Olix raises $312M to scale optical interconnects for AI inference

The investment signals a shift in AI infrastructure toward solving the interconnect bottleneck, where moving data between chips currently consumes more energy and time than the arithmetic itself. For streaming platforms and cloud providers, Olix’s rack-scale optical approach offers a potential path to bypass the high-bandwidth memory (HBM) shortages that have throttled AI deployment. As Nvidia and Cerebras also explore deterministic scheduling and specialized memory, the success of Olix’s light-based architecture could redefine the cost-per-token for high-volume inference. Watch for the delivery of the first DX-1 customer systems in H2 2027 to validate performance claims against incumbent GPU-based clusters.

The Olix funding arrives as the AI industry shifts focus from raw compute power to infrastructure-level connectivity. Per Adtek-Fiber (May 2026), the focus of the AI boom is transitioning toward GPU-to-GPU communication and rack density, as connectivity increasingly determines cluster utilization and training efficiency. This has spurred a broader movement toward silicon photonics and co-packaged optics (CPO) to overcome the physical throughput limits of copper wiring. In February 2026, Mesh Optical Technologies launched with a mission to mass-produce American-made transceivers to address these specific infrastructure build-outs, which are expected to scale significantly by 2027.

Olix’s reliance on on-chip static random-access memory (SRAM) instead of high-bandwidth memory (HBM) addresses a persistent market pain point. According to reports from UncoverAlpha (January 2026), HBM suppliers including Micron, Samsung, and SK Hynix have already sold out their capacity for the 2026 calendar year due to overwhelming demand from Nvidia and other ASIC developers. While SRAM offers significantly higher bandwidth than HBM, it typically occupies five to ten times more silicon area, making it difficult to deploy at large capacities. This trade-off is central to the "slow and wide" rack architecture Olix is pursuing to manage 10-trillion-plus parameter models.

The inclusion of Nick McKeown on the board provides Olix with significant B2B pedigree. As a pioneer of software-defined networking (SDN), McKeown’s previous ventures, Nicira and Barefoot Networks, were acquired for $1.26 billion and $6.6 billion, respectively. His involvement, alongside the UK government’s Sovereign AI venture fund participation (per TrendingTopics, August 2026), suggests a strategic effort to build a vertically integrated European alternative to the dominant US-based GPU ecosystem. This mirrors a wider trend of large-scale European semiconductor investments, including a $220 million round for Bristol-based Fractile in May 2026.