Point2 Technology, a startup developing a new class of RF-based interconnect for artificial intelligence data centers, has closed a $136 million Series B financing round. The funding, led by LB Investment with new participation from Arm and continued support from Maverick Silicon, brings together a roster of strategic investors that includes Nvidia, UMC Capital, Molex, and Bosch Ventures. The company is betting that its e-Tube platform—which transmits signals using radio frequency over plastic waveguides—can replace both copper cables and optical transceivers in the tightly packed world of rack-scale AI compute.
The Wiring Bottleneck
Modern AI systems are built on massive clusters of GPUs and custom accelerators. Training a frontier model involves synchronizing tens of thousands of processors, each exchanging data with its neighbors at speeds measured in terabits per second. The interconnect—the wiring that carries those exchanges—has become as critical as the compute itself. When chip designers talk about scaling AI infrastructure, they are increasingly constrained not by transistor counts or memory bandwidth, but by the physical links between devices.
Copper has been the workhorse of data center networking for decades. It is inexpensive, reliable, and consumes very little power at short distances. But as data rates climb, copper's reach shrinks dramatically. At 112 Gbps or 224 Gbps per lane, a well-engineered copper cable may only carry a signal 1 to 2 meters before attenuation becomes a problem. In a rack where switches and accelerators are spread across several meters, that limitation forces architects to add repeaters, active cables, or more complex topologies—all of which add cost, power, and reliability concerns.
Optical interconnect solves the reach problem. Light can travel hundreds of meters without degradation, and modern optical modules are fast enough to keep up with the highest-speed serdes. But optics come with their own tradeoffs. Every optical link requires a laser driver, a photodetector, a TIA, and a DSP to encode and decode the signal. These components draw significant power and generate heat, which is a painful penalty in power-constrained data centers. Lasers also age, shift wavelength, and fail with enough frequency that hyperscalers budget for ongoing replacements. And the latency added by optical conversion—while measured in nanoseconds—can compound across thousands of hops.
The e-Tube Approach
Point2's e-Tube platform is designed to avoid both sets of problems. Rather than pushing electrons through copper or photons through glass, the technology uses RF signals guided by a plastic waveguide. The waveguide is a dielectric tube—hence the name e-Tube—that confines electromagnetic waves and carries them from transmitter to receiver without the need for a continuous metallic conductor.
The company says the approach delivers a 10x reach improvement over copper at a comparable cost point. That reach means servers in adjacent racks can be connected directly, without optical conversion or active retiming. The cables are also 5x lighter and take up 2x less volume than copper equivalents, which matters significantly in data centers where airflow, cable trays, and physical space are at a premium.
Compared to optics, Point2 claims 3x lower power, 3x lower cost, and 1,000x lower latency. That last number is striking, but it refers to the elimination of the serdes-to-optical conversion chain. Because the RF signal is generated and received directly by CMOS transceivers, there is no need to drive a laser or synchronize a photodetector. The reduction in analog and digital processing steps translates into a dramatic drop in per-link latency.
Reliability is another factor. Optical modules have a mean time between failures (MTBF) that is often limited by laser degradation, particularly in high-temperature environments. RF over plastic waveguides has no laser to burn out, no fiber to break, and no dirt to accumulate on a lens. The passive waveguide is essentially a piece of plastic cabling, which can be manufactured at scale with similar processes used to make standard dielectric materials.
Strategic Investors
The composition of Point2's Series B is notable for the breadth of the industrial network it represents. Nvidia is the dominant supplier of AI accelerators, and its interest in Point2 signals that interconnect is a constraint it wants to see solved. Arm is the architect behind nearly every mobile and embedded CPU, and increasingly the core technology in data center processors; its venture arm's participation gives Point2 access to a broad ecosystem of system-on-chip designers. Molex is a major manufacturer of connectors and cabling solutions, while Bosch Ventures brings an automotive perspective, where radar and RF technology are already widely used.
That diversity is a strong vote of confidence in the underlying physics. If the e-Tube were a niche research project, it is unlikely that such a wide range of companies would invest. The fact that chipmakers, interconnect suppliers, and automotive sensor companies all see value in RF-based waveguide transmission suggests the technology has the potential to become a standard building block.
Lumilens, an optical interconnect startup, raised $700 million earlier this month to push photonic solutions for AI. That indicates the market is large enough to support multiple competing approaches. But Point2's positioning is distinct: it is aiming at the middle distance—between a few hundred meters and a few meters—where copper is too short and optics is overkill.
Commercialization Roadmap
The new funding will accelerate the rollout of three product lines. The first is an Active RF Cable, a pluggable cable assembly that can replace existing direct-attach copper cables (DACs) at longer reaches. The second is a near-packaged e-Tube solution, in which the waveguide is connected to a module that sits close to the ASIC. The third is a co-packaged e-Tube, where the RF transceiver is integrated directly into the same package as the compute die, minimizing signal path length and power.
Co-packaging is particularly interesting for AI accelerators. In co-packaged optics, the optical engine is placed adjacent to the GPU or switch chip to eliminate the electrical trace to a front-panel module. Point2 is applying the same idea to RF. By embedding a tiny RF transceiver in the package and routing signals over an e-Tube that connects to the backplane, the technology could offer the bandwidth and reach of optics but with the power efficiency of copper.
CEO Sean Park emphasized that interconnect has become the defining bottleneck in AI systems. "As AI systems scale and bandwidth demands reach terabit-per-second speeds, interconnect has become the defining bottleneck," he said. His comment reflects a growing consensus among data center architects that simply adding more processors is no longer sufficient. The network fabric must evolve in step with compute density.
Arm's Paul Williamson, SVP of strategic ventures, framed it as a full-stack challenge: "The next generation of AI infrastructure will rely on advances across the entire technology stack, from compute and memory to interconnect." Arm's presence in the funding round suggests that chip designers will need to account for the physical layer in their architecture decisions, switching from designing on-chip interconnects to designing for an integrated RF front-end.
Power and Economics
The economic case for RF versus optics is compelling. Optics require expensive lasers, modulators, and photodetectors, as well as the DSP to compensate for signal impairments. Point2's RF approach uses standard CMOS transceivers, which are already optimized for high-yield and low cost. The waveguide itself is a simple plastic extrusion, which is far cheaper than a pair of optical fibers with precision-polished connectors.
Power savings are perhaps the most critical factor. A modern data center spends a significant fraction of its electricity on networking equipment and the cooling needed to remove waste heat. Reducing interconnect power by 3x means more of the facility power budget can go directly to compute. That can mean adding more GPUs per rack or reducing the number of cooling units needed. In a hyperscale facility, a 3x power savings across hundreds of thousands of links could translate into a meaningful reduction in total cost of operation.
Weight and volume also matter for mechanical design. A large AI cluster can have tens of thousands of cables between nodes. Copper cables at high data rates are thick and heavy, and they exert strain on connectors and cable trays. Reducing weight by 5x allows denser packaging and simpler cable management, which directly improves serviceability and airflow. The smaller cross-section of the e-Tube also enables higher cable counts within the same cable tray width.
Scalability and Adoption
The key question for Point2 is whether the e-Tube technology can scale to the volumes hypercalers need. The company has not disclosed manufacturing details, but the waveguide is a fundamentally simpler component than an optical module. It can be cut to length and terminated with connectors in a process similar to that used for standard coaxial or RF cables. Point2 may need to work with established cable manufacturers to produce at scale, but the raw material cost is low.
Adoption will also depend on standards. Data center interconnect relies on well-defined interface standards such as QSFP-DD, OSFP, and CFP. For a new transmission medium to be accepted, it must be packaged in a form factor that plugs into existing switch and server interfaces. Active RF Cables that mimic DAC form factors could be adopted without changes to the host system. Near-packaged and co-packaged solutions will require more integration effort, but any standard that allows multiple vendors to interoperate could accelerate adoption.
Another consideration is signal integrity at higher frequencies. At terabit-per-second speeds, every discontinuity in a waveguide can cause reflections and crosstalk. Point2 has not published detailed specs on connector loss, skew, or electromagnetic interference, but the company's investor list suggests that those engineering problems are manageable. The involvement of Molex, a company that specializes in high-speed connectors, adds credibility to the claim that the physical interface can be produced with the necessary tolerances.
The competitive landscape will be intense. Optical interconnect is not standing still. Co-packaged optics are being developed by many players, and the cost of optical components continues to decline. But Point2's value proposition is not about beating optics at long distances. It is about offering a superior alternative in the rack-scale domain—the place where most AI compute actually happens. If the e-Tube can deliver the claimed 10x copper reach and 3x lower optic power in volume, it could become the default interconnect for the next generation of AI data centers.