Today I saw a report mentioning that by 2030, 75% of future AI computing demands will come from inference, which will create huge opportunities for the second half of AI investment. Let's briefly discuss it below. 🧐

In the past two years, everyone has been crazily hoarding GPUs to train models, burning money to stack computing power. But recently, have you noticed that the situation has changed? Especially after the emergence of the small lobster 🦞, various Agents have exploded, with token demand increasing exponentially. A brand new era of inference is coming, and the biggest narrative point for AI in the future is to bring AI applications into the hands of billions of users worldwide and make real use of them!

This turning point, I call the switch from the "training era" to the "inference era." In this new stage, CPUs and custom chips will be the real protagonists of the future.

👇 Let’s break it down and talk about why I’ve started focusing on these previously inconspicuous guys:

🎯 AI now has a "brain."

Previously, AI was just a tool; I asked it questions, and it answered, but it couldn't perform specific tasks. Now things are different; Agentic AI has arrived, and it can help me book flights, edit documents, write code, and even plan entire workflows.

This kind of work requires a lot of logical judgment and sequential arrangement. Previously, GPUs were the musclemen, excelling in heavy lifting; but now, for these execution tasks, CPUs are the commanders, skilled at handling complex logic. The smarter AI becomes, the more it needs CPUs to orchestrate and direct.

🎯 Big companies also need to save money.

Looking at the financial reports, major companies' capital expenditures in the AI field are enormous. Among the seven giants, excluding Apple, the total capital expenditure for the whole year of 2026 exceeded 600 billion US dollars. Don’t think that giants like Google and Meta are short on cash. On the contrary, they also spend money wisely. If they continue using Nvidia's general-purpose GPUs for inference, electricity and hardware costs will eat them alive.

Thus, big companies have started to design their own custom ASIC chips, just like we use a kitchen knife for cutting vegetables and an axe for chopping wood; specialized chips for specialized tasks, which are more efficient and cost-effective. This is a tangible reduction in costs and an increase in efficiency.

Based on the above logic, we have filtered five core targets:

• #AMD: An all-rounder in the inference field, the GPU MI300/325 series has industry-leading memory bandwidth (256GB HBM3E memory, bandwidth reaching 6TB/s, outperforming Nvidia's H200), capable of easily handling real-time inference for ultra-large-scale LLMs, and it's fast. The key point is that it is also the second in server CPUs; when agent-based AI needs strong CPUs, AMD benefits on both ends.

• #ARM: The king of underlying architecture, Google and Amazon’s self-developed chips are basically running on ARM architecture; the ARM Neoverse platform has become the preferred choice for over 90% of self-developed AI server CPUs. So as big companies develop in-house, ARM is like sitting back and collecting tolls.

• Broadcom (#AVGO): The chief engineer of custom chips, as an important partner of Google TPU, Meta MTIA, and ByteDance, Broadcom has absolute dominance in high-speed interconnect and packaging technology, making it the preferred partner for cloud giants to reduce inference costs. Big companies wanting to save money and develop in-house cannot do without Broadcom's technology.

• Marvell (#MRVL): Deeply involved in chip research and development for Amazon AWS and Microsoft Azure, these two giants must rely on Marvell to break away from dependence on Nvidia. Currently, Marvell's custom silicon business is transitioning from a "research and development phase" to an "explosive harvest phase," and future business will only grow larger.

• Intel (#INTC): Although it has been facing difficulties recently, it is the only designer with its own wafer fabrication plant and possesses the globally scarce IDM 2.0 foundry capacity. As inference drives a surge in global CPU demand and supply chain tensions, Intel's internal synergy and domestic manufacturing advantages will gradually become prominent, and it is still the leader in server CPUs.

After reading the entire report, I personally believe several major trends will become inevitable:

First, the shift from the previous "brute force computing" to future "refined operations." Currently, the entire AI investment logic is changing; it's no longer about which chip is the strongest, but whose solution is the cheapest and most practical. This indicates that under the true commercialization and full implementation of AI, energy efficiency ratio is the key.

Secondly, the competition in the supply chain. Major cloud giants are developing their own chips, and the future AI hardware market will shift from "Nvidia's dominance" to "a multitude of contenders." I believe this will be a good thing for the entire industry chain.

Finally, electricity will still become a long-term bottleneck. Although CPUs and ASICs are optimizing efficiency, the overall explosion in computing power still leads to an infinitely large demand for electricity. While keeping an eye on chips, the energy sector also deserves attention.

In summary, in the second half of #AI, don’t just focus on the most expensive GPUs; those CPUs that are responsible for command and those ASICs that help big companies save money might hold the next wave of opportunities.

Currently, the companies mentioned above are basically on the #MSX platform. For trading U.S. stocks, I choose to use the #RWA U.S. stock tokenization platform #MSX to invest and participate in the U.S. stock market: http://msx.com/?code=Vu2v44

Early U.S. stock investment fans and partners can private message me. After filling out the form, you can enter our U.S. stock communication and discussion community for free (currently limited to 10 people per week, approved by assistants, may take some time, thank you 🙏)!

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。