At last week’s GTC conference, Nvidia provided a glimpse of its upcoming AI chips, including the Rubin Ultra GPU, set to launch in late 2026. This next-generation AI accelerator is expected to consume up to 1.8kW per unit and 600kW per rack, a significant jump from previous models. For comparison, the current GB200 draws 1.2kW per unit and 120kW per rack, while the soon-to-be-released GB300 is projected at 1.4kW per unit and 140kW per rack.
To mitigate this sharp rise in power consumption, Nvidia announced that future AI chip setups will integrate co-packaged silicon photonics networking switches (CPO), branded as Spectrum-X and Quantum-X. These switches are expected to significantly cut power consumption from transceivers linking GPUs, potentially saving tens of megawatts per data center.
As a reminder, data movement within data centers is one of the biggest bottlenecks in unlocking full computing potential. Optical transmission not only reduces latency but also significantly lowers energy demands, which account for roughly 25% of a data center’s total power consumption.
As we detailed in a report last week, recent breakthroughs in semiconductor packaging techniques that allow for a smooth integration of photonic and electronic circuits are poised to boost the power efficiency of Silicon Photonics integrated circuits by up to 50%, cut latency, and, importantly, make large-scale production feasible as they can be fabricated using standard manufacturing processes.
Nvidia’s move could accelerate CPO adoption among other chipmakers such as Broadcom and Intel. More insights are expected in the coming days at the Optical Fiber Communications Conference.
While testing companies like Teradyne and optical component suppliers that find the right partnerships such as Lumentum and Coherent (which will provide transceivers for Nvidia) stand to benefit, others could be left behind in an industry increasingly shifting toward fully integrated solutions.
Another key takeaway from Nvidia’s roadmap is that despite these advances in silicon photonics, data center power consumption will continue rising—especially once Rubin Ultra ships in 2026. And over the longer term, as AI progresses from simple prompts to reasoning models requiring exponentially more compute, demand will only intensify.
This suggests that hyperscalers will have no choice but to explore new power supply solutions, including direct utility agreements, fuel cells, and small modular reactors, to prevent bottlenecks.
In conclusion, while the AI and Powering AI trades have been affected by recent macro concerns, we believe there is still a massive infrastructure buildup ahead (both computing and electric power), with even some early signs of spending acceleration in China following the recent capex hikes from Tencent and Alibaba.
