A post-2000s school dropout develops AI chips, with a valuation of 22.3 billion yuan.

According to a report from TechCore on August 3, London-based AI chip startup OLIX in the UK has just announced a financing round of $312 million (approximately RMB 2.1 billion), with a valuation reaching $3.3 billion (approximately RMB 22.3 billion). Arm and Reed Hastings, co-founder of Netflix are among the investors participating in this round.

The new round of funding for OLIX will be used to support the delivery of AI inference chip DX-1 to its first batch of customers in the second half of 2027, build its custom chip platform, and fulfill the manufacturing and supply chain commitments required to scale up its inference chips.

Founded in March 2024, OLIX focuses on developing AI inference chips with high throughput and high interactive performance for ultra-high-load inference tasks based on new photonic interconnection technologies, with a plan to develop AI chips that are faster and cheaper than NVIDIA GPUs.

Its core idea is to integrate the SRAM storage architecture and optical interconnection technology and promote large-scale implementation, so that it can outperform the traditional architecture built on HBM high-bandwidth memory in the two indicators of throughput per unit power consumption (throughput/MW) and total cost of ownership (TCO). Meanwhile, compared with the pure silicon-based SRAM solution, it can take a leading edge in interactive response capability and data latency performance.

In February this year, OLIX secured a $220 million (approximately RMB 1.49 billion) financing round, with its valuation exceeding $1 billion (approximately RMB 6.75 billion) to become a unicorn. Six months later, its valuation has risen by more than 2 times. According to the disclosure from the UK's Financial Times, OLIX's total accumulated financing has exceeded $562 million (approximately RMB 3.8 billion) so far.

The founder of OLIX is 25-year-old James Dacombe. He dropped out of high school at the age of 16, and founded CoMind, a UK-based brain monitoring startup, at the age of 18. The startup has raised a total of $100 million (approximately RMB 680 million) to date.

James Dacombe, founder of OLIX, source from the UK's Financial Times

01. Deliver the first chip in the second half of next year, the multi-rack architecture can adapt to models with 10-trillion parameters

According to OLIX's official website, its researchers believe that the industry's existing way of building AI inference chips is approaching the efficiency limit. A data center is like a factory that produces tokens, and producing one token requires hundreds of processes, each of which has different hardware requirements.

Other factories will equip each process with a dedicated machine, but the token factory runs all processes on the same general-purpose chip. Each generation of new chips improves performance by increasing the specifications of a single chip, which makes the chips more versatile but never professional-level chips. Based on this, the system developed by OLIX allows each stage of the token production process to use a dedicated chip.

In OLIX's X-1 platform, the model is disassembled and distributed to a massive number of chips for parallel operation, and each chip only needs to be responsible for processing a specific segment of the model's computing tasks. In addition, the iteration speed of AI model architectures is extremely fast, and all chips are equipped with programmable flexible computing arrays, which will not solidify specific model architectures in the chip hardware in the form of hard coding.

The above entire system takes the low-speed and high-bandwidth optical interconnection architecture as the core, abandons the traditional copper wire transmission method, and relies on optical signals to directly complete data interaction between chips, so as to achieve low latency and low power consumption.

OLIX adopts a rack-level collaborative integrated design for all components of the entire transmission link, and its platform is equipped with a fully deterministic compiler responsible for scheduling computing power tasks across racks. Researchers believe that this solution can reduce the usage cost of existing cutting-edge AI models, increase the supply of computing power, and more importantly, clear obstacles for the deployment of future super-large models with stronger computing power.

At present, researchers are developing the first chip DX-1 on the X-1 platform, which is specially built for large model inference operations and content output stages.

According to the data released by OLIX, for large models with hundreds of billions of parameters, DX-1 can achieve Pareto optimal inference performance, with the single-user token generation rate reaching more than 10,000 tokens per second. Meanwhile, compared with general-purpose chips running large-batch tasks, it has better throughput performance of token output per unit power consumption.

In addition, OLIX's multi-rack scale-up expansion architecture can adapt to super-large models with 10-trillion parameters or above.

02. Appoint co-inventor of software-defined networking and name new Chief Financial Officer

OLIX also officially announced personnel changes, announcing that Nick McKeown has joined its board of directors, and Matt Briers has been appointed as Chief Financial Officer.

McKeown is the co-inventor of Software-Defined Networking (SDN), the OpenFlow protocol and the P4 programming language. He currently serves as Professor Emeritus of Computer Science and Electrical Engineering at Stanford University.

McKeown co-founded Nicira, a software-defined networking (SDN) startup, and Barefoot Networks, a programmable gate company. In 2012, Nicira was acquired by cloud computing company VMware, and after Barefoot Networks was acquired by Intel in 2019, McKeown joined Intel, and served as Senior Vice President and General Manager of Intel's Network and Edge Group (NEX) from 2021 to 2023.

Briers previously served as Chief Financial Officer of UK fintech company Wise for nine years. He joined Wise from Google in 2015, when Wise had about 500,000 customers and was in a loss-making state. In 2021, Wise successfully transformed and listed on the London Stock Exchange, which also set a new size record for listed companies on the exchange in the same period.

03. Conclusion: Infrastructure cannot keep up with the growth of computing power, the key constraint lies in the token generation method

Nowadays, the growth rate of computing power demand has exceeded the expansion capacity of its underlying infrastructure. OLIX's researchers believe that the restricting factor lies in the token generation method.

Generating a token involves multiple operations, each of which has different hardware requirements. To promote the continuous development of AI and make it benefit everyone, it is necessary to rebuild the data center like all large factories, build it into a production line, equip it with dedicated machines, and realize fast and low-cost resource transfer between machines.

This article is from the WeChat Official Account "TechCore", author: Cheng Qian, editor: Panken, published by 36Kr with authorization.