潛在影響GB200 NVL36/72出貨的短期因素分析 / Analysis of short-term factors potentially affecting GB200 NVL36/72 shipments

AI伺服器未來數年的成長趨勢仍正向，這邊僅提出潛在影響GB200 NVL36/72出貨的短期因素以供思考。當相關股票已大部分反映GB200 NVL36/72對供應鏈貢獻之樂觀預期並大漲之際，至少短期的投資決策就需要考慮更多層面而非僅依靠供應鏈調查。

1. 根據集邦科技預測，AI伺服器佔整體伺服器出貨量在2024與2025年的出貨占比分別約8.8%與12.2%。GB200 NVL36的每個機櫃耗電約80kW，而根據AMAX今年四月的調查，目前全球少於5%的資料中心可以支援每機櫃50kW伺服器。所以，購買GB200 NVL36前，需先確保有沒有足夠空間安裝。
2. GB200 NVL72的單一機櫃版本，每機櫃耗電130kW，短期內無法量產。雖有提供雙NVL36機櫃的NVL72方案，但NVL72提案時對客戶的關鍵賣點在於用最小的空間提供最大的AI算力，故須觀察更佔空間的雙NVL36機櫃方案能否得到客戶青睞。
3. Blackwell Ultra與Vera Rubin分別在3Q24與1H25投片，客戶在可見未來會不會改買效能更好/更划算的Blackwell Ultra或Vera Rubin AI伺服器，而影響GB200的短期需求？如果客戶改買Blackwell Ultra與Vera Rubin AI伺服器，Nvidia無疑還是最大贏家，但供應鏈受益者跟現在的GB200還是同樣的廠商嗎？
4. 採用Blackwell晶片的AI伺服器有許多機型，不只GB200 NVL36/72，如果客戶因AI算力需求先購買其他先出貨的Blackwell的AI伺服器機型 (如GPU少於NVL36的GB版本，或x86方案)，短期內會否影響GB200 NVL36的需求？
5. GB200 NVL36的算力優勢無庸置疑，但也面臨許多前所未見的設計與生產挑戰，故能否確保如期大量出貨 (市場共識為9–10月)？以組裝而言，目前L10與L11的EVT均約8週 (耗時最久分別是可靠性驗證與安全測試流程)，均較iPhone的組裝EVT短。AI伺服器對系統開發要求理應顯著高於消費電子，如此緊湊的開發時程能否解決前所未有的設計與生產挑戰並確保如期量產？

The growth trend for AI servers will remain positive in the coming years. Here, I present some potential short-term factors that could impact the shipment of the GB200 NVL36/72 for consideration. As related stocks have already reflected the optimistic expectations of the GB200 NVL36/72’s contribution to the supply chain and have surged significantly, short-term investment decisions should consider more aspects rather than relying solely on supply chain investigations.

1. According to TrendForce, AI servers are expected to account for approximately 8.8% and 12.2% of total server shipments in 2024 and 2025, respectively. The GB200 NVL36 consumes 80kW per rack, but according to a survey by AMAX in April this year, less than 5% of data centers worldwide can currently support servers consuming 50kW per rack. Therefore, before purchasing the GB200 NVL36, it is necessary to consider whether there is enough space for installation.
2. The single-rack version of the GB200 NVL72 consumes 130kW per rack and can’t go to mass production in the short term. Although there is a dual NVL36 rack version of the NVL72, the key selling point of the NVL72 proposal to customers is to provide maximum AI computing power in the smallest space. Therefore, it remains to be seen whether the dual NVL36 rack solution of NVL72 that takes up more space will appeal to customers.
3. With Blackwell Ultra and Vera Rubin slated for tape-out in 3Q24 and 1H25, respectively, will customers in the foreseeable future opt for the more powerful and cost-effective Blackwell Ultra or Vera Rubin AI servers, affecting the short-term demand for the GB200? Suppose customers switch to Blackwell Ultra and Vera Rubin AI servers. In that case, Nvidia will undoubtedly still be the biggest winner, but are the supply chain beneficiaries the same as those currently benefiting from the GB200?
4. Many AI server models use Blackwell chips, not just the GB200 NVL36/72. If customers purchase other Blackwell AI server models (such as those with fewer GPUs than the NVL36 or x86 solutions) due to AI computing demand, will this impact the short-term demand for the GB200 NVL36?
5. While the computing power advantage of the GB200 NVL36 is undeniable, it also faces unprecedented design and production challenges. Can it be ensured that mass shipments (market consensus is September to October) can be achieved on schedule? Currently, in terms of assembly, the EVT for L10 and L11 takes about eight weeks, with the longest processes being reliability testing and safety testing, both shorter than the assembly EVT for iPhones. The system development requirements for AI servers should be significantly higher than for consumer electronics. Can such a tight development schedule solve unprecedented design and production challenges and ensure mass production on schedule?