Tenstorrent’s $1.5B Series E Hits a $6.2B Valuation — and Keller Said No to a Sale

Tenstorrent's Series E just closed: $1.5 billion at a $6.2 billion post-money valuation — more than double where the chip startup stood in late 2024. The most interesting part isn't the money, though. It's that CEO Jim Keller walked away from acquisition interest from Intel and Qualcomm at around $5 billion to take it. The cash goes into scaling Galaxy servers built for serving open-weight AI models and video generation, aimed squarely at Nvidia's inference business.

This is one of the largest chip-startup raises of the year, and it lands at a telling moment: investors have decided the next fight in AI hardware isn't about training giant models — it's about serving them cheaply. Tenstorrent has spent a decade building an open, RISC-V-based alternative to Nvidia's stack. Now it finally has the product demos and the order book to raise like it means it.

Inside Tenstorrent's Series E

The round was led by Trace Capital — a new firm started by Quiet Capital co-founder Ben Mahdavi — and Innovation X. The rest of the cap table reads like a cross-section of everyone who wants Nvidia to have competition:

• Asset managers: Fidelity, accounts advised by T. Rowe Price, ARK Invest, Morgan Stanley
• Chip and storage industry: SanDisk, Cadence Ventures, Maverick Silicon
• Trading and crypto money: Tether Investments, XTX Markets, Virtu Financial
• Venture: Eclipse, Protagonist

The $6.2 billion figure is post-money, so it includes the new cash — and it sits comfortably above the ~$5 billion valuation that Intel and Qualcomm were reportedly circling at five months ago. For comparison, the company's last raise was $693 million-plus in December 2024 at roughly a $2.6 billion valuation, led by Samsung Securities and AFW Partners with Bezos Expeditions participating. That's a 2.4x markup in under two years, which tells you how fast the open-inference story is repricing.

The sale that didn't happen

The acquisition subplot is what makes this round worth reading closely. Back in May, Bloomberg reported that Tenstorrent had held preliminary takeover talks with Intel and Qualcomm while also courting investors. Keller told Bloomberg the company chose independence after a May demonstration in which its Galaxy servers ran high-performance models from DeepSeek — the moment, in his telling, when the financing conditions, the product performance, and actual shipments all lined up at once.

"When we had the financing conditions, showed product performance and started shipping, investor interest came," Keller said, adding that many investors — himself included — believe the company still has substantial room to grow, and that growing it would deliver the best return for backers and employees.

Keller's calculus is easy to follow. If you believe open-weight models are going to eat a growing share of AI workloads — and everything about the last year points that way — then the company that serves those models cheapest is worth far more than $5 billion as a standalone business. Selling now would mean handing that upside to Intel or Qualcomm at a discount. The risk, of course, is that belief has to survive contact with Nvidia's pricing department.

Galaxy servers and the open-weight bet

The money has a specific destination: scaling production of Galaxy servers, which bundle dozens of Tenstorrent's AI chips into racks aimed at chatbots, coding assistants, and video generators. The pitch is full-stack openness — RISC-V processors with no licensing fees and no dependence on Arm's ecosystem, chip designs the company also licenses to customers like Samsung and LG, and a software stack that's open source. As Keller puts it: "We built one architecture and sell it at every size," from single cards and workstations up to racks and superclusters.

Two details from the raise deserve attention. First, video generation: Protagonist's George Bousis said the firm's video generation results came in ahead of its expectations, and the company is explicitly earmarking capital for video workloads — one of the most compute-hungry inference categories there is. Second, the company is now around 1,000 to 1,150 people across Santa Clara, Austin, Toronto, Belgrade, Tokyo, and Bangalore, which is real-company scale, not lab scale.

None of this is free of caveats. The announcement names no revenue figures, no hyperscaler customers, and no ship dates for the scaled-up silicon. Keller's résumé — Apple's A4 and A5, AMD's Zen, Tesla's self-driving chip — is genuinely elite, but résumés don't ship servers. The honest read: the demo worked, the money is real, and now the company has to turn a working demonstration into a hardware business that stands on its own.

The Nvidia challengers are raising real money now

Tenstorrent isn't alone. Earlier this month, TechCrunch reported that fellow AI chip startup Etched was fielding investment offers at valuations of $40 billion or more. The pattern is clear: capital is piling into inference, the part of the AI stack where Nvidia's moat is thinnest and where open-weight models have already broken the software lock-in.

That makes this a useful moment to check the startup funding climate — Tenstorrent's round is the biggest signal in it — and to compare notes with the other big AI infrastructure bets getting funded right now. The through-line across all of them: the money is flowing to whoever can make running AI cheaper, not whoever can train the biggest model.

Watch shipments, not valuations. Tenstorrent chose independence over a $5 billion exit because it believes the open-weight serving market is worth far more. With $1.5 billion in fresh capital and a $6.2 billion valuation to justify, it now has to prove it — one Galaxy rack at a time.