Scaling advanced AI models ranging from DeepSeek V4 and Kimi K2.6 to the latest NVIDIA Nemotron series, requires infrastructure capable of moving unprecedented volumes of data with minimal latency. As model parameters reach the trillions and users as well as agents want tokens generated instantly, the physical limitations of traditional copper wiring is fast becoming a primary constraint on performance. Pushing electrical signals at 200 gigabits per second across circuit boards results in signal degradation and high power consumption. To sustain the trajectory of AI capabilities in the world of power-constrained grids, infrastructure providers must look to transition to optical interconnects and silicon photonics for greater efficiency.
With the introduction of the NVIDIA Vera Rubin Ultra architecture at GTC 2026, Co-Packaged Optics (CPO) moves from a theoretical option to a production necessity. By integrating optical conversion directly alongside the compute silicon through heterogeneous integration, CPO enables the bandwidth and energy efficiency required for gigawatt-scale AI deployments. This blog post examines the architectural shift toward optical networking, the engineering challenges involved, and how NVIDIA Rubin establishes a new foundation for hyperscale AI workloads.
Rubin Ultra is Breaking the Copper Barrier
NVIDIA’s interconnect strategy for its upcoming platform, Vera Rubin Ultra can be summarized as: use copper where you can, and optics where you must.
For single-rack deployments, NVIDIA is pushing copper to its absolute physical limits. The Rubin Ultra NVL144 (housed in the new "Kyber" rack) connects 144 GPUs using an all-copper scale-up network. By utilizing dense flyover cables and extreme mechanical proximity, NVIDIA manages to keep data moving electrically within the rack without major signal loss.
But agentic AI requires clusters far larger than a single rack. To create the Rubin Ultra NVL576, a massive 576-GPU AI supercomputer spanning eight racks, copper simply cannot bridge the physical distance fast enough. Driving high-speed electrical signals across racks consumes much more power and introduces unacceptable latency. Co-packaged optics for scale-up networking between racks is a huge factor in making the NVL576 configuration a reality.
What the Shift From Pluggable to Co-Packaged Optics Means
Historically, optical networking relied on pluggable transceivers. Data traveled electrically from a switch ASIC across a motherboard to the front panel, where a pluggable module converted the electrical signal into light. At the extreme bandwidths required by Rubin Ultra (1.6 Tb/s to 3.2 Tb/s per port), this legacy architecture breaks down.
Co-Packaged Optics (CPO) fundamentally rewrites this design. By integrating the photonic integrated circuits (PICs) and electronic integrated circuits (EICs) directly onto the same advanced package substrate as the switch ASIC, CPO eliminates the long electrical traces entirely.
In the context of Rubin Ultra, the benefits of CPO are significant:
- Energy Efficiency: CPO cuts interconnect power consumption by up to 3.5x, achieving sub-5pJ/bit energy efficiency. NVIDIA's CPO implementations reduce per-port power consumption to just 9W compared to 30W from pluggable optics.
- Bandwidth Density: By replacing bulky electrical I/O pins with optical fiber arrays, CPO enables massive aggregate bandwidths of up to 409.6 Tb/s on a single switch, allowing the NVL576 to maintain a non-blocking, all-to-all fabric across eight racks.
- Resiliency: Reducing the number of discrete active components, along with eliminating failure-prone transceivers, increases uptime and operational reliability by 10x.
Solving the Thermal Challenge for NVL576
While the benefits of CPO enable the NVL576 to exist, manufacturing it at scale is a complex multi-physics challenge.The primary hurdle is thermal reliability. Photonic components, particularly the lasers and microring resonators used to generate and modulate light, are incredibly sensitive to temperature fluctuations. Too much heat can shift optical wavelengths, degrade laser performance, and cause catastrophic system failures.
In a co-packaged optics architecture, the delicate optical engines are placed millimeters away from switch ASICs that can dissipate 400W to 800W+ of heat. Furthermore, the Rubin Ultra NVL576 pushes an estimated 600 kW total power footprint. Solving this thermal crucible requires a revolution in advanced packaging and materials science:
- Next-Gen Materials: The industry is rapidly adopting high-thermal-conductivity adhesives, low-CTE (Coefficient of Thermal Expansion) underfills, and phase-change thermal interface materials to manage the extreme thermal crosstalk and mechanical stress between the silicon and the optics.
- Advanced Cooling: Liquid-cooled cold plates and micro-channel heat sinks are no longer optional; they are mandatory to maintain the strict temperature control loops required by the optical engines.
- External Laser Sources (ELS): To protect the most heat-sensitive component, the laser, many CPO designs decouple the laser from the package entirely, placing it on the front panel where it can be easily cooled and replaced, while piping the raw light into the CPO package via fiber.
What Rubin Ultra Integration Means for Facility Design
The thermal and optical engineering challenges of Vera Rubin Ultra do not stop at the chip package. They propagate outward to every layer of the physical facility. For infrastructure operators, this means four distinct and deeply interconnected challenges.
Power Density Per Rack
The NVL576's power requirements represent a categorical break from prior generations. Average rack density across the industry has more than doubled in just two years and is projected to keep climbing, yet the NVL576 sits at a level that dwarfs even the most aggressive current projections. Most colocation facilities built in the last decade were designed around a fraction of this power draw. Deploying Rubin Ultra is not a matter of provisioning more breakers; it requires a wholesale reimagining of power distribution architecture, including NVIDIA's new 800 VDC facility-to-rack delivery model, which demands new transformers, switchgear, and busbar infrastructure.
Liquid Cooling Requirements and Lead Times
At this power density, air cooling is not a consideration. The Vera Rubin NVL72 already moves to fully liquid-cooled compute trays, with dedicated cold plates on every module, internal manifolds, and liquid-cooled busbars managing the extreme current loads. The operational challenge is as much logistical as technical. The supply chain for liquid cooling infrastructure, including coolant distribution units, manifolds, and facility-level piping, cannot ramp as quickly as GPU supply. Operators who have not already begun facility preparation will face meaningful deployment delays regardless of when their hardware allocation arrives.
Structural Load Implications
A fully populated liquid-cooled rack of this generation can weigh well in excess of two metric tons when coolant, cabling, and compute modules are accounted for. Most legacy data center environments were simply not engineered for this. Concrete slab assessments, raised-floor removal, and point-load calculations will be prerequisites for any operator seeking to deploy Rubin Ultra in an existing facility. Taken together, these requirements explain why NVIDIA frames Rubin Ultra not as a product to be purchased, but as an infrastructure program to be planned
NVIDIA is Pushing the Scaling Envelope with Kyber NVL1152
While Vera Rubin Ultra introduces CPO to the data center, the subsequent Feynman generation will push this technology even higher with the Kyber NVL1152. Slated for 2028, this system represents the realization of NVIDIA's optical interconnect strategy.
The NVL1152 is constructed by linking eight high-density Kyber racks together. Connecting 1,152 GPUs in a single, coherent, all-to-all network domain is a feat that copper wiring simply cannot achieve. By utilizing NVLink switches equipped with co-packaged optics, the NVL1152 allows every GPU in the massive cluster to communicate with any other GPU at maximum bandwidth and near-zero latency.
This architecture effectively blurs the line between a single server and a data-center-scale supercomputer. For enterprises delivering low-latency inference at scale, the Kyber NVL1152 means the ability to produce tokens without the traditional penalties of network overhead.
Beyond Rubin Ultra: The Feynman Horizon
The transition to CPO is being aggressively accelerated by NVIDIA's roadmap. Recognizing that photonics is the next gating factor for AI scaling, NVIDIA has made a $4 billion strategic investment in optical connectivity providers such as Lumentum and Coherent to secure manufacturing capacity.
Vera Rubin Ultra is the critical bridge in this transition—utilizing copper in-rack and CPO out-of-rack. But the generation following Rubin, codenamed Feynman (2028), will be the true optical inflection point. Feynman will feature NVLink switches with co-packaged optics, enabling native optical NVLink scaling across the system (up to NVL1152), fundamentally blurring the lines between a single server and a data-center-scale computer.
NVIDIA's Optical Interconnect Roadmap
To understand how this transition unfolds, we can look at NVIDIA's projected scale-up interconnect roadmap across its upcoming generations. Jensen Huang announced at GTC 2026 that NVIDIA would be offering both copper and optical connections for scale up for both Rubin and Feynman.
Source: SemiAnalysis
What Comes Next for CPO
Vera Rubin Ultra is expected to be available in the second half of 2027. The integration of Co-Packaged Optics in the Vera Rubin Ultra architecture is merely the first phase of a broader transformation. While current implementations focus on scale-out network switches, the trajectory of CPO points toward even deeper integration. The next frontier involves embedding optical I/O directly onto the compute silicon itself.
As the industry moves toward the Feynman generation and beyond, we can expect optical connections to replace copper entirely, even for short-reach, intra-rack communication. This will enable much higher bandwidths directly from the processor package, drastically reducing latency and power consumption. Furthermore, advancements in materials science and automated photonic packaging are expected to drive down costs, making CPO the default standard for AI computing.
Ultimately, the shift to light will likely dissolve the traditional boundaries between servers, racks, and clusters. By eliminating the penalties of data movement, CPO will allow entire gigawatt data centers to operate as singular, optically unified compute engines.
Deploy Next-Generation AI in Radiant's AI Factories
Deploying the next generation of AI requires more than just procuring advanced compute. It demands a fundamental reimagining of physical infrastructure. Legacy hyperscale data centers are simply not equipped to handle the power footprint, specialized liquid cooling loops, and structural reinforcements required by systems like the Rubin Ultra NVL576.
This physical reality is the driving force behind Radiant. By uniting deep capital and data center expertise with a powerful AI cloud, Radiant delivers a true end-to-end AI infrastructure platform. We provide the gigawatt-scale power, purpose-built facilities, and intelligent software layer necessary to support the Co-Packaged Optics era. Step into next-generation AI with Radiant’s cutting-edge AI factories.
