SambaNova has closed a funding round that values the company at $11 billion, according to Bloomberg. The figure marks a jump from the roughly $10 billion reported when the round was first taking shape in late June, and it represents a near quintupling in worth over a matter of months. The data confirms growing investor appetite for hardware alternatives to Nvidia, at a moment when the AI rush pushes enterprises and governments to consider proprietary, on-premise stacks for reasons of control, latency, and data sovereignty.

SambaNova's thesis is not about competing on the brute-force GPU terrain. The California-based company offers integrated systems (DataScale) built on a reconfigurable dataflow architecture that optimizes the execution of LLMs without relying on CUDA libraries. In practice, it sells ready-to-use appliances that combine custom silicon, networking, and a software layer that abstracts deployment complexity. For those evaluating bringing inference into their own data centers — or into air-gapped installations for regulatory compliance — this approach lowers the Total Cost of Ownership and simplifies operational management.

The valuation surge must be read in parallel with other signals. Demand for H100 GPUs remains insatiable, but the bottleneck is pushing more and more organizations to explore alternatives. The issue is not just the price of the cards or lead times: it's the realization that entrusting the entire stack to a single vendor creates a systemic risk that is hard to justify in regulated sectors. Banks, healthcare providers, public administrations, and defense need to run LLMs in controlled environments, with predictable latency and without data ever leaving the corporate perimeter. SambaNova's appliances address exactly this constraint, promising performance of hundreds of tokens per second on models with hundreds of billions of parameters, without requiring the extreme cooling of a GPU cluster.

Who gains from this scenario? First, early adopters that have already chosen to diversify their AI infrastructure. Second, the entire ecosystem of chip startups (Cerebras, Graphcore, Groq) that sees its existence legitimized precisely when Nvidia seems untouchable. Who stands to lose, paradoxically, could be the public cloud market for hypersensitive inference workloads: if the on-premise appliance becomes competitive in cost per token, the calculus that pushed everything toward AWS or Azure could reverse for a significant share of workloads.

The execution hurdle remains. Being valued at $11 billion means investors are betting on rapid revenue growth and on production capacity to match. So far SambaNova has announced partnerships with government entities and research institutions, but scaling to the broad enterprise requires sales channels, global support, and a partner ecosystem that Nvidia has presided over for years. Financial validation is an important step, but it does not replace the thousands of developers who have been programming on CUDA for a decade.

For those evaluating on-premise deployment, the SambaNova story offers a useful yardstick: the market is seriously pricing in independence from the Nvidia monopoly. Whether through dataflow chips, dedicated ASICs, or hybrid solutions, the message is clear: AI hardware is entering a maturity phase where control, more than raw power, dictates architectural choices.