Wallstreetcn
2024.06.09 02:11
portai
I'm PortAI, I can summarize articles.

Broadcom, winning big!

Broadcom emerges as the second largest AI winner in the semiconductor industry, surpassing Intel and AMD, relying on its custom chip business for success. Google chose Broadcom as a partner for its self-developed chips, launching the first generation TPU. Broadcom's custom chips are widely used in cloud service companies, meeting future computing needs. Broadcom's achievements demonstrate the fierce competition in the AI chip field, with NVIDIA still leading in the AI chip race

With NVIDIA's market value surpassing Apple, AI chips have once again become a hot topic of discussion. From 1 trillion to 2 trillion, and then to 3 trillion, AI seems to have cast a magical spell on NVIDIA's stock, causing it to rise continuously, far exceeding other players in the semiconductor industry.

Does NVIDIA have the opportunity to surpass Microsoft and become the number one in the US stock market? The answer to this question is still unknown, but what can be certain is that in this AI chip race ignited by NVIDIA, there will never be a shortage of competitors.

In the AI frenzy, it's not just NVIDIA with a market value of 3 trillion that is winning, but also a company that has climbed to second place through custom chips.

Broadcom, the Second in AI

For Broadcom, custom chips have long been its specialty.

In 2013, the head of Google AI calculated meticulously and found that for a huge cloud service company like Google, with the continuous development of the Internet, relying solely on off-the-shelf general-purpose CPUs and GPUs would inevitably be unable to meet the future massive computing demands. The best solution was to customize chips for specific applications. To this end, Google set a goal: to build a domain-specific computing architecture for machine learning and reduce the overall total cost of ownership (TCO) for deep neural network inference to one-tenth of the original.

As an internet software company, Google itself did not have the capability to develop a chip from scratch and had to seek external help. After several discussions, Google chose Broadcom as its partner for self-developed chips.

At the 2016 Google I/O developer conference, Google's CEO officially showcased the self-developed chip TPU v1 to the world. The first-generation TPU was manufactured using 28-nanometer process technology, with a operating frequency of 700MHz, a power consumption of 40W during operation. Google packaged the processor into an external accelerator card, installed it in the SATA hard drive slot for plug-and-play use, and connected the TPU to the host via a PCIe Gen 3x16 bus, providing an effective bandwidth of 12.5GB/s.

According to reports from Morgan Stanley analysts, starting from TPU v1, Google has had deep cooperation with Broadcom. They jointly designed all TPUs that have been publicly announced so far, and Broadcom's revenue in this area has soared from $50 million in 2015 to $750 million in 2020 due to Google's success. In addition to chip design, Broadcom also provided Google with crucial intellectual property and was responsible for manufacturing, testing, and packaging new chips, thereby supporting Google's new data centers.

The advent of the AI wave has made the already profitable Broadcom even more lucrative. According to Broadcom's first-quarter 2024 financial report, semiconductor revenue for the quarter was $7.39 billion, a year-on-year increase of 4%, accounting for 62% of total revenue. Network revenue was $3.3 billion, a year-on-year increase of 46%, accounting for 45% of semiconductor revenue. This growth was mainly driven by the custom DPU chip demand from two major customers, with an expected year-on-year growth of over 35% in 2024.

In the financial report, Broadcom classified AI ASIC and AI-focused network solutions as AI accelerators. As of 2023, the total sales of this business accounted for 15% of the annual semiconductor revenue, approximately $4.2 billion. In the first quarter of 2024, AI revenue was around $2.3 billion, accounting for 31% of semiconductor revenue, quadrupling compared to the same period last year. It is expected that the total AI revenue share will be above 35% in 2024, indicating an AI revenue scale exceeding $10 billion (previously estimated at $7.5 billion). Custom DPU chips are expected to contribute around $7 billion, with 20% from switch/router chips, 10% from optical chips, and interconnect chips, with an expected year-on-year growth of approximately 133%.

What does $10 billion mean? It means that Broadcom has become the second-largest AI winner in the semiconductor industry, surpassing even established processor manufacturers like Intel and AMD, second only to NVIDIA. Broadcom's achievement is largely attributed to its custom chip business that started in 2016.

More to Win?

However, Broadcom seems to aim for even more victories.

During the March earnings conference, Broadcom's CEO showed optimism towards AI and custom chips. He mentioned that by the 2024 fiscal year, Broadcom's network revenue is expected to grow by 30%, driven by accelerated network deployments and the expansion of AI accelerators in large-scale enterprises. It is projected that generative AI revenue will account for over 25% of semiconductor revenue.

In April's Investor Day event, Broadcom showcased its latest custom chips. They highlighted the powerful computing capabilities, large HBM capacity, high-speed on-chip interconnects, and top-notch external networks of their low-latency chips. Notably, Broadcom's most advanced custom chips feature 12 HBM stacks, surpassing NVIDIA's Blackwell with 8 HBM stacks, exceeding the industry leader by 50% in HBM, showcasing its strength in chip design.

After what we have seen, Broadcom seems to have secured more orders following its muscle-flexing in March and April.

Recently, analysts from Morgan Stanley reported that Broadcom has obtained the design contract for Google's next-generation AI chip TPU v7. The report stated: "This marks Broadcom's assistance in launching the seventh generation of AI processor chip series for Google." This long-term partnership allows Broadcom to significantly benefit from Google's continuous investment in AI technology.

Morgan Stanley analysts project that Google's TPU project will bring Broadcom over $8 billion in revenue this year, a 125% year-on-year increase. They expect this number to exceed $10 billion by 2025, attributed to the upcoming TPU v5 and v6 chips.

The report also highlighted Broadcom's success with Meta, with analysts stating: "Meta collaborated with Broadcom to jointly design Meta's first and second-generation AI training processors." They anticipate Broadcom to accelerate the production of Meta's third-generation AI chip (MTIA 3) in the second half of 2024 and 2025.

Overall, analysts estimate Broadcom's AI revenue to reach $11-12 billion this year and surpass $14-15 billion by 2025.

Just three months later, Broadcom's originally estimated $10 billion in AI-related revenue has increased by another 10-20%. This period may have caught many optimistic investors off guard, considering it's only halfway through the year. Who can guarantee that this revenue won't rise again in the remaining half of the year?

Morgan Stanley's analysts emphasized Broadcom's dominant position in the cutting-edge custom chip design market (7nm, 5nm, 3nm, 2nm). According to their estimates, Broadcom holds a 55-60% market share in this segment, with clients including major tech companies like Apple, Microsoft, and Meta, making it a dominant player.

Of course, such a lucrative market inevitably attracts other companies. Marvell, responsible for customizing chips for Amazon, and Nvidia, which recently announced plans to enter the custom chip market, are potential competitors for Broadcom. However, given Broadcom's extensive IP portfolio, rich experience, and large customer base, it is unlikely that any company will emerge in the short term to challenge its position.

Is VMware a Threat?

At the end of last year, Broadcom successfully acquired virtualization vendor VMware for $61 billion. However, it subsequently announced significant adjustments to its future product sales and service policies, completely terminating perpetual licensing in favor of a subscription model. This means that VMware's virtualization platform has officially transitioned to a subscription model, no longer offering perpetual licenses.

This move caused a stir, with many VMware customers expressing dissatisfaction. Some believe that they now have to purchase products they may not want or are not yet ready to implement in order to access the products they currently use and rely on. There are also concerns about price increases, with reports indicating that prices for some products have even risen by 500% or 600%, with licensing costs jumping from $17,000 to $470,000 and from $8 million to $100 million Broadcom has not revoked or withdrawn its plan to implement a subscription system despite the objections raised. This has led many companies to threaten to switch to other virtualization platforms, with some actually taking action. For example, Computershare, an Australian company with 24,000 virtual machines, may abandon VMware and focus on Nutanix products. It is reported that the company is currently undergoing a large-scale IT service migration, moving 24,000 virtual machines from "other" virtual machine management programs to Nutanix AHV.

This has cast a shadow over Broadcom, which is currently thriving in AI.

However, the impact may not be as significant as expected. According to a study released by CloudBolt, after Broadcom's $61 billion acquisition of VMware, the prediction that a large number of customers would switch to other virtualization technology providers may have been exaggerated.

CloudBolt surveyed 300 IT executives responsible for VMware-related acquisitions last month and found that although 76% of them were very or somewhat concerned about the impact of this acquisition on their business, most of them were taking time to evaluate their options. The top choices were retaining some VMware products (43%), continuing to invest fully in VMware (40%), moving more workloads to the public cloud (38%), migrating to alternative virtual machine management programs (34%), and fully migrating to the public cloud (33%). Recognizing a choice does not necessarily mean intending to take action.

Only 5% of companies have made a decision, while 87% plan to make a decision in the next 12 months. The study did not indicate which option these 15 companies chose.

Against the backdrop of the booming AI industry and the impact on virtualization businesses, it remains to be seen whether Broadcom will maintain its winning position.

Source: Semiconductor Industry Observation, original article titled "Broadcom, Winning Too Much!"