Olix plans to ship its first photonic chips in 2027, targeting AI inference workloads as the infrastructure market moves beyond general-purpose GPUs. The startup joins a wave of custom silicon companies raising capital to address latency and efficiency bottlenecks in enterprise AI deployments.
Nio's semiconductor unit GeniTech closed a $330M Series A in February 2026 for autonomous driving chips. The fundraising surge reflects enterprise demand for application-specific processors as AI valuations reach historic levels—OpenAI at $840B and Anthropic at $380B.
Photonic chips use light instead of electricity to transmit data, reducing power consumption and heat generation in data centers. Early photonic systems achieved 10x improvements in energy efficiency for matrix multiplications, the core operation in neural network inference. Olix's 2027 timeline positions it to capture demand from hyperscale operators expanding AI capacity.
Language Processing Units represent another specialization path, using SRAM-centric designs to accelerate transformer models. These chips store model weights in on-chip memory, eliminating the memory bandwidth constraints that throttle GPU performance on large language models. SRAM-based architectures deliver 50-100x lower latency for inference compared to GPU solutions.
HPE is integrating specialized accelerators into its enterprise infrastructure offerings. The company's AI systems combine custom silicon with advanced packaging techniques that stack memory dies directly on processing units. This 3D integration reduces data movement, the primary energy cost in AI workloads.
Advanced packaging technologies enable chipmakers to combine photonic components with traditional CMOS logic on a single substrate. TSMC's CoWoS and Intel's EMIB platforms support these heterogeneous designs, creating silicon photonics integration paths that didn't exist three years ago.
The infrastructure buildout accelerates as foundation model providers race to deploy real-time AI applications. Inference represents 80% of AI compute costs in production, creating a $50B+ market for specialized processors by 2028. Custom silicon targeting specific neural network architectures captures this demand more efficiently than general-purpose alternatives.
Autonomous vehicle chips face similar specialization pressures, requiring real-time sensor fusion and perception processing. GeniTech's funding validates the autonomous driving semiconductor market as distinct from datacenter AI, with different latency, power, and safety requirements.