SambaNova raises $1 billion at an $11 billion valuation as demand for AI inference surges

SambaNova Systems, which makes specialized chips for artificial intelligence, has raised $1 billion at an $11 billion valuation. The round signals investor confidence in growing demand for AI inference infrastructure. CEO Rodrigo Liang discussed the raise in an interview with Ed Ludlow on Bloomberg Tech.

What we know about the round

SambaNova is a Palo Alto startup founded in 2017 by Stanford alumni, including professor Kunle Olukotun. The company designs its own chip architecture, the Reconfigurable Dataflow Unit (RDU), as an alternative to Nvidia GPUs, which currently dominate the market for neural network training and inference. The idea behind the RDU is to let data "flow" through the chip along a route optimized for a specific neural network, instead of pushing it through the general-purpose GPU architecture. According to the company, this reduces latency and power consumption when running inference on large models.

• Round: $1 billion in funding
• Post-money valuation: $11 billion
• CEO: Rodrigo Liang
• The interview aired on Bloomberg Tech with Ed Ludlow

Why inference has become the main source of demand

Liang stressed that demand for high-speed inference — fast, cheap execution of already-trained models in real products — is growing at an enormous pace. A few years ago, most of the industry's capital spending went into training large models. Now companies are rolling AI into their services at scale, and they need infrastructure that can respond to users with minimal latency and at an acceptable price per token.

The shift is driven by how businesses use AI. Instead of one-off chatbot queries, they increasingly deploy AI agents and copilots that call models many times within a single task, from analyzing a document to automating a workflow. Each of those calls needs a model response in a fraction of a second. That makes inference cost and speed, not just model quality, the deciding factor when choosing infrastructure.

Who is competing for the inference market

This shift in focus from training to inference has become one of the defining themes of 2026 for the entire chip industry. Nvidia still holds the bulk of the AI accelerator market thanks to its CUDA ecosystem, but others are moving into specialized inference chips: hyperscalers with in-house designs — Google's TPU, Amazon's Trainium and Inferentia — as well as chip startups Groq and Cerebras, which compete with SambaNova for enterprise customers looking for an alternative to expensive GPU clusters. For these customers, the key metrics are not peak performance on paper but the real cost per token of model output and latency under heavy load.

What it means

SambaNova's large round, set against the broader boom in AI infrastructure investment, shows that the market is increasingly betting not only on model developers but also on companies building the hardware to run those models cheaply and quickly. That hardware, not just the quality of the neural networks themselves, increasingly determines the real economics of scaling AI in products.