Long|AI ASIC part I: Broadcom, Google, AMD

Overview of Major CSP’s ASICs and Their Attempts to Break Free from the NVIDIA Ecosystem

We primarily focuses on Broadcom, Google, and AMD, providing an in-depth analysis of their self-developed ASICs, efforts to move away from the NVIDIA ecosystem, and related financial and technological developments.

Broadcom

Key Highlights

VMware Business:

• We anticipate VMware revenue to maintain a slight quarter-over-quarter growth trend, reaching over USD 4 billion in Q4. GP margin is expected to increase further, primarily due to business integration in Europe.

• In terms of customer structure, financial clients (banks and insurance) are the largest, followed by government and manufacturing sectors. The first two categories account for over 50% of revenue. Migration costs are substantial, and the willingness to migrate among these customers is very low, resulting in a favorable overall competitive landscape.

ASIC Chip Business:

• Broadcom forms deep partnerships with clients through self-developed IP and active involvement in the front-end design of ASIC chips. In 2024, ASIC revenue will primarily be driven by Google TPU. Starting next year, Meta is expected to gradually scale up, and ByteDance is projected to begin small-scale shipments by Q4 2025 (though this project is currently stalled). OAI is expected to contribute small-scale revenue in 2026. ASIC business revenue is projected to reach USD 13.7 billion in 2025, mainly from contributions by Google and Meta.

• We forecast Broadcom’s AI-related revenue to approach USD 19 billion in 2025; overall, non-GAAP EPS for 2025 is USD 6.5.

• Based on current CoWoS booking trends for major chipmakers in 2025, NVIDIA, AVGO, and MRVL show significant growth, while AMD and Intel booking volumes continue to decline.

Broadcom Business Analysis

Firstly, we analyze Broadcom focusing on two major business segments: the AI ASIC business and the VMware business. We believe that Broadcom’s long-term strategic logic is clear, and its competitive barriers are steadily increasing.

AI ASIC Business Analysis

• Customized IP Development: Broadcom customizes existing IP based on customer requirements. For example, it collaborates deeply with its largest client, Google, in the design of several generations of chips, jointly participating in front-end chip design to establish a strong partnership with the client.

• Competitive Advantage: As a leader in the communications industry, Broadcom already has ready solutions available when new standards are introduced. In contrast, other ASIC customers opting for third-party suppliers may need to wait for them to develop or co-develop solutions, increasing time costs. Thus, choosing Broadcom can effectively shorten the overall chip development cycle for customers.

• Development Efficiency: Broadcom emphasized at the GS conference that it can compress a three-year development cycle into one year. Additionally, Broadcom’s self-developed IP spans multiple product lines, is thoroughly validated, and has undergone multiple tape-outs, reducing the risk of repeated tape-outs for customers. Customers can potentially complete the tape-out phase in one attempt, significantly enhancing development efficiency and success rates.

• ASIC Pricing: Broadcom’s ASIC products are typically priced through multiple components, including NRE fees, fees for using Broadcom’s self-developed IP, and TSMC tape-out costs. Through this pricing strategy, Broadcom maintains high gross margins of around 60%-70% (especially for clients like Google).

Sources of Broadcom’s High Gross Margins in ASIC Business:

• Technical Expertise Advantage: For example, when Google needed to reduce chip specifications to meet MediaTek’s technical requirements for TPU V6E production, Broadcom’s core strengths in chip design, IP provision, tape-out, and production, particularly in high-speed interconnects and high-computing power chips, enhanced Broadcom’s gross margins.

• Client Collaboration Advantage: Broadcom assists clients in rapidly iterating chips and increasing tape-out success rates, ensuring smoother production processes. This flexible solution adjustment and high tape-out success rate are critical factors in maintaining Broadcom’s high gross margins.

VMware Business Analysis

• Customer Structure: Broadcom’s VMware customers are primarily from the financial, government, and manufacturing sectors, with financial and government clients accounting for nearly 50%. These customers have extremely high data security requirements and low familiarity and dependence on upper-layer software interfaces, resulting in high debugging costs and security risks during migration. The top four confirmed customers are Google, ByteDance, Meta, and OpenAI. There are rumors that Apple and Microsoft are also negotiating with Broadcom for design solutions. In fact, any customer looking to produce chips consults Broadcom for solutions.

• Competitive Advantage: Compared to competitors like Nutanix, Broadcom’s VMware offers greater hardware flexibility. Nutanix’s virtual machine business requires customers to bind to specific server hardware, whereas Broadcom’s VMware does not have this limitation. Additionally, AWS and Microsoft’s solutions are more oriented towards public cloud solutions, while Broadcom’s VMware primarily serves privately deployed customers, offering advantages in interface familiarity, ease of management, and hardware flexibility.

• Post-Merger Benefits: Broadcom was one of Nutanix’s key customers before the merger. Post-merger, the framework was migrated to VMware, reducing internal costs by over 50%. This was mainly due to the elimination of the need for deep hardware binding and VMware’s significant superiority in energy efficiency management compared to competitors.

• Both the AI ASIC business and the VMware business have clear long-term strategic logic and significant competitive advantages for Broadcom. Next, we will separately analyze the AI-related revenue.

Broadcom AI-Related Revenue Analysis

For 2025 and 2026, we expect AI-related revenue to be approximately USD 12.3 billion and USD 18.9 billion, respectively, representing a YoY growth of over 50%. The proportion of AI-related revenue within Broadcom’s total revenue is expected to increase from 23% this year to around 30% next year. Specifically, in terms of sales volume and pricing, based on private confirmations with Google and Broadcom’s sales channels, the total ASIC business volume is expected to be around USD 13.7 billion next year, with USD 12.9 billion coming from Broadcom itself. Currently, Meta’s contribution to Broadcom remains relatively small, primarily due to the low value of the current second-generation inference chips. Therefore, future growth relies more on the launch of next-generation chips. As mentioned earlier, improvements in GPU architecture will significantly enhance growth potential, potentially introducing new growth curves.

Broadcom’s AI chip revenue also includes switch chip segments. We employ a bottom-up approach based on customer capital expenditures to estimate:

Core Assumptions:

• In 2024, AI-related capital expenditures (CAPEX) are expected to increase from 50% this year to 60%. Within total capital expenditures, AI-related switch CAPEX is estimated to account for 3%.

• Assuming the four major cloud service providers (CSPs) account for approximately 50% of the total market share, based on observations from NVIDIA’s financial reports, where last quarter’s GPU sales were about 45% from CSPs, and this quarter they have increased to 50%.

Market Estimation:

• Based on the above assumptions, we calculate the total market CAPEX for switches and assume Broadcom’s market share is 45%. This market share is derived from the changes in the market share of Ethernet and InfiniBand, where last year the ratio was 7:3, and this year it is expected to reverse to 6:4 or higher.

• The average value of switch chips accounts for approximately 20%-25%. Based on conservative estimates, we derive next year’s switch chip revenue and convert it to an equivalent of 51.2T switches, estimated to be about 170,000 units.

Other AI Chip Business

AI chips also involve optical chips, DSPs, and PCIe components. Although we have detailed this data, some discrepancies may exist. Therefore, we refer to numbers provided by Broadcom’s sales channels. In 2023, revenue ranged between USD 3-3.2 billion, and Broadcom expects approximately 30% growth in 2024, bringing total AI-related business revenue to about USD 18.9 billion.

Impact of Broadcom Exiting UA Link

Broadcom is recently removed from the list of founding members of UA Link, which may reflect internal disagreements regarding future development directions:

• Scenario One: If Broadcom continues to promote the UA Link protocol and opts to open-source it, the company would need to share its semi-finished technological成果. This could potentially open up incremental markets in the short term. However, in the long run, competitors catching up would erode Broadcom's competitive advantage.

• Scenario Two: Broadcom chooses a full-stack customization path, focusing on the design of AI chips and internal server switch chips. Although this would require higher R&D investments in the short term and result in lower ROI due to insufficient customer chip volumes, this strategy would further strengthen its competitive barriers in the long term.

Impact of Google's In-House Development on Broadcom's Gross Margins

With TPU V6, some core modules and algorithms are now being developed by Google's own chip design team. However, Broadcom's strengths in network interconnects and SerDes remain robust. We believe that in 2025, Broadcom might experience a slight decline in gross margins due to this shift, but the impact will not be significant since Broadcom still handles the majority of TPU V6 production.