What Happened? Following reports earlier this month that OpenAI and GOOGL were finalizing a deal for OpenAI to use Google Cloud's compute capacity, new reports suggest OpenAI will also be renting GOOGL's TPUs to power its inference workloads as part of the agreement. This comes as OpenAI looks to meet its surging inference demand and manage inference costs as well as possible. Note that OpenAI would not have access to GOOGL's most powerful TPUs, which GOOGL is reserving to train its own Gemini models.
事件经过：继本月初有报道称 OpenAI 与谷歌母公司 Alphabet（GOOGL）即将达成协议，使用谷歌云的计算资源后，最新消息显示作为协议的一部分，OpenAI 还将租用谷歌的 TPU 芯片来运行其推理工作负载。此举源于 OpenAI 需要应对激增的推理需求，并尽可能控制推理成本。值得注意的是，OpenAI 将无法使用谷歌最强大的 TPU 芯片，这些芯片被谷歌预留用于训练其自研的 Gemini 模型。

This Would Be a Significant Endorsement of GOOGL's TPU Tech Infrastructure: Following the initial report of Google and OpenAI's partnership, we highlighted (see here) two positive takeaways: 1) it is a potential driver of a Google Cloud acceleration which is not priced into GOOGL's stock and 2) it is a positive read across to GOOGL's confidence of its position within search. The report on Friday that OpenAI will be using GOOGL's TPUs, if confirmed, would be another positive takeaway to add to this list. We view OpenAI as the most notable TPU customer to date (others include Apple, Safe Superintelligence and Cohere) and this agreement would be a significant endorsement of GOOGL's AI infrastructure capabilities which have been in development for a decade (initial TPUs were used by GOOGL in 2015). This would also be the first time OpenAI has used non-NVIDIA chips in a meaningful way, particularly interesting given OpenAI would opt to use TPUs despite the fact that they will not have access to the most advanced versions, speaking again to GOOGL's leading position within the broader ASIC ecosystem.
这将成为对 GOOGL 张量处理单元技术基础设施的重要背书：在谷歌与 OpenAI 合作的首份报告发布后，我们曾强调（参见此处）两个积极信号：1）这可能是尚未反映在 GOOGL 股价中的谷歌云业务加速催化剂；2）这对谷歌在搜索领域保持地位的信心具有积极传导效应。若周五关于 OpenAI 将采用 GOOGL 张量处理单元的报道属实，将成为该清单上新增的第三个利好。我们认为 OpenAI 是迄今为止最引人注目的张量处理单元客户（其他用户包括苹果、安全超级智能公司和 Cohere），该协议将有力印证 GOOGL 历时十年研发的人工智能基础设施实力（首批张量处理单元于 2015 年由 GOOGL 投入使用）。这也将是 OpenAI 首次大规模采用非英伟达芯片——尤其值得注意的是，尽管无法获得最先进版本，OpenAI 仍选择使用张量处理单元，这再次印证了 GOOGL 在专用集成电路生态系统的领先地位。

And Another Potential Under-Appreciated Call Option for Google Cloud Growth: While we expect NVDA's GPUs to remain the primary chip set used by tech and non tech enterprises through this GenAI era (see below), in our view, rising developer familiarity and work with the TPUs and resulting further adoption of TPUs by companies outside of Google could be a further driver for GOOGL's Cloud business. For context, our current bottom up chip model assumes that $243bn/$258bn is spent on NVDA GPUs in '27/'28...but only ~$21bn/$24bn is spent on TPUs (many of which are used by GOOGL internally). This speaks to the share shift (or TAM expansion) opportunity for GOOGL.
谷歌云增长的另一个潜在被低估的催化剂：虽然我们预计英伟达 GPU 仍将是科技与非科技企业在这一生成式 AI 时代使用的主要芯片组（详见下文），但我们认为，开发者对 TPU 的日益熟悉与应用，可能促使更多谷歌外部企业采用 TPU，从而进一步推动谷歌云业务增长。作为背景参考，我们当前自下而上的芯片模型预测 2027/2028 年英伟达 GPU 支出将达 2430 亿/2580 亿美元...而 TPU 支出仅约 210 亿/240 亿美元（其中多数为谷歌内部使用）。这凸显了谷歌面临的市场份额转移（或总可用市场扩张）机遇。

NVDA Capacity Constraints Likely Also a Factor in OpenAI's Decision:From NVIDIA's perspective, we would continue to highlight ongoing progress at Google; our estimates have NVIDIA growing more than 3x this year with Google as a customer, to more than $20 bn, outgrowing ASIC materially and growing to close to 65% processor share. While ASIC vendors will say that NVIDIA shipments are only to the cloud portion of their business, our sense is that NVIDIA is also growing its footprint in the non cloud portion of the business. But NVIDIA is sold out, particularly rack-scale products. In our view, the strength for alternative architectures is driven at least in part by an inference shortage that needs to be solved with some urgency, rather than a shifting competitive dynamic; that said, that alone reflects significant differentiation for Google vs. peers.
英伟达产能限制或成 OpenAI 决策关键因素：从英伟达视角来看，我们仍需强调谷歌的持续进展；据我们估算，今年英伟达来自谷歌客户的业务将增长超 3 倍，突破 200 亿美元，显著超越 ASIC 业务，处理器市场份额将逼近 65%。虽然 ASIC 厂商声称英伟达出货仅涉及其云业务板块，但我们观察到英伟达在非云业务领域的市场份额也在扩张。不过英伟达目前产能告罄，尤其是机架级产品。我们认为替代架构的崛起至少部分源于亟待解决的推理算力短缺问题，而非竞争格局的根本转变；即便如此，这一现象本身已凸显谷歌相较同行的显著差异化优势。

What Does the OpenAI Report Mean for AWS and Trainium? With the OpenAI/GOOGL partnership (if confirmed), OpenAI would now be running AI workloads across most major cloud providers including Google Cloud, Azure, Oracle and CoreWeave…with AMZN the notable player missing from the list. While we don't know the exact reason why OpenAI has not reached an agreement with AWS, this could speak to AMZN's (still) larger than appreciated capacity constraints (and an inability to meet OpenAI's demands). But in our view it is also a negative read across to AWS's Trainium custom silicon chips where we do not believe there are as significant of constraints. In this regard, the most notable aspect is the fact that OpenAI is reportedly choosing to use a prior generation of TPUs over Trainium. Expect this dynamic to keep investors highly focused on AWS growth (and expected accelerating growth in 2H).
OpenAI 报告对 AWS 及 Trainium 芯片意味着什么？随着 OpenAI 与 GOOGL 达成合作（若消息属实），OpenAI 的 AI 工作负载将覆盖包括谷歌云、Azure、Oracle 和 CoreWeave 在内的大部分主流云服务商…而 AMZN 成为这份名单中显著的缺席者。虽然我们不确定 OpenAI 未与 AWS 达成协议的具体原因，但这可能反映出 AMZN（仍然）存在超出预期的容量限制（且无法满足 OpenAI 需求）。但我们认为，这对 AWS 自研芯片 Trainium 同样传递出负面信号——我们并不认为该芯片存在同等程度的产能限制。最值得注意的细节在于，据报道 OpenAI 选择使用上一代 TPU 而非 Trainium 芯片。预计这一动态将促使投资者持续高度关注 AWS 的增长态势（以及下半年预期中的加速增长）。

Morgan Stanley is acting as financial advisor to Wiz, Inc. (“Wiz”) in connection with its definitive agreement to be acquired by Google LLC, as announced on March 18, 2025. The proposed transaction is subject to customary closing conditions including regulatory approvals. Wiz has agreed to pay fees to Morgan Stanley for its financial services, including fees that are contingent upon the consummation of the transaction. Please refer to the notes at the end of this report.
摩根士丹利担任 Wiz 公司（"Wiz"）的财务顾问，协助其与谷歌有限责任公司达成最终收购协议，该交易已于 2025 年 3 月 18 日对外公布。本次拟议交易需满足包括监管审批在内的常规交割条件。Wiz 已同意就摩根士丹利提供的财务顾问服务支付相关费用，其中部分费用将以交易完成为支付前提。详情请参阅本报告末尾注释。

We would greatly appreciate your 5-star vote in the Internet Large Cap and Internet SmidCap categories in this year's All-America Extel survey. We hope our work is helpful to your research process and are very grateful for you support. Thank you!
我们诚挚邀请您在本年度全美 Extel 调查中，为互联网大盘股及互联网中小盘股类别投出五星评价。若我们的工作对您的研究有所助益，我们将不胜感激。感谢您的支持！

How to vote: To request a ballot, please go to https://www.extelinsights.com/voting and select “All-American Research Team”.
投票方式：如需获取选票，请访问 https://www.extelinsights.com/voting 并选择"全美研究团队"选项。