Qualcomm is in talks to acquire Tenstorrent, a startup focused on artificial intelligence chips. The price is estimated at somewhere between eight and ten billion dollars. Qualcomm shares fell by roughly one percent following the report.
Negotiations are ongoing, the price could still change, and the entire deal could theoretically fall through. No one has yet confirmed whether the amount also includes payments tied to meeting performance targets, a structure Qualcomm has used in past acquisitions of chip companies. Both companies declined to comment, and Reuters was unable to independently verify the report.
What Qualcomm does and where its gap lies
You probably know Qualcomm for its Snapdragon processors and mobile modems. But this business has matured, is cyclical, and is currently shrinking. Revenue from the mobile division fell 13 percent year over year due to expensive memory and reduced phone production in China.
The company has therefore been trying for years to diversify. Its automotive division, for example, has been successful, growing 38 percent annually. But the center of gravity of the entire chip industry has shifted elsewhere, toward infrastructure for generative artificial intelligence. And Qualcomm’s position there remains weak. It is preparing the AI200 and AI250 accelerators for data centers, but they are based on Hexagon neural processing units designed from the ground up for mobile devices. Scaling such an architecture to data-center racks will be expensive. You are fighting the physics of what the chip was originally optimized for.
Qualcomm CEO Cristiano Amon declared 2026 the “year of agents” at Computex 2026 and introduced the Dragonfly brand for data-center AI chips. But the brand announcement so far appears to be merely a marketing move, because the technology that would power it has not yet been unveiled. That is why the company spent all of 2025 and the beginning of 2026 acquiring the pieces it was missing. In December 2025, it bought RISC-V processor specialist Ventana Micro Systems and completed its $2.4 billion acquisition of Alphawave Semi for its interconnect technologies for AI clusters.
What Tenstorrent is and why everyone is targeting it
The startup was founded in 2016 and is led by Jim Keller, a very well-known name in chip circles. Keller worked on Apple’s A-series processors, AMD’s Zen architecture, and oversaw the development of Tesla’s autonomous driving chip. He then spent eight years at Tenstorrent building an AI accelerator architecture that deliberately does the opposite of Nvidia. He said so himself: “Whatever Nvidia does, we’ll do the opposite.”
In practice, this means that the basic compute unit of every chip is a Tensix core, a self-contained tile combining three components: a RISC-V processor for moving data, a matrix engine for tensor operations, and a vector unit for computations such as activation and normalization. Each core gets roughly 1.5 MB of its own high-speed memory. The trick is that data movement and computation itself run in parallel. In conventional graphics cards, these components are separate, and coordinating them takes time. Tenstorrent saves that time instead.
The company has already shipped three generations of silicon. Grayskull proved that the concept worked. Wormhole advanced the architecture and added Ethernet switching directly to the chip. Blackhole is the current product, manufactured using a six-nanometer process. It features 120 Tensix++ cores, along with 16 full-fledged RISC-V processors directly on the chip, 180 MB of on-chip memory, and 12×400 Gbps Ethernet ports. Those sixteen processors solve one problem. In a conventional system, the accelerator is connected to a host processor via PCIe, and every small batch of work must travel back and forth over that connection. Blackhole eliminates that overhead by managing the data locally.
Cheaper memory and standard Ethernet
Almost the entire industry has converged on expensive HBM, meaning memory chips stacked on top of one another right next to the compute core. Nvidia achieves bandwidth of around 3.35 TB/s with it in its H100 chip. But HBM is difficult to manufacture, reduces yields, and is made by only a handful of companies. Its shortage was behind the H100 supply problems in 2023 and 2024. Analysts estimate that GDDR memory costs roughly half as much per gigabyte.
Tenstorrent is betting on GDDR6. Blackhole therefore offers around 512 GB/s of bandwidth, which is about 15 percent of what the H100 can handle. That looks like a huge gap. But Tenstorrent compensates by packing as much high-speed memory as possible directly onto the chip and trying to keep active data there instead of constantly pulling it from main memory. When the model weights fit on the chip, main-memory bandwidth ceases to be a bottleneck.
The result is the Wormhole n150 card, which costs $999 and delivers 262 FP8 TFLOPS. The n300 model nearly doubles performance to 466 TFLOPS for $1,399. Nvidia’s H100 offers around 2,000 TFLOPS, but costs roughly $30,000. On a per-chip basis, Nvidia wins. On a per-dollar basis, the picture is completely reversed. And the market validated this bet more dramatically than anyone expected. At GTC 2026, Nvidia completely canceled the Rubin CPX project, and its replacement strategy for running models relies on an architecture based on on-chip memory. In other words, it reached the same conclusion Tenstorrent had arrived at years earlier.
The second pillar is interconnecting chips. Nvidia uses its proprietary NVLink and NVSwitch technologies for this, locking you into its ecosystem. Tenstorrent put Ethernet switching directly onto the silicon. Its chips can therefore form an interconnected network using standard networking equipment, scaling from a single chip all the way to superclusters that the company calls Galaxy.
The Ascalon processor and RISC-V
Tenstorrent has another ace up its sleeve. The Ascalon processor core is a 64-bit RISC-V processor developed with contributions from people who worked on Apple’s A-series team, AMD Zen, Arm, and Tesla’s Autopilot. For Qualcomm, this component is strategically valuable. Its entire portfolio is built on Arm technology, and that is not without risk. Arm is now a publicly traded company that needs to grow, is raising fees, and is increasingly competing with its own customers. The dispute over Nuvia demonstrated this clearly. Qualcomm acquired Nuvia in 2021 to build its own processor cores, and Arm sued it, arguing that the license did not survive the acquisition. Qualcomm won at the end of 2024. Its own RISC-V technology from Tenstorrent would give it independence in data centers, where it could start building from scratch.
Software
This is where it will become clear whether the whole plan succeeds. Nvidia’s challengers have usually won on hardware and lost on software. The fifteen years during which developers became accustomed to CUDA are the real obstacle, not the silicon.
Tenstorrent is addressing this through open source. All of its software is available on GitHub, from the low-level TT-Metalium for writing custom compute routines to the TT-Forge compiler, which supports PyTorch, TensorFlow, and ONNX. The company claims that 90 percent of models from Hugging Face run on its hardware without modification. It also offers the TT-QuietBox workstation with eight Wormhole processors for $1,500, making it cheaper than a single H100. The question is whether enough people will try the platform for a self-sustaining developer ecosystem to emerge around the hardware. After all, “the model runs on it” and “our team routinely uses it in production” are two very different things.
What Qualcomm will actually get and what it will not like
When all the pieces are put together, the acquisition makes sense. Hexagon at the network edge, Tensix accelerators as the backbone of Dragonfly in data centers, Ascalon as an in-house high-performance RISC-V processor, and Alphawave to interconnect the racks. There is also international reach: Tenstorrent has development centers in Serbia, Germany, Poland, India, and Japan. The automotive aspect is often underestimated. Automotive is Qualcomm’s fastest-growing business, and Tenstorrent is already working on automotive silicon. That is precisely the market where Qualcomm has the relationships and sales infrastructure needed to bring what it is acquiring to customers.
But the risk will not disappear once the money reaches the account. Qualcomm will suddenly be managing Arm-based processors, RISC-V chiplets from Ventana, and Ascalon cores, as well as two incompatible accelerator families, Hexagon and Tensix. Every decision about direction will also be a political decision about whose work survives. These are exactly the kinds of internal struggles that have already buried acquisitions that looked excellent on paper.
Then there is regulatory pressure. In 2018, Qualcomm abandoned its $44 billion acquisition of NXP after Chinese regulators allowed the deadline to expire amid trade tensions between the United States and China. Most major governments now view AI silicon as a matter of national security, so regulators in the United States, Europe, and China will each have their own views on the deal, on their own timelines.
And amid all this is the person Qualcomm is betting on. Jim Keller has an impressive résumé, but he joined Intel in 2018 and left in 2020 before any of his work reached the market. His reputation rests on designs that succeeded only after he had moved elsewhere. Whether performance-based payments will keep him at Qualcomm throughout the entire Dragonfly development cycle is a question the deal price does not answer.
Sources: tomshardware.com and medium.com
