Tenstorrent has raised $1.5 billion at a $6.2 billion valuation, TechCrunch AI reports. The 10-year-old chip startup builds hardware for inference, which means running AI models rather than training them. Five months ago, Intel and Qualcomm reportedly looked at buying it for as much as $5 billion. Instead of selling, Tenstorrent raised a bigger round and stayed independent.
⚡ The Short Version
- The deal: $1.5 billion raised at a $6.2 billion valuation.
- The jump: The valuation more than doubled from the $2.6 billion it reached in late 2024.
- The leads: Trace Capital, a new firm from Quiet Capital co-founder Ben Mahdavi, and Innovation X.
- Other backers: Cadence Ventures, Eclipse, Fidelity, SanDisk, and T. Rowe Price.
- The plan: Scale chips built to serve open-weight AI models and video generation.
- The big picture: Another sign that investors will pay up for real alternatives to Nvidia.
🧠 Why Tenstorrent Said No to Intel and Qualcomm
CEO Jim Keller is one of the best-known chip designers in the industry. He told Bloomberg the company chose to stay independent and raise more money after it showed its Galaxy servers could run DeepSeek models.
That detail matters. DeepSeek’s models are among the most widely used open-weight models, meaning anyone can download the weights and run them on their own hardware. If Tenstorrent’s systems can serve those models well, the company has a working product for a market that’s growing fast, not just a promising roadmap.
The math worked out too. Turning down a buyout at up to $5 billion and raising at $6.2 billion a few months later is a strong outcome for Keller and his team. It also tells you where investors think AI chip money is heading.
🎯 Why Open-Weight Models Are the Bet
What stands out is how focused the strategy is. Tenstorrent isn’t trying to beat Nvidia at everything. It’s going after one growing slice of the market: companies that run open-weight models on their own hardware.
That slice is getting bigger for a few reasons:
- Cost: Running open models on your own hardware can cost less than paying per token for closed APIs.
- Control: Companies with strict data rules want models they can host in-house.
- Quality: Open-weight models like DeepSeek’s have closed much of the gap with the top closed models.
Video generation is the other target. It’s one of the most compute-hungry jobs in AI right now, and the demand for cheaper inference there is real.
🏁 Nvidia’s Challengers Are Getting Expensive
Tenstorrent isn’t alone. A wave of AI hardware startups is going after Nvidia’s dominance, and the valuations keep climbing. TechCrunch AI reported earlier this month that Etched, another AI chip maker, was fielding investment offers at valuations of $40 billion or more.
The status quo has been simple for years: if you want to run serious AI workloads, you buy Nvidia. That’s still mostly true for training. Inference is a different fight. It rewards cost per token, power use, and specialization, and those are areas where smaller, focused chip companies can actually compete.
The presence of SanDisk and big asset managers like Fidelity and T. Rowe Price in this round is worth noting. This isn’t just venture money chasing a hot sector. Strategic and institutional investors are betting that inference hardware gets split among several winners.
🔍 What Practitioners Should Watch
If you run AI workloads or make infrastructure decisions, keep these on your radar:
- More hardware options for open models. Expect more vendors pitching servers tuned for DeepSeek-class models. Compare cost per token, not just raw chip specs.
- Software maturity. Nvidia’s real moat is its CUDA software ecosystem. Any challenger has to show its tooling works without a painful migration.
- Pricing pressure. More serious competitors in inference should push prices down over time, for cloud GPU rentals and on-premise hardware alike.
- Consolidation risk. Intel and Qualcomm were already circling. Big chip companies may come back with bigger offers, or go after other startups instead.
🚀 What Comes Next
Tenstorrent now has the cash and a clear lane. The next question is whether it can ship Galaxy servers at scale and win real customers away from Nvidia hardware. If open-weight models keep improving, specialized inference chips could become one of the most competitive corners of AI. Full details are available in the original TechCrunch AI report.