SambaNova Systems

Overview

SambaNova sells complete AI systems rather than chips. Its Reconfigurable Dataflow Unit maps a neural network onto a configurable dataflow fabric, so that instead of repeatedly fetching instructions and shuttling intermediate results through a memory hierarchy, the computation is laid out across the device as a pipeline. The company pairs that silicon with its own software stack and sells the result as an appliance, a rack, or hosted capacity.

It was co-founded in November 2017 by Rodrigo Liang, a former Oracle executive who serves as chief executive, and Stanford professors Kunle Olukotun and Christopher Ré. Olukotun's academic work on multi-core and reconfigurable architectures is the direct intellectual lineage of the RDU.

SambaNova's recent history is the most volatile of any company tracked on this site, and it is worth setting out plainly. Having raised roughly $1.1 billion between 2018 and 2021 and reached a $5.1 billion valuation in April 2021, the company was reported in October 2025 to be exploring a sale after failing to raise new funding. Intel signed a non-binding term sheet to acquire it at a valuation of roughly $1.6 billion. That deal lapsed. SambaNova then raised more than $350 million in February 2026 alongside the launch of its SN50 chip, and on 8 July 2026 announced the first close of a $1 billion Series F at an $11 billion valuation.

From exploring a sale at a reported $1.6 billion to an $11 billion valuation in roughly eight months is an extraordinary reversal, and any assessment of the company has to account for how quickly sentiment moved in both directions.

Chip and product families

The 5× speed, 3× total cost of ownership and tokens-per-watt figures are SambaNova's own published claims, measured against its prior generation or unnamed competitive parts. They have not been independently verified by this publication.

Role in inference

SambaNova has consistently sold to a different buyer than most inference specialists. Rather than competing for hyperscaler capacity or developer API traffic, it targets enterprises, governments and national programmes that want to run frontier-class models on infrastructure they control — often on-premises, often for data-residency or sovereignty reasons, and usually as a complete system rather than components to integrate.

That positioning has real advantages. It avoids direct price competition with Nvidia in the cloud, it commands higher margins, and sovereign AI budgets have grown substantially. The SN50's framing around agentic workloads fits the same logic: agentic systems chain many sequential model calls, so per-call latency compounds, and an architecture with predictable low-latency execution is well matched to it.

The corresponding difficulty is that enterprise and government systems sales are slow, lumpy and relationship-driven. Revenue arrives in large irregular blocks, which is precisely the pattern that made the company's 2025 funding difficulties possible despite a technically credible product.

Customers and partners

SoftBank Corp. is the first announced customer for the SN50, deploying it in next-generation AI data centres in Japan to power low-latency inference for sovereign and enterprise customers across Asia-Pacific. SoftBank is also a long-standing investor, having led the company's 2021 round, so the relationship is both commercial and financial.

Intel's involvement is unusual and worth stating carefully. Intel Capital was an early investor, leading the $150 million Series B in 2019. Intel chief executive Lip-Bu Tan chairs SambaNova, a position that predates his appointment at Intel and originates in his time at Walden International, which led SambaNova's Series A. Intel holds a minority stake of roughly 9% following an additional investment in early 2026, and the two companies announced a strategic collaboration in February 2026 to deliver cost-efficient inference. Intel's attempt to acquire the company outright did not complete.

Readers assessing the Intel relationship should note the overlapping roles: Intel is simultaneously an early investor, a minority shareholder, a commercial partner, a competitor through its own Crescent Island inference GPU, and a would-be acquirer whose bid lapsed. Its chief executive chairs SambaNova's board.

Strengths and key questions

Funding history

SambaNova is private and does not disclose revenue. Any revenue figure attributed to the company should be treated as unverified.

What to watch

• Series F second close. Whether it completes at the same terms, and the final round size.
• SN50 shipments. Delivery against the H2 2026 commitment, and the first independent benchmarks of a fifth-generation RDU.
• SoftBank deployment scale. How much capacity is actually installed in Japan, as distinct from announced.
• Customer diversification. Whether new sovereign or enterprise names appear, or whether concentration remains high.
• The Intel collaboration. What it produces commercially, given Intel is simultaneously building competing inference silicon.

Sources

1. SambaNova — First close of $1B financing at an $11B valuation. Primary announcement of the Series F.
2. General Atlantic — SambaNova $1B first close at $11B. Lead investor statement and participant list.
3. TechCrunch — SambaNova draws $1B at an $11B valuation (8 July 2026). Round context and the lapsed Intel approach.
4. SambaNova — SN50 launch, Intel collaboration and $350M+ raise (February 2026). SN50 positioning, performance claims and SoftBank deployment.
5. SambaNova — Introducing the SN50 RDU. Technical description of the fifth-generation part.
6. Data Center Dynamics — SN50, Intel partnership and $350m fundraise. Independent account.
7. Wikipedia — SambaNova Systems. Founding dates and early funding history, used for background only.

Performance figures are company claims and are labelled as such. The October 2025 sale process and the $1.6B Intel term sheet were reported by Bloomberg and others rather than announced by the companies. See the editorial methodology.

Related pages