TITLE: Radio Cables Eyed to Overcome Copper Limits in AI Data Centers

AI data centers are pushing the boundaries of copper cabling as demands for faster GPU interconnects intensify. Startups Point2 Technology and AttoTude are developing radio-based alternatives that promise longer distances, lower power use and narrower profiles than traditional copper wires, while avoiding the high costs of optical fiber.

The challenge stems from scaling AI systems, which requires two strategies: scaling out, linking multiple computers over long distances using optical fiber, and scaling up, packing more graphics processing units (GPUs) into a single system for parallel processing. Scaling up demands dense, short-range connections, typically handled by copper cables spanning one to two meters. However, as GPU-to-GPU data rates approach terabits per second, copper encounters physical constraints.

At high speeds, the skin effect confines electrical current to the wire's surface, increasing resistance and signal loss. This necessitates thicker cables and shorter lengths, exacerbating congestion in crowded server racks. Nvidia, a dominant player in AI hardware, plans to increase maximum GPUs per system from 72 to 576 by 2027, amplifying these issues. Industry experts describe this as the 'copper cliff,' where current infrastructure struggles to keep pace.

Radio Waveguides as a Hybrid Solution

Point2 Technology addresses this with e-Tube cables, polymer waveguides that transmit data via radio waves at 90 GHz and 225 GHz frequencies. Each cable comprises eight fibers, each supporting over 200 gigabits per second, for a total of 1.6 terabits per second. Plug-in modules at each end convert electronic signals to modulated radio waves and back.

These cables occupy half the cross-sectional area of a 32-gauge copper cable and extend up to 20 meters without significant degradation—far beyond copper's practical limits at similar speeds. Point2 claims the system uses one-third the power of optical interconnects, costs one-third as much and delivers latency as low as one-thousandth that of photonics. Manufacturing begins later this year, targeting integration into data center racks.

AttoTude pursues a similar approach but employs terahertz frequencies with flexible, slender cables. Both technologies blend copper's simplicity and reliability with optical fiber's slimness and range, positioning radio as a bridge for intra-rack and board-level connections.

Advantages Over Existing Technologies

Copper remains viable for lower-speed, short-haul links, but its limitations grow acute in AI environments. Optical solutions, while excellent for scaling out over hundreds of meters, involve complex photonic chips and lasers that drive up costs and power demands when scaled to dense GPU clusters.

Radio cables sidestep these by leveraging mature semiconductor fabrication for transceivers, avoiding optics' precision alignment needs. Proponents argue this enables direct GPU integration, potentially replacing copper traces on printed circuit boards. Such advancements could simplify cooling, as radio modules generate less heat than dense copper wiring or optical components.

In benchmarks, Point2's design maintains signal integrity over 10 to 20 meters, sufficient for Nvidia's expanded systems. AttoTude's terahertz links aim for even higher bandwidths, with prototypes demonstrating flexibility for routing in tight spaces.

Industry Implications for AI Efficiency

As AI models grow larger, training times depend on efficient data movement. Bottlenecks in scaling up could slow progress in fields like natural language processing and image recognition. By extending reach and reducing power, radio interconnects could lower data center operating costs, which already consume vast electricity for cooling and computation.

The transition aligns with broader efforts to optimize AI infrastructure. While optical fiber dominates inter-rack links, intra-system connections remain a pain point. Radio's potential to 'unclog' racks supports denser deployments, enabling hyperscalers like Google and Microsoft to maximize GPU utilization.

Challenges persist, including regulatory hurdles for higher-frequency radio use and ensuring electromagnetic compatibility in shared environments. However, both startups report progress in shielding and interference mitigation.

Path to Adoption

Point2 plans commercial shipments in 2026, initially for 800-gigabit links before scaling to 1.6 Tbps. AttoTude targets similar timelines, focusing on customizable cable lengths for varied rack designs. Partnerships with chipmakers could accelerate integration, as radio transceivers leverage existing silicon processes.

If successful, these innovations may redefine data center architecture, shifting from copper dominance to a radio-optical hybrid. This evolution is critical as AI workloads demand ever-higher interconnect speeds, projected to reach 3.2 Tbps by 2030.