SPARC processor development and ASIC engineering before co-founding SambaNova, giving him direct credibility in the custom-silicon market the company is targeting. He is the primary voice in SambaNova's external communications and customer relationships.

Kunle Olukotun (Co-Founder and Chief Technologist) is simultaneously one of the world's most cited computer architecture researchers and SambaNova's technical architect. His earlier career includes the chip-multiprocessor work that became the foundation for modern multi-core processors, and his Afara Websystems startup (acquired by Sun Microsystems in 2002) gave him the hands-on experience of taking custom silicon from concept to acquisition. At SambaNova he oversees the RDU architecture and SambaFlow compiler research.⁶

Christopher Ré (Co-Founder) is a Stanford ML professor and co-creator of the Snorkel data-programming framework. His current operational involvement at SambaNova is less visible than Liang and Olukotun's, and his ongoing academic affiliations suggest he operates more as a technical advisor and co-founder figurehead than an operating executive.

Lip-Bu Tan was appointed Executive Chairman in May 2024. He is a well-known figure in the semiconductor industry and also serves as CEO of Intel — a dual role that makes his relationship with SambaNova more complex following the February 2026 Intel collaboration agreement and the collapse of Intel's earlier acquisition attempt.¹⁴

The board also includes Amarjit Gill (venture investor), Dr. Edward H. Frank, and Catriona Mary Fallon (appointed August 2024). Headcount is not publicly disclosed.

  Funding, Ownership & Business

SambaNova has raised approximately $1.49 billion in total funding across seven rounds since its November 2017 founding.

The Series D in April 2021 established SambaNova's peak valuation of $5.1 billion and made SoftBank Vision Fund the dominant financial backer.⁹ The subsequent period saw SambaNova navigate a more difficult environment: competition intensified, the private funding market cooled, and the valuation compressed substantially, reaching a reported low of approximately $1.6 billion during Intel acquisition negotiations in late 2025.

The Series E (February 2026) raised $350 million-plus co-led by Vista Equity Partners and Cambium Capital, with Intel Capital as a new strategic investor, T. Rowe Price Associates, Battery Ventures, Gulf Development Public Company, Saudi First Data, and the Qatar Investment Authority (reported), alongside existing investors including BlackRock, GV, Atlantic Bridge, and Redline Capital.⁴ The post-money valuation is estimated at approximately $2.24 billion based on disclosed per-share prices ($30.99 for E-1 shares, $21.70 for E-2 shares), though this is an estimate rather than an official company statement. The valuation remains roughly 57% below the 2021 peak.

Intel signed a term sheet to acquire SambaNova at approximately $1.6 billion in late 2025; those talks collapsed without a transaction, and Intel instead entered the company as a strategic investor (reported ownership approximately 9%) through the Series E alongside a multi-year co-selling and collaboration agreement.¹⁴

SambaNova's revenue model has three streams: hardware sales (DataScale and SambaRack systems to national labs, governments, and enterprises); professional services (reported at approximately 25–33% of new customer engagements); and subscriptions (SambaNova Cloud API per-token billing and Dataflow-as-a-Service enterprise subscriptions). The company reported record bookings and revenue in 2025, though specific ARR figures are not officially disclosed (an estimated $100 million ARR figure has been cited in industry analysis but is unverified).

  Customers & Partnerships

SambaNova's customer base divides between national-laboratory and government deployments — which require on-premises, air-gapped, data-sovereign infrastructure — and commercial sovereign-AI and enterprise engagements.

National laboratories and government represent SambaNova's most prestigious and technically demanding deployments. Los Alamos National Laboratory (LANL) operates DataScale systems as part of its "Darwin" heterogeneous computing environment for quantum chemistry workloads, with expanded deployment confirmed.¹⁶ Lawrence Livermore National Laboratory (LLNL) deployed DataScale through funding from the NNSA's Advanced Simulation and Computing program.¹⁷ Argonne National Laboratory (ANL) operates both an SN30 training cluster and an inference-optimized cluster at the ALCF AI Testbed, funded by the Department of Energy Office of Science.¹⁸ A broader DOE/NNSA strategic partnership agreement covers multiple DataScale installations across NNSA lab facilities.

Sovereign AI partners are national or regional cloud providers deploying SambaNova hardware to serve AI inference with full data sovereignty. Current deployments include OVHcloud (European sovereign AI inference endpoints), Argyll Data Development (Scotland/UK; described as the first fully renewable-energy AI deployment in the UK), Infercom (Germany/EU sovereign inference), and SouthernCrossAI (Australia).¹⁵

SoftBank Corp. (Japan) is the first announced customer for the SN50 fifth-generation chip, with deployments planned for next-generation Japanese AI data centers. This makes SoftBank both a major historical financial backer (via SoftBank Vision Fund) and SambaNova's lead hardware customer for the next product generation.

SambaNova solutions are also listed on AWS Marketplace, enabling enterprise procurement through existing cloud purchasing agreements. The February 2026 Intel collaboration adds co-selling through Intel's enterprise sales channels and joint reference architectures with system integrators.

Verticals served include financial services, telecommunications, energy, life sciences (drug discovery), and public sector.

  Competitive Position

SambaNova occupies a distinct position in the custom-AI-silicon landscape, differentiated from its closest peers by architecture, go-to-market, and model-size focus.

vs. Groq: Groq's Language Processing Unit (LPU) architecture targets ultra-low latency for single-model inference through a deterministic, fixed-function design. The LPU excels at consistent, predictable latency for a single model at a time. SambaNova's RDU is reconfigurable and supports simultaneous multi-model hosting, larger models (serving 671B-parameter models in full precision), and deeper on-premises and sovereign AI deployment. Groq operates primarily as an API-first developer service; SambaNova has a broader enterprise and government go-to-market.

vs. Cerebras: The Cerebras WSE-3 is a monolithic wafer-scale processor (46,000+ mm²) optimized for training very large models. Cerebras has been more training-centric; SambaNova has pivoted more decisively toward inference, particularly for the agentic AI use cases that dominate 2025–2026 customer demand. Their memory architectures are fundamentally different: Cerebras achieves extreme on-chip SRAM density; SambaNova uses a three-tier hierarchy to support trillion-scale model parameters that would not fit on any monolithic chip.

vs. NVIDIA: NVIDIA holds an estimated 85% share of the AI accelerator market and an enormous ecosystem advantage in CUDA software, tooling, and developer familiarity. SambaNova's claims of 5x speed and 8x cost savings over the Blackwell B200 for specific agentic inference workloads are SambaNova's own benchmarks, pending independent verification on shipping SN50 hardware.⁴ The RDU's air-cooling, multi-model concurrent hosting, and sovereign deployment capabilities target use cases where NVIDIA's ecosystem advantages matter less than raw efficiency and data-control guarantees.

vs. hyperscaler custom silicon (AWS Trainium, Google TPUs): Hyperscaler-built accelerators are optimized for internal workloads and offered to customers as cloud services, meaning data leaves the customer's control. SambaNova's on-premises and sovereign-AI products directly address the segment of the market — national governments, defense agencies, and regulated enterprises — for which this is not acceptable.

SambaNova's four primary differentiators are: the three-tier memory hierarchy enabling simultaneous multi-model hosting and microsecond model switching; single-rack efficiency for the largest open-weight models; a proven sovereign and on-premises go-to-market with national-lab validation; and full-stack vertical integration from silicon to managed cloud service.

  Notable Events

The most consequential single event in SambaNova's recent history is the Intel acquisition drama of 2025. Intel signed a term sheet to acquire SambaNova at approximately $1.6 billion — a figure that implied roughly a 69% discount to the 2021 peak — in what would have been a significant semiconductor consolidation deal. The talks collapsed, reportedly reflecting either valuation disagreement or Intel's own financial constraints, and SambaNova instead completed the Series E with Intel as a strategic investor rather than an acquirer.¹⁴ The episode simultaneously validated SambaNova's strategic importance (Intel wanted the technology) and underscored the valuation compression the company had experienced over four years.

The DeepSeek moment in February 2025 served as an unexpected commercial catalyst for SambaNova. When DeepSeek released the R1 671B model and it rapidly became the most-discussed open-weight model in the industry, SambaNova was uniquely positioned: the RDU's three-tier memory could serve the full-precision 671B model from a single rack, and SambaNova Cloud was among the first providers to make it available at production speed. The independently measured 198–255 tokens per second from Artificial Analysis — far above any GPU-based provider — drove substantial developer traffic to SambaNova Cloud and generated significant press coverage.²

  Outlook & Roadmap

SambaNova's 2026 roadmap centers on three overlapping pillars. The SN50 production ramp in H2 2026 is the most technically significant: if independent benchmarks confirm SambaNova's claims of 5x agentic inference speed and 8x cost advantage versus NVIDIA Blackwell B200, it would meaningfully shift the competitive calculus in enterprise inference procurement. SoftBank Japan's deployment will be the first large-scale real-world test. The Intel collaboration opens SambaNova's enterprise go-to-market through Intel's substantial distribution network, adds Xeon-based compute capacity to SambaNova Cloud for complementary workloads, and brings joint reference architectures that reduce integration complexity for system integrators. The third pillar is agentic AI positioning: all of SambaNova's 2026 product marketing emphasizes the multi-step, low-latency, high-throughput inference patterns of AI agents, voice assistants, and enterprise copilots — a workload profile that favors the RDU's deterministic, multi-model, high-throughput characteristics over GPU alternatives.

The strategic risks are real. SambaNova's valuation remains roughly 57% below its 2021 peak despite record 2025 revenues, reflecting investor skepticism about whether custom silicon companies can scale fast enough to matter against NVIDIA's ecosystem dominance. The Intel partnership is simultaneously a commercial opportunity and a competitive complication, since Intel also develops and markets its own AI accelerators (Gaudi series). SambaNova's cloud inference offering depends entirely on open-source model releases from Meta, DeepSeek, Google, and others — the company has no proprietary models, meaning its differentiation is entirely on hardware speed and platform cost, not model capability. And the SN50's headline performance claims will remain benchmarks rather than shipped-hardware reality until H2 2026 deliveries and independent testing.

The company that emerges from the SN50 ramp and Intel collaboration in 2027 will have a clearer answer to the central question its entire existence poses: whether a purpose-built dataflow architecture, sold through a vertically integrated platform, can carve out a durable position in enterprise AI inference against an incumbent whose software moat is measured in millions of developer-years.

---

  References

1. SambaNova Cloud — Artificial Analysis benchmarks
2. SambaNova blog: Fastest DeepSeek-R1 671B
3. SambaNova Cloud pricing and plans
4. SambaNova press release: SN50, Intel collaboration, Series E (BusinessWire)
5. SambaNova — About page
6. Kunle Olukotun background — University of Michigan CSE
7. SambaNova Series A and stealth emergence — tracxn funding history
8. SambaNova Series C — SoftBank
9. SambaNova Series D — Sacra profile
10. Los Alamos National Laboratory DataScale deployment — internal references
11. LLNL and SambaNova collaboration
12. Argonne ALCF SambaNova inference cluster
13. SN40L — IEEE ISSCC 2024
14. Intel and SambaNova — CNBC (Feb 2026)
15. SambaNova sovereign AI solutions
16. Los Alamos DataScale deployment details — SambaNova
17. LLNL SambaNova collaboration
18. Argonne ALCF AI Testbed — SambaNova inference cluster

  References

1. Artificial Analysis independent benchmarks of SambaNova Cloud providers — artificialanalysis.ai/providers/sambanova. ↩ ↩² ↩³

2. SambaNova blog: fastest DeepSeek-R1 671B deployment and Artificial Analysis measurements — sambanova.ai/blog/sambanova-cloud-launches-the-fastest-deepseek-r1-671b. ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Pricing comparison is a company-reported figure; competitive API pricing shifts frequently — cloud.sambanova.ai/plans. ↩ ↩²

4. SambaNova SN50 announcement, Series E, and Intel collaboration — BusinessWire, 24 Feb 2026. SN50 performance vs. B200 is SambaNova's own benchmark; independent verification pending H2 2026 shipments. Post-money valuation (~$2.24B) is estimated from disclosed share prices, not an official company statement. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

5. SambaNova company background — sambanova.ai/company/about. ↩

6. Kunle Olukotun biography and Afara Websystems history — University of Michigan CSE. ↩ ↩²

7. Funding history — Tracxn; Wilson Sonsini Series E advisory. ↩

8. SoftBank Series C participation — Sacra. ↩

9. Series D $5.1B valuation — Sacra. ↩ ↩²

10. Los Alamos DataScale — SambaNova company communications. ↩

11. LLNL and SambaNova — llnl.gov. ↩

12. Argonne ALCF SambaNova cluster — alcf.anl.gov. ↩

13. SN40L IEEE ISSCC 2024 publication — ieeexplore.ieee.org. ↩ ↩²

14. Intel acquisition talks collapse and strategic investment — CNBC, 24 Feb 2026. Intel ownership stake (~9%) is reported, not officially confirmed in public filings. ↩ ↩² ↩³ ↩⁴

15. Sovereign AI deployments — sambanova.ai/solutions/sovereign-ai. ↩ ↩²

16. LANL Darwin system — SambaNova company communications. ↩

17. LLNL NNSA program deployment — llnl.gov. ↩

18. Argonne ALCF inference cluster — alcf.anl.gov. ↩