UBS Research Report Interpretation: Broadcom AI Guidance Adjusted, Nvidia Extends to Computing Infrastructure, Terafab Begins Construction

CN
1 hour ago
UBS maintains buy ratings on Broadcom and NVIDIA, believing that the competitive focus of AI computing power supply is shifting from the chips themselves to upstream factors.

Written by: Rita

The AI computing supply chain is undergoing a restructuring on the supply side. Broadcom, NVIDIA, and Terafab, a subsidiary of SpaceX, are pushing the boundaries of AI hardware infrastructure from different directions.

On August 24, UBS released a SemiBytes industry commentary covering the latest developments from three key players: Broadcom's second-quarter preview, NVIDIA's vertical integration extending into land and electricity, and the construction progress of the Terafab project. UBS maintains buy ratings on Broadcom and NVIDIA, believing that the competitive focus of AI computing power supply is shifting from the chips themselves to upstream factors.

Broadcom's AI revenue expectations slightly adjusted, Google's share controversy does not affect core judgment

Broadcom will release its second-quarter report on September 2. UBS expects moderate revenue to exceed expectations in the third quarter, with guidance for the fourth quarter at about $34 billion. Third-quarter AI revenue is expected to be around $16 billion, further rising to $21 to $22 billion in the fourth quarter, with the growth rate of custom computing chips continuing to exceed that of AI networks.

Anthropic's TPU deployment adds new volume. UBS estimates a contribution of about $2.5 billion in the third quarter, increasing to $3.3 billion in the fourth quarter, with an annual contribution of about $9.3 billion in 2026. Google Cloud's related AI revenue increased by about $2 billion quarter-over-quarter in the third quarter, and another increase of about $1.5 billion is expected in the fourth quarter.

There is controversy in the market regarding Broadcom's share with Google. UBS determines that Broadcom remains a major chip supplier for Google, even if Google's partnership with MediaTek on custom chips progresses, it will not encroach on Broadcom's core share. Broadcom's custom computing shipments to Meta are expected to remain flat in 2026, with more significant growth anticipated only in 2027. Concerning new customers, OpenAI is expected to contribute about $1.25 billion and $12 billion in 2026 and 2027, respectively, while Apple's pace is slower.

UBS has lowered its 2027 AI revenue forecast from about $135 billion to around $130 billion due to overly optimistic shipment assumptions for TPUs, now expecting 6.4 million units to be shipped in 2027. The management is unlikely to update the previous guidance framework of $100 billion, and may instead guide the market's focus towards 2028, when AI revenue is expected to leap to about $200 billion. The target price has been slightly adjusted from $485 to $470, maintaining a buy rating.

NVIDIA extends into the infrastructure layer, evolving from chip vendor to computing power integrator

NVIDIA CEO Jensen Huang elaborated on the company's business model transformation in a blog this week, participating more deeply in the procurement and financing of land, electricity, and facilities. Currently, most AI capacity is procured through hyperscale vendors or specialized cloud providers, and AI labs are increasingly inclined to control their own long-term infrastructure.

The PORTS-Pike project is a flagship case of this transformation. NVIDIA is investing $1.5 billion in SB Energy and providing up to $105 billion in residual value and credit support to help OpenAI secure land, electricity, and facility infrastructure. The initial deployment is 4.25GW, scalable to about 8GW. NVIDIA estimates that each generation of AI factory systems will deploy about 1.5 million GPUs at this site, corresponding to about $150 billion to $200 billion in revenue, with multiple upgrade cycles expected over the entire park's lifecycle.

UBS believes that NVIDIA's recent investments in SB Energy, Lancium, and Cloverleaf Infrastructure all form part of a larger strategy to acquire "computing power channels," which may even lay the foundation for NVIDIA ultimately becoming the developer and operator of AI factory assets. Land, electricity, and facilities are becoming strategic input factors in the AI supply chain.

Terafab starts construction, signals of internal memory manufacturing emerge

UBS is tracking the progress of the Terafab project through Evidence Lab data. In July, job listings related to Terafab reached 44 on Tesla's recruitment portal, the vast majority for senior engineers with mass manufacturing and wafer fabrication capabilities. The positions cover all major semiconductor process steps, including photolithography, deposition, dry etching, and epitaxy, indicating that SpaceX is building a comprehensive engineering organization rather than selectively hiring around a few modules.

Multiple positions have appeared simultaneously in Austin, Texas, and Hsinchu, Taiwan, clearly showing SpaceX's intent to attract talent from TSMC. The recruitment focus has recently expanded to include memory design engineers and memory integration engineers, confirming Terafab's intention to initiate internal memory manufacturing. UBS points out that it remains an open question where Terafab will acquire memory IP, as existing memory suppliers lack motivation for licensed IP.

Regarding construction, since the official announcement on August 6, jobs related to cleanroom construction have become the largest category. Satellite imagery shows that the Terafab campus in Grimes County, Texas is still in a very early stage, with only limited site clearing and logging activities visible. Large-scale construction has not yet truly commenced.

The three companies are driving changes in the AI supply side from three different dimensions. Broadcom maintains a dominant position in custom computing chips; NVIDIA is evolving from a chip supplier to a computing power infrastructure integrator; Terafab is attempting to establish autonomous semiconductor manufacturing capabilities from scratch. These three lines point in the same direction: The supply side of AI computing power is shifting from a mere chip capacity race to a comprehensive competition for land, electricity, talent, and manufacturing autonomy.

Disclaimer

This article is a compilation and interpretation of third-party broker research reports (UBS, August 24, 2026) by潮向研究, combined with整理 of publicly available market information. The ratings, target prices, earnings forecasts, and related judgments quoted in the text are all opinions of the broker analysts and represent only the position of their respective institutions, not the views of潮向研究, nor do they constitute any investment advice.

The market has risks, and decisions should be made independently. This article should not be used as the basis for buying or selling any securities.

免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到support@aicoin.com,本平台相关工作人员将会进行核查。

Share To
APP

X

Telegram

Facebook

Reddit

CopyLink