Tenstorrent valuation jumps to $6.2 billion to expand inference chips and video generation
Tenstorrent’s valuation jumped to $6.2 billion after closing a $1.5 billion financing round, a shift that underscores the growing investment bet on decoupling large-model training from inference and directing liquidity toward inference-specialized processors.
The new round more than doubled the company’s valuation, which is a ten-year-old specialist in inference chip manufacturing, compared with its $2.6 billion valuation at the end of 2024. The move comes five months after the company received informal acquisition offers from chip-industry peers Intel and Qualcomm at valuations of around $5 billion, but the management chose to remain independent and attract capital to support its standalone growth.
Running DeepSec models on Galaxy servers was the technical proof that led the company to reject the acquisition and maintain its independence.That is what Jim Keller, the prominent microprocessor engineer and CEO of Tenstorrent, explained in a statement to Bloomberg, noting that demonstrating the servers’ efficiency in running open-weight models settled the decision to continue direct competition. The company intends to use the new capital to expand production of chips specifically designed for open-weight AI models and to handle video-generation workloads that demand increasing compute capacity.
The round was led by Tris Capital, a newly founded investment firm created by Ben Mahdawi, co-founder of Quiet Capital, together with Innovation X. It also attracted a broad roster of investors, including Kidness Ventures, Eclipse, Fidelity, Sandisk and The Rowe Price. The move is part of a growing wave of AI-hardware startups seeking to curb Nvidia’s dominance over global supply chains, at a time when other rivals such as Etched are also receiving investment offers at valuations of $40 billion or more.
The inference cost of running open models has become the governing equation for regional data-center and infrastructure budgets.In our region, specifically the Gulf markets, Egypt and the Arab Levant, this trajectory has a direct impact on engineering teams and technology managers. The rush of financial institutions and government bodies to host open models locally to protect data sovereignty faces a steep operating bill when relying on general-purpose processors, making the emergence of silicon alternatives that focus solely on lowering inference cost and handling multimedia a factor that reshapes procurement calculations and cloud contracts. Diversifying hardware manufacturers gives local data centers negotiating leverage against a single-supplier monopoly and allows engineering staff to improve energy-consumption efficiency for each software call.
We see in this round an explicit signal that inference-specific hardware architecture is no longer a marginal engineering experiment but has become an independent chip-market segment aiming to break price constraints, which requires engineers in local data centers to reassess infrastructure viability before committing to long-term hardware obligations.