The demand for hardware to train AI models is skyrocketing. According to McKinsey, AI chips are projected to make up nearly 20% of the $450 billion semiconductor market by 2025. Meanwhile, The Insight Partners anticipate that AI chip sales will surge to $83.3 billion by 2027, compared to just $5.7 billion in 2018, reflecting an impressive compound annual growth rate of 35%. This rate is almost ten times higher than that expected for non-AI chips.
A notable example is Tenstorrent, the AI hardware startup led by engineering expert Jim Keller. This week, the company announced it has raised $100 million through a convertible note funding round, co-led by Hyundai Motor Group and Samsung Catalyst Fund. Of the total, $50 million was contributed by Hyundai's automotive divisions — Hyundai Motor ($30 million) and Kia ($20 million) — which intend to collaborate with Tenstorrent to co-develop chips, including CPUs and AI co-processors, for advanced mobility solutions and robotics. The remaining $50 million came from Samsung Catalyst and several venture capital firms, including Fidelity Ventures, Eclipse Ventures, Epiq Capital, and Maverick Capital.
A convertible note is a type of short-term debt that can convert into equity upon a specific event, although Tenstorrent’s choice of debt over equity and the company’s post-money valuation remain somewhat unclear, despite the startup labeling it an “up-round” in its announcement. Previously, Tenstorrent secured $200 million at a valuation exceeding $2 billion.
With this convertible note round, which saw participation from Fidelity Ventures among others, Tenstorrent’s total funding has now reached $334.5 million. Keller indicated that these funds will advance product development, focusing on the design of AI chiplets and Tenstorrent’s machine learning software roadmap.
Based in Toronto, Tenstorrent specializes in AI processors and licenses AI software solutions and intellectual property related to RISC-V, the open-source instruction set architecture designed for customized processors across diverse applications.
The company’s custom-designed hardware for AI processing is a testament to its innovative approach. Founded in 2016 by Ivan Hamer, Ljubisa Bajic, and Milos Trajkovic, who previously worked at AMD, Tenstorrent initially concentrated on developing its proprietary infrastructure. In 2020, Tenstorrent unveiled Grayskull, an integrated system aimed at accelerating AI model training across various platforms, including data centers and edge servers, incorporating Tenstorrent’s cutting-edge Tensix cores.
However, in response to competition from established players like Nvidia, Tenstorrent pivoted towards licensing and services, with Bajic transitioning into an advisory role. In 2021, the company launched DevCloud, a cloud service that enables developers to run AI models without having to invest in hardware upfront. More recently, Tenstorrent has partnered with India-based Bodhi Computing and LG to integrate its products into Bodhi's servers and LG's automotive solutions and TVs. As part of the collaboration with LG, Tenstorrent aims to enhance video processing capabilities in its forthcoming data center offerings.
Ambitiously, Tenstorrent expanded its global reach by opening a Tokyo office in March, supplementing its existing locations in Toronto, Austin, and Silicon Valley. The pressing question is whether it can contend with the industry giants in the AI chip arena.
Prominent players like Google have developed processors — the TPU (tensor processing unit) — to train expansive generative AI systems such as PaLM-2 and Imagen. Amazon equips AWS customers with proprietary chips for both training (Trainium) and inference (Inferentia). Additionally, Microsoft is reportedly collaborating with AMD to create an in-house AI chip dubbed Athena.
Nvidia has experienced remarkable growth, briefly achieving a market capitalization of $1 trillion this year, thanks to soaring demand for its GPUs in AI training, holding an impressive 80% share of the discrete GPU market as of Q2 2022. While GPUs may not possess the specific capabilities of custom-designed AI chips, they excel at handling numerous computations simultaneously, making them ideal for training complex models today.
The current environment presents challenges for both startups and established tech companies. Last year, AI chipmaker Graphcore faced a $1 billion valuation drop after a deal with Microsoft collapsed and announced plans for job cuts because of the difficult economic landscape. Similarly, Habana Labs, Intel's AI chip subsidiary, had to reduce its workforce by approximately 10%.
Compounding these challenges is a supply shortage of essential components required to manufacture AI chips. Only time will reveal which vendors will emerge victorious in this evolving landscape.
Ranking
