从百美元到数千美元，AI互联为什么越卖越贵？

TL;DR - Astera Labs argues that high-speed interconnects are becoming a core determinant of AI system performance and per-token cost as clusters scale to thousands of GPUs. Its strategy is expanding from PCIe retimers into higher-value switching, memory, copper, and optical connectivity products.

• Interconnect design is increasingly governed by bandwidth, GPU scale, power, latency, and total token economics—not physical distance alone.
• Copper should remain cost-effective around 200G per lane, while optical links are expected to penetrate racks more deeply at 400G per lane and larger GPU scales.
• Astera expects near-package optics in scale-up systems around 2027, while broad co-packaged-optics deployment may not occur until 2029–2030 or later.
• Larger scale-up domains could reduce traffic through higher-latency scale-out networks; meanwhile, inference growth and increasingly customized AI systems should diversify connectivity requirements.