By SambaNovaLearning Objectives
- Understand what SambaNova is and how its RDU chip differs from a general-purpose GPU
- See where SambaNova fits among inference-focused Nvidia challengers like Groq, Cerebras, and Etched
- Recognize why "inference," not training, is the battleground these chip startups are fighting over
What Is SambaNova?
SambaNova is a US AI chip company that designs processors and a full software stack for running large AI models efficiently. Instead of competing with Nvidia on general-purpose graphics processors, SambaNova built a fundamentally different chip around a Reconfigurable Dataflow Unit (RDU) — an architecture that maps the structure of an AI model directly onto the silicon so it can run large models with fewer chips and less power than a comparable GPU cluster.
Founded in 2017 in Palo Alto, California by Stanford computer scientists Kunle Olukotun and Chris Ré together with CEO Rodrigo Liang — a former SPARC processor engineering leader at Sun Microsystems and Oracle — SambaNova is one of a handful of well-funded startups betting they can carve out a piece of the AI hardware market that Nvidia dominates.
💡Key Concept
Training vs. inference: Training is the expensive, one-time work of building a model; inference is running that finished model to answer real user queries, over and over, for as long as the product exists. As AI moves from research into everyday products, inference becomes the larger and more cost-sensitive workload — which is exactly the market SambaNova and its rivals are targeting.
The RDU Architecture and Full Stack
SambaNova sells a full-stack platform rather than a bare chip. Three pieces fit together:
- RDU hardware — the reconfigurable dataflow chips that do the actual computation, designed for high efficiency on large-model inference
- SambaNova Suite — the enterprise software layer for deploying and managing models on that hardware, including in a customer's own data center
- SambaNova Cloud — a hosted inference service that lets developers call open and custom models through an API without owning any hardware
Its flagship SN40L processor, launched in 2023, was designed to run models with up to five trillion parameters on a single rack. The next-generation SN50, unveiled in February 2026, begins shipping in the second half of 2026, with SoftBank as its first deployment partner.
On-Premises and Regulated Enterprises
SambaNova leans hardest into markets where its efficiency and its ability to run entirely on-premises are worth the most: regulated industries and inference at scale. In 2026 JPMorgan Chase selected SambaNova as an inference-infrastructure partner to run secure, in-house AI inside the bank — a marquee reference for the argument that not every enterprise wants to send sensitive workloads to a public cloud.
That on-premises story is SambaNova's clearest wedge against both Nvidia and the big cloud providers: a bank, hospital, or government agency that cannot ship data off-site can still run large models on SambaNova hardware behind its own firewall.
SambaNova vs. Competitors
| Provider | Approach | Best For |
|---|---|---|
| SambaNova | RDU chips + full stack, on-premises friendly | Efficient inference for regulated, private deployments |
| Nvidia | General-purpose GPUs, dominant ecosystem | Training and inference across nearly every workload |
| Groq | LPU inference chips, very low latency | Fast token generation for real-time apps |
| Cerebras | Wafer-scale chips | Large-model training and high-throughput inference |
| Etched | Transformer-specialized ASIC | Maximum efficiency on transformer models only |
The common thread among the challengers is a bet that the next phase of the AI boom is defined less by training ever-larger models and more by running them cheaply and reliably at scale. Each attacks that from a different angle; SambaNova's is a reconfigurable chip plus a full enterprise stack aimed at private deployments.
Pricing
- Hosted inference API
- Pay per token for open and custom models
- No hardware to manage
- On-premises deployment
- RDU hardware plus management software
- For regulated and private workloads
- Full rack-scale deployments
- Bespoke capacity and support
- For banks, governments, and large enterprises
SambaNova is priced like enterprise infrastructure: a usage-based cloud tier for developers who just want inference through an API, and contracted hardware-plus-software arrangements for organizations that need models running inside their own walls. Exact rates are quoted per deal.
Company Details
| Attribute | Detail |
|---|---|
| Founded | 2017, Palo Alto, California |
| Leadership | Rodrigo Liang (CEO), Kunle Olukotun, Chris Ré (co-founders) |
| Type | Private |
| Valuation | 11 billion dollars (Series F, July 2026) |
| Key backers | General Atlantic, Intel, BlackRock, T. Rowe Price, Qatar Investment Authority |
| Notable partners | JPMorgan Chase, SoftBank |
Strengths
- Genuinely different architecture. The RDU is a real alternative design, not a GPU clone, aimed squarely at efficient large-model inference.
- On-premises strength. Running fully behind a customer's firewall is a strong fit for banks, hospitals, and governments that cannot use public clouds.
- Well-capitalized. A 1 billion dollar Series F at an 11 billion dollar valuation, with blue-chip backers, funds the expensive work of designing and manufacturing custom silicon.
Limitations and Considerations
- Nvidia's ecosystem is the moat. Most AI software is written and tuned for Nvidia's CUDA platform; any challenger has to overcome that gravity, not just match raw performance.
- Capital and manufacturing risk. Custom chips take years and enormous capital to design and fabricate, and depend on scarce advanced foundry capacity.
- Not a consumer tool. SambaNova is infrastructure — its value shows up in contracted enterprise and on-premises deployments, not in a self-serve app.
Key Takeaways
- SambaNova is an AI chip company whose Reconfigurable Dataflow Unit (RDU) targets efficient inference — running finished models cheaply — rather than competing head-on with Nvidia GPUs on training.
- It sells a full stack: RDU hardware, the SambaNova Suite for on-premises deployment, and SambaNova Cloud for hosted inference; the SN50 chip ships in the second half of 2026 with SoftBank as first partner.
- Backed by a July 2026 round of 1 billion dollars at an 11 billion dollar valuation and an on-premises deal with JPMorgan Chase, SambaNova is one of several challengers wagering that inference at scale — not training — is where a piece of Nvidia's market can be won.
