|

楼主 |
发表于 2025-3-20 10:57:50
|
显示全部楼层
HERE'S WHAT ANALYSTS HAVE TO SAY AFTER $NVDA'S GTC KEYNOTE 2025Citi (Buy, PT: $163): "Jensen Huang (CEO of NVIDIA) delivered the GTC keynote today. Three key points jumped out to us: 1) NVIDIA is adding more color to its TAM expectations with total annual capex reaching $1T by 2028 as both inference and training continue to require more compute. 2) Blackwell is not only back on track, it is outperforming expectations with units (individual dies) from top 4 US hyperscalers already reaching 3.6M in 2025, 2.8x vs. Hopper’s peak year. 3) The company reminded investors that it is leading inference and is not stepping its foot off the gas with a blisteringly fast compute roadmap (B300, Rubin, Rubin Ultra), software leadership (e.g. Dynamo), and networking innovation (CPO). Net-net, we came out of the keynote reassured in NVIDIA’s leadership which, if anything, seems to be expanding. We view positively NVIDIA’s push for inference, which per company comments now requires significantly more compute. Maintain Buy."
Morgan Stanley: "Our preliminary conclusions are positive. We aren't sure what's going to shift the market from taking a glass-half-empty view of this situation, but the demand picture will stay strong through the visibility window. The concerns we hear that ASICs are causing competitive and margin pressure are going to vanish fairly quickly when they materially outgrow ASIC competition in 2H, as gross margins return to the mid-70s. The stock trades at a P/E discount to direct peers. We expect the company to have a very confident tone in the Q&A in the morning, in contrast to the somewhat more restrained management view at CES when there were supply chain and Hopper challenges. The risk, of course, is government export controls, which appear likely. However, we remain optimistic that the limitations will be mostly China-specific and that requiring licenses in so-called 'tier 2 countries' will be limited. We remain OW, stock remains our top pick in semis."
BofA (Buy, PT: $200): "We maintain Buy, $200 PO following a slate of product/partner announcements at the flagship GTC conference in addition to a post-keynote meeting with the CFO that demonstrated NVDA continuing to deepen its competitive moat in a $1T+ infrastructure/services TAM. We were excited by: 1) Structural uplift in demand in part fueled by over 100x greater compute intensity for inference evidenced by 1.8M Blackwell packages (each package = 2 GPUs) shipped/ordered to date vs. 1.3M Hopper units total in 2024 across top 4 CSPs. 2) New Blackwell Ultra, Rubin, and Rubin Ultra silicon/server/system platforms pointing to an unmatched roadmap as Rubin advances AI performance 900x (scale-up FLOPs) over Hopper (Blackwell is 68x) in a TCO-optimized way. 3) New CPO-based Spectrum-X and Quantum-X switches bolster networking edge and advance scale-out to +1M GPU clusters. 4) Physical AI TAM expansion progress with key announcements in robotics and AVs. 5) CFO discussions suggested confidence in GM recovery to mid-70% in 2H as Blackwell ramps. Overall, NVDA continues to dominate the AI value chain with its full-stack turnkey (hardware, software, systems, services, developers) model."
KeyBanc (Overweight, PT: $190)"Today, NVDA CEO Jensen Huang kicked off GTC with a keynote speech with announcements largely in line with expectations heading into the event. Key takeaways: 1) Announced Blackwell Ultra (GB300) NVL72, which is expected to be 1.5x performance of GB200 NVL72 and is expected to be available in 2H25. 2) Announced Vera Rubin NVL144, with Vera being the next-gen ARM-based CPU and Rubin being the next-generation GPU, with 144 GPUs per rack and performance expected to be 3.3x GB300 NVL72 and expected 2H26. 3) Announced Co-packaged Optical (CPO) at 1.6TB, expected to ship in 2H25. NVDA continues to push the envelope on performance with its annual cadence roadmap, such that it remains the clear leader in AI. Maintain Overweight."
Stifel (Buy, PT: $180): "The GTC Keynote featured updates on next-gen Blackwell Ultra, Rubin, and Rubin Ultra architectures. The Keynote also featured the launch of Dynamo – NVDA's inferencing software designed for the optimization of reasoning models within data centers, which was described as the operating system for AI Factories. As was widely expected, Mr. Huang also touched on scale-out networking with the formal announcement of silicon photonics/co-packaged optics Quantum-X and Spectrum-X switches. On the enterprise infrastructure front, NVDA announced the DGX Spark and DGX Station PCs, offering AI capabilities and performance in a desktop form factor. Continued full-stack infrastructure development is targeted at evolving reasoning models and agentic AI, which NVDA believes will drive 10-100x compute intensity."
Wells Fargo (Overweight, PT: $185): "While much of what was announced had been somewhat anticipated, we think NVIDIA's continued full stack/platform innovation was once again showcased; NVDA is solidly in a league of its own. NVDA's scale-up capabilities with networking innovations (co-packed optics/SiPho; NVLink evolution)—enabling racks of 576 GPUs per rack—is a key/significant competitive advantage that could drive incremental improvements in perf/tokens/$, broadening end-market applications."
Cantor Fitzgerald (Neutral, PT: $200): "It’s tough not to be impressed. The excitement at GTC is palpable, and advancements being made in AI are truly amazing. The absolute key focus was on the extreme computing needed for inference at scale and the work NVDA is doing to enable the proliferation of greater intelligence (i.e., reasoning) across applications/verticals. At the crux of everything here is NVDA’s work in combining its best-in-class hardware with software initiatives, such as the newly announced Dynamo software stack that acts as the OS for AI factories, and when combined with Blackwell offers ~40x inference improvement vs. the prior-generation Hopper."
Raymond James (Strong Buy, PT: $170): "During the keynote, NVDA’s CEO sounded highly confident about data center capex growth sustaining (to exceed $1T by 2028), and highlighted the company’s opportunities in Inferencing, Agentic AI, and Robotics. Blackwell demand remains strong, driven by top 4 CSPs. Blackwell Ultra is on track for 2H25 shipments and management sees a smooth transition. Vera Rubin NVL144 is on track for 2H26 and Rubin Ultra NVL576 for 2H27, with performance specs that would be hard to match, in our view. Overall, we walked away comfortable with long-term AI demand and continue to be impressed with NVDA’s roadmap & technology innovation. Reiterate Strong Buy."
Bernstein SocGen Group (Outperform, PT: $185): "Nothing hugely surprised given all the pre-event speculation, but we still thought it sounded good. The roadmap looks really solid, and their capability gap vs competitors across their entire massive stack continues to widen. And the company still seems positive on datacenter growth, calling for $1T+ in datacenter capex by 2028 while capturing an increasing share of it. It is still NVIDIA’s game to lose, and they don’t appear to be losing..."
Deutsche Bank (Hold, PT: $145): "NVDA remains highly bullish about the future of AI compute spend, with its expectations for inference-related compute needs having 100x'd since last year's GTC. NVDA sees a continued need for AI compute solutions to 'scale-up', and introduced several technological innovations which should drive this trend forward (sophisticated GPU roadmap, disaggregated NVLink switches, NVDA 'Dynamo' OS for AI factories, co-packaged silicon photonics options, etc.)."
UBS: "The first thing that hit us is the sheer volume of people competing to get into each panel at this event - it is, in our experience, extraordinary. NVDA's edge AI panel highlighted the sheer breadth of its offerings spanning service provider, enterprise, industrial, embedded, and physical AI. Investors have lately often said 'nobody makes much money yet from AI' but IBM's panel on ROI of accelerated computing showcased $3.5 billion in annual savings exiting C2024 via improved productivity."
由翻译自 英语
以下是$NVDA 2025 年 GTC 主题演讲后分析师的评价
花旗银行(买入,预计:163 美元):“黄仁勋(NVIDIA 首席执行官)今天发表了 GTC 主题演讲。我们注意到三个关键点:1)NVIDIA 正在为其 TAM 预期增添更多色彩,到 2028 年,年度总资本支出将达到 1 万亿美元,因为推理和训练都需要更多的计算。 2) Blackwell 不仅重回正轨,而且表现超出预期,美国前四大超大规模制造商的单位(单个芯片)在 2025 年已经达到 360 万,2. 8x与 Hopper 的巅峰年份相比。 3)该公司提醒投资者,它在推理方面处于领先地位,并且不会因为极快的计算路线图(B300、Rubin、Rubin Ultra)、软件领导力(例如 Dynamo)和网络创新(CPO)而松懈。总而言之,我们从主题演讲中对 NVIDIA 的领导地位感到放心,如果说有什么不同的话,那就是这种领导地位似乎正在扩大。我们认为对 NVIDIA 推动推理的举措持积极态度,根据公司评论,推理现在需要更多的计算。维持“买入”评级。
摩根士丹利:“我们的初步结论是积极的。我们不确定什么因素会改变市场对这种情况的悲观看法,但需求前景将通过可见性窗口保持强劲。我们听到的有关 ASIC 造成竞争和利润压力的担忧将在 ASIC 下半年大幅超过 ASIC 竞争对手时很快消失,因为毛利率将回到 70% 左右。该股票的市盈率低于直接同行。我们预计该公司在早上的问答环节中会非常自信,这与 CES 上供应链和 Hopper 挑战时管理层的略显克制的看法形成鲜明对比。当然,风险是政府的出口管制,这似乎是有可能的。然而,我们仍然乐观地认为,这些限制将主要针对中国,并且在所谓的“二线国家”需要许可证的情况将有限。我们仍然保持增持,股票仍然是我们在半导体行业的首选。”
美国银行(买入,预计:200 美元):“我们维持买入,200 美元的 PO 评级,因为在旗舰 GTC 会议上发布了一系列产品/合作伙伴公告,此外,在与 CFO 的主题演讲后,NVDA 还展示了其在 1 万亿美元以上基础设施/服务 TAM 中的竞争护城河。我们感到兴奋的是:1)需求的结构性提升,部分原因是推理计算强度提高了 100 倍以上,迄今为止 Blackwell 封装(每个封装 = 2 个 GPU)的发货/订购量为 180 万个,而 2024 年前四大 CSP 的 Hopper 单元总数为 130 万个。 2)新的 Blackwell Ultra、Rubin 和 Rubin Ultra 硅/服务器/系统平台指向无与伦比的路线图,因为 Rubin 以 TCO 优化的方式将 AI 性能提高到 Hopper 的 900 倍(扩大 FLOPs)(Blackwell 为 68 倍)。 3)新的基于 CPO Spectrum-X 和 Quantum-X 交换机增强了网络优势,并推进了向 +1M GPU 集群的扩展。4) 物理 AI TAM 扩展取得进展,机器人和 AV 领域发布了重要公告。5) CFO 讨论表明,随着 Blackwell 的加速发展,GM 的恢复率有望在下半年达到 70% 左右。总体而言,NVDA 凭借其全栈交钥匙(硬件、软件、系统、服务、开发人员)模式继续主导 AI 价值链。”
KeyBanc(增持,PT:190 美元)“今天,NVDA 首席执行官黄仁勋在 GTC 上发表主题演讲,宣布的事项与活动前的预期基本一致。主要内容:1) 宣布推出 Blackwell Ultra (GB300) NVL72,预计其性能是 GB200 NVL72 的 1.5 倍,预计将于 2025 年下半年上市。 2)宣布推出 Vera Rubin NVL144,其中 Vera 是下一代基于 ARM 的 CPU,Rubin 是下一代 GPU,每机架有 144 个 GPU,性能预计是 GB300 NVL72 的 3.3 倍,预计于 2026 年下半年上市。 3)宣布推出 1.6TB 的 Co-packaged Optical (CPO),预计将于 2025 年下半年出货。NVDA 继续通过其年度节奏路线图突破性能极限,从而继续保持 AI 领域的领先地位。维持增持评级。”
Stifel(买入,PT:180 美元):“GTC 主题演讲介绍了下一代 Blackwell Ultra、Rubin 和 Rubin Ultra 架构的最新进展。主题演讲还介绍了 Dynamo 的发布——NVDA 的推理软件,旨在优化数据中心内的推理模型,被称为 AI 工厂的操作系统。正如人们普遍预期的那样,黄先生还谈到了横向扩展网络,并正式宣布了硅光子/共封装光学 Quantum-X 和 Spectrum-X 交换机。在企业基础设施方面,NVDA 宣布推出 DGX Spark 和 DGX Station PC,以台式机形式提供 AI 功能和性能。持续的全栈基础设施开发旨在发展推理模型和代理 AI,NVDA 认为这将推动 10-100 倍的计算强度。”富国银行(增持,PT:185 美元):“虽然大部分宣布的内容都已在意料之中,但我们认为 NVIDIA 的持续全栈/平台创新再次得到展示;NVDA 稳居一流水平。NVDA 的扩展能力与网络创新(共封装光学/SiPho;NVLink 演进)相结合,使每架机架可容纳 576 个 GPU,这是一项关键/重要的竞争优势,可以推动性能/代币/美元的逐步改进,拓宽终端市场应用。”
Cantor Fitzgerald(中性,PT:200 美元):“很难不被打动。GTC 的兴奋之情显而易见,人工智能的进步确实令人惊叹。绝对关键的焦点是大规模推理所需的极端计算,以及 NVDA 为实现更高智能(即推理)在应用程序/垂直领域的普及所做的工作。这里一切的关键是 NVDA 将其一流的硬件与软件计划相结合,例如新发布的 Dynamo 软件堆栈,它充当人工智能工厂的操作系统,与 Blackwell 结合使用时,与上一代 Hopper 相比,推理能力提高了约 40 倍。”
Raymond James(强力买入,PT:170 美元):“在主题演讲中,NVDA 的首席执行官对数据中心资本支出持续增长(到 2028 年将超过 1 万亿美元)充满信心,并强调了公司在推理、Agentic AI 和机器人领域的机遇。在前四大 CSP 的推动下,Blackwell 的需求依然强劲。Blackwell Ultra 有望在 2H25 出货,管理层预计过渡将顺利进行。Vera Rubin NVL144 有望在 2H26 出货,Rubin Ultra NVL576 有望在 2H27 出货,我们认为其性能规格很难匹敌。总体而言,我们对长期 AI 需求感到满意,并继续对 NVDA 的路线图和技术创新印象深刻。重申强力买入。”
伯恩斯坦法兴集团(优于大盘,预期:185 美元):“考虑到所有事前猜测,这并不令人感到意外,但我们仍然认为这听起来不错。路线图看起来非常可靠,他们在整个庞大堆栈中与竞争对手的能力差距继续扩大。该公司对数据中心的增长仍持乐观态度,预计到 2028 年数据中心资本支出将超过 1 万亿美元,同时占据越来越大的份额。英伟达仍是输家,但他们似乎没有输的迹象……”
德意志银行(持有,预期:145 美元):“NVDA 对 AI 计算支出的未来仍然非常乐观,自去年的 GTC 以来,其对推理相关计算需求的预期增长了 100 倍。NVDA 认为 AI 计算解决方案将继续需要‘扩大规模’,并推出了多项技术创新,这些创新将推动这一趋势向前发展(复杂的 GPU 路线图、分解的 NVLink 交换机、用于 AI 工厂的 NVDA‘Dynamo’操作系统、共同封装的硅光子学选项等)。”
瑞银:“首先让我们感到震惊的是,参加本次活动的每个小组讨论的人数之多——根据我们的经验,这是非同寻常的。NVDA 的边缘 AI 小组强调了其产品范围之广,涵盖服务提供商、企业、工业、嵌入式和物理 AI。投资者最近经常说‘目前还没有人从 AI 中赚到很多钱’,但 IBM 关于加速计算投资回报率的小组展示了通过提高生产力,到 2024 年每年可节省 35 亿美元。”
翻译得准确吗?请提供反馈,以便我们加以改进:

下午9:02 · 2025年3月19日
·
2.7万
查看
|
|