$NVIDIA (NVDA.US)$ The allure of ASICs (Application-Specific Integrated Circuits) is hard to resist.

After NVIDIA CEO Jensen Huang confirmed in June 2024 that NVIDIA would begin ASIC chip design, recent industry news revealed that NVIDIA has selected Taiwan as its ASIC design and research and development center and has started aggressive talent acquisition efforts.

This move has put chip design companies like MediaTek and ChipMOS Technologies on high alert to prevent NVIDIA from poaching staff.

ChipMOS Technologies, officially ChipMOS ELECTRONICS INC., is a leading global ASIC design company without its own fabrication plants, founded in 2003 and listed on the Taiwan Stock Exchange on October 28, 2014, with stock code '3661'.

Currently, NVIDIA's GB200 production is facing difficulties. According to the established plan, the GB300 is set to launch in Q3 2025, making the GB200 likely to become a transitional product, leading to market concerns about NVIDIA's performance; NVIDIA's competitors are also making full efforts to promote 'de-NVIDIA' strategies by implementing AI training using ASICs.

For example, $Apple (AAPL.US)$ In July 2024, the first preview version of iPhone AI will be released, followed by a paper stating that the AI model was trained on Google's TPU (Tensor Processing Unit). Recently, Apple announced at the AWS Reinvent Conference that it will use Amazon's self-developed ASIC AI Chip for model training. $Amazon (AMZN.US)$ At the AWS Reinvent Conference, it was announced that Amazon's self-developed ASIC AI Chip will be used for model training.

TPU is also a category of ASIC.

Therefore, the production of its own general-purpose AI accelerator card GB200 is not going smoothly, while competitors like $Broadcom (AVGO.US)$ not to mention, Amazon and $Alphabet-C (GOOG.US)$ Self-developed ASICs are also "poaching" NVIDIA's market, so Jensen Huang is feeling more anxious.

Recent market news indicates that NVIDIA's layout for ASICs is accelerating. In Taiwan, NVIDIA continues to recruit senior engineers in the ASIC field, including talent with front-end design verification, IP integration, and physical layer design experiences, with engineers being continuously poached.

In fact, whether it's Amazon, or $Alphabet-A (GOOGL.US)$ or $Meta Platforms (META.US)$ , or even Microsoft's Maia or Cobalt projects, there are **** ASIC chip design engineers participating in guidance.

As a result, this group of people has become the target of NVIDIA's poaching.

This is because NVIDIA offers RSUs (Restricted Stock Units), and from 2023 to 2024, NVIDIA's stock price increased by 239%+ and 180%+ respectively. Engineers who are usually poached typically receive RSUs, where the profit from the stock price increase far exceeds their salary.

With the GenAI (Generative AI) technology brought by ChatGPT, NVIDIA has elegantly transformed from a leader in gaming chips to a 'computing power lord' of GenAI.

This high level of market dominance has sparked a sense of 'crisis awareness' and 'unified resistance' among global technology giants, which is why these companies, including Cloud Service Providers (CSP), are seeking computing power sources for AI training that can serve as alternatives to general-purpose AI acceleration cards.

Clearly, these companies looking to band together have set their sights on ASICs, which lays a solid foundation for Broadcom's impressive profit performance for ASIC chips to be announced on December 12, 2024.

Naturally, NVIDIA cannot overlook such market performance and collective behavior.

Jensen Huang previously revealed his personal thoughts regarding the ASIC business: he believes that NVIDIA's push for ASIC business development will further expand the customer base. Although CSPs are seeking to escape the embarrassing situation of relying solely on NVIDIA, and even designing their own ASICs, if NVIDIA participates in ASIC design, then CSPs will still become NVIDIA customers through ASIC services.

Does this logic seem a bit convoluted? Jensen Huang is aware of this, so he provided three points of explanation.

Overall, Jensen Huang believes that whether it is a General AI Accelerator Card or ASIC, the systematic capabilities of NVIDIA lead to lower costs and a richness of ecosystems, which presents a clear advantage compared to CSP directly entering ASIC design.

Due to these two advantages, Jensen Huang is very confident about NVIDIA establishing an ASIC department.

Of course, there are different opinions in the industry; some believe that General AI Accelerator Cards and ASIC are not actually opposing or substituting each other, but rather complementary forms that enhance each other.

General AI Accelerator Cards (GPUs) have numerous processing cores, excel at large-scale parallel computing, and are particularly suitable for deep learning. Their production costs are relatively low and development time is short, making them suitable for diverse market demands.

And what about ASIC? This is custom-designed for specific tasks like AI inference, requiring extensive preliminary research, design, verification, and production, leading to higher initial costs and longer development cycles.

Alphabet and Amazon began ASIC development in 2013 and 2015 respectively, while Microsoft and Meta started relatively later, only beginning in 2019 and 2020.

In fact, before Broadcom released its stunning Earnings Report for the fiscal year 2024, the industry did not realize the computational support value of ASIC for GenAI technology, and the stock price increase of related stocks in the secondary market was essentially a realization of expected differences: the greater the difference, the more astonishing the stock price increase.

As a result, Broadcom achieved an astonishing increase of 24.51% the day after releasing its Earnings Report, propelling its market value to over 1 trillion dollars.

Broadcom is confident about the shipment volume of customized AI accelerator cards (ASICs), and the company expects that the shipment volume in the first quarter of fiscal year 2025 will double compared to the previous quarter.

In terms of A-shares, direct ASIC designers have little advantage, but in the field of active cables supporting short-distance data transmission for the IDC (Datacenter) consisting of ASICs, this direct mapping relationship will still be effective now and in the future.

Editor/Somer