TL;DR

• Rapid GPU compute density and terabit-bandwidth demands are pushing traditional short-reach copper cables to physical limits in reach, routing thickness, power consumption, and thermal management.
• While optical fiber addresses reach, using it for dense, short-reach GPU-to-GPU connections introduces exponential reliability issues while consuming roughly 10% of total compute cost.
• Point2 Technology offers an e-Tube solution using RF over polymer waveguide to transmit millimeter-wave signals through thin, flexible polymer tubes for 5-10 meter reaches.

# # #

By Sean Park, Founder, President and CEO at Point2 Technology 

GPU and XPU compute density is increasing dramatically, and interconnect bandwidth is rising faster than traditional data center wiring can support. As AI systems pack more devices into each rack and push link speeds into multi‑terabit territory, the physical infrastructure that connects these components is becoming a critical bottleneck. The copper cables that once served short‑reach interconnects reliably are now hitting fundamental limits in reach, thickness, system power consumption, and routability. These are constraints that directly cap how far rack‑level compute can scale.

The Copper Cliff: A Physical Limit on GPU Expansion

Copper twin‑ax has been the default for short‑reach connectivity because it is inexpensive, predictable, and operationally simple. But copper’s physics does not scale with AI’s bandwidth trajectory. As signaling rates push past 200G/lane and into terabit‑class aggregates, copper links must become shorter, thicker, heavier, and more power‑hungry on the system level to maintain acceptable BER and signal integrity.

The NVIDIA NVL72 illustrates this cliff with uncomfortable clarity:

• Over 5,000 copper twin‑ax cables inside one system
• 5 kilometers of copper routed within a single rack‑scale design

As hyperscalers and neo-cloud service providers attempt to expand compute density, adding more GPUs per rack, copper becomes the limiting factor. You cannot increase GPU count if you cannot physically route or install the copper required to connect them.

This is why Point2 Technology (Point2) focuses on the short‑reach domain (5–10 meters). Inside a rack or between adjacent racks, copper is not viable at the bandwidths and densities demanded by future AI clusters.

Why Not Just Use Fiber Optics? The Reliability Curve Breaks at Scale

Optical fiber solves distance and bandwidth but introduces scaling behavior that becomes problematic in dense AI systems.

At modest scale, fiber reliability issues, including link flap, contamination sensitivity, and connector alignment tolerances, are manageable. But as interconnect density increases, especially in AI clusters where scale‑up (more GPUs per rack) is 10× those of scale‑out use cases, optical reliability issues grow exponentially, not linearly.

Engineering teams see:

• Link flap events multiplying with port count
• Contamination sensitivity increases with connector density
• Rising operational overhead for servicing, monitoring, and remediation
• Higher failure rates in high‑vibration, high‑thermal‑flux environments

Any of these interruptions is cascaded downstream and causes delays for multi-million-dollar training runs. Additionally, optical interconnects can represent ~10% of total compute cost in large AI clusters. Fiber is excellent for long‑reach, but for short‑reach GPU‑to‑GPU traffic, its cost, fragility, and scaling behavior are mismatched to the environment.

Point2 e‑Tube: RF Over Waveguide as the Best Copper Replacement

Point2’s e‑Tube technology introduces a fundamentally different short‑reach interconnect: RF over polymer waveguide. Instead of electrons through copper or photons through glass, e‑Tube transmits modulated millimeter‑wave signals through a thin, flexible polymer tube.

Each e‑Tube cable contains eight waveguide cores and delivers:

• 5–10 meter reach with minimal signal loss
• ~1/3 the power of optical solutions
• ~1/3 the cost of comparable fiber links

Critically, e‑Tube preserves the operational characteristics engineers value in copper:

• Copper‑class user experience
• Copper‑class reliability (stable, no reliability issues that comes with optical components)
• Copper‑class economics (built in the same facilities as copper with similar manufacturing techniques)

This combination (scalable bandwidth with copper‑class behavior) makes e‑Tube the strongest candidate to replace copper in the short‑reach domain.

Why Short‑Reach Is the New Battleground

AI racks are becoming thermally and physically constrained. Liquid cooling is now standard, and copper’s bulk and routing complexity intensify those constraints. Short‑reach interconnects determine:

• How many GPUs can fit in a rack
• How cooling loops can be arranged
• How power distribution must be engineered
• How flexible GPU placement can be

If GPUs can be distributed more flexibly, connected by longer, lighter, more efficient RF waveguide cables, hyperscalers can:

• Scale rack‑level GPU counts and compute density
• Avoid copper’s physical limitations (routing congestion, weight, and reach)
• Scale with reliability and economics of copper

Short‑reach is where the copper cliff hits first, and where Point2’s e‑Tube provides immediate relief.

The Road Ahead: From Pluggables to Near-Packaged and Co‑Packaged Form Factors

e-Tube is extending copper experience at different stages, each unlocking new architectural freedom.

1. Pluggable Active RF Cables (ARC) – The first step is already shipping: Point2’s 1.6T OSFP Active RF Cable (ARC). These pluggable cables plug directly into existing OSFP cages in switches and servers, delivering terabit‑class bandwidth with copper‑class reliability and economics. ARC provides an immediate, low‑friction path to break through the copper cliff today.
2. Near‑Packaged Waveguide Routing – The next phase brings e‑Tube RF IOs closer to the GPU modules themselves. Near‑packaged designs reduce connector count, lower power overhead, simplify routing in rack-scale backplane connectivity to increase GPU count in data center racks. This decouples GPU count from copper’s constraints.
3. Co‑Packaged RF Waveguide Interfaces – The long‑term destination is co‑packaged RF IO interfaces, where e‑Tube connects directly at the GPU package level. This eliminates external connectors, minimizes latency, and dramatically improves power efficiency at each hop while enabling new rack‑scale architectures to maximize GPU count and compute inside the same rack

Beyond the 72‑GPU Ceiling

Today’s best‑in‑class systems, such as the NVL72, top out at 72 GPUs per rack, which is a limit imposed not by compute, but by copper. As e‑Tube progresses from pluggable ARC to near‑packaged and co‑packaged designs, it unlocks the ability to scale GPU count beyond 72, enabling larger, more flexible, and more power‑efficient rack‑scale AI systems.

Conclusion

The copper cliff is not theoretical. It is already here in modern AI systems. Fiber solves some problems but introduces others, especially as interconnect density scales. Point2’s e-Tube RF‑over‑waveguide approach offers a balanced, engineering‑sound path forward: scalable performance, copper‑class reliability, and copper‑class economics in the short‑reach domain where AI systems are most constrained.

For data center architects and hardware engineers designing next‑generation AI clusters, the message is clear: Short‑reach interconnects must evolve, and e‑Tube is the technology that unlocks the next wave of GPU expansion.

# # #

About the Author

Sean Park is a seasoned executive with over 25 years of experience in the semiconductors, wireless, and networking market. Throughout his career, Sean has held several leadership positions at prominent technology companies, including IDT, TeraSquare, and Marvell Semiconductor. As the CEO, CTO, and Founder at Point2 Technology, Sean was responsible for leading the company’s strategic direction and overseeing its day-to-day operations. He also served as a Director at Marvell, where he provided invaluable guidance and expertise to help the company achieve its goals. He holds a Ph.D. in Electrical Engineering from the University of Washington and also attended Seoul National University.

The post The Copper Cliff: Why AI Data Centers Need a New Kind of Cable appeared first on Data Center POST.