Tenstorrent raises $1.5bn to scale its alternative to Nvidia’s AI computing
Image: Tenstorrent
Think of Tenstorrent as a company that designs engines for artificial intelligence, and also sells complete machines built around them. Its processors perform the calculations that AI models need to answer questions, generate images or produce video. Customers can buy computing systems or license processor designs to incorporate into their own chips.
That gives Tenstorrent two routes into the AI market. A business wanting to run an AI service can purchase its hardware, while a semiconductor developer can use its designs as building blocks for a customised processor.
With the recently announced series E funding, the company can now support larger-scale deployments of its Galaxy servers and further development of its licensable AI and central processing unit technology.
Why Tenstorrent is raising $1.5 billion
The size of the investment reflects the resources needed to turn processor designs into computing infrastructure supplied at substantial scale.
For a hardware business, that means financing production, securing components, assembling and testing systems, and supporting installations worldwide. Customers need confidence that equipment will arrive, operate reliably and remain supported as their requirements change.
Tenstorrent explicitly identifies volume manufacturing and global commercial expansion as purposes of the funding. The round supports a scaling phase for a company that already sells hardware, rather than its first transition from designs to physical products.
Building a credible alternative to Nvidia requires this wider operation to work alongside the silicon. It also requires manufacturing capacity, although that can be provided through partners. The announcement does not describe plans to construct Tenstorrent-owned semiconductor factories.
The company behind Jim Keller’s AI processors
Tenstorrent was founded in 2016 by Ljubisa Bajic, Ivan Hamer and Milos Trajkovic, before generative AI brought widespread attention to the processors behind chatbots.
Its best-known figure today is chief executive Jim Keller. His career includes senior engineering roles at AMD, Apple, Tesla and Intel, with contributions to processor families including AMD Zen and Apple’s A4 and A5. His experience lends significant credibility to Tenstorrent’s ambitions.
A major target for Tenstorrent is AI inference: the work an already trained model performs when someone asks it to respond. Training develops a model’s capabilities; inference puts them to use. For a popular service, those repeated calculations become a continuing operating expense.
Why Blackhole uses GDDR6 memory
One consequential design choice concerns how processors obtain the information needed for those calculations.
Tenstorrent’s Blackhole AI processor uses GDDR6 external memory, while many high-end AI accelerators use high-bandwidth memory, or HBM. Blackhole also contains substantial local memory on the chip. Its published card specifications list 180 megabytes of on-chip SRAM and up to 32 gigabytes of GDDR6.
Think of local memory as a small workspace beside a worker, with external memory providing a larger store further away. Keeping frequently needed information nearby reduces trips, but the workspace has limited capacity. Efficient operation depends on deciding what stays close and what must be fetched.
Tenstorrent’s software partner Moreh describes a strategy combining relatively lightweight processors, GDDR memory and conventional packaging. The intended benefits include supplier diversity and cost efficiency, with software managing data movement to reduce unnecessary external transfers.
Blackhole’s external memory bandwidth is lower than that of leading HBM-equipped accelerators. Its architectural approach must compensate through effective use of local memory, communication and distributed computation.
Lower component costs therefore do not automatically produce a cheaper installation. Additional processors, networking, electricity and engineering work all enter the calculation. What matters is how much useful AI work the complete system delivers at the response time customers require.
How Tensix processors and Galaxy systems connect
Tenstorrent’s AI processors are organised around Tensix cores, combining programmable RISC-V processors, specialised computing resources and local memory.
RISC-V is an open instruction-set architecture: a specification of instructions that compatible processors understand. It provides a shared foundation for different implementations, while Tenstorrent’s particular designs remain commercial products.
Networking operates at distinct levels. Within a chip, a network-on-chip links processing resources. Between chips, integrated Ethernet interfaces enable direct exchanges without routing every transfer through a host processor.
This supports the distribution of AI workloads across connected processors and Galaxy systems. Software must still divide the work effectively, keeping computation and communication in balance as installations grow.
Open Chiplet Atlas offers another building option
Tenstorrent’s modular ambitions also extend inside semiconductor packages.
A chiplet is a smaller semiconductor die that can be combined with others to build a larger system. Designers can assemble components serving different functions instead of implementing everything on one large die.
In October 2025, Tenstorrent launched its Open Chiplet Atlas ecosystem, proposing an interoperability architecture, reusable design framework and compliance arrangements.
As an illustration, a company developing a processor for industrial equipment might combine established computing components with its own specialised functions. Reusing components could let engineers concentrate on the features that distinguish their product.
Achieving that requires compatible communication protocols, software, packaging and power delivery. The initiative aims to reduce integration effort; its practical value will depend on adoption and components demonstrated to work together.
Competing through AI hardware and processor licensing
Alongside finished systems, Tenstorrent licenses TT-Ascalon RISC-V CPU designs and Tensix AI technology. It identifies Samsung and LG among its licensing customers.
These are distinct elements of its portfolio. Licensing a CPU design allows a customer to incorporate that implementation into its own silicon; buying a Galaxy system supplies computing equipment for running workloads.
In systems, customers compare Tenstorrent’s offering with alternative hardware and software platforms. Nvidia’s strength includes CUDA’s extensive libraries and development tools, while AMD supplies accelerators supported by ROCm. Compatibility, optimisation effort and ongoing support weigh heavily in purchasing decisions.
In licensing, the decision concerns the foundations of a future chip. Customers must assess performance, integration work, development risk and how well the design fits their intended application.
Tenstorrent reports progress in deploying its own hardware: in June, it said Japanese AI platform ai& had deployed more than 120 Galaxy systems. Next for the company will be customers expanding installations and bringing licensed designs into production. Those outcomes would show that its architectural choices are becoming infrastructure and products customers are prepared to keep investing in.
Further reading on MoveTheNeedle.news:
What has Ilya Sutskever built that Nvidia wants to scale?
Verda bets on NVIDIA Rubin as demand for AI computing power accelerates