Nvidia Crackdown, AI Chip War Explodes

● GPU Monopoly Cracks

Big Tech AI Chip War in Full: The Real Reason Nvidia’s Dominance Is Shaking and the Opportunity for Korea’s AI Semiconductor Industry

The core point of the AI chip war is not simply “who can beat Nvidia.”

The truly important point in today’s market is that AI training remains GPU-centered, while AI inference is rapidly diversifying into NPU, TPU, and proprietary AI chips.

Google, Amazon, Microsoft, Meta, and Tesla are all making their own AI chips, while startups such as Cerebras, Groq, Tenstorrent, and SambaNova are also carving out niches.

In Korea, Rebellions and FuriosaAI have emerged as leading players in the Korean AI semiconductor ecosystem.

In this article, we will connect AI data center investment, Nvidia stock, the semiconductor industry cycle, the foundry market, and sovereign AI strategy all at once and organize them clearly.

1. What Is an AI Chip: You Need to Start with the Difference Between GPU and NPU

To understand AI chips, you first need to know the difference between CPU, GPU, and NPU.

CPU is a general-purpose processor that is good at complex calculations.

Simply put, it is like a “PhD-level calculator” that thinks deeply and solves difficult problems.

GPU, on the other hand, was originally created for game graphics processing.

Because graphics require calculating vast numbers of pixels simultaneously, GPUs are strong at parallel processing.

But AI models are also fundamentally structured around matrix multiplication, meaning they repeat multiplication and addition an enormous number of times.

That is why GPUs became the core semiconductor of the AI era.

NPU is a concept one step further.

If a GPU is a versatile high-performance chip that can handle many tasks, an NPU is designed to process the multiplication and addition needed for AI more cheaply, with less power, and more efficiently.

Google calls this TPU, and each company uses different names such as NPU, LPU, and AI accelerator.

The core takeaway is the same.

It is about processing only the calculations essential for AI quickly and at low cost.

2. The Separation of AI Training and Inference: The First Crack in Nvidia’s Monopoly

The most important distinction when looking at the AI chip war is training and inference.

Training is the process of building an AI model.

It is the stage where massive amounts of data are fed in and the model is trained.

In this area, Nvidia GPUs are still strong.

That is because high-performance GPUs such as the H100, H200, B100, and B200, along with the CUDA ecosystem, are overwhelmingly dominant.

Inference, by contrast, is the process in which an already built AI model actually generates answers.

When we ask ChatGPT, Claude, or Gemini a question and receive an answer, that is inference.

In this inference area, many people say GPUs are overkill.

That is because there are many tasks that do not require an expensive, power-hungry GPU.

So Big Tech started building their own inference AI chips.

This is why it is often mentioned in the semiconductor peak-out debate.

It does not mean GPU demand will collapse completely, but rather that the growth benefits of the AI semiconductor market may not be concentrated in Nvidia alone.

3. Even If GPU Share Falls, Nvidia Revenue Can Keep Growing

According to market research cited in the original text, GPU share in the data center AI market was estimated at about 87% in 2024.

There is also a forecast that it could fall to around 75% in 2025.

On the surface, that sounds like bad news for Nvidia.

But what matters here is that the entire market is exploding in size.

Even if market share falls, Nvidia revenue can still rise if the market grows even faster.

In fact, investment in AI data centers remains strong.

Cloud companies, Big Tech, national AI projects, and sovereign AI buildout demand are all increasing at the same time.

So even if Nvidia’s stock shakes in the short term, its earnings are still likely to remain strong.

However, the market’s perspective is changing.

In the past, “AI semiconductors = Nvidia.” Now it is expanding to “AI semiconductors = Nvidia + proprietary chips + ASIC + NPU + foundry + HBM ecosystem.”

4. Why Big Tech Is Making AI Chips Directly

The biggest reason Big Tech makes AI chips directly is to reduce dependence on Nvidia.

Nvidia GPUs are powerful, but expensive.

They are also hard to secure in supply.

Being too dependent on one company’s supply chain creates risks for cloud businesses and AI service operations.

So Big Tech is aiming for three effects through proprietary AI chips.

Cost reduction: They can lower GPU purchase costs and power costs.
Supply chain stabilization: They can be less vulnerable to Nvidia shortages.
Service optimization: They can make chips tailored to their own AI models and data center architecture.

The important thing is that Big Tech is not directly building semiconductor fabs.

Google, Amazon, Microsoft, and Meta lead the planning and architecture design of the chips, while actual mass-production design is handled by specialized fabless companies or design partners such as Broadcom.

Manufacturing is handled by foundry companies such as TSMC and Samsung Electronics.

As the semiconductor industry has become more specialized, an environment has been created in which both Big Tech and startups can enter the AI chip market.

5. Google TPU: Nvidia’s Most Realistic Rival

Among Big Tech’s proprietary AI chips, the most noteworthy right now is Google.

Google has been developing TPU for a long time, and is now considering selling it externally beyond internal use.

In particular, Google mentioned that it actively used TPU during the development of Gemini 3.

This means Google is no longer just a company good at AI services, but is becoming a company that vertically integrates AI models, cloud, and semiconductors.

Google TPU does not simply mean it is always faster than GPUs.

The key is better cost and energy efficiency in large-scale system operations.

In AI data centers, the efficiency of the whole system, heat, network, and operating costs are becoming more important than the performance of a single chip.

According to Morgan Stanley estimates, selling 500,000 TPUs could generate about 1.8 trillion won in new revenue for Google.

If Google aims for production or utilization of 5 million TPUs by 2027, it could theoretically become an even larger new revenue source.

Of course, not all of those units will be sold externally, and a large amount will likely be used internally.

Still, for Google, this creates both cost savings from not needing to buy GPUs and revenue from external sales.

6. Amazon AWS: The Trainium and Inferentia Strategy

Amazon is aggressively using AI chips through AWS.

The main chips are Trainium and Inferentia.

Trainium is a training chip.

Inferentia is an inference chip.

However, in the original text, Inferentia was described as fading out without a next version, with a shift toward Trainium handling both training and inference.

Amazon is not necessarily ahead of Google or OpenAI in proprietary LLM competitiveness.

But it has AWS, one of the world’s largest cloud infrastructures.

So Amazon’s AI chip strategy is clear.

The strategy is to sell through AWS cloud instances rather than selling chips directly.

Customers can choose Nvidia GPU instances on AWS or Trainium-based instances.

This structure is favorable for Amazon.

That is because customers get more choices, dependency on Nvidia decreases, and cloud margins can improve.

Another important connection is that Anthropic is a major investment target for Amazon, and Claude operations use AWS infrastructure and Trainium.

7. Microsoft, Meta, and Tesla: Proprietary AI Chips Are Not a Choice but a Survival Strategy

Microsoft is developing its own AI chip called Maia.

The initial version was not as smooth as expected, but through Maia 200 the company is expanding AI workload adoption.

This can be seen as an effort to use proprietary chips in some Microsoft 365 Copilot and OpenAI-related services.

Meta is developing an AI chip called MTIA.

Meta is a company with enormous inference demand across Facebook, Instagram, Reels, ad recommendations, and AI chatbots.

It is not a cloud provider, but since its internal AI compute volume is so large, it has plenty of incentive to build its own chips.

Tesla’s demand for proprietary AI chips may also grow when considering autonomous driving, robots, and the xAI ecosystem.

In Tesla’s case, market expectations tied to Samsung Electronics foundry are often mentioned.

Ultimately, for Big Tech, proprietary AI chips are not just a way to cut costs.

They are a core infrastructure strategy for maintaining competitiveness in AI services.

8. The Opportunity the AI Chip War Brings to the Foundry Market

As Big Tech makes more proprietary AI chips, new opportunities arise in the foundry market.

The biggest beneficiary is of course TSMC.

That is because most high-performance AI chips require advanced processes and advanced packaging.

However, if too much volume flows to TSMC, that could create opportunities for Samsung Electronics’ foundry as well.

Unlike traditional mobile APs or CPUs, AI chips often have less direct competitive overlap with customers.

So for Samsung Electronics, this becomes an important market where it can secure outsourced AI semiconductor production volume.

In particular, Korea has Samsung Electronics foundry, SK Hynix HBM, and AI chip startups such as Rebellions and FuriosaAI.

If this combination works well, it could develop into a Korean AI semiconductor industry ecosystem.

Of course, the reality is not easy.

There are many challenges to solve, including the technology gap with TSMC, yields, customer trust, and packaging competitiveness.

9. Overseas AI Chip Startups: Cerebras, Groq, Tenstorrent, SambaNova

It is not only Big Tech that is making AI chips.

Startups are also targeting the gaps in a market dominated by Nvidia, each in different ways.

9-1. Cerebras: A Company That Uses an Entire Wafer as One Chip

Cerebras is the most symbolic company among AI semiconductor startups.

Normally, semiconductors are made by creating many chips on a wafer, then cutting and selling them individually.

Cerebras reversed that approach.

It uses the entire wafer as a single giant chip.

In simple terms, it is like selling an entire whole pizza as one product rather than selling slices.

This structure has the advantage of handling enormous amounts of compute.

It also places SRAM memory inside the chip, reducing dependence on HBM.

It is technically very unique and not easy to imitate.

Cerebras gained major attention after signing a large supply contract with OpenAI.

Based on the original text, it was also mentioned that revenue and net profit improved rapidly.

However, the reason the stock has swung so much is clear.

The customer base is too concentrated in a very small number of buyers.

Because it is such an oversized chip, the pool of potential customers is limited.

There were also concerns that customers related to the United Arab Emirates accounted for most of the revenue.

So Cerebras is moving toward a cloud service model rather than simply selling chips.

That is because even if chips cannot be split, cloud usage can be sold in smaller units.

9-2. Groq: The LPU and AI Compute Cloud Strategy

Groq is a company that drew attention with the concept of LPU.

Unlike the traditional approach of combining GPU and HBM, it emphasized the direction of rapidly performing AI language model processing on a single chip.

The original text mentioned that Groq secured funding through a large technology licensing deal and then aimed to transition into a data center and cloud provider specialized in AI compute.

This direction is important.

That is because AI chip startups are unlikely to survive by selling only chips, and will likely need to expand into AI infrastructure service businesses.

9-3. Tenstorrent: A Card-Type AI Chip Targeting Smaller Operators

Tenstorrent is an AI chip startup led by talent from AMD and Intel.

The company’s distinctive feature is that it aims for a card-shaped AI chip that can be easily installed in ordinary servers and workstations.

Its strategy is not limited to giant data centers; it also targets mid-sized AI operators and enterprise customers.

It is also diversifying revenue by selling not only chips but also design IP.

Its potential collaborations with LG, Hyundai Motor, and Samsung were also noted.

9-4. SambaNova: A National-Level Strategy Aiming at Sovereign AI

SambaNova is a company targeting national-level AI infrastructure contracts.

Germany, the UK, Australia, Japan, and France are uneasy about relying entirely on American Big Tech for AI infrastructure.

So these countries are pursuing sovereign AI strategies to build their own AI models and infrastructure.

SambaNova aims to position itself as an independent AI infrastructure partner aligned with this trend.

However, it is still hard to say that large-scale commercial validation has clearly been completed, unlike Cerebras.

10. Korea’s Leading AI Chip Player: Rebellions’ Competitiveness and Risks

Among Korean AI chip startups, the one getting the most attention is Rebellions.

Rebellions has focused on inference NPU from the start.

Domestically, KT and SKT are often mentioned as key supporters.

It is also receiving strong government support.

Based on the original text, Rebellions’ valuation was discussed at around 3 trillion won, and there was also mention of significant support from the National Growth Fund.

Rebellions’ strengths are clear.

As Korea’s representative AI semiconductor company, it has strong policy symbolism.
It has secured a customer base in telecom and cloud through SKT and KT.
It has a domestic supply chain that can connect to Samsung Electronics foundry.
Its strategy aligns with the growth of the inference AI chip market.

But the risks are also significant.

Based on the earnings figures mentioned in the original text, revenue is in the tens of billions of won, while operating losses are much larger.

For AI semiconductor startups, heavy R&D spending means losses are not unusual.

The issue is how quickly commercial mass-production customers can be secured.

Based on the current revenue scale, it seems more like a middle stage beyond PoC and early validation rather than a global mass-production stage.

For Rebellions to become a true global company, it must go beyond domestic telecom references and secure customers in the Middle East, Japan, the United States, and European data centers.

In the end, it must become not just “a company recognized in Korea,” but “a chip company used by global AI data centers.”

11. FuriosaAI: The Korean AI Chip Unicorn Revived After Meta Acquisition Rumors

FuriosaAI is also a major Korean AI chip startup.

It received a great deal of market attention after rumors of a Meta acquisition.

There were various interpretations regarding the truth and specifics of the acquisition rumors, but the important thing is that Big Tech showed interest.

After that, FuriosaAI successfully raised investment, and its valuation rose significantly.

The original text mentioned that it was valued at around 1 trillion won and later pushed forward with a pre-IPO, aiming for an even higher valuation.

FuriosaAI’s strength is its focus on inference AI chips.

It has also been linked to possible collaborations with LG AI Research and Samsung SDS.

However, some say it is at an earlier stage than Rebellions.

The revenue base is still small, and securing real customers after mass production remains the key challenge.

There is also risk in the fact that the audit report mentioned going-concern concerns, indicating that fundraising is critical.

For FuriosaAI to survive, it must prove three things.

It must prove that the chip delivers usable performance in real service environments.
It must secure overseas customers beyond large domestic enterprises.
It must secure long-term R&D funding through a pre-IPO and eventual listing.

12. The Most Important Point That Other News Rarely Emphasizes: AI Chips Do Not Succeed Just Because the Chip Is Good

There is a very important point in the AI chip war that often gets overlooked.

AI chips do not sell just because semiconductor performance is good.

For customers to actually use a chip, they need software ecosystems, developer tools, model compatibility, cloud infrastructure, operational stability, and maintenance systems.

The reason Nvidia is strong is not simply because of GPU performance.

It is because of the overwhelming strength of the CUDA ecosystem.

AI developers are already accustomed to building, tuning, and deploying models on CUDA.

To use a new AI chip, existing code often has to be changed or optimized.

This process is cumbersome and expensive.

That is why even if a startup AI chip shows good benchmark numbers, it takes time for real customers to adopt it at scale.

This is also where Korean AI chip companies face their real problem.

Even if Rebellions or FuriosaAI make excellent chips, if Korea lacks large-scale LLM services and inference traffic, it is hard to create references.

AI chips ultimately sell when lots of AI services are running.

In other words, for Korean AI semiconductors to succeed, it is not enough for semiconductor companies alone to do well.

Domestic LLMs, cloud, data centers, sovereign AI policy, and enterprise AI adoption must all grow together.

That is the core takeaway that is relatively underemphasized in other news.

13. Will Nvidia Actually Be Hurt?

In the short term, it is difficult for Nvidia’s dominance to collapse.

Nvidia has everything: GPU performance, CUDA ecosystem, networking, server references, and customer trust.

It remains the strongest option in the AI training market.

But in the medium to long term, its market share could decline.

Especially in the inference market, chips such as Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia, and various NPUs and ASICs can take on increasingly important roles.

For Nvidia, two things are likely to happen at the same time.

Its market share may gradually decline.
However, as the overall AI semiconductor market grows, revenue may continue to expand.

This is similar to AWS.

AWS’s cloud share is lower than it was in the past, but because the entire cloud market grew, AWS revenue kept increasing.

Nvidia may enter a similar phase.

Still, from an investor’s perspective, we should move beyond the era of looking only at Nvidia.

We need to look at AMD, Broadcom, TSMC, Samsung Electronics, SK Hynix, Micron, AI chip startups, and cloud companies together.

14. How Should We View the Semiconductor Peak-Out Debate?

AI chip mass production is often cited as one of the reasons for semiconductor peak-out.

The logic is this.

If Big Tech makes its own AI chips, purchases of Nvidia GPUs could decline.

If GPU demand falls, HBM demand could also be affected.

Then the high-growth cycle of the semiconductor industry could slow down.

But this argument is only half right.

Even if proprietary AI chips increase, investment in AI data centers is still growing.

Inference demand is only at the beginning stage.

If enterprise AI, search, ads, shopping recommendations, video generation, robots, autonomous driving, and on-device AI expand, the total compute volume can grow much more.

Therefore, rather than peak-out, it is more accurate to view this as a phase in which growth is shifting from a single GPU axis to a diversified AI semiconductor ecosystem.

From an investment perspective, it is less “AI semiconductors are over” and more “the beneficiaries of AI semiconductors are changing.”

15. The Opportunity for Korea’s AI Semiconductors: Sovereign AI and a Third Choice

Korean AI chip companies clearly have an opportunity.

Amid the technological rivalry between the United States and China, many countries do not want to depend entirely on American Big Tech or Chinese companies.

At that point, Korea can become a third choice.

Just as K-pop and K-content grew globally as an alternative that was neither American nor Chinese, Korea can also become an alternative in certain AI infrastructure markets.

In particular, some countries in the Middle East, Southeast Asia, Japan, and Europe are highly interested in building sovereign AI.

If Korea can provide AI chips, cloud, LLMs, and data center packages together in this market, there is an opportunity.

But in reality, competition is extremely fierce.

In the Middle East, overseas startups such as Cerebras are already present.

American Big Tech has overwhelming capital strength and customer networks.

China is also strongly pushing its own AI chip ecosystem centered on Huawei.

For Korean companies to survive, they cannot remain in the protected domestic market.

They must secure real customers in the global market.

16. Key Checkpoints Investors Should Watch

When looking at the AI chip war from an investment perspective, you should not simply ask which company made the better chip.

The following checkpoints are much more important.

First, is there a real large-scale customer?
Second, is the business moving beyond PoC into mass-production revenue?
Third, is the software ecosystem and developer tooling sufficient?
Fourth, is the foundry and packaging supply chain stable?
Fifth, is there enough cash and runway?
Sixth, is there solid backing from Big Tech or a national-level partner?
Seventh, is the valuation excessive relative to revenue?

In particular, AI chip startups cannot survive with technology alone.

Semiconductors require high development costs, high mass-production costs, and long customer validation periods.

Unlike platform companies, they cannot easily survive by cutting marketing expenses.

That is why fundraising ability and strategic investors are extremely important.

17. This Is How the AI Chip Market Is Likely to Reshape Going Forward

The AI chip market is likely to split into three branches going forward.

First, the ultra-large training market will likely remain centered on Nvidia.

For large LLM training, GPUs and the CUDA ecosystem are still strong.

Second, the inference market will likely diversify rapidly around proprietary AI chips and NPUs.

Google TPU, AWS Trainium, Meta MTIA, and Microsoft Maia chips belong here.

Third, sovereign AI and specialized markets can become opportunities for startups.

Companies such as Cerebras, Tenstorrent, SambaNova, Rebellions, and FuriosaAI are targeting this area.

However, not all startups will survive.

Just as many companies disappeared in the graphics card market and Nvidia emerged as the winner, the AI chip market is also likely to see repeated M&A and bankruptcies.

Ultimately, the companies that survive will be those that secure not only technology, but also customers, capital, ecosystem, and supply chains.

< Summary >

The core point of the AI chip war is not completely replacing Nvidia GPUs, but the separation of the AI training and inference markets.

Training is still dominated by Nvidia GPUs, but inference is rapidly diversifying into proprietary AI chips such as Google TPU, AWS Trainium, Meta MTIA, and Microsoft Maia.

Google is vertically integrating AI models, cloud, and semiconductors through TPU, and has emerged as Nvidia’s most realistic rival.

Amazon is improving cloud profitability and supply chain stability by offering its own AI chip instances on AWS.

Startups such as Cerebras, Groq, Tenstorrent, and SambaNova are targeting specialized AI chips and the sovereign AI market.

In Korea, Rebellions and FuriosaAI are the leading players, but securing global customers and funding remains the key challenge.

Nvidia’s market share may decline over the long term, but as the AI semiconductor market grows, earnings growth is likely to continue.

Investors should look at the semiconductor industry, AI data centers, the foundry market, HBM, and the sovereign AI ecosystem together rather than focusing only on Nvidia.

[Related Articles…]

*Source: [ 티타임즈TV ]

– 빅테크 어느 곳 하나 AI칩 안 만드는 곳이 없네? AI칩 전쟁

NextGenInsight.Net

Like this: