Tenstorrent Raises $1.5B to Challenge Nvidia

On Monday, October 19, AI chip startup Tenstorrent announced it had closed a $1.5 billion Series E funding round at a $6.2 billion post-money valuation. Reuters first reported the raise, which more than doubles the company’s valuation from December 2024, when it raised over $693 million at roughly $2.7 billion.

The round was led by Trace Capital and Innovation X, with participation from Fidelity, T. Rowe Price, Cathie Wood’s ARK Invest, SanDisk, Morgan Stanley, and Tether Investments. That list matters. This is not a specialist-only chip bet. Some of the biggest generalist investors on the planet just agreed that a six-year-old startup is worth more than twice what it was worth eighteen months ago, in a funding environment that has not been generous to everyone.

What Tenstorrent is actually building

Tenstorrent, based in Santa Clara and led by chip architect Jim Keller, designs AI chips and software on the open-source RISC-V architecture. The key detail: RISC-V means no Arm license fees, which changes the economics of designing and selling a chip. But Tenstorrent does not just sell silicon. It also sells Galaxy servers, full racks of its own chips built to run chatbots, coding assistants, and video generation workloads.

The money will go to scaling Galaxy production, developing new generations of AI and RISC-V CPU intellectual property, and hiring globally. And the company is not pitching vaporware. Named deployments so far include Virtu Financial, which runs on-prem AI agents on Tenstorrent hardware, Equinix’s Distributed AI Hub, and Japan’s ai&, which ordered more than 120 Galaxy systems for its sovereign AI program. That last one is worth noting: sovereign AI buyers are exactly the customers looking for non-Nvidia options.

It turned down the giants

The most revealing detail came from TechCrunch’s coverage. About five months ago, Intel and Qualcomm reportedly showed acquisition interest in Tenstorrent at a price of up to $5 billion. The startup chose independence instead, after demonstrating Galaxy servers running DeepSeek’s models.

A startup turning down up to $5 billion from incumbents is a statement. It says the founders and their backers believe the standalone path is worth more than being absorbed. It also says something about how acquisition interest functions in this market: the giants are not just buying capacity, they are trying to buy alternatives to Nvidia before those alternatives become too expensive to acquire.

The most credible Nvidia challenger yet

Nvidia’s moat was never just its chips. It is CUDA, the software stack that a generation of AI developers grew up on and that an entire ecosystem of tools assumes. Every challenger has to beat not the hardware but the habit. That is why most “Nvidia killers” have failed: they shipped silicon and discovered the software gap the hard way.

Tenstorrent’s approach is interesting because it aims at a gap in that moat rather than the moat itself. By building on open-source RISC-V, the company avoids one layer of licensing cost. By designing for open-weight models, it rides the fastest-growing segment of inference demand, the same demand created by open models from Meta, DeepSeek, and others. And by selling full server racks with its own software, it offers buyers a complete system, not a chip that requires six months of integration work.

That is a fuller answer than most challengers have managed. It does not guarantee success. But it explains why Fidelity and Morgan Stanley are in this round.

The timing says everything

Here is what makes this raise genuinely interesting: it landed in a week when the AI infrastructure trade looked shaky. The Information reported that executives now expect only 7 to 10 of the hundreds of GPU-rental “neoclouds” to survive, with CoreWeave down roughly 25% this week and Firmus pulling its planned IPO. I covered a parallel story yesterday: Microsoft terminated its agreement with British cloud provider Nscale for the UK’s biggest planned AI data center, after grid officials said the site would have no power until the 2030s.

Put together, the picture is not “AI spending is collapsing.” It is “money is getting choosy.” Investors are pulling back from GPU-rental middlemen, whose debt-heavy business models look fragile, while doubling down on the company trying to build a genuine hardware alternative. The market is separating the stack: commodity capacity gets repriced, real differentiation gets funded. Tenstorrent is on the funded side of that split.

What it means for builders

The direct consequence of more credible inference competition is pricing pressure. Today, serving AI models at scale means paying Nvidia prices, directly or through a cloud provider. If Tenstorrent’s Galaxy systems reach volume, open-weight model serving gets a real second option, and every real second option pulls prices down.

For builders, two practical takeaways. First, keep your inference layer portable. If your product only runs economically on one vendor’s hardware and software stack, you are exposed to whatever that vendor charges. The RISC-V and open-model ecosystems are the most realistic hedge against that. Second, watch where the sovereign AI demand flows. Governments and large institutions buying “sovereign” capacity are actively seeking non-Nvidia suppliers, and Tenstorrent’s deal with Japan’s ai& shows that demand is real, not theoretical.

Three things to watch

First, execution. Chip startups live or die on manufacturing and delivery. Tenstorrent now has the money to scale Galaxy production; the question is whether it can ship at volume, on time, with competitive performance per dollar.

Second, the paper-valuation question. At $6.2 billion, Tenstorrent is being valued like a company that is already winning, while it is still early in its revenue story. Nscale filed for a US IPO last month targeting around a $35 billion valuation while running a billion-dollar net loss in the first half of the year. The AI infrastructure market is assigning public-company valuations to private companies at record speed. That works until it doesn’t.

Third, open-model demand itself. Tenstorrent’s bet only pays off if serving open-weight models keeps growing fast enough to need a whole new hardware ecosystem. The trend line points that way, but it is the assumption underneath everything else in this story.