Tenstorrent hits $6.2B valuation to scale open-weight AI chips

Tenstorrent secured an unusually large $1.5 billion funding round. The investment valued the chip startup at $6.2 billion. This is a major jump for a company that was valued at $2.6 billion only months earlier.

The round was led by Trace Capital and Innovation X. Trace Capital was launched by Ben Mahdavi, the co-founder of Quiet Capital. Other investors included Cadence Ventures, Eclipse, Fidelity, SanDisk, and T. Rowe Price.

The new valuation also exceeded the amounts reportedly discussed in possible acquisition interest from Intel and Qualcomm. Those companies had considered valuations of as much as $5 billion. Tenstorrent instead chose to remain independent and use the capital to expand its AI-chip business.

An AI inference chip is hardware designed to run a trained model and generate outputs. It might produce text, classify an image, recommend content, or create video. Training chips perform a different job: they help the model learn patterns by processing huge datasets and repeatedly adjusting its internal settings. This distinction uses established computing knowledge; the article focuses on inference rather than defining it.

For example, when Tenstorrent says its Galaxy servers can run Pramana models, the servers are serving those already-built models. They accept a request, process it through the model, and return a result. They are not described as training Pramana from scratch.

Inference often happens repeatedly for users, so speed, energy use, and operating cost matter. Specialized chips can target the calculations and memory needs of model serving. Tenstorrent is raising money to scale chips tailored for open-weight models and video generation.

Tenstorrent plans to put the new funding into expanding its inference-chip business. The company is targeting hardware tailored for serving open-weight AI models. It is also pursuing chips that can support video generation. These areas require companies to run models repeatedly for users, making practical serving hardware important.

The key demonstration involved Tenstorrent’s Galaxy servers. CEO Jim Keller told Bloomberg that the systems could run Pramana models. That result helped Tenstorrent decide to remain independent and raise more money rather than pursue a sale. The article does not give technical performance figures for the servers.

The company’s next step is scaling products, not merely showing a prototype. More capital can support production, deployment, and broader availability of its targeted chips, although the article does not specify timelines. Tenstorrent is competing in a wider group of startups seeking alternatives to Nvidia.

Tenstorrent reportedly attracted acquisition interest from Intel and Qualcomm, with possible valuations reaching $5 billion. Instead of selling, the company decided to remain independent. The decision came after it demonstrated that its Galaxy servers could run Pramana models, according to CEO Jim Keller.

That demonstration appears to have given Tenstorrent confidence in its technology and business direction. The company then raised $1.5 billion at a $6.2 billion valuation. The article does not identify a single formal reason for rejecting either potential buyer, so the clearest stated explanation is the successful server demonstration and the choice to raise more capital.

Independence lets Tenstorrent pursue its own expansion plan. The new money will support chips designed for serving open-weight AI models and video generation. Its valuation more than doubled from $2.6 billion in late 2024.

Nvidia’s dominant position in AI chips creates an opening for startups that promise different or more targeted hardware. Companies may seek alternatives for specific workloads, suppliers, or deployment needs. The article describes Tenstorrent as an Nvidia challenger and places it among startups trying to reduce Nvidia’s control of the market. The broader market explanation uses established industry knowledge.

Tenstorrent focuses on inference, or serving trained models. Its Galaxy servers demonstrated that they could run Pramana models. The company is now raising money for chips tailored to open-weight AI models and video generation. Those focused products give it a concrete way to compete rather than copying every part of Nvidia’s business.

Etched is another example. TechCrunch reported that it was fielding investment offers at valuations of $40 billion or more. Together, these companies show investor interest in alternatives to Nvidia, although the article does not measure their market share or technical advantage.

Open-weight AI models are models whose learned weights are made available for others to obtain, inspect, modify, or run, depending on their license. This definition uses established AI knowledge; the article names open-weight models but does not define them. They differ from models available only through a provider’s private service.

A company using an open-weight model can deploy it on its own infrastructure or with a chosen hardware supplier. Tenstorrent says it wants chips tailored for serving these models. Its Galaxy servers also demonstrated that they could run Pramana models, providing a concrete example of model-serving hardware.

Specialized chips may be designed around the repeated calculations and memory demands of inference. That can help target speed, energy use, or cost, though the article provides no benchmark results. Tenstorrent’s funding will support this product direction, along with hardware for video generation.

When a user sends a request, a computer converts it into numerical data and moves that data through an AI model. The model applies learned weights through many mathematical operations. The resulting numbers are converted into an answer, such as generated text, an image, or a video. This explanation uses established computing knowledge, because the article does not describe chip mechanics.

For instance, Tenstorrent’s Galaxy servers can run Pramana models. The servers receive a request, process it through the model, and return the model’s output. Inference chips perform this serving process repeatedly. They must move data between memory and processing units while completing large numbers of calculations.

Specialized hardware can include circuits and memory arrangements suited to those operations. That focus may reduce unnecessary work, improve speed, or lower energy and operating costs. Tenstorrent is applying this approach to open-weight models and video generation, but the article gives no performance comparisons.