OLIX has raised $312 million in a Series B funding round at a $3.3 billion valuation, two years after the London-based artificial intelligence infrastructure company was founded.
The financing included participation from Fundomo, Arm and Hudson River Trading, along with angel investors including Netflix co-founder Reed Hastings.
Existing investors Hummingbird Ventures, Crane, Plural, Creandum, Phoenix Court and Transition also increased their commitments to OLIX.
The funding will support OLIX’s effort to deliver its first DX-1 inference systems to customers during the second half of 2027. The company will also expand the broader custom silicon platform behind DX-1 and make the manufacturing and supply chain commitments required to produce advanced inference hardware at scale.
OLIX is hiring across silicon, photonics, compiler and systems engineering. The company is expanding its teams in London, Bristol, Austin, Toronto and San Francisco.
Alongside the financing, OLIX appointed Professor Nick McKeown to its board of directors. McKeown is a co-inventor of software-defined networking, OpenFlow and the P4 programming language.
McKeown is Professor Emeritus of Computer Science and Electrical Engineering at Stanford University and received the 2025 Marconi Prize. He also co-founded networking companies including Nicira, which VMware acquired, and Barefoot Networks, which Intel acquired.
Following Intel’s acquisition of Barefoot Networks, McKeown led Intel’s networking business. OLIX expects his experience in networking, programmable infrastructure and company building to support the development of its rack-scale AI systems.
OLIX also appointed former Wise executive Matt Briers as chief financial officer.
Briers served as Wise’s CFO for nine years after joining the financial technology company from Google in 2015. At the time, Wise had approximately 500,000 customers and was operating at a loss.
He built the finance organization that supported Wise’s transition from a private company to the public markets while the business remained profitable. Wise completed a direct listing on the London Stock Exchange in 2021 at a valuation of £8.75 billion, or approximately $12 billion.
OLIX is developing hardware intended to address what it sees as the efficiency limits of using general-purpose chips for AI inference.
The company describes a data center running AI models as a factory that produces tokens. Generating each token requires hundreds of different operations, with each stage placing different demands on the underlying hardware.
Current AI systems generally run those stages on the same type of general-purpose processor. OLIX believes performance and cost can be improved by assigning different stages of token production to specialized chips designed for their particular workloads.
Its X-1 platform distributes, or fully unrolls, AI models across a large number of chips. The architecture operates as a production line in which each chip focuses on a particular portion of the model.
The chips retain a flexible compute fabric rather than permanently embedding the architecture of a particular AI model. This is intended to allow the system to adapt as model architectures continue to change.
The X-1 platform also uses what OLIX calls a “slow and wide” optical interconnect. The system moves data directly between chips using light instead of copper, with the goal of reducing latency and energy consumption.
OLIX said the optical system is made possible through rack-scale co-design across the entire connection. A deterministic compiler schedules workloads across the racks.
The company believes this approach could reduce the cost of running existing frontier AI models while enabling the deployment of substantially larger and more capable models in the future.
DX-1 is the first chip OLIX is developing for the X-1 platform. The decode accelerator is designed for the inference stage during which an AI model reasons and generates its output.
For models with approximately 100 billion parameters, OLIX said DX-1 can simultaneously deliver more than 10,000 tokens per second per user and higher output-token throughput per watt than general-purpose processors operating with large batch sizes.
OLIX said the architecture can also scale to models with 10 trillion parameters or more through its multi-rack scale-up domain design.
DX-1 stores the model in fast on-chip static random-access memory, or SRAM, to improve energy efficiency and reduce latency.
The chip does not require advanced packaging or high-bandwidth memory, two components facing significant supply constraints across the AI hardware industry. OLIX designed the system this way to improve its ability to expand production despite broader supply chain shortages.
Founded in London in 2024, OLIX develops complete rack-scale systems for demanding AI inference workloads. The company designs the chips, lasers and networking technology used throughout its platform.
