英伟达AI机架转型加速；TrendForce称NVL72产值2027年将超7100亿美元
TrendForce报告显示，GB300机架级解决方案将成为英伟达2026年出货的主要驱动力，需求预计持续强劲。NVL72机架产值预计2027年超过7100亿美元。该预测涉及AI服务器及机架产品，与元器件供应链相关，但为未来预测。
原文深度解读
TrendForce于2026年8月28日发布的AI服务器产业报告指出，英伟达的AI机架过渡加速，NVL72机架产值预计在2027年超过7100亿美元。报告涉及GB300、VR200、VR300等机架产品，以及客户结构变化和芯片竞争扩展到数据中心基础设施。关键数字包括NVL72出货量2027年同比增长超50%，产值同比增长214%，北美CSP采购份额从2025年约70%降至2026年60%，英伟达数据中心业务占营收93%，OpenAI租赁8 IT-GW容量，英伟达支付上限1050亿美元，谷歌承诺440亿美元。该报告与电子元器件供应链相关，涉及AI服务器、机架、HBM、晶圆等上游组件。
- TrendForce报告显示，GB300机架级解决方案将成为英伟达2026年出货的主要驱动力，需求预计持续到2027年上半年。
- 下一代VR200机架将在2027年增加产量并占据最大出货份额，而Rubin Ultra平台（VR300）因HBM配置和服务器架构仍在最终确定，份额相对较小。
- TrendForce估计，包括GB和VR平台在内的NVL72机架出货量在2027年将同比增长超过50%。
- Vera Rubin系统的平均售价预计约为GB300的两倍，晶圆和HBM等关键上游组件价格可能上涨，导致GB300、VR200和VR300 NVL72机架系统总产值在2027年超过7100亿美元，同比增长214%。
- 五大北美CSP仍是主要买家，但其采购份额预计从2025年约70%降至2026年60%，原因是特斯拉（包括xAI和SpaceX）和CoreWeave等公司需求上升。
- 英伟达最新财务业绩显示数据中心业务占总营收近93%，neocloud和主权云提供商成为额外增长驱动力。
供应链影响
- NVL72机架出货量增长可能带动上游组件（如晶圆、HBM）需求增加，但价格涨幅取决于供需平衡。
- Vera Rubin系统ASP翻倍可能影响整体服务器成本，进而影响下游采购决策，具体取决于客户预算。
- 客户结构变化（CSP份额下降）可能使供应链更加多元化，但取决于新兴客户（如特斯拉、CoreWeave）的持续需求。
- 英伟达和谷歌扩展至数据中心基础设施可能改变供应链格局，涉及土地、电力、冷却等，但具体影响取决于项目执行。
- 英伟达提供残值担保和谷歌租赁担保可能增加财务风险，但可能稳定长期需求，对供应链的影响取决于这些安排的可持续性。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
英伟达AI机架过渡加速；TrendForce称NVL72产值2027年将超过7100亿美元
2026年8月28日——TrendForce最新的AI服务器产业报告显示，GB300机架级解决方案将成为英伟达2026年出货的主要驱动力，需求预计持续强劲至2027年上半年。此后，下一代VR200机架将增加产量，并在2027年占据最大出货份额。与此同时，较新的Rubin Ultra平台（VR300）预计份额相对较小，因其HBM配置和服务器架构仍在最终确定中。
TrendForce估计，包括GB和VR平台在内的NVL72机架出货量在2027年将同比增长超过50%。同时，Vera Rubin系统的平均售价预计约为GB300的两倍，而晶圆和HBM等关键上游组件价格也可能上涨。因此，GB300、VR200和VR300 NVL72机架系统的总产值预计在2027年超过7100亿美元，同比增长214%。这一增长预计将为更广泛的AI数据中心供应链提供显著动力，包括ODM系统制造商以及电力和冷却解决方案提供商。
机架级解决方案的客户结构也在转变。尽管五大北美CSP仍是主要买家，但其采购份额预计从2025年约70%降至2026年60%。这主要归因于特斯拉（包括xAI和SpaceX）和CoreWeave等公司需求的上升。
英伟达最新财务业绩还显示，其数据中心业务接近总营收的93%，而neocloud和主权云提供商正成为额外的增长驱动力。
芯片竞争扩展到数据中心基础设施
英伟达最近在财报电话会议上概述了其AI基础设施战略，俄亥俄州的PORTS-Pike园区是一个显著例子。该园区将由SB Energy开发、拥有和运营，OpenAI租赁8 IT-GW的容量。
英伟达将作为整个园区的独家计算提供商，并为最初的4.25 IT-GW提供残值担保，总支付义务上限为1050亿美元。该安排超越了芯片，涵盖土地、电力和数据中心设施，标志着英伟达将其竞争版图从xPU级别扩展到数据中心基础设施。
TrendForce指出，ASIC生态领导者谷歌正在采取类似策略，为Anthropic使用的数据中心担保租赁付款，总承诺达440亿美元。作为回报，谷歌可获得相关数据中心和电厂项目约20%的股权，同时确保Anthropic成为其TPU的主要旗舰客户。与谷歌共同设计TPU的博通也承诺回购芯片，并在产品无法售出时承担价格差异。
总体而言，谷歌和英伟达都在将其战略从芯片和系统供应扩展到数据中心基础设施。通过将自有芯片纳入融资安排，这些公司寻求在更长期内确保其产品的需求。
NVIDIA’s AI Rack Transition Accelerates; NVL72 Output Value to Exceed US$710 Billion in 2027, Says TrendForce
Aug. 28, 2026 ---- TrendForce’s recent AI server industry report reveals that GB300 rack-scale solutions are set to be NVIDIA’s leading driver of shipments in 2026, with demand anticipated to remain robust through the first half of 2027. Following that, the next-gen VR200 racks will increase production and account for the largest portion of shipments in 2027. Meanwhile, the newer Rubin Ultra platform (VR300) is expected to hold a relatively small share, as its HBM configuration and server architecture are still being finalized.
TrendForce estimates that shipments of NVL72 racks, including GB and VR platforms, will grow by more than 50% YoY in 2027. At the same time, the ASP of the Vera Rubin systems is expected to be roughly double that of GB300, while prices for key upstream components such as wafers and HBM could also rise. As a result, the combined output value of GB300, VR200, and VR300 NVL72 rack systems is projected to exceed US$710 billion in 2027, surging 214% YoY. This growth is expected to provide significant momentum for the broader AI data center supply chain, including ODM system manufacturers and power and cooling solution providers.
The customer mix for rack-scale solutions is also shifting. Although the five major North American CSPs remain the primary buyers, their share of procurement is expected to decline from approximately 70% in 2025 to 60% in 2026. This is primarily due to rising demand from companies such as Tesla (including xAI and SpaceX) and CoreWeave.
NVIDIA’s latest financial results also show that its Data Center business is approaching 93% of total revenue, while neocloud and sovereign cloud providers are emerging as additional growth drivers.
Chip competition expands into data center infrastructure
NVIDIA recently outlined its strategy for AI infrastructure during its earnings call, with the PORTS-Pike campus in Ohio serving as a notable example. The campus will be developed, owned, and operated by SB Energy, with OpenAI leasing 8 IT-GW of capacity.
NVIDIA will serve as the exclusive compute provider for the entire site and has provided a residual value guarantee for the initial 4.25 IT-GW, with aggregate payment obligations capped at $105 billion. The arrangement extends beyond chips to encompass land, power, and data center facilities, signaling NVIDIA’s expansion of its competitive presence from the xPU level into data center infrastructure.
TrendForce notes that Google, a leader in the ASIC ecosystem, is pursuing a similar strategy by guaranteeing lease payments for data centers used by Anthropic, with total commitments reaching $44 billion. In return, Google can obtain approximately 20% equity stakes in related data center and power plant projects while securing Anthropic as a major flagship customer for its TPUs. Broadcom, which co-designs Google’s TPUs, has also committed to repurchasing chips and covering price differences if the products cannot be sold.
Overall, both Google and NVIDIA are extending their strategies beyond chip and system supply into data center infrastructure. By incorporating their proprietary chips into financing arrangements, the companies are seeking to secure demand for their products over the longer term.