AI Data Centers Demand More Than Copper Can Deliver

Why It Matters

Radio‑based interconnects could unlock higher GPU density and faster AI training while reducing power and cooling costs, reshaping data‑center architecture.

Key Takeaways

• Copper hits "skin effect" limit at terabit speeds
• Radio‑frequency cables offer 10‑20 m reach with lower power
• Point2’s 1.6 Tb/s ARC costs and latency far below optics
• AttoTude targets terahertz waveguides for flexible, low‑loss links
• Direct‑GPU radio transceivers could simplify cooling and packaging

Pulse Analysis

Scaling AI workloads demands two complementary strategies: adding more machines (scale‑out) and packing more GPUs into each node (scale‑up). While optical fiber handles long‑haul scale‑out links, the dense, short‑range connections inside racks still rely on copper. At data rates approaching a terabit per second, copper suffers from the skin effect, forcing wires to thicken and power consumption to rise—an untenable path as Nvidia plans to increase GPU counts per system eightfold by 2027.

Enter radio‑frequency interconnects. Point2 Technology’s e‑Tube cable uses eight polymer waveguides to transmit 1.6 Tb/s across 10‑20 meters, consuming roughly one‑third the power of comparable optics and offering latency an order of magnitude lower. AttoTude pursues a terahertz‑band approach, leveraging flexible waveguides that lose only 0.3 dB per meter and can carry hundreds of gigabits per second. Both solutions retain copper’s low cost and manufacturing simplicity while achieving the reach and form factor traditionally reserved for fiber, positioning them as viable successors to active electrical cables.

The broader market impact could be profound. By moving high‑speed links closer to GPUs, radio cables reduce the need for aggressive liquid cooling and simplify PCB layouts, potentially enabling co‑packaged transceivers that bypass the optical alignment challenges of photonics. With venture backing and partnerships with established connector makers like Molex and Foxconn, Point2 and AttoTude are poised to challenge entrenched copper and optical ecosystems. If adopted at scale, these technologies may become the new standard for AI‑centric data centers, accelerating training cycles and lowering total cost of ownership.