今天有許多前輩提出 2025 NVIDIA GTC Preview，重點在於 CPO與 GB300 的推出。CPO 主要與 T 公司 合作，因此今年將推出採用 Hybrid Bonding 技術結合的 PIC（Photonic Integrated Circuit） 與 EIC（Electronic Integrated Circuit） 矽光晶片。NVIDIA 的 CPO 產品預計 2026 年才會推出，目前光通訊市場仍由 AVGO 主導。因此，在良率的關鍵環節 FAU（Fiber Array Unit） 方面，我認為 Browave 相較於 FOCI 更具優勢。根據 WT 研究，後者尚未獲得市場採納，儘管 EC（Edge Coupling） 是一種理想的方法，但前提是能夠成功量產。在短期內，GC 仍是較佳的解決方案。 GB300 於 2025 年 2 月初陸續投片，基本上在 2025 Q4 會少量出貨，而 Hon Hai 可能成為最大贏家。 近期市場也討論到 NVL288，但目前我沒有太多看法。我認為短期內無法實現 Blade Server 形式的設計，而 Rubin 的架構與我先前提及的概念相符（wt6191.substack.com/p/r…）。目前一個 Rubin 內含 3 顆 die，因此 TDP 預估約 2,000W，可能透過2個 NVL72 機櫃串聯，再透過話術表達其擁有 288 顆 GPU。不過，由於 Rubin 尚未 tap-out，所有規格仍有可能變動。

Many industry experts recently shared insights on the 2025 NVIDIA GTC Preview, with a focus on the launch of CPO (Co-Packaged Optics) and GB300. CPO is primarily developed in collaboration with T Company, and this year, they plan to introduce silicon photonic chips integrating PIC (Photonic Integrated Circuit) and EIC (Electronic Integrated Circuit) using Hybrid Bonding technology. NVIDIA's CPO products are expected to be released in 2026, while the optical communication market is still dominated by AVGO.

Regarding FAU (Fiber Array Unit), which is a critical factor in yield improvement, I believe Browave holds a stronger position compared to FOCI. According to WT Research, the latter has not yet been widely adopted in the market. Although EC (Edge Coupling) is a promising approach, successful mass production remains a prerequisite. In the short term, GC remains the more practical solution.

GB300 has started wafer production in early February 2025, with limited shipments expected in 2025Q4. Hon Hai is likely to be the biggest winner.

Recently, there has been market discussion about NVL288, but I don’t have many thoughts on it at this stage. I believe that achieving a Blade Server design in the short term will be challenging. The Rubin architecture aligns with my previous insights (wt6191.substack.com/p/r…). Currently, a single Rubin contains three dies, resulting in an estimated TDP of around 2,000W. It may be possible to interconnect two NVL72 racks and then market it as a system containing 288 GPUs. However, since Rubin has not yet taped out, all specifications remain subject to change.