Intel and Qualcomm lose their shot at Tenstorrent as it raises $1.5 billion at a $6.2 billion valuation
Tenstorrent chose new money over reported sale talks with Intel and Qualcomm, more than doubling its valuation. The round funds volume manufacturing of Galaxy, its open RISC-V server line for AI workloads.
Richard Tang · 3 min read
Tenstorrent~$2.6B — Tenstorrent, private, latest valuation ~$2.6B closed a $1.5 billion Series E on Monday 1. The round values the Toronto-founded AI-computer maker at $6.2 billion 1. Intel+3.12% — Intel, up 3.12 percent today and Qualcomm+1.39% — Qualcomm, up 1.39 percent today held talks earlier this year to buy it, at prices reported up to $10 billion 2. Jim Keller, the designer of Apple's−0.16% — Apple, down 0.16 percent today A4 and A5, AMD's+2.23% — AMD, up 2.23 percent today Zen and Tesla's+3.78% — Tesla, up 3.78 percent today self-driving chip, made the call 3. Growing the company is "the best return for the investor and the employees," he told Bloomberg 4. The bet pays off only if Galaxy becomes the standard hardware for running open-weight AI models.
New money prices Tenstorrent below the reported bid ceiling
- Estimate
Data
| Value | |
|---|---|
| Series D, Dec 2024 | $2.6 |
| Takeover talks, Jun 2026 (estimate) | $10 |
| Series E, Oct 2026 | $6.2 |
The chips were built for inference, and May showed they work
Tenstorrent's Tensix cores each carry one to two megabytes of their own on-chip memory and reach one another through a direct mesh of hardware routers 2. When a model's weights fit inside that pooled memory, inference runs without the slow trips to external memory that a GPU hides by keeping thousands of threads in flight 2. The chips are built for that job, answering one user's request to a trained model, rather than for the giant training runs where Nvidia's+1.56% — Nvidia, up 1.56 percent today design is strongest 2. Galaxy, the server line built on them, reached general availability on April 28 2. In May it ran DeepSeek's open-weight models, which anyone can download and modify 5, and Keller called the demonstration the turning point: once the company could show performance and ship, "the investor interest came in" 4.
The round pays to turn shipped servers into volume production
The money funds volume manufacturing of Galaxy servers, next-generation hardware and licensable chip designs, and hiring 1. It deepens the bet on RISC-V, the open chip architecture its processors run on, with the software stack published as open source 3. Customers already sit in that middle: the trading firm Virtu Financial runs on-premises AI trading agents on Galaxy servers, and Equinix−1.23% — Equinix, down 1.23 percent today, a data-center colocation provider, offers them to enterprises 6. Samsung and LG license the designs, and the same architecture sells from a $999 card to rack-scale superclusters 18. The weak link is software: TT-Metalium, the open-source rival to Nvidia's CUDA, makes programmers manage data movement themselves, and independent production-scale benchmarks do not yet exist 2.
New investors set the price, and early ones have banked returns
Trace Capital, the Los Angeles firm former NFL linebacker Ben Mahdavi started this year, and New York-based Innovation X led the round 57. Fidelity, Morgan Stanley−0.81% — Morgan Stanley, down 0.81 percent today, T. Rowe Price−0.32% — T. Rowe Price, down 0.32 percent today, Sandisk+6.07% — Sandisk, up 6.07 percent today and Tether Investments joined 34. Real Ventures, an early Montreal investor, sold its entire stake last year and made back its whole $88-million fund 5. Keller and the employees keep the upside, and the IPO the company began preparing with its 2023 move from Toronto to Santa Clara stays open 1. Nvidia, past $5 trillion in market value, keeps its grip for now 5.
Buyers gain an independent supplier, and one benchmark to wait for
For buyers of AI compute, the result is an independent supplier. Intel and Qualcomm could have bought the alternative; now they have to compete with it. The signal to watch is an independent, production-scale benchmark of TT-Metalium against CUDA, because until one exists the new valuation rests on Tenstorrent's own performance claims 2.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.



