RISC-V AI chips just got their biggest vote of confidence yet.
By Blumefield | October 20, 2026
A funding round that doubles as a declaration
On October 19, Tenstorrent closed a $1.5 billion Series E at a $6.2 billion post-money valuation, and the size of the round is almost beside the point. What matters is the message. The company led by Jim Keller, the engineer behind Apple's early mobile processors, AMD's Zen revival and Tesla's self-driving silicon, has decided that RISC-V AI chips are ready to compete for the most contested real estate in technology, the racks inside AI data centers.
The round was led by Trace Capital and Innovation X, with a roster of participants that reads like a cross-section of the money now chasing AI infrastructure. Fidelity, T. Rowe Price, Morgan Stanley and ARK Invest sit alongside Sandisk, Virtu Financial, XTX Markets and Tether Investments. Existing backers include Bezos Expeditions, Samsung, LG Electronics and Hyundai Motor Group. The valuation is more than double the roughly $2.6 billion mark Tenstorrent carried at the end of 2024, a jump that shows how fast sentiment around alternative silicon has turned.
Keller's framing is deliberately plain. "We built one architecture and sell it at every size," he said, describing a product line that runs from a single developer card up to multi-rack superclusters. The pitch underneath is sharper. "The future of compute is open, scalable, and owned by the people building on it." That is not neutral language. It is a direct challenge to the closed, proprietary model that has made one company the gatekeeper of the entire AI boom.
Why RISC-V is the whole argument
The reason investors are willing to underwrite RISC-V AI chips at this scale comes down to the instruction set itself. RISC-V is an open standard, which means there are no licensing fees and no single vendor controlling the roadmap. For a chip designer, that is the difference between renting the foundations of your product and owning them outright. The case for RISC-V AI chips is partly a licensing case: Tenstorrent sells its TT-Ascalon CPU cores and Tensix AI cores to other companies, including Samsung and LG, which gives it a second business line that most accelerator startups simply do not have.
The flagship demonstration is the Galaxy system. A 16-server cluster packs 512 chips and, according to the company, runs the full 671-billion-parameter DeepSeek model at around 350 tokens per second per user. Tenstorrent pairs the hardware with an open-source software stack built to run frontier models on-premises rather than through a rented cloud. For enterprises worried about data leaving their walls, and for governments building sovereign AI capacity, that on-prem story is a real selling point rather than a slogan.
The caution is just as real. Independent benchmarks against Nvidia's newest systems have not been published, and total-cost-of-ownership claims are notoriously sensitive to how you draw the boundaries. Keller has said publicly that RISC-V will take over the data center within five to ten years. That is a prediction, not a fact, and the gap between a convincing demo and volume deployment at hyperscaler scale is exactly where most challengers have stalled.
Turning down Intel and Qualcomm
The most revealing part of this story is what Tenstorrent chose not to do. Earlier in 2026, both Intel and Qualcomm explored buying the company. Reporting put a Qualcomm-led deal in the range of $8 billion to $10 billion, comfortably above the valuation Tenstorrent just accepted from its new investors. The company turned both suitors away, held talks with investment banks, and raised money instead.
Keller has said a demonstration in May changed his thinking. When Galaxy servers ran a full-size model at production scale, selling the company began to look like leaving the best part of the story on the table. Walking away from a premium acquisition to stay independent is the kind of decision that only makes sense if you believe the standalone outcome is far larger. It also keeps the architecture out of the hands of a bigger rival that might have folded it into an existing roadmap and quietly blunted its edge.
There is a strategic logic to independence beyond pride. An acquirer would have wanted Tenstorrent's intellectual property for its own products first. As a standalone licensor, the company can sell its cores to many customers at once, including the very hyperscalers that are increasingly designing their own parts. Being everyone's supplier is worth more than being one company's division, provided you can survive long enough to prove it. Independence also lets Tenstorrent keep pricing its RISC-V AI chips aggressively, the one lever a smaller player can always pull.
The Nvidia math is finally moving
For years, the case against Nvidia was theoretical. The company still holds somewhere between roughly 75 and 86 percent of the AI accelerator market depending on whose estimate you trust, and its software moat has frustrated every would-be competitor. What has changed is the shape of demand. Training enormous models remains Nvidia territory, but the center of gravity is sliding toward inference, the work of actually running models for users, and inference is where the economics start to favor specialized, cheaper chips.
The numbers behind that shift are striking. Custom inference silicon, the application-specific parts built by and for large operators, is growing shipments at around 44.6 percent a year, while merchant GPUs are growing closer to 16.1 percent. Google and Amazon are leading the move to in-house chips, and every watt and dollar they save compounds across millions of queries. That is the opening Tenstorrent is aiming at. RISC-V AI chips do not have to beat Nvidia at everything, they have to be cheaper and good enough at the one workload that is scaling fastest.
This is also why the investor list matters. Trading firms like Virtu and XTX live and die by latency and cost per operation, and their presence signals that the people most obsessed with squeezing compute think the on-prem, open-architecture pitch holds up. Sandisk and the memory-adjacent backers understand the supply chain intimately. A bet on RISC-V AI chips is a bet that this coalition is reading the market correctly, and it includes operators who would benefit directly if the technology actually works. The broader reshaping of AI infrastructure now runs through decisions like this one.
What to watch now
The money removes the most immediate risk, which was that Tenstorrent might run out of runway before its architecture matured. It does not remove the hardest one. Nvidia's advantage has never been only silicon, it is the CUDA software ecosystem that millions of developers already know by heart. Open hardware is necessary but not sufficient, and the open-source stack behind these RISC-V AI chips has to become something teams genuinely want to use, not merely something they are permitted to.
The signals to track over the next year are concrete. Watch for independent benchmarks that test the Galaxy claims against Nvidia's current generation rather than its last one. Watch the licensing business, because every major customer that commits to TT-Ascalon or Tensix cores validates the model more than any single hardware sale could. And watch whether the hyperscalers, who have both the volume and the incentive to diversify away from one supplier, start treating RISC-V AI chips as a serious second source rather than a science experiment.
Tenstorrent has already done the thing that was most in doubt. It has convinced a wide field of sophisticated investors to fund independence at a moment when selling would have been easy and lucrative. For a market that has spent three years assuming Nvidia's dominance was permanent, a $1.5 billion vote for RISC-V AI chips is a meaningful crack in the story. Whether it widens is now a question of execution, and the clock on Keller's five-to-ten-year prediction has officially started. Readers following the shift away from proprietary silicon can track continuing coverage at Blumefield.