Track Hyper | Broadcom takes down NVIDIA:沃尔核材 benefits from it

Author: Zhou Yuan / Wall Street News

AI-specific chips (ASIC) and related industry chain companies have become dark horses at the end of the year.

The discussion topics related to this stem from the value extension direction of AI-specific chips for data centers: AEC, a special short-distance data transmission communication cable between servers and between servers and workstations.

The relevant sectors of this cable saw significant increases in May and August of this year. After that, these sectors entered a prolonged adjustment period. Now, with the performance proof of ASIC, AEC has once again become the focus.

On December 16, multiple AEC concept stocks hit the daily limit. Among them, Walden Materials continued to rise during intraday trading on December 17 (as of 11:01).

The Extended Value of Dedicated AI Chips

Broadcom, listed on the US stock market, released astonishing results for fiscal year 2024 on December 12: for fiscal year 2024, Broadcom's AI ASIC (custom or dedicated chips) business revenue increased by 220% year-on-year, reaching $12.2 billion.

At the same time, Broadcom announced that it is currently developing AI ASIC chips with three major clients, expecting the market size for its AI chips to be between $15 billion and $20 billion in 2025.

This achievement and the forecast for next year's market size have shocked the global market.

As of the close of US stocks on December 13, Broadcom recorded an astonishing increase of 24.43%, with a market capitalization exceeding $1 trillion, joining the global semiconductor market capitalization trillion-dollar club (along with NVIDIA, TSMC, and Broadcom); on December 16, Broadcom opened high, with intraday gains exceeding 6%, increasing its market capitalization to $1.12 trillion; on December 17, Broadcom closed at $250, up 11.21%, with its market capitalization touching $1.2 trillion.

From the release of the financial report on December 12 to the close on December 17, Broadcom's stock price increased by nearly 36.5%.

The emergence of GenAI (generative AI) (ChatGPT) not only allowed NVIDIA CEO Jensen Huang to achieve a turnaround with GPUs, becoming the "Huang Master," but also gave rise to endless imaginative space in the market for FPGAs and ASICs.

However, before Broadcom released its financial report, the market had not seen performance validation for ASICs or FPGAs, thus all AI accelerator card industry chains revolved around NVIDIA GPUs.

The difference between GPUs, FPGAs, and ASICs compared to CPUs mainly lies in the technical architecture. Traditional CPUs are not as fast as GPUs, FPGAs, and ASICs when processing massive data calculations.

GPUs are standard (general-purpose) computing chips, representing companies including NVIDIA, AMD, and Intel.

Compared to standard GPU chips, ASICs are custom chips, also known as dedicated chips, primarily designed for specific user requirements and scenarios; FPGAs, formally known as field-programmable gate arrays, are semi-custom chips that lie between GPUs and ASICs.

Broadcom's impressive performance comes from ASIC's AI-specific chips.

ASIC chips can also be divided into three categories: TPU, DPU, and NPU Given that the academic names of these three types of chips are quite convoluted, it's easier to summarize their respective uses: TPU is mainly used for machine learning; DPU provides computing engines for data centers (IDC).

NPU, this one should be familiar, right?

Huawei first launched the Da Vinci architecture in 2018, along with the AI accelerator card based on this architecture—Ascend 310; in 2019, Huawei launched the SoC chip integrated with the first NPU based on the Da Vinci architecture—Kirin 810.

The next generation of the "God U" Kirin 990 (the first SoC with an integrated 5G baseband launched by a Chinese company) also reached the global peak for Huawei smartphones due to its outstanding computing performance.

So what is the main "job" of the NPU? It is evident that the NPU is mainly used for processing data in deep learning instruction sets, which can be understood as primarily providing computing power.

Compared to GPUs, ASIC chips have stronger performance and lower energy consumption, but they have a high research and development threshold and a long design cycle, making them more expensive than GPUs. However, if mass-produced, the cost issue can be resolved.

As a result, the "new star" in AI, Broadcom, is experiencing a surge, while the previous "old-school" AI giant, NVIDIA, appears somewhat desolate: on December 16, NVIDIA's stock price hit a two-month low since October 15 during intraday trading.

Currently, the leading companies in the global ASIC chip market mainly include Google, Intel, and Broadcom, while companies like IBM, Samsung, Apple, and NVIDIA are also laying out plans for ASICs.

With ASIC chips, the data processing speed of data centers has correspondingly become faster.

However, this brings about a problem: the data transmission speed between the servers and workstations that make up the data center also needs to be synchronized to fully leverage the speed advantages of ASIC chips.

In other words, the demand for short-distance high-speed data transmission (in contrast, optical modules mainly address long-distance high-speed data transmission) will grow rapidly based on ASICs.

At this point, it's time for AEC (Active Electrical Cable) to make its appearance.

Domestic ASICs are weak, but AEC is acceptable

The reason it seems a bit verbose and has only now gotten to the main topic is that without the above background, it would be hard to understand why Wolong Nuclear Materials has become the subject of discussion.

AEC is a type of cable that differs from traditional common cables (wires) in that AEC is an active power cable.

Inside traditional cables, there may only be some metal (like copper), but AEC cables integrate some electronic components (such as signal amplifiers and equalizers) that can enhance signal transmission in addition to copper.

What is the role of this cable?

The core function is singular: to support high-speed data transmission. How large is the market size for AEC active cables?

According to Light Counting data forecasts, the global high-speed copper cable market will continue to expand at a compound annual growth rate of 25% during the period from 2023 to 2027; it is expected that by the end of 2027, the global shipment of high-speed copper cables will reach a scale of 20 million units The basic signal transmission path for AEC is copper cables; in addition, there is a core module specifically designed for connecting active copper cables, known as copper cable interconnect.

Compared to some other high-speed transmission solutions, the cost of copper cable interconnect is relatively low. When constructing data centers or other network systems, using copper cable interconnect for AEC can meet performance requirements while reducing overall construction and operational costs.

Previously, NVIDIA introduced at GTC 2024 that NVL72 significantly saves costs by using copper cable interconnect instead of optical modules.

Domestic A-share companies with AEC-related products mainly include: Walden Materials, Broadcom Technology, Zhaolong Interconnect, and Xinke Materials.

Walden Materials' 2024 semi-annual report shows the technical performance and application fields of the company's high-speed communication cables as shown in the figure:

On August 3, 2023, Walden Materials stated on Interactive Easy that "the 400G and 800G high-speed communication cables produced by its subsidiary Letin Intelligent Connection are DAC copper cables, used for data signal transmission in data centers, servers, and switches/industrial routers."

Walden Materials' AEC products are mainly produced by its subsidiary Letin Intelligent Connection, including SAS series, PCIe series, and QSFP series; major customers include Amphenol Corporation, Volex, Molex, and TE Connectivity.

What is DAC?

It is a type of dual-axis copper cable (passive copper cable) that requires corresponding connectors at both ends to achieve connection with devices.

Therefore, the AEC cable (active) copper connector is a key component for DAC to achieve device connection, and the two work together to complete data transmission between devices.

In addition to Walden Materials, Broadcom Technology also has high-speed cable products, which are divided into two types: Active Optical Cable (AOC) and Active/Passive Copper Cable (passive DAC, ACC, and active AEC):

Broadcom Technology mainly focuses on optical modules, and the company also has ACC (Active Copper Cable), which is also a type of active copper cable similar to AEC, but the latter has relatively more comprehensive performance, such as certain AEC performances being closer to Active Optical Cable (AOC).

Xinke Materials is a newcomer in the high-speed copper connection field: on September 28 this year, the company initiated an "Investment Cooperation Agreement for High-Speed Copper Connection Projects," just beginning to lay out high-speed copper connection projects.

Zhaolong Interconnect has a similar product layout breadth in high-speed cable business as Broadcom Technology: it offers high-speed cables (DAC and ACC), connectors, and connection modules