Report: SambaNova is raising $350M+ from Intel-backed consortium

Chip startup SambaNova Systems Inc. is in talks to raise more than $350 million from investors, Reuters reported today.

The deal is expected to be led by Vista Equity Partners. That’s notable partly because the private equity firm has until now invested solely in enterprise software companies. Its reported change of focus may be motivated by this week’s broad selloff in software stocks.

Reuters’ sources said that venture capital firm Cambium Capital and Intel Corp. will also contribute to the round. The chipmaker reportedly plans to invest up to $150 million. The valuation that SambaNova could receive in the round wasn’t specified.

In December, word emerged that Intel had offered to acquire the chipmaker for $1.6 billion including debt. That’s about one-third what SambaNova was valued at following its most recent funding round in 2021. The company reportedly started seeking new financing last month after the acquisition talks with Intel stalled.

SambaNova’s flagship product is an artificial intelligence inference chip called the SN40L. It contains 1,040 cores made using Taiwan Semiconductor Manufacturing Co.’s five-nanometer process. They can perform 638 trillion calculations per second on data stored in the BF16 format.

The SN40L’s cores are spread across two separate silicon dies. Those dies sit on a base layer, or interposer, made using a TSMC technology called CoWoS. The interposer also hosts a high-speed pool of HBM memory that the chip uses to store AI models’ data.

A neural network comprises multiple modules called kernels. Those kernels, which are programs optimized to run across a large number of cores at one, regularly exchange data with one another. The longer it takes the data to flow from one program to another, the more time is needed to complete calculations.

The SN40L installs kernels on chip sections that are immediately adjacent to one another. That enables data to move between them faster than in configurations where kernels run on remote corners of the chip. According to SambaNova, the resulting reduction in data movement not only speeds up processing but also saves power.

The chip uses a technique called operator fusion to skip unnecessary computations. It works by combining hundreds of calculations into a single operation. The technique can boost processing efficiency, but it produces a significant amount of on-chip data traffic that can cause bottlenecks. SambaNova says the SN40L’s onboard network automatically fixes those bottlenecks.

The company ships its silicon as part of an air-cooled system (pictured) that contains 16 inference accelerators. The SambaStack, as it’s called, also includes networking hardware and power management components. It can run AI models with up to 5 trillion parameters.

Last July, SambaNova introduced an offering called SambaManaged that reduced the amount of time needed to deploy its hardware to 90 days. Over the next six months, the company inked hardware deals with four different data center operators. One of the contracts will see SambaNova supply chips for a sovereign cloud campus in Scotland with more than 2 gigawatts of planned capacity.

According to Reuters, the chipmaker exceeded its 2025 sales goals. SambaNova’s recent customer momentum may be one of the reasons it has drawn interest from institutional investors and Intel.

Image: SambaNova