AI, Chip, Supply, Chain, Battle, Nvidia, AMD, Broadcom, Marvell, Surge

● AI Chip Supply Chain BATTLE AMD Broadcom Marvell vs Nvidia

The AI semiconductor investment landscape is changing: the 700 trillion won order war of AMD, Broadcom, and Marvell that you miss the moment you only watch Nvidia

What matters now in the AI semiconductor market is not the obvious story that “Nvidia is number one.”

The real key point is that even though Nvidia dominates the AI chip market, AMD and Marvell have moved far more strongly in stock performance over the past year.

Based on the original text, Nvidia rose about 26%, while AMD rose about 189% and Marvell about 221%.

To understand why this is happening, you need to see how GPU, CPU, ASIC, switch chips, HBM, and CoWoS packaging are connected inside AI data centers.

In particular, AMD is being linked to data center investment demand from mega-scale customers such as OpenAI, Meta, and Anthropic, with order expectations reaching the 600 trillion to 700 trillion won range.

Broadcom and Marvell are growing not by directly replacing Nvidia GPUs, but by becoming core components that solve bottlenecks in custom AI chips designed by big tech and in data center networks.

Today’s core takeaway is simple.

AI semiconductor investment is no longer a game of choosing one GPU company; it has become a game of identifying who controls the bottlenecks across the entire AI data center supply chain.

1. Why AMD and Marvell rose more in an Nvidia-dominated market

Nvidia’s position in the AI semiconductor market is still overwhelming.

The original text explains that Nvidia sells about 90% of all AI chips.

But if you look only at stock performance, an unexpected picture appears.

Nvidia: about 26% rise over the past year
AMD: about 189% rise over the past year
Marvell: about 221% rise over the past year

Nvidia’s 26% gain is not a small return.

But AMD and Marvell posted much higher gains.

This difference means the market is reacting more sensitively to who will see explosive volume growth in the future than to who is currently number one.

The stock market prices in future revenue growth and profit growth before current market share.

So the most important question in AI semiconductor investing is this.

“Who will actually build more chips next year, and who will win more data center orders?”

2. To understand AI semiconductor companies, you first need to know fabless and foundry

Nvidia, AMD, Broadcom, and Marvell are all fabless companies.

Fabless refers to companies that design semiconductors without owning their own manufacturing plants.

“Fab” means semiconductor factory, and “less” means without.

In simple terms, fabless companies draw the semiconductor blueprint and outsource actual production to other companies.

So where are these chips made?

Production is typically handled by foundry companies such as TSMC in Taiwan.

A foundry is a factory that manufactures semiconductors designed by its customers.

That is why it is very important to look at how many orders an AI semiconductor company placed with TSMC when estimating next year’s results.

Order volume becomes a leading indicator of next year’s shipments.

3. The real bottleneck in the current AI semiconductor supply chain is CoWoS

One of the most important keywords in AI semiconductors these days is CoWoS.

CoWoS is TSMC’s advanced packaging process.

It is a core process that attaches AI GPUs, HBM memory, and other semiconductors onto a single substrate so they can operate at high performance.

In the past, improving semiconductor performance meant the chip’s own process node was the most important factor.

But in the AI era, how efficiently GPU and HBM are connected plays a huge role in performance.

That is why CoWoS has emerged as the key bottleneck in back-end AI semiconductor manufacturing.

Every AI chip has to go through this process, but production capacity is limited.

In the end, companies like Nvidia, AMD, Broadcom, and Marvell are lining up to secure TSMC’s CoWoS capacity.

This creates an important investment point.

A company that secures a lot of CoWoS capacity has that much product to sell next year.

In particular, the company that stands out most in Morgan Stanley’s tracked CoWoS wafer allocation is AMD.

4. Why AMD surged: CoWoS volume rises from 130,000 wafers to 530,000 wafers

According to the original text, AMD is expanding from about 130,000 CoWoS wafers this year to about 530,000 next year.

That is nearly a fourfold increase.

In an AI semiconductor supply chain where everyone struggles to secure CoWoS, AMD increasing volumes by this much is highly significant.

It can be interpreted not merely as “we can make them,” but as “there are already places to sell them.”

For a semiconductor company to secure massive production capacity in advance is a strategy only possible when visibility into large customer orders already exists to some extent.

That is why AMD’s stock rise is hard to explain by simple optimism alone.

The market has begun to see AMD as capable of taking part of Nvidia’s dominant structure.

5. AMD’s core weapon: AI data center racks combining GPU, CPU, and networking chips

AMD is a company that competes directly with Nvidia.

It sells GPUs for AI computation and CPUs that serve as the brain of the computer.

The recent shift in the AI semiconductor market is moving away from selling chips one by one.

Now the trend is increasingly toward selling at the data center level and rack level.

AMD is preparing a rack-scale AI system called Helios.

Based on the original text, one rack the size of a refrigerator contains 72 AI GPUs.

It also contains 18 CPUs.

It additionally includes networking chips that allow chips to exchange data.

AMD’s Pensando chips handle this networking role.

In other words, AMD is not just selling GPUs; it is pursuing a strategy to sell an AI data center package that bundles GPU, CPU, and networking chips together.

This is similar to the strategy Nvidia has been pushing strongly since the Blackwell series.

It means the AI infrastructure market is shifting from competition among individual parts to competition among entire systems.

6. AMD’s point of differentiation versus Nvidia: HBM capacity

One of AMD’s differentiating points is memory.

AI GPUs are paired with HBM, or high-bandwidth memory.

As AI models get larger, memory capacity and bandwidth become just as important as GPU compute performance.

According to the original text, each AMD GPU is equipped with 432GB of HBM4.

By contrast, Nvidia’s next-generation Rubin chip is mentioned at around 288GB.

On a simple comparison, AMD is structured to carry about 1.5 times more HBM.

In a situation where HBM supply is already tight, securing more memory can become a powerful competitive advantage for AMD.

As memory bottlenecks grow in AI model training and inference, HBM capacity becomes an increasingly important differentiator.

7. The background behind AMD’s order explosion: mega-scale data center investment from OpenAI, Meta, and Anthropic

The reason AMD can confidently raise volumes is that large customers have appeared.

The original text explains that OpenAI and Meta each began orders on a 6GW scale.

6GW can be viewed as roughly the power output of six nuclear power plants.

From a data center perspective, that means building several massive AI data centers.

OpenAI’s 6GW plus Meta’s 6GW equals 12GW in total.

Based on the original text, AMD chip cost per 1GW is mentioned at about 4 billion dollars, or roughly 6 trillion won.

Nvidia’s Vera Rubin-based infrastructure is mentioned at about 8 trillion won per 1GW.

AMD can be seen as relatively cheaper, but when multiplied by 12GW, the amount becomes enormous.

According to this calculation, AMD is interpreted as already having order expectations on the scale of roughly 600 trillion to 700 trillion won.

If you add the detail that Anthropic is placing an additional 2GW order, AI data center investment could grow even further.

This is the strongest background behind AMD’s stock surge.

In AI semiconductors, it is no longer just a matter of “there is demand.” It is an era in which data center investment is calculated in power units.

8. The spread of AI assistants could explode CPU demand

Another positive catalyst for AMD mentioned recently is CPUs.

In the past, AI investment was all about GPUs.

But as AI assistants spread, CPU demand could also increase significantly.

The original text says Meta released its AI assistant “Muse” on the 8th of last month, and that it is ranking number one in U.S. downloads while user activity is rising rapidly.

These AI assistants must keep working in the cloud even after the user closes the app.

They open browser windows, organize documents, handle reservations, and operate multiple services on the user’s behalf.

In simple terms, it is like a personal secretary running 24 hours a day in the cloud.

These tasks do not use only GPUs.

In fact, tasks such as user session management, browser execution, general computation, and service connections require a lot of CPUs.

So as AI assistants become mainstream, CPU demand can surge.

In the past, it was common for one CPU to be paired with eight GPUs, but in the future the CPU-to-GPU ratio could move closer to one-to-one.

This shift is very important for AMD.

That is because AMD is a company with strong competitiveness not only in GPUs but also in server CPUs.

Because of this trend, the original text says AMD has, for the first time, surpassed a 1 trillion dollar market cap.

9. Broadcom makes money differently from Nvidia: custom AI chip ASICs

Broadcom grows in a somewhat different way from Nvidia and AMD.

It is easy to understand if you think of Nvidia GPUs as ready-made clothing sold in a store.

They are standard products that anyone can buy and use.

By contrast, Broadcom designs dedicated semiconductors tailored to how a customer works.

Such custom semiconductors are called ASICs.

Broadcom is like a custom suit shop.

It co-designs dedicated AI chips to fit the customer’s AI model, data processing method, and power efficiency goals.

A representative example is Google TPU.

Broadcom is said to have co-developed Google’s TPU from the first one in 2015.

It is also mentioned as participating in Meta’s AI chip and OpenAI’s own chip design.

Companies like Anthropic can also connect with the Broadcom ecosystem through Google TPU or Amazon’s custom chips.

As big tech tries to reduce dependence on Nvidia and expand its own AI chips, Broadcom’s opportunity grows.

10. Broadcom’s hidden core business: switch chips and CPO

If you think of Broadcom only as a custom AI chip company, you are seeing only half of it.

Broadcom is also strong in switch chips, which are central to data center networks.

Inside AI data centers, thousands or tens of thousands of chips exchange data with one another.

A chip is needed to direct where that data should go.

That role is performed by switch chips.

If switch chips are insufficient or underperforming, data bottlenecks occur no matter how many GPUs you have.

Broadcom’s representative product Tomahawk 5 is mentioned.

As AI data centers grow larger, demand for such high-performance switch chips keeps increasing.

Recently, CPO technology has also been drawing attention.

CPO stands for Co-Packaged Optics, a technology that integrates optical communication components close to the switch chip.

At present, the common method is to separately attach components that convert electrical signals into light next to the switch chip.

But CPO integrates optical components more closely with the chip.

Shorter wiring reduces power consumption and improves signal quality.

In a situation where AI data center power costs are exploding, CPO is a very important technology.

Broadcom is regarded as a company with a strong position in this CPO-based switch technology.

11. Marvell is a smaller Broadcom: Amazon custom AI chips and optical DSP

Marvell recently drew even more attention when Jensen Huang mentioned it as a future 1 trillion dollar market cap candidate.

Looking at the business structure alone, Marvell is quite similar to Broadcom.

However, it is smaller in scale than Broadcom.

Marvell also designs custom AI chips for customers.

It is connected with custom AI semiconductors for companies such as Amazon.

It also sells switch chips that connect chips to one another.

Another core Marvell business is DSP chips.

DSPs are key semiconductors needed in the process of converting electrical signals to light and light back to electrical signals.

In AI data centers, data must move at enormous speed.

Because there are limits to electrical signals over copper wires alone, the share of optical communications is increasing.

In this trend, demand for Marvell’s DSP chips can increase.

In the end, Marvell has growth opportunities both in AI chips themselves and in data center network infrastructure.

12. Why Marvell projected 10x growth over five years

Marvell recently set a very aggressive target at its investor day.

Based on the original text, it set its fiscal 2030 revenue target at 80 billion dollars.

Considering that the previous fiscal year revenue was about 8.2 billion dollars, this is nearly 10x growth.

Raising revenue from 8.2 billion dollars to 80 billion dollars within five years is a very bold forecast.

If profit margins improve as well, net income could increase by 12 to 13 times, which is also a plausible expectation.

That is why the market says that if Marvell really achieves this goal, a 1 trillion dollar market cap would not be impossible.

Of course, this outlook still needs verification.

A target is only a target, and actual orders, revenue recognition, production capacity, and competitive conditions all have to line up.

Still, there is no doubt that Marvell is emerging as a core company in AI data center infrastructure.

13. Valuation comparison: companies that look cheap and companies with high growth rates

The original text compares major AI semiconductor companies based on next year EPS growth and expected PER.

PER is the price-to-earnings ratio, calculated by dividing stock price by earnings per share.

Generally, the lower the PER, the cheaper the stock looks relative to earnings.

Of course, for growth stocks, judging only by PER is risky.

You have to look at revenue growth, profit margins, market dominance, and supply chain risk together.

Company	Core business	Investment point in the original text	Valuation perspective
Nvidia	AI GPU, AI data center systems	Overwhelming market dominance, Blackwell and Rubin ecosystem	Expected PER mentioned as about 14.4x, relatively cheap
AMD	AI GPU, server CPU, rack-scale systems	CoWoS volume rises 4x, expectations for large data center orders	EPS growth about 108%, expected PER about 38x
Broadcom	ASIC, switch chips, CPO	Google TPU, custom AI chips for big tech, solving network bottlenecks	Expected PER mentioned as about 16.8x, cheap relative to growth
Marvell	Custom AI chips, DSP, optical communications, switch chips	2030 revenue target of 80 billion dollars, benefiting from data center optical communications	EPS growth about 75%, expected PER about 37x
Intel and ARM	CPU, IP, semiconductor ecosystem	Potential benefit from AI infrastructure expansion	Valuation burden is mentioned as relatively high versus growth

Based on the original text, Nvidia and Broadcom look relatively attractive if you want both stability and price appeal.

On the other hand, if you approach it by focusing on explosive growth, AMD and Marvell stand out more.

In particular, AMD could benefit greatly from a phase of rising CPU demand because it has both CPU and GPU businesses.

Marvell’s high-growth target is very appealing, but that also means execution risk must be watched closely.

14. The real key point rarely discussed in other news: AI data centers have become a “power-unit” industry

Most news focuses on Nvidia GPU prices or new product performance.

But the real change is that the AI semiconductor market is now being calculated not by chip count but by power units.

Expressions like OpenAI 6GW, Meta 6GW, and Anthropic 2GW are not simple order volumes.

They mean the AI industry has become a massive capital expenditure industry combining power, real estate, cooling, networks, semiconductors, memory, and cloud infrastructure.

If tens of trillions of won worth of semiconductors are going into each 1GW, then AI data center investment becomes a very large variable even in the global economic outlook.

Grid investment will increase, nuclear and natural gas power demand may be re-evaluated, and demand for copper and power equipment could also rise.

In other words, AI semiconductor investment is not just a technology stock theme; it is a huge infrastructure investment cycle.

From this perspective, what investors should watch is not only “who makes the best GPUs.”

It has become more important to see who secured CoWoS capacity, who can attach more HBM, who can solve data center network bottlenecks, and who can improve power efficiency.

15. Key checkpoints to watch from an investment perspective

CoWoS secured volume
This is a key leading indicator for AI semiconductor shipments.

AMD increasing from 130,000 wafers to 530,000 wafers is a very strong signal.
HBM supply capability
AI GPU performance is heavily influenced by HBM capacity and bandwidth.

AMD equipping 432GB of HBM4 is a differentiating point versus Nvidia.
Rising CPU demand
If AI assistants spread, CPU demand can rise sharply along with GPU demand.

This structure favors AMD.
ASIC expansion
Big tech wants to expand its own AI chips to reduce dependence on Nvidia.

Broadcom and Marvell are the main beneficiaries of this trend.
Data center network bottlenecks
As AI clusters grow, switch chips, CPO, DSP, and optical communication components become more important.

This is where the growth logic for Broadcom and Marvell becomes stronger.

16. Risks must also be clearly considered

The growth potential of the AI semiconductor market is very large, but there is no guarantee every forecast will be realized as written.

First, there is a time lag between order expectations and actual revenue recognition.

Data centers need power access, land, cooling, server installation, and network buildout.

Second, shortages in CoWoS and HBM supply could instead lead to shipment delays.

No matter how strong demand is, if bottlenecks are not cleared, revenue recognition slows down.

Third, if AI infrastructure investment grows too quickly, big tech’s capital expenditure burden increases.

Depending on interest rates, cash flow, regulation, and power grid issues, the pace of investment can be adjusted.

Fourth, valuation pressure cannot be ignored.

AMD and Marvell have high growth rates, but much of that expectation may already be priced into the stock.

Fifth, Nvidia and Broadcom may look relatively cheap, but because their market caps are so large, the pace of further gains could be slower than in the past.

17. Conclusion: If you are looking for the next move after Nvidia, you should watch the AI data center supply chain

The period when the answer to AI semiconductor investment was fixed as Nvidia alone is passing.

Nvidia is still the overwhelming number one.

But as AI data centers get bigger, not only GPUs but also CPUs, HBM, CoWoS, ASICs, switch chips, optical communications, and CPO all become important.

AMD is a company that competes directly with Nvidia while taking both GPU and CPU share.

Broadcom has a strong position in custom AI chips for big tech and in resolving network bottlenecks.

Marvell is relatively smaller, but its custom AI chips and optical communication DSP growth potential is very large.

In the end, the investment point in AI semiconductors going forward is not “who is the most famous.”

It is “who solved the bottlenecks in data center expansion and secured actual orders?”

From this perspective, it becomes much clearer why AMD, Broadcom, and Marvell are growing together with Nvidia.

< Summary >

The AI semiconductor market may look like Nvidia’s monopoly, but AMD, Broadcom, and Marvell are also growing rapidly in their own areas.

AMD is increasing CoWoS volume by roughly 4x and is linked to expectations for mega-scale data center investment from OpenAI, Meta, and Anthropic.

Broadcom has strengths in custom AI chip ASICs such as Google TPU, as well as data center switch chips and CPO technology.

Marvell has set an 80 billion dollar revenue target for 2030 through growth in custom AI chips, DSP, and optical communications infrastructure for customers such as Amazon.

The core of investing is not one GPU, but the entire AI data center supply chain.

Watching CoWoS, HBM, CPU, ASIC, switch chips, and optical communications together is the most important strategy for AI semiconductor investing going forward.

[Related Articles…]

*Source: [ 월텍남 – 월스트리트 테크남 ]

– 5년간 1200%더 오른다? 700조 주문 폭주하는 ‘이 곳’

NextGenInsight.Net

Like this: