UBS has released a specialized research report on Broadcom (AVGO.US), maintaining its "Buy" rating and $475 price target. The report centers on the surging demand for Tensor Processing Units (TPUs), noting that LLM developers are accelerating their push toward custom ASIC roadmaps. TPUs are experiencing significant growth as an intermediate alternative to GPUs, becoming a core driver of Broadcom's performance growth. Founded in 1961 and headquartered in San Jose, California, Broadcom is a semiconductor giant with products spanning data center networking, communications, and smartphones, among other areas. The explosive growth of its TPU business not only solidifies its position in the semiconductor industry but also opens a new growth cycle in the AI era, establishing the company as a key bellwether for hardware innovation in the global tech sector.

The report states that as large language model developers advance plans for custom Application-Specific Integrated Circuits (ASICs), many manufacturers are adopting TPUs as a transitional alternative to Graphics Processing Units (GPUs). UBS believes demand for this product is rising significantly. Following a series of supply chain analyses, the bank has refined its bottom-up custom ASIC model and now forecasts that Broadcom will ship over 5 million TPUs in 2027, up from approximately 3.7 million units in 2026. Before the v8ax (Sunfish) model becomes the primary volume driver in 2028, slightly more than half of the 2027 shipments will consist of the v7 (Ironwood) model. Both products are built on TSMC's 3-nanometer process, and UBS believes Broadcom is well-positioned to capitalize on this demand growth, thanks to TSMC's ample wafer supply allocation.

UBS currently projects Broadcom's AI-related revenue will reach approximately $60 billion in fiscal year 2026, representing year-over-year growth of about 200%. This is expected to increase to roughly $106 billion in fiscal 2027, up about 80% year-over-year, and further rise to about $150 billion in fiscal 2028. Within custom compute revenue, Google is projected to contribute around $30 billion to Broadcom this year, growing to $56 billion by 2027. Of the remaining TPU-related revenue, the majority is expected to come from Anthropic, with Meta contributing a smaller portion. The report anticipates Broadcom will begin shipping custom ASICs to OpenAI in the second half of this year and to Anthropic in the second half of 2027.

TPUs offer significant technical advantages over GPUs. The report highlights the rapid development of Mixture of Experts (MoE) models, a machine learning technique that reduces cost per token, which is driving an inflection point in TPU demand. This is because TPUs' built-in hardware Matrix Multiply Units (MXUs), combined with hardware-defined sparse compute engines, drastically reduce data read-write interactions with memory. This feature gives TPUs an advantage over GPUs in both tokens processed per watt and cost per token processed. UBS suggests this is part of the reason for NVIDIA's recent acquisition of Groq, which possesses a unique software compiler and an on-chip interconnect protocol that allows its Low-Power Processors (LPUs) to work collaboratively as a single core.

In discussions with industry experts, respondents consistently noted that TPUs, with their architecture centered on tensor operations/matrix multipliers and the addition of hardware-defined sparse compute engines, offer highly attractive cost efficiency for inference in both dense LLMs and sparse MoE models. Besides Google, core TPU clients like Anthropic and Meta differ significantly from traditional cloud service customers. These companies have full control over their software stacks, making them far less dependent on NVIDIA's CUDA unified computing architecture than traditional enterprise cloud clients. Consequently, CUDA's relative advantage is much less pronounced in the LLM developer segment of the cloud data center market compared to the broader accelerator market. These LLM developers are directly and indirectly driving an increasing share of global hyperscale data center capital expenditures.

Addressing market concerns about Google's collaboration with MediaTek on a Customer-Owned Tooling (COT) model, the report concludes the negative impact on Broadcom will be limited. Due to its SerDes technology barriers, Broadcom is expected to continue supplying I/O chiplets to Google even under a COT model, accounting for 40-50% or more of the total chip cost. This segment boasts gross margins exceeding 75%, higher than the 60% margins for XPUs. In a base-case scenario, the COT model would dilute Broadcom's EPS by only 8%, which is significantly lower than market fears. Furthermore, MediaTek's TPU project is relatively behind schedule, making it unlikely to challenge Broadcom's market share in the near term.

For valuation, the report employs a Sum-of-the-Parts methodology, applying a 25x EV/FCF multiple to the infrastructure software business and a 30x EV/FCF multiple to the semiconductor business for fiscal 2027. An upside scenario suggests a price target of $560, implying 63% potential upside, while a downside scenario points to $290, indicating 16% potential downside. With the current stock price at $343.94, this implies 38% upside to the base target. The report forecasts total potential return of 39% and excess return of 30.5%.

UBS has raised its forecasts for several of Broadcom's financial metrics. Revenue for fiscal years 2026, 2027, and 2028 is now projected at $105.8 billion, $155.5 billion, and $200.2 billion, representing year-over-year growth of 65.6%, 47.0%, and 28.7%, respectively. Diluted EPS is forecast at $11.05, $16.68, and $21.53, representing upward revisions of 1.3%, 17.9%, and 12.6% from previous estimates. AI business performance is particularly strong, with AI revenue expected to surge 205% year-over-year to $60.2 billion in fiscal 2026 and grow a further 76.9% to $106.4 billion in fiscal 2027.