● AI Hype Shift

The Real Reason AI Semiconductors Are Hitting New Highs: The More Important Story Than Marvell, AMD, and Nvidia Is ‘Network, Power, and CPU’

The real core point to watch in today’s market is not simply that “AI semiconductors rose again.”
U.S. stocks are hitting new highs even though Treasury yield pressure remains heavy, and at the center of that move are AI infrastructure companies such as Nvidia, Micron, Marvell, AMD, and Broadcom.
In particular, Marvell suggested that future AI-related revenue could rise explosively, and AMD continues to see talk of higher price targets on expectations that CPU demand will increase in the era of AI agents.
On top of that, Google signed a 890MW power contract with nuclear company Constellation, bringing data center power infrastructure back to the center of the market.

In this article, let’s review U.S. stock highs, the AI semiconductor rally, Treasury yield risks, data center power demand, and nuclear investment trends all at once.
And we will separately highlight the core point that is relatively undercovered in other news and YouTube coverage: “The next bottleneck after GPUs is network and power.”

1. Today’s core point in U.S. stocks: prices are hitting highs even though rates are high, and why that combination is uncomfortable

U.S. stocks generally rose around 0.5% today.
The S&P 500 and Nasdaq continued their high-price trend, and on the surface there was strong risk-on sentiment.

But there is an uncomfortable point.
The U.S. 10-year Treasury yield is still stuck in the upper 5.2% range.
In the original text, the 10-year yield was mentioned at around 5.269%, and the area near 5.3% is a level the market watches very closely.

Normally, when Treasury yields are this high, they weigh on growth and technology stocks.
Because the present value of future earnings falls, valuation pressure builds especially on names like AI semiconductors that depend on high growth expectations.

Even so, U.S. stocks are rising for a simple reason.
Earnings growth is overpowering rate pressure.
In particular, earnings growth in technology, energy, and AI infrastructure companies is lifting the overall market.

• The 10-year Treasury yield remains at a high level.
• But GDP growth expectations and rising corporate earnings are partly justifying high rates.
• Some easing in global bond yields, including in Germany and France, has temporarily reduced rate pressure.
• Still, because the market saw major turbulence around the 5.3% area in 2023, this is not a zone to become complacent about.

2. Only a handful of Nasdaq 100 names are at highs: this is not a market rising broadly

The most important part of this rally is market breadth.
It was said that only about 6 names in the Nasdaq 100 are hitting new highs.
And the common trait among those names is that most are AI infrastructure-related companies.

In other words, the current U.S. stock market is not a “everyone rises together” market.
It is a market where only companies that are actually generating revenue and cash flow from AI infrastructure are moving strongly.

This part is really important.
If you only look at the index, the U.S. market seems easy, but in reality stock selection difficulty has increased significantly.
It is not a market where everything with the AI label rises; only companies that are actually benefiting from the data center investment cycle are rising selectively.

• Companies with actual AI infrastructure revenue growth are strong.
• Companies supported by free cash flow, or FCF, are being chosen by the market.
• Companies that only have an AI theme but no earnings are likely to be left behind.
• This is a phase where the gap between index gains and perceived returns widens.

3. Nvidia and Micron concentration: what it means that one-quarter of AI growth comes from these two companies

It is also important to note how large the share of AI infrastructure growth held by Nvidia and Micron is.
The original text says that nearly one-quarter of AI infrastructure-related growth comes from Nvidia and Micron.

This is both good news and risk.
It is good news because it means the earnings growth scale of those two companies is overwhelming.
On the other hand, it is a risk because it means the market’s rise is concentrated in too few companies.

In particular, Nvidia and Micron are already large-cap companies.
The fact that such large companies are still responsible for a significant portion of earnings growth shows how strong the AI infrastructure cycle is.
But at the same time, it is a burden that the whole market may be overly dependent on the earnings expectations of these two names.

• Nvidia is the key supplier in the AI GPU market.
• Micron is a core beneficiary of the HBM and memory cycle recovery.
• The scale of EPS growth at these two companies is so large that it can distort the earnings growth rate of technology stocks overall.
• For the AI semiconductor rally to stay healthy, it needs to broaden to Broadcom, Marvell, AMD, TSMC, and networking companies.

4. Are technology valuations really expensive? In fact, P/E is compressing

This year, the IT sector’s stock price is said to have risen about 22%.
But net income growth is faster, at around 57%.

The meaning is simple.
Earnings are growing faster than stock prices.
So even though technology stocks look as if they have surged a lot, valuations are actually moving lower.

The original text says that the technology sector’s P/E change rate has fallen about 21%, and that the current technology sector P/E is about 19% below its 5-year average.
This is a point many people miss.

In other words, even though AI semiconductor stocks are near new highs, it is hard to say valuation is automatically a bubble.
That is because earnings are growing that quickly.

• Stock prices have risen, but EPS has increased even faster.
• High rates and oil prices are weighing on technology valuations.
• If rate pressure eases, technology stocks could rebound strongly again.
• However, if earnings fall short of expectations, profit-taking could be severe.

5. Marvell Investor Day: the next bottleneck after GPU is ‘connectivity’

One of today’s main characters is Marvell.
Through Investor Day, Marvell strongly presented growth expectations tied to AI infrastructure.
The original text mentions expectations that revenue could increase more than fourfold over the next few years, along with the possibility of even larger long-term growth.

Marvell’s core is not just semiconductors.
Marvell is a company with strengths in ASICs, networking chips, retimers, and data center connectivity solutions.
If Broadcom is the representative company in custom semiconductors, Marvell can be seen as a strong runner-up following behind it.

So far, the biggest bottleneck in AI infrastructure has been GPUs.
The next bottleneck was HBM and memory.
But going forward, network and connectivity are likely to become the new bottleneck.

Inside AI data centers, countless GPUs, CPUs, memory modules, storage devices, and server racks must be connected to each other.
Chips must connect to chips, racks to racks, data centers to data centers, and one data center to another.
In this process, demand is growing both for copper-based connections and optical communication-based connections.

• Scale-up is the area that connects GPUs and CPUs within a rack.
• Scale-out is the area that connects multiple racks to build a large cluster.
• Scale-across means connecting across data centers.
• In all of these stages, demand for network chips, optical transceivers, and retimers increases.

Personally, I think the core keyword for AI infrastructure after 2027 is likely to be “network.”
The market already knows a lot about GPUs and memory, but the connectivity bottleneck still has a relatively weak mainstream investment narrative.
That is why it may actually be the more important phase.

6. AMD price target raised: CPUs matter again in the era of AI agents

AMD is also showing very strong momentum lately.
It was mentioned that Citi raised AMD’s price target from 575 dollars to 800 dollars.
The core reason is the expectation that CPU demand, not only GPU demand, could strengthen again.

Until now, the center of AI investment has been GPUs.
But when the AI agent era arrives, CPU demand can grow substantially.
Personal AI agents handle user requests, browse, make reservations, shop, create documents, and connect multiple apps.
That process requires large-scale data center CPU resources.

For example, Meta’s personal AI agent service reportedly surpassed 5 million users, with a goal of 100 million within a year.
If each user is allocated between 2 and 8 virtual CPU cores, the number of required CPU cores rises enormously.

If personal AI agent services from OpenAI, Meta, and xAI’s Grok spread widely, the CPU market could grow much faster than previously expected.
The original text suggests it would not be surprising for the CPU market to grow at a 30% annual rate through 2030.

• AMD sells both CPUs and GPUs.
• The strategy of supplying CPUs and GPUs together for AI servers could become stronger.
• The spread of AI agents stimulates data center CPU demand.
• AMD’s expected EPS growth rate is described as very high, and its valuation is seen as attractive relative to growth.

7. Semiconductor valuation comparison: the differences between AMD, Nvidia, Broadcom, and Marvell

The original text compares EPS growth rates and expected P/Es for major AI semiconductor companies.
The core point is that a stock is not automatically expensive just because it has risen; it must be viewed together with earnings growth.

Company	Core Point	Market Interpretation
AMD	Dual benefit from CPUs and GPUs, expectation of AI agent expansion	Strong view that it is attractive relative to growth
Nvidia	Dominant position in AI GPUs, overwhelming absolute EPS growth	Multiple compression despite the stock’s rise
Broadcom	Custom ASICs, networking, and big tech AI chip benefits	There is still a valuation case relative to earnings growth
Marvell	Beneficiary of ASICs, networking, retimers, and connectivity bottlenecks	A key candidate for the network cycle after 2027
Arm	Low-power architecture and AI edge expectations	Valuation burden exists because P/E is high
TSMC	Dominant position in advanced foundry manufacturing	Stable, but upside depends on the equipment cycle
Intel	Challenges in foundry transition and restoring CPU competitiveness	Investor sentiment is relatively weak

In particular, AMD is argued to be the most attractive on a growth-adjusted valuation basis even after its stock has risen substantially.
Broadcom and Nvidia were also companies that in the past received 30 to 40 times multiples, and as earnings rise quickly their valuations are compressing instead.

8. Optical communication stocks surge: why Ciena, Fabrinet, and Lumentum moved

When Marvell emphasized connectivity as the next bottleneck in AI infrastructure, optical communication-related stocks reacted strongly as well.
Ciena is a company with strengths in DCI, or data center interconnect.
Fabrinet is an optical transceiver assembly company, and Applied Optoelectronics and Lumentum are classified as companies related to optical sources.

This field is difficult for ordinary investors to explain.
It is not intuitive like GPUs, because it mixes electrical, electronic, and optical concepts.
That is why public attention is weak, but it is an area absolutely necessary for AI data center expansion.

• Ciena is a beneficiary of DCI, which connects one data center to another.
• Fabrinet plays an important role in the assembly and production of optical communication components.
• Lumentum is highly regarded from the perspective of optical source technology.
• As AI cluster size grows, optical communication demand increases structurally.

9. Meta’s AI strategy: inference monetization may be more important than cloud

Wells Fargo reportedly set a 1,000-dollar price target for Meta.
But the stock did not react much.
That is because it had already risen significantly.

Since early this year, a common burden for big tech has been massive AI capital spending.
In the process of building data centers, buying GPUs, and securing power, FCF is drained heavily.
That is why the argument has been strong that companies supplying AI infrastructure are more favorable than companies consuming AI.

That said, Meta is a company that can be viewed relatively positively among big tech.
That is because once recommendation algorithms, advertising, commerce, and personal AI agents are connected, there is a strong chance of monetizing inference costs.

There was talk that Meta has secured 10GW of data center power, and the market became interested in whether excess computing resources would be sold to the cloud.
But in reality, the more likely uses appear to be its own AI inference services, personal agents, ad efficiency improvements, and better commerce conversion rates.

• Data centers take more than about two years to build.
• A full-scale AI investment recovery phase could begin starting in 2028.
• Meta’s core goal is not to become a cloud company, but to connect AI inference to advertising and commerce.
• If AI agents improve user experience, Meta’s monetization ability could become even stronger.

10. Memory and storage face short-term pullbacks: Micron and Seagate issues

Memory and storage companies were relatively weak today.
Micron was weak after news emerged that it would pay a 600 million dollar settlement to its competitor Netlist.
By contrast, Netlist surged sharply.

Seagate fell as the possibility of Toshiba expanding supply in the hard drive market became a burden.
Hard drive companies still have a long-term hope tied to AI data center storage demand, but in the short term they are reacting sensitively to news of supply growth.

• Micron is a core beneficiary of the HBM and memory cycle.
• However, litigation and settlement issues weigh on the stock in the short term.
• Seagate and hard drive companies reacted weakly to supply expansion news.
• AI storage demand remains valid over the long term, but short-term volatility can increase.

11. AI security is a quiet but powerful growth sector

The software sector is generally mixed, but the security sector should be viewed differently.
As AI agents begin to be used as attack tools, cyber security demand is structurally increasing.

The original text says that the cost of hacking attacks using AI could fall to around 2 dollars.
By contrast, defense costs could rise to around 40,000 dollars per month.
The structure is one in which attack costs collapse and defense costs explode.

This is a very scary change for companies.
In the past, hackers had to code and infiltrate directly, but now AI agents can find vulnerabilities and automate attacks.

• Automated AI attacks expand the cyber security market.
• Security budgets at financial institutions, large corporations, and public agencies are likely to increase further.
• Security software is showing stronger differentiation than general software.
• In the AI era, “falling attack costs” becomes a core investment thesis for security spending.

12. Google and Constellation’s 890MW contract: the AI data center power war has begun

Google signed a 890MW power contract with Constellation Energy, one of America’s leading nuclear companies.
The contract period is mentioned as 20 years.

Since 890MW is smaller than 1GW, it may feel like it is not that big a deal.
But in data center power contracts, it is a very large scale.
It is especially meaningful because stable baseload nuclear power was secured for the long term.

AI data centers must run 24 hours a day without stopping.
Solar and wind alone have limits in providing stable power supply.
That is why power infrastructure such as nuclear, natural gas, fuel cells, small modular reactors, and onsite generation is gaining attention again.

• Google is expanding long-term power contracts to operate AI data centers.
• Constellation is a representative company in U.S. nuclear infrastructure.
• The U.S. Department of Energy showed support for nuclear power by offering Holtec a conditional 4.2 billion dollar loan.
• This can be interpreted as a signal that the U.S. government is strategically linking AI infrastructure and nuclear power.

13. The real core of power infrastructure: behind-the-meter onsite generation, not the existing grid

When looking at power infrastructure, it is not enough to think only about utilities.
The core point emphasized in the original text is BTM, or Behind The Meter.
In simple terms, this means onsite generation that produces electricity right next to the data center.

AI data centers do not have time to wait for grid connections.
Expanding transmission lines takes a long time for permits and construction.
That is why directly attaching generation facilities next to data centers is becoming increasingly important.

Companies mentioned as 대표 examples in this area are NuScale Power and Bloom Energy.
NuScale has expectations tied to small modular reactors, and Bloom Energy is drawing attention as an onsite generation solution based on fuel cells.

• Existing grid companies are stable, but their growth pace may be limited.
• BTM onsite generation can solve the speed problem of AI data centers.
• Bloom Energy has risen sharply this year, but structurally the demand case remains intact.
• Power infrastructure could become the next key investment theme after AI semiconductors.

14. The Fed and rates: weaker expectations for an October rate cut are a burden

On rates, the possibility that the Fed will slow its pace has increased.
The probability of an October rate cut is said to have fallen from around 70% to the 25% range.

Fed officials are sending the message that there is no need to rush right now.
A cut by year-end is still possible, but it may not be the rapid easing the market was hoping for.

Two-year Treasury yields are sensitive to Fed policy.
The original text says the 2-year yield fell from 4.83% to around 4.785%.
The fact that short-term rates have stabilized somewhat is positive, but the fact that the 10-year yield remains high is a burden.

• If expectations for Fed rate cuts weaken, it is a burden for growth stocks.
• But if earnings are strong, stocks can hold up even in a high-rate environment.
• The core question is whether earnings growth is faster than rates.
• In the next earnings season, AI infrastructure companies must meet expectations.

15. Next week’s earnings season: if companies fail to beat expectations, profit-taking could follow

The market’s attention is now moving to earnings season.
The original text says that about 93% of companies that have recently reported earnings beat expectations.

The problem is that expectations are already high.
When stock prices have risen a lot, “good earnings” may not be enough.
The market wants “very good earnings” and “even stronger guidance.”

In particular, AI semiconductors, data centers, and power infrastructure companies already have a lot of expectations priced in.
Therefore, in earnings releases, revenue growth, EPS growth, backlog, CapEx outlook, and power procurement plans will all be checked carefully.

• If earnings beat expectations, the AI infrastructure rally could continue.
• If earnings are good but guidance is weak, profit-taking could emerge.
• If rates and oil prices both rise, technology multiples could come under pressure again.
• It will also be important to see how much oil price increases are reflected in the consumer price index.

The most important point that is less discussed in other news

First, the next bottleneck in AI infrastructure is likely to be network, not GPU.
Many people only watch Nvidia GPUs, but as real data centers get bigger, the problem of connecting chips to chips, racks to racks, and data centers to data centers becomes even larger.
That is why companies like Marvell, Broadcom, Ciena, Fabrinet, and Lumentum could receive not just a thematic boost but structural benefits.

Second, CPUs matter again in the era of AI agents.
Until now, AI investment has been centered on GPUs, but when personal AI agents become mainstream, vast numbers of users will need virtual CPU cores assigned to them.
This trend could be a very important opportunity for AMD.

Third, the final bottleneck for AI data centers is power.
Google’s 890MW nuclear power contract is not just a simple power deal; it is a signal that securing stable energy has become a core asset in the competition for AI dominance.
Going forward, it is likely to become difficult to separate AI semiconductor investment from nuclear investment, power grid investment, and onsite generation investment.

Fourth, the market looks good, but in reality the rally is very narrow.
The fact that only a tiny number of Nasdaq 100 names are at highs shows how selective this market is.
The gap between companies that are actually making money in AI infrastructure and those that are not may widen further.

Fifth, cyber security is becoming a hidden essential consumer good of the AI era.
If AI agents drive attack costs down dramatically, companies will find it difficult to reduce defense spending.
Unlike general software, security stocks can have a structural growth thesis.

< Summary >

U.S. stocks are continuing to hit new highs thanks to expectations for AI infrastructure earnings growth even though Treasury yields are high.
But the rally is not broad, and it is concentrated in a small number of AI infrastructure names such as Nvidia, Micron, Marvell, AMD, and Broadcom.
Marvell emphasized that the bottleneck after GPUs is network and connectivity, which is an important signal for optical communication and data center interconnect companies.
AMD is benefiting from expectations that CPU demand could rise again as AI agents spread.
Google’s 890MW nuclear power contract shows that securing power has become a key variable in the competition over AI data centers.
Going forward, investors should watch not only AI semiconductors, but also network, power infrastructure, nuclear power, onsite generation, and cyber security.

[Related Articles…]

• AI Semiconductor Supercycle and Data Center Investment Trends
• AI Data Center Power Demand and Nuclear Infrastructure Investment Strategy

*Source: [ 월텍남 – 월스트리트 테크남 ]

– AI반도체 신고가 ..마벨 AMD “매출 4배 증가”ㄷㄷ

---