Zhi东西（WeChat official account: zhidxcom）
compiled | Sun Xiaowen
edited | Cheng Xi
Zhi东西 reported on November 12, that yesterday, the UK AI inference chip unicorn Fractile announced that it has completed a $720 million C round of financing. This round was led by Lightspeed Venture Partners and Dragonfly Capital, with participation from Google's CapitalG and the UK National AI Venture Fund Sovereign AI. The company's post-money valuation reached $7 billion, which is more than six times its valuation of about $1 billion in May this year.
According to founder Walter Goodwin, the new funds will support the company in delivering its first chips and systems by 2027. The company aims to make AI faster in generating responses, completing complex tasks, and reducing operational costs by redesigning the hardware for running large models. It has already achieved a speed improvement of about 20 times on DRAM memory.
Fractile was founded in the summer of 2022, with its headquarters in London, UK. It mainly develops AI inference chips and systems for running pre-trained models, enabling them to generate text, code, etc., based on user needs. The company revealed that it is building its first AI chip, aiming to achieve up to 100 times faster performance than existing hardware, while reducing costs to 1/10.
According to the company's official information, prior to this round, it had disclosed its seed round and B round financing, totaling about $235 million. Accel, NATO Innovation Fund, and Oxford Science Enterprises were among the participants. In May this year, the company completed a $220 million B round of financing, with a post-money valuation of about $1 billion, thus entering the unicorn category.
Currently, the company's team has exceeded 100 people. CapitalG, one of the investors, revealed that the team consists of engineers from leading tech companies such as NVIDIA, Broadcom, Meta, and Intel.
▲Walter Goodwin, founder and CEO of Fractile (second from the right), with CapitalG and other related personnel (Photo source: CapitalG)
Founder Goodwin established Fractile while pursuing his PhD at the University of Oxford, with his main research focus on artificial intelligence and robotics. During his research, Goodwin realized that the time, memory bandwidth, and energy consumption required to run AI models would limit the large-scale application of AI. Therefore, Fractile has been dedicated to developing AI hardware since its establishment.
▲Walter Goodwin, founder and CEO of Fractile (Photo: Entrepreneur UK)
According to the lead investor Lightspeed Venture Partners, Fractile is developing inference accelerators and rack-level systems based on stacked DRAM, which directly packages DRAM on top of logic chips. The goal is to increase memory bandwidth while providing greater capacity than SRAM solutions, allowing large models to maintain fast generation speeds when handling long contexts.
Fractile has revealed that it is developing a machine learning compiler to optimize the operation of large models on its chips. The team will co-design hardware and software to improve the inference speed of large models.
In terms of customer collaboration, according to a report by Bloomberg on August 20, citing insiders, Fractile has reached a preliminary agreement with the U.S. large model company Anthropic to sell approximately $250 million (about 1.679 billion RMB) in chips, with an expected deployment in 2027 and plans to expand the contract size in the future.
The UK government stated in this investment announcement that supporting Fractile is aimed at enhancing the country's capabilities in key technologies, reducing reliance on foreign technology and supply chains, and keeping more job and economic benefits created by technological breakthroughs within the UK.
Conclusion: Optimizing from data transmission, speed and cost for long-range complex tasks are key
Fractile's funding reflects capital's attention to specialized inference hardware. The actual user experience of large models is influenced by generation speed, memory capacity, and cost. The company has chosen to start from data transmission to find new optimization opportunities.
Its research and development covers chips, compilers, and systems, indicating that improving inference efficiency requires collaboration across multiple areas. For customers, the appeal of such solutions lies in maintaining acceptable response speed and operational cost when large models handle longer and more complex tasks.
Source: Reuters, Fractile
