SambaNova

SambaNova builds dataflow AI accelerators and full-stack software for enterprise AI training and inference at scale. Raised $1B+ total; SN40L chip announced 2024. Customers include DoE labs and Saudi Aramco. Competing in on-prem enterprise.

SambaNova's reconfigurable dataflow architecture and enterprise focus differentiate it from GPU-centric competitors.

sambanova.ai AI AcceleratorsUSAprivate

At a glance

Total raised

$1.5B

last round series e

Headcount

—

no public estimate on record

Headquarters

Palo Alto

USA

Founded

2017

9 years active

Status

Private

AI Accelerators

Latest valuation

$2.2B

Series E · 2026

For subscribers

Subscribe to see the full profile

Including Funding rounds, Bull / Bear thesis, Stock + earnings, Roster changes, Patents, News, and Open roles.

Founding

50% off

/month

91 of 100 spots left

Full

$10

/month

Available once all 100 Founding Member spots are claimed.

Get full access

Already subscribed? Sign in →

Funding history

Total raised

$1.5B

Rounds tracked

Latest stage

Series E

Latest valuation

$2.2B

Date	Stage	Amount	Valuation	Lead investors
Feb 1, 2026	Series E	$350M	$2.2B	Vista Equity Partners, Cambium Capital
Apr 13, 2021	Series D	$676M	$5.1B	SoftBank Vision Fund 2
Feb 1, 2020	Series C	$250M	—	BlackRock
Apr 1, 2019	Series B	$150M	—	Intel Capital
Mar 1, 2018	Series A	$56M	—	Walden International, GV

How they make money

Hardware + recurring servicesPattern

SambaNova runs a hardware-plus-recurring-services model: customers buy its dataflow accelerators and then pay for ongoing software and support. The company sells the SN40L chip and a full-stack AI platform — including the SambaNova Suite and DataScale systems — to large enterprises and government labs that want to run training and inference on-prem rather than in the cloud. Revenue is recognized at two layers: up-front hardware sales for the accelerator and system infrastructure, and recurring annual contracts for software licenses, updates, and support. The structural economics of the hardware layer are capital-intensive — fab costs, packaging, and system assembly are incurred before any sale, which compresses near-term margins until volume scales. The recurring software and services layer, by contrast, carries the higher incremental margins typical of enterprise software and provides revenue predictability after the initial hardware deployment. SambaNova subsidizes the hardware qualification cycle (long sales cycles, bespoke integration) to land anchor customers like DoE labs and Aramco, betting that follow-on license renewals and expanded deployments improve lifetime value per account. As a private company, SambaNova does not disclose specific margin or revenue figures, but the model structurally depends on converting initial hardware wins into sticky software renewals.

Generated by deepseek-v4-flash · June 2, 2026Analytical summary · not investment advice

Founder backgrounds

Rodrigo Liang

Co-founder & CEO

Holds B.S. and M.S. degrees in electrical engineering from Stanford University. Began his career at Hewlett-Packard, then joined Afara Websystems (a Kunle Olukotun spinout) to work on multi-core processor design; Sun Microsystems acquired Afara in 2002. At Sun he became Director of Engineering for UltraSPARC processor development, instrumental in the world's first scalable multi-core Niagara processors. At Oracle he served as SVP overseeing SPARC processor and ASIC development, releasing 12 major SPARC chips over 15 years. Co-founded SambaNova Systems in 2017 to build a purpose-built AI computing platform, having reunited with Kunle Olukotun. SambaNova has raised over $1 billion and offers full-stack AI inference solutions.

Kunle Olukotun

Co-founder & Chief Technologist

The Cadence Design Professor of Electrical Engineering and Computer Science at Stanford University and widely credited as a 'father of multicore processors.' Earned his Ph.D. from the University of Michigan. Led the Stanford Hydra chip multiprocessor (CMP) research project and founded Afara Websystems, whose Niagara multi-core processor architecture was acquired by Sun Microsystems and powers Oracle's SPARC servers. Co-founded SambaNova Systems in 2017 with Rodrigo Liang and Chris Ré to commercialize his research on domain-specific, dataflow-based hardware for ML. A member of the National Academy of Engineering and Fellow of both ACM and IEEE.

Christopher Ré

Co-founder

Professor of computer science at Stanford University, affiliated with the Stanford AI Lab. His research spans database theory, systems, and machine learning, and work from his group has been incorporated into Google Ads, Gmail, YouTube, and Apple products. Won the Jim Gray Dissertation Award, NSF CAREER award, Alfred P. Sloan Fellowship, and a MacArthur Genius Fellowship. Co-founded SambaNova Systems in 2017 with Rodrigo Liang and Kunle Olukotun; has also co-founded Snorkel AI and two companies acquired by Apple (Lattice/DeepDive and Inductiv/HoloClean), building a track record of applied ML systems research that reaches production at hyperscale.

AI bull / bear

Bull case

SambaNova's bet on reconfigurable dataflow architecture is finally paying off as the market shifts from monolithic training runs to complex agentic AI workflows. The company's SN40L chip, announced in 2024, was designed from the ground up for exactly this paradigm -- dynamic computation graphs, branching logic, and tool-use loops that starve traditional GPU pipelines. Unlike Nvidia's fixed-ISA approach, SambaNova's architecture lets enterprises reconfigure the data path on the fly, eliminating the memory-bandwidth bottlenecks that plague agentic inference at scale. With over $1B in backing, customers at DoE labs and Saudi Aramco, and a focused on-prem enterprise sales motion, SambaNova has the capital and credibility to capture a meaningful slice of the agentic AI inference market before the hyperscalers consolidate it. The company doesn't need to beat Nvidia at training. It needs to own the enterprise inference tier where GPU efficiency drops and software complexity spikes -- and its engineering optimizations for agentic workloads show it understands that fight.

Bear case

SambaNova has raised over $1B and still operates as a niche player selling to a handful of government and oil-and-gas customers. That capital intensity without broad commercial traction suggests a product that wins benchmarks but loses in the marketplace. The reconfigurable dataflow architecture is elegant on paper, but it forces enterprises to rewrite their software stack from scratch -- a non-starter when CUDA's ecosystem moat deepens with every new Nvidia release. The SN40L arrived in 2024, but Nvidia's annual hardware cadence and Blackwell's inference optimizations threaten to obsolete SambaNova's differentiation in a single product cycle. Meanwhile, AMD and Intel are competing aggressively for the same on-prem enterprise wallets with open-source software compatibility. Agentic AI workloads are real, but they are also early and fragmented. SambaNova is shipping engineering demos, not platforms that IT buyers can trust. Without a clear path to revenue scaling beyond a handful of lighthouse accounts, the company risks becoming a cautionary tale about brilliant architecture that could not escape the gravitational pull of the CUDA ecosystem.

Generated by deepseek-v4-flash · June 2, 2026Refreshes quarterly · see disclosures

In the news

SambaNova · last 30 days

No articles ingested yet for SambaNova. Once the hourly news pipeline is live, every article the classifier tags as mentioning this company appears here with its one-line AI summary and sentiment.

Watch

Leadership

Rodrigo Liang

Co-founder & CEO & Chief Executive Officer

Kunle Olukotun

Co-founder & Chief Technologist

—

Christopher Ré

Co-founder

—

Abhi Ingle

Chief Product & Strategy Officer

—

Harry Ault

Chief Revenue Officer

—

Penny Li

EVP, Hardware

—

Products

SN40L Reconfigurable Dataflow Unit (RDU)
SambaNova's third-generation AI chip, built on TSMC's 5nm process with a Reconfigurable Dataflow Architecture and a unique three-tier memory system (on-chip SRAM, HBM, and large external DRAM). The tiered memory lets a single eight-socket node host and serve models of up to five trillion parameters, giving SambaNova a differentiated angle on large-model inference versus GPU clusters that must shard models across many devices.
SambaNova Cloud / DataScale Systems
SambaNova's full-stack platform, delivered as turnkey DataScale rack systems for on-premises deployment and as SambaNova Cloud, a high-speed inference service. Built on SN40L RDUs and the SambaFlow software stack, it offers record-setting token-generation throughput on open models like Llama and DeepSeek, targeting enterprises, government labs, and sovereign-AI customers that want fast inference without managing GPU infrastructure.

Most recent patents

113 patents on file, but none with both an extractable figure and an abstract on Google Patents yet.