Nvidia unveiled an updated data center roadmap at its GPU Technology Conference, revealing a strategy to refresh its AI GPU architecture every two years and its AI accelerator family annually. The presentation highlighted the upcoming 2028 Feynman GPU and Rosa CPU architectures, alongside significant advancements in memory and interconnect technologies. For 2027, Nvidia plans to launch the Rubin Ultra AI accelerators. These chips will feature four compute chiplets and 1 terabyte of HBM4E memory, offering substantial performance gains over current models. These accelerators will pair with Groq LP35 LPUs that support the NVFP4 data format to further enhance efficiency. Additionally, the company will introduce the Kyber NVL144 rack-scale solution, utilizing 144 Rubin Ultra GPUs via an NVLink 7 switch. This configuration promises at least four times the performance of current Oberon NVL72 racks. The 2028 roadmap marks a qualitative leap with the introduction of the Feynman GPU, the Rosa CPU, and the LP40 LPU. CEO Jensen Huang described Feynman as a transformative generation that integrates Groq's hardware capabilities with Nvidia's scale. Key innovations for Feynman include die stacking to boost performance, custom high-bandwidth memory likely exceeding 1 terabyte per package, and support for the LP40 LPU which utilizes the NVLink protocol for system integration. A major shift in the 2028 lineup is the adoption of co-packaged optics for NVLink switches. This technology enables optical interconnections, allowing Nvidia to scale rack solutions to 576 or even 1,152 GPU packages using Oberon and Kyber chassis, respectively. This addresses the physical limitations of copper interconnects and strengthens Nvidia's competitive position against AMD's Instinct and custom hyperscaler accelerators. Simultaneously, Nvidia announced the development of the Rosa CPU, formerly known as Rosalyn. This processor focuses on ultimate single-thread performance and represents a strategic move to shorten the CPU development cycle from four years to two. By accelerating its release cadence, Nvidia aims to align its CPU innovation speed with industry leaders like AMD and Intel. The Rosa CPU will connect with the new BlueField-5 DPU and the ConnectX-10 SuperNIC, all supported by 7th Generation SpectrumX Ethernet with co-packaged optics. The roadmap demonstrates Nvidia's commitment to both quantitative improvements in 2027 through increased GPU density and qualitative breakthroughs in 2028 via new architectures and optical networking. These developments signal an aggressive push to maintain dominance in the AI data center market as demand for massive computational power continues to grow.