Tenstorrent Raises $1.5 Billion at $6.2 Billion Valuation to Scale Open-Weight AI Chips

Posted on 20th Oct 2026 12:03:33 in Artificial Intelligence, Machine Learning

Tagged as: Tenstorrent, AI chips, open-weight models, RISC-V, AI inference

Tenstorrent, the AI chip company led by veteran microprocessor architect Jim Keller, has closed a $1.5 billion Series E round at a $6.2 billion post-money valuation. The raise, announced on Monday, more than doubles the $2.6 billion the company was worth after its Series D in December 2024, and it answers a question that had been hanging over the business since spring: whether Tenstorrent would sell itself to an incumbent chipmaker or stay independent and scale.

The answer was independence. Tenstorrent had drawn acquisition interest from Intel and Qualcomm earlier this year at valuations of as much as $5 billion, as Bloomberg reported. Keller's explanation for walking away from those talks is more instructive than the headline number. A May demonstration in which Tenstorrent's Galaxy servers ran high-performance models from DeepSeek, he said, convinced the company it had a product worth scaling on its own terms.

That detail points to where the AI hardware market is actually moving. Tenstorrent is not chasing frontier training runs against Nvidia's biggest customers. It is building for inference, the business of serving finished models to users, and specifically for open-weight models that customers can download, fine-tune and run themselves.

What the $1.5 Billion Round Actually Contains

The round was led by Trace Capital and Innovation X, with participation from a long list of new and existing investors including ARK Invest, Cadence Ventures, Eclipse, Fidelity Management and Research, Maverick Silicon, Morgan Stanley, Protagonist, Sandisk, T. Rowe Price, Tether Investments, Virtu Financial and XTX Markets. Trace Capital is a young firm, founded in 2026 by Ben Mahdavi, who previously co-founded Quiet Capital. Innovation X is a late-stage venture investor founded in 2016 by Roger Monteforte.

The company said the capital will go to three things: deploying Galaxy servers at scale, continuing to serve open-weight models across large language models and video generation, and advancing the AI and CPU designs it licenses to customers who are building their own silicon. It also plans to expand its global workforce. Tenstorrent has around 1,150 employees, according to Bloomberg, and operates from Santa Clara, Austin, Toronto, Belgrade, Tokyo and Bangalore.

The valuation jump is steep even by the standards of the current AI funding cycle. Tenstorrent raised roughly $693 million in December 2024 at a valuation above $2.6 billion, with Jeff Bezos' Bezos Expeditions, Samsung Securities and AFW Partners among the investors. The company's existing backers also include Samsung, LG Electronics and Hyundai Motor Group. Doubling a valuation in under two years is now routine for AI infrastructure companies, but doing it while remaining independent is not, particularly when two of the largest chipmakers in the world have already kicked the tyres.

Why Open-Weight Models Are the Wedge

The strategic logic rests on a shift in how companies consume AI. Frontier labs still spend enormous sums training models on Nvidia hardware, but a growing share of real-world work happens at inference, where a finished model answers queries, generates images or renders video for paying users. That work is repetitive, latency-sensitive and extremely price-sensitive. It is also increasingly served by open-weight models, where the customer controls the weights, the fine-tuning and, critically, where the data sits.

That is the market Tenstorrent is aiming at. The company positions Galaxy as a way for developers to run frontier-class open models locally and for enterprises to deploy models on-premises at competitive latency and throughput. Its pitch to buyers is straightforward: you do not have to rent capacity from a hyperscaler or a GPU neocloud, and you do not have to accept whatever price the accelerator monopoly sets that quarter.

The DeepSeek demonstration matters in this context because it proved the stack could handle a serious open-weight model, not a benchmark toy. Keller's argument to investors, as reported, is that Tenstorrent can compete on price and accept orders that are too small to interest the largest chipmakers, a segment the incumbents have little incentive to chase.

Inside Galaxy: RISC-V, Tensix and 23 Petaflops of FP8

The hardware itself is unusual. Galaxy servers, which began shipping in April 2026, pack 32 of Tenstorrent's RISC-V Blackhole processors with Tensix AI cores, delivering 23 petaflops of FP8 compute per system. Tenstorrent sells the same architecture at every size, from a single card or workstation up through racks, superclusters and custom silicon designs.

"We built one architecture and sell it at every size," Keller said in the company's announcement. "The future of compute is open, scalable, and owned by the people building on it."

Two things distinguish the approach. First, the processors are built on RISC-V, the open instruction set architecture, rather than a proprietary design. Second, Tenstorrent publishes its software stack as open source, which lowers the cost of evaluating the hardware and reduces the fear of being stranded on an orphaned toolchain. The company also licenses its TT-Ascalon RISC-V CPU and Tensix AI cores to chip designers, with Samsung and LG named as customers. That licensing business is a meaningful hedge: it earns revenue from companies that never buy a Tenstorrent server.

Investor commentary suggests the video generation workload has been a specific point of differentiation. George Bousis, co-founder and managing partner at Protagonist, said the company designs both the chip and the software that runs on it, and that its video generation results beat industry benchmarks ahead of expectations.

The Road Not Taken: Intel and Qualcomm

The counterfactual is worth sitting with. Tenstorrent was founded in 2016 and spent years as a technically respected but commercially unproven challenger. When Intel and Qualcomm came calling, a sale at up to $5 billion would have been a respectable outcome for most venture-backed chip companies. Qualcomm was still reported to be exploring an acquisition as recently as June 2026.

Instead, Tenstorrent raised at a higher valuation than either suitor reportedly offered. That is either a sign of genuine conviction that independent scale is achievable, or a sign that the private market is pricing AI silicon more generously than strategic buyers are. Both readings can be true at once, and the next two years will decide which one dominates.

What This Means for Businesses Buying AI Compute

For companies that buy AI capacity rather than build chips, the practical significance is choice. Every credible alternative to Nvidia hardware exerts downward pressure on inference pricing, and inference is where most operating AI budgets actually go once a product is live. Tenstorrent's on-premises pitch is also relevant to organisations with data residency obligations, including Indian enterprises in regulated sectors, where moving customer data to a foreign cloud is a compliance problem rather than an engineering preference.

The software layer is the harder question, and it is moving in the same direction. On the same day as the Tenstorrent announcement, Amazon Web Services released a tool called TorchNeuron that lets PyTorch applications run on its Trainium chips by changing a single line of configuration. Amazon says about 85% of AI applications are written with PyTorch, and it is contributing upstream changes to make the framework less tied to any one vendor's hardware. Nvidia's moat has never been just silicon; it is CUDA, the software layer that makes its chips the path of least resistance. Multiple companies are now attacking that layer simultaneously.

The caveat for buyers is that hardware portability rarely delivers the savings promised on the slide. Moving a workload to a new accelerator usually means re-tuning kernels, re-validating accuracy and rebuilding operational tooling. A one-line configuration change gets a model running; it does not make it fast. Anyone evaluating alternative silicon should insist on benchmarks from their own workload, not the vendor's, and should price in the engineering time.

What Could Go Wrong

A $6.2 billion valuation on $1.5 billion of fresh capital sets a high bar. Tenstorrent now has to convert manufacturing capacity into volume deployments and turn pilot installations into sustained customer use. Scaling hardware production is a notoriously unforgiving exercise, and the company is doing it while competing against Nvidia, AMD and Broadcom, each with vastly larger supply chains and software ecosystems.

The competitive field is also getting expensive. Etched, an inference chip startup, was reportedly fielding investment offers at valuations of $40 billion or more earlier this month. Capital is flooding into the same thesis, which validates the market and simultaneously guarantees that Tenstorrent will not be the only well-funded alternative on a buyer's shortlist.

What the round does establish is that the open-weight thesis now has a funded hardware champion with a shipping product, a licensing business and a founder-level engineer running it. For an industry that has spent three years worrying about a single point of failure in its supply chain, that is a meaningful development, even if the volume numbers are still to come.

Sources

• Tenstorrent — Tenstorrent Announces a $1.5B Series E at a $6.2B valuation
• TechCrunch — Tenstorrent hits $6.2B valuation to scale open-weight AI chips
• Data Center Dynamics — AI chip startup Tenstorrent raises $1.5bn at $6.2bn valuation
• The Next Web — Tenstorrent raises $1.5B at a $6.2B valuation to scale its Galaxy servers
• Investing.com — Chipmaker Tenstorrent raises $1.5 billion at $6.2 billion valuation
• TechCrunch — Amazon is taking its biggest shot at Nvidia yet with new TorchNeuron dev tool