Click the “blue text” above to follow the latest investment research news
Recently, many investors have noticed that Broadcom’s stock price has outperformed NVIDIA. Behind this lies a significant shift in the AI computing field: represented by Broadcom, ASIC (Application-Specific Integrated Circuit) chip manufacturers have quietly risen.
The global ASIC chip market is expected to reach approximately $12 billion in 2024, with projections to exceed $30 billion by 2027,achieving a compound annual growth rate of 34%.
In previous articles, we have discussed chip-related topics multiple times, including the fact that Shannon Semiconductor has seen a cumulative increase of over 100% since its article published on August 20. Interested investors can check past articles.
Tongfu Microelectronics: The “AMD leg” of the packaging and testing leader and the ambition for domestic substitution, with a 300% profit growth just the beginning?
Shannon Semiconductor’s joint venture layout in intelligent computing, with storage leaders welcoming new growth points
[Industry Decoding] The core links and key enterprises in chip manufacturing
Understanding the “storage chip” industry chain and key enterprises in one image
01 Major Players Betting: Why Are Tech Giants Turning to ASIC?
The competition for AI computing power has shifted from general-purpose GPUs to customized ASIC chips. Broadcom’s third-quarter financial report shows that its AI chip revenuesoared by 63% year-on-year, with the company securing a $10 billion custom AI chip order from its fourth major client.
Later media reports revealed that this mysterious client is OpenAI.
Cloud computing companies have ample reasons for their collective preference for ASICs:
- AWS’s Trainium2 completes inference tasks faster than NVIDIA’s H100 under the same budget,improving cost-effectiveness by 30% to 40%.
- Google’s seventh-generation TPU Ironwood supports 10MW-level liquid-cooled cabinets, with FP8 computing power surpassing NVIDIA’s latest B200 chip.
- Meta’s MTIA series ASIC plans to use a 170kW high-power liquid-cooled rack, optimized for short video recommendation algorithms.
A chip leader from a cloud company vividly explained this shift: “It’s like a takeout restaurant needing a dedicated wok during peak hours, rather than a home frying pan. General-purpose GPUs can do everything, but they are inefficient for specific tasks.”
02 Performance Comparison: What Makes ASIC Challenge GPU?
The core difference between ASIC and GPU lies in their design philosophy. GPUs are like multifunctional Swiss Army knives, capable of handling various scenarios but not extreme; ASICs are like custom kitchen knives, optimized for specific tasks, with far superior sharpness.
In the AI large model inference scenario, this difference is significantly amplified:
1. Computing Power Density: AWS Trainium2’s TOPS/W (TOPS per watt) is 40% higher than NVIDIA’s H100, allowing it to handle more requests under the same power consumption.
2. Cost Control: Google’s TPUv4 has a three-year total cost of ownership (TCO) that is 55% lower than that of GPUs, mainly saving on electricity and cooling.
3. Algorithm Adaptation: ASICs optimized for the Transformer architecture have 30% lower latency in natural language processing tasks compared to GPUs.
In the past, ASIC design cycles were long and upfront investments were high, with a chip taking 18 to 24 months from design to mass production, and the tape-out cost reaching tens of millions of dollars. However, recent technological advancements have compressedthe development cycle to 6 to 12 months, with costs reduced by over 60%.
03 Market Landscape: How ASIC Restructures the AI Chip Industry Chain?
According to analysis from Guojin Securities, by 2025, Google’s TPU shipments are expected to reach 1.5 to 2 million units, while Amazon’s AWS T2 is expected to reach 1.4 to 1.5 million units. In contrast, NVIDIA’s AI GPU supply will exceed 5 to 6 million units.
Currently, the total shipment of AI TPU/ASIC from Google and AWS has reached40-60% of NVIDIA’s AI GPU shipment. As Meta begins large-scale deployment of its self-developed ASIC solutions in 2026 and Microsoft starts large-scale deployment in 2027, it is expected that total ASIC shipments will surpass NVIDIA GPU shipments sometime in 2026.
The rise of ASICs not only changes the chip design field but also drives the development of the entire industry chain:
Key Links and Representative Enterprises in the ASIC Industry Chain
| Industry Chain Link | Main Content | Representative Enterprises | Technical Features/Market Position |
|---|---|---|---|
| Design Side | ASIC chip design | Broadcom, Cambricon, Rockchip | Broadcom holds 60% market share, Cambricon’s revenue is expected to increase by 4347% in H1 2025 |
| Manufacturing Side | Chip manufacturing | SMIC | 14nm FinFET process supports over 50% of domestic ASIC demand |
| Packaging Side | Chip packaging and testing | JCET, Tongfu Microelectronics | 2.5D packaging technology increases interconnect bandwidth by 3 times |
| Supporting Side | Liquid cooling/Optical interconnect | Inspur, Zhongji Xuchuang | The value of liquid cooling is 5 times that of air cooling |
04 Investment Logic: What Investment Opportunities Exist in the ASIC Sector?
For investors, the investment logic in the ASIC field can be summarized as “three considerations”:Look at order visibility, look at technical barriers, look at supporting flexibility.
Design Side Opportunities
ASIC design is the core link of the industry chain and also the area with the highest technical barriers. It mainly consists of three types of players:
- IDM Giants: Broadcom, with its accumulation in high-speed interfaces, holds 60% market share in data center interconnect scenarios with its XPU products.
- Cloud Vendors’ Self-Development: Amazon, Google, etc., act as both demand and supply sides, with self-developed ASICs accounting for 25% of their computing power procurement in 2024.
- Specialized Design Companies: Domestic companies like Cambricon and Rockchip focus on vertical fields.
Cambricon, as a representative domestic ASIC enterprise, achieved a revenue of 2.881 billion yuan in the first half of 2025,an astonishing increase of 4347.82%, with a net profit of 1.038 billion yuan, compared to a loss of 530 million yuan in the same period last year.
Supporting Industry Chain Opportunities
The growth of these supporting links often outpaces that of ASIC chips themselves. The cost share of liquid cooling systems in ASIC servers accounts for 15% to 20%, three times that of ordinary servers.
The value of optical interconnect components also grows exponentially with the increase in computing power, forming a positive cycle of “ASIC volume → power consumption increase → supporting upgrades”.
05 Risk Warning: Potential Risks of ASIC Investment
Despite the broad prospects for ASICs, investors should also pay attention to the following risks:
1. Technology Iteration Risk: AI algorithms change rapidly, and if future algorithm paradigms undergo a dramatic shift, the ASICs already invested may not adapt to new demands as quickly as GPUs.
2. Ecological Dependency Risk: Some companies depend on specific technological ecosystems (e.g., Haiguang Information’s reliance on the x86 instruction set), and caution is needed regarding the loosening of compatibility barriers.
3. Capacity Expansion Risk: Industries such as Low-Dk electronic cloth/copper foil may face risks of rapid capacity expansion, affecting the supply-demand balance in the industry.
4. Capital Expenditure Below Expectations Risk: If large tech companies reduce capital investment in ASICs, it will directly impact the development speed of the entire industry chain.
06 Future Outlook: How Will ASIC Reshape the AI Computing Landscape?
Broadcom CEO Chen Fuyang predicts that the global AI ASIC market will reach$60-90 billion in the 2027 fiscal year. Marvell has also revised its forecast for the global ASIC market in 2028 to $55.4 billion.
In the future, the AI chip industry will form a hybrid architecture of “GPUs for training, ASICs for inference.” The training side will be dominated by GPUs due to the rapid iteration of algorithms and the variety of model types, which require high flexibility and maturity of the chip ecosystem.
On the other hand, the inference side will increasingly adopt ASIC solutions, as in AI inference scenarios, when algorithms converge and models stabilize, ASICs can be customized once and reused long-term, effectively spreading non-recurring expenses as deployment scales up, resulting in a total cost of ownership per unit of computing power that is far superior to GPUs.
Google has already engaged with major small cloud service providers renting NVIDIA chips to discuss hosting Google’s AI chips in their data centers simultaneously. This indicates that ASICs are transitioning from self-use to commercial use, further expanding market space.
The future is here. While NVIDIA continues to dominate the AI training field, ASICs are paving a more cost-effective path in the inference side. Tech giants like Google, Amazon, Meta, and Microsoft are all building their own ASIC solutions.
Broadcom holds$100 billion in orders, with an additional $10 billion order locked in for delivery in 2026. These figures not only signify orders but also herald the structural changes occurring in the AI computing market.
Will investors continue to bet on the king of general computing or embrace the new nobility of dedicated computing? The answer may determine the greatest dividends in AI investment over the next few years.
Disclaimer
This content is provided by Chengqi Investment Research.
The content of this article does not constitute any investment advice, guidance, or commitment. All opinions are for reference and learning purposes only. Any mention of specific stocks is for case analysis only; investors should make their own decisions!
Relevant information is sourced from publicly available data, company financial reports, industry research reports, institutional research, official media, etc. If there are any discrepancies, please refer to the latest official information, and corrections are welcome.
The stock market has risks; investment requires caution!
Share valuable investment research information to help investors achieve long-term gains 📈
Click the link below to follow ⬇️⬇️
Chengqi Investment Research