What If You Could Mix CPUs, Accelerators & Chiplets? Tenstorrent at CES 2026

Tenstorrent arrived at CES 2026 with a clear message for engineers, AI developers, and silicon teams. The future of compute should be more open, more scalable, and easier to build around. Rather than focusing on one narrow part of the market, the company is positioning itself across the full stack, from licensable IP to PCIe cards, workstations, and compact AI systems. That matters because many engineering teams are now looking for alternatives to closed hardware ecosystems, especially as AI workloads spread across cloud, edge, embedded, and local development environments.

One of the biggest ideas shown in the video is Open Chiplet Atlas, or OCA. Tenstorrent describes this as an open chiplet ecosystem designed to make multi-vendor silicon more practical. In simple terms, it aims to let customers mix and match chiplets and IP blocks from different suppliers rather than being forced into a single vendor stack. That is a big shift. Chiplet-based design promises better flexibility and potentially lower development cost, but only if the interfaces, software layers, and system architecture can work together cleanly. Tenstorrent says OCA goes beyond the physical layer and covers transport, protocol, system, and software layers too, which is what makes the idea meaningful for real product development. The company also says the ecosystem already involves more than 50 partners.

The second major theme is TT-Ascalon, Tenstorrent’s high-performance RISC-V CPU IP. This is not being pitched as a niche embedded core. It is aimed at serious compute. Tenstorrent says TT-Ascalon is RVA23 compliant, uses a 64-bit out-of-order superscalar design, includes 256-bit vector data paths, and is backed by industry-standard SPEC CPU benchmarks. According to the company, it achieves more than 22 SPECint 2006 per GHz, more than 2.3 SPECint 2017 per GHz, and runs at more than 2.5 GHz on Samsung SF4X. That helps explain why Tenstorrent is framing it as a high-performance RISC-V option for servers, AI infrastructure, automotive high-performance compute, and advanced driver assistance systems. In the video, that story is made more tangible by the Linux and Doom demo running on TT-Ascalon hardware.

Tenstorrent is also pushing hard on the AI side of the portfolio. Its Tensix Neo IP is presented as silicon-tested AI IP for both training and inference, built around RISC-V and tuned for performance per watt. That matters because many customers no longer just want a finished chip or a full server. They want licensable building blocks they can integrate into their own silicon roadmap. Tenstorrent’s pitch is that it can provide CPU IP, AI IP, and chiplet technology together, which gives system designers more room to shape their own architecture rather than starting from a fixed platform.

Beyond IP, the company is also turning that technology into hardware that developers can buy now. The Blackhole family is Tenstorrent’s current PCIe AI hardware platform. Officially, Blackhole boards offer up to 120 Tensix cores, 16 big RISC-V cores, and up to 32 GB of GDDR6 memory per chip, with PCI Express 5.0 x16 connectivity. For engineers who want a ready-to-run desktop system, Tenstorrent’s TT-QuietBox line packages its hardware into workstation-style form factors aimed at local AI development and testing. On top of that, CES 2026 also featured a first-generation compact AI accelerator device built with Razer, designed to connect over Thunderbolt 4 or Thunderbolt 5 and scale up to four units for larger workloads. That opens the door to more portable local AI workflows for developers who do not want to jump straight to rack-scale infrastructure.

The roadmap story is just as important. Tenstorrent has said it is expanding its portfolio with higher-performance and higher-efficiency CPU cores, growing its AI IP offering, pushing further into automotive with partners such as BOS and AutoCore, and deepening its chiplet strategy after acquiring Blue Cheetah Analog Design. Put together, that makes the CES 2026 message quite straightforward. Tenstorrent does not just want to sell accelerators. It wants to give customers the pieces to own more of their silicon future, whether that means licensing CPU IP, integrating AI cores, experimenting with chiplets, or building local AI systems around open hardware and software. For engineering teams that want flexibility without vendor lock-in, that is what makes Tenstorrent worth watching.