Tesla-Waymo-China-AI-Driving-Clash

● Autonomous-Driving War, Tesla, Waymo, China, Korea’s Last Shot, 3rd-Gen AI Race

Tesla, Waymo, and China are advancing simultaneously in the autonomous driving race; for Korea, the last viable move is not a second-generation catch-up strategy but a third-generation first-mover strategy

Autonomous driving competition is no longer simply a matter of who drives better.
The core issue is now competition across data, AI models, AI semiconductors, manufacturing AX, and physical AI within the future mobility industry.
The most important point in this discussion is that Korea cannot stop at catching up with Tesla.
It must pursue second-generation autonomous driving quickly while simultaneously securing leadership in third-generation physical AI-based autonomous driving.
If successful, this strategy could transform Korea’s automotive industry from a conventional manufacturing base into a new export-growth pillar for the Korean economy.

1. Autonomous driving has shifted from an automotive technology contest to an AI industry competition

In the past, autonomous driving was closer to a competition in vehicle control technologies among carmakers.
That is no longer the case.
Autonomous driving has moved into a physical AI competition in which AI understands, judges, and acts in the real world.

Byeongwook Jeon, head of the AI Autonomous Driving Technology Research Institute at the Korea Automotive Technology Institute, divides autonomous driving into three generations.
This framework clarifies where Tesla, Waymo, China, and Hyundai Motor are competing.

First-generation autonomous driving: rule-based autonomous driving
Second-generation autonomous driving: end-to-end AI-based autonomous driving
Third-generation autonomous driving: physical AI and world-model-based autonomous driving

This trajectory is similar to how generative AI systems such as ChatGPT and Gemini process language.
In the past, engineers tried to build language models by encoding grammar rules one by one, but that approach had clear limits.
With the emergence of large-scale data-trained AI models, machines began to understand and generate language more naturally.
Autonomous driving is moving in the same direction.
The era of manually coding every road scenario has ended, and AI is now learning to drive by training on large volumes of driving data.

2. Why first-generation autonomous driving failed: real roads cannot be fully captured by rules

First-generation autonomous driving relied on rules defined directly by humans.
Examples include: “stop when the vehicle ahead stops,” “follow the lane when lane markings are present,” and “slow down when pedestrians are detected.”

The problem is that real roads are far more complex than such rules can capture.
Rain, construction zones, motorcycles appearing suddenly, illegally parked vehicles, pedestrians using ambiguous gestures, and snow-covered lanes create situations that are difficult to define through fixed logic.

As a result, first-generation systems were vulnerable to edge cases.
Edge cases are rare scenarios that may not occur often but can directly lead to accidents.
The real challenge in autonomous driving is not routine straight-line travel, but the safe handling of such rare and complex situations.

3. The center of current global competition is second-generation end-to-end AI autonomous driving

Second-generation autonomous driving is based on an end-to-end AI architecture.
Camera and sensor inputs are learned directly by the AI model, which then produces driving decisions appropriate to the situation.

In practical terms, the earlier approach was for humans to define rules such as “do this in that situation,” whereas the current approach is for AI to learn patterns from large datasets and infer how driving should typically be performed in similar conditions.

Tesla is widely regarded as the leading company in this area.
Tesla has already deployed a large fleet of vehicles on public roads and is improving FSD through the data generated by those vehicles.
Chinese companies are also advancing rapidly.
China has strong advantages in domestic market scale, data collection speed, government support, and the EV supply chain.

However, second-generation systems also have limitations.
Training on more data does not guarantee full understanding of all scenarios.
Jeon described this as analogous to solving many practice problems without necessarily achieving a perfect score on the exam.
Familiarity with patterns improves performance, but truly novel scenarios still require deeper understanding.

4. The core of third-generation autonomous driving is physical AI and world models

Third-generation autonomous driving is physical AI-based autonomous driving.
Here, physical AI does not simply mean AI operating in the physical world.
Its key capability is that AI understands the real world, predicts the consequences of its actions, and selects the best response.

The central concept is the world model.
A world model is an internal representation that allows AI to understand the structure of the real world.
For example, an autonomous vehicle should be able to anticipate how the vehicle behind will react if it changes lanes, whether a pedestrian may suddenly enter the road, or whether a maneuver at the current speed is safe.

This is also closely linked to robotics.
When a robot picks up a cup or folds laundry, simple reactive behavior is not enough.
It must predict object position, weight, movement, and the consequences of the next action.
An autonomous vehicle is ultimately a robot on wheels.
Accordingly, future mobility competition is evolving into a convergence of robotics, manufacturing AX, AI semiconductors, and data platforms.

5. Global rankings must be assessed separately for Level 2 and Level 4

A common mistake in evaluating autonomous driving companies is comparing Tesla and Waymo on the same basis.
Jeon described this as competing in different leagues.

Tesla is closest to supervised autonomous driving with a human driver on board, and is therefore strongest in the Level 2 or Level 2+ segment.
Waymo, by contrast, is closer to Level 4 autonomous driving through unmanned robotaxi operations in designated areas.
Its operating domain is limited and depends on high-definition maps, control systems, and multiple sensors.

Level 2 perspective: Tesla remains the leader, with Chinese companies rapidly closing the gap
Level 4 robotaxi perspective: Waymo remains the leader, with Chinese robotaxi firms in pursuit
Korea: it must both catch up in Level 2 and prepare for Level 4 testing and third-generation physical AI

This distinction matters because investment priorities and industrial policy differ by level.
Level 2 depends on data generated by mass-market vehicles.
Level 4 depends on city-scale demonstrations, control infrastructure, regulatory sandboxes, and safety verification systems.

6. Korea’s largest weakness in autonomous driving is data

Autonomous driving competitiveness begins with data.
Because AI is a data-driven technology, high-performance autonomous driving models cannot be built without sufficient data.

Korea faces a structural contradiction.
To build autonomous vehicles, it needs data.
To collect data, autonomous vehicles must already be widely deployed in the market.
This is the classic chicken-and-egg problem in autonomous driving data.

Tesla solved this problem early.
It sold vehicles in large volumes before fully autonomous driving became mainstream and collected data from those vehicles.
Tesla’s strength comes not only from building a strong AI model, but from establishing a data flywheel first.

China has a similar advantage.
It benefits from rapid EV adoption, large-scale domestic road data, government-led industrial policy, and a fast-paced commercialization culture.

7. Korea’s practical data strategy: use taxis, rental cars, and production vehicles

Korea should not wait for autonomous vehicles to mature before collecting data.
It should begin with data collection from conventional vehicles.

One policy direction under consideration is installing data collection devices in taxis and rental cars.
This is a meaningful approach.
Taxis and rental cars operate for long hours and pass through a wide range of road environments, making them well suited for collecting high-quality driving data.

Hyundai Motor and Kia’s production vehicles can also be used.
Hyundai Motor and Kia produce more than 7 million vehicles annually on a global basis.
Given that Tesla has sold roughly 9 million vehicles cumulatively since its founding, Hyundai Motor and Kia have a scalable foundation for rapidly narrowing the data gap if they commit to doing so.

Installing data collection equipment raises unit cost.
However, this should not be viewed only as a cost passed on to customers.
It should be treated as a strategic AI investment for the autonomous driving industry.

8. Data should not remain a private asset: publicization and open innovation are required

One of the most important issues in this discussion is the publicization of autonomous driving data.
When multiple companies conduct testing in an autonomous driving demonstration city such as Gwangju, substantial data accumulates.
If each company keeps this data entirely to itself, national competitiveness may weaken.

At this stage, competition among firms remains important, but the broader contest is between nations.
As the United States, China, Europe, and Japan compete for leadership in future mobility, Korea must build a structure that enables academia, industry, and research institutions to use data jointly.

For example, if one company contributes 10 units of data, a public platform could enable access to 100 units of usable data.
Such a structure would give companies a stronger incentive to contribute data.
Individual datasets may be limited, but when aggregated at the national level they become a much larger training asset.

This requires three foundations.

Data standardization: driving data in different formats must be integrated
Data governance: ownership, use rights, and accountability must be clearly defined
Security and privacy protection: sensitive data such as vehicle location, driver information, and road video must be handled safely

9. Real-world data alone is insufficient: simulation data is a key battleground

Autonomous driving data does not have to come only from public roads.
In fact, dangerous scenarios are difficult to collect in the real world.
This is where simulation-based synthetic data becomes essential.

For example, scenarios such as a child suddenly running into the road, a vehicle sliding on icy roads, or temporary sensor instability after entering a tunnel are difficult to reproduce repeatedly on public roads.
In a virtual environment, however, they can be trained thousands or even tens of thousands of times.

Global autonomous driving solution providers, including Nvidia, also place strong emphasis on simulation-based learning.
In robotics, the most effective approach is often to learn basic capabilities in simulation and then refine them with real-world data.
Autonomous driving is no different.

For Korea to lead third-generation autonomous driving, it must establish a structure in which real-road data, public data platforms, and simulation data operate together.

10. Autonomous driving is impossible without AI semiconductors

An autonomous vehicle contains far more semiconductors than a conventional internal combustion vehicle.
An internal combustion vehicle typically uses hundreds of automotive semiconductors.
Electric vehicles require more, and autonomous electric vehicles drive demand for high-performance AI semiconductors even further.

More important than the number of chips is the need for high-performance AI inference semiconductors.
Even the best autonomous driving AI model cannot be used effectively unless it can run properly on in-vehicle hardware.
This is the core of on-device AI.

Autonomous vehicles must make decisions in real time.
They cannot send all data to the cloud and wait for a response.
Recognition, judgment, and control must all occur inside the vehicle.
Therefore, the final competitive frontier in autonomous driving is full-stack on-device physical AI.

11. Customized semiconductors matter more than general-purpose AI chips

Many assume that AI semiconductors can simply be purchased from suppliers such as Nvidia.
In autonomous driving, that may not be sufficient.

General-purpose AI semiconductors are designed to support a wide range of AI models.
That makes them versatile, but not necessarily optimized for a specific autonomous driving system.
Because autonomous driving places strong emphasis on power efficiency, thermal management, response speed, reliability, and cost, customized semiconductors are likely to become more valuable.

Tesla already has its own FSD chip.
Chinese companies are also pursuing vertical integration in proprietary chips.
Xpeng has a proprietary chip strategy associated with Turing chips, and Huawei is building a strong automotive AI semiconductor ecosystem.
Li Auto, Momenta, and others are also working to align autonomous driving solutions with chip optimization.

The implication is clear.
Future mobility competition is becoming a vertically integrated contest across vehicles, software, AI models, and semiconductors.
This is why Korea should not treat AI semiconductors and the automotive industry as separate sectors.

12. Korea’s strategy must be a two-track approach

Korea’s strategy is straightforward in principle.
Execution, however, will be difficult.

First, it must rapidly catch up in second-generation end-to-end AI autonomous driving.
Second, it must design a leadership strategy for third-generation physical AI-based autonomous driving now.

Focusing only on second-generation catch-up while discussing third-generation leadership is risky.
But pursuing only second-generation technology risks leaving Korea permanently in a follower position relative to Tesla and China.
Korea must therefore pursue both catch-up and first-mover positioning simultaneously.

This strategy also connects to manufacturing AX.
The Ministry of Trade, Industry and Energy’s M·AX initiative is not limited to factory automation.
It refers to an industrial transformation in which AI is embedded across manufacturing, robotics, semiconductors, automotive, and data infrastructure.
Autonomous driving is one of the most consequential sectors within that transformation.

13. From a Korean economy perspective, autonomous driving is not optional; it is a survival issue

Semiconductors and automobiles are Korea’s most important export pillars.
If semiconductors are the top export item, automobiles and auto parts are the next critical pillar.
The automotive industry is not only a matter for OEMs.
It is a core manufacturing ecosystem linking numerous parts suppliers, equipment makers, material companies, and SMEs.

If Korea loses leadership in future mobility, the impact will extend beyond a few carmakers.
The supply chain, regional employment, export competitiveness, and the broader manufacturing base could all weaken.
Conversely, if Korea secures leadership in autonomous driving, AI semiconductors, and physical AI, it can create new growth engines amid global supply chain reconfiguration.

In a global environment of prolonged high interest rates and low growth, Korea needs AI transformation in its existing manufacturing base to lift growth.
Autonomous driving is a leading example of that transition.

14. The key issue not emphasized enough elsewhere: the real question is not who collects more data, but who enables broader use of data

Most coverage stops at the observation that Tesla is ahead, China is catching up, or Waymo is operating robotaxis.
The more important questions are different.

How will Korea collect data?
What incentives will cause companies to share data?
Who will operate the public data platform?
How will real-road data and simulation data be combined?
How will AI models and automotive semiconductors be co-designed?

In particular, a data-sharing structure could become a distinctive Korean strength.
Korea has the advantage of fast infrastructure deployment at the national level and a dense cluster of automotives, parts, telecom, semiconductor, and software companies.
If the government establishes proper data governance, the national platform can compensate for the data limitations of individual firms.

Another critical area is customized AI semiconductors.
The ability to efficiently run autonomous driving AI models inside vehicles is as important as building the models themselves.
This issue is still under-discussed in public markets, but it is likely to become a major barrier to entry in future mobility.

15. Korea must operate like one large startup

Jeon argued that Korea must function at the national level like one large startup.
This is important because speed matters enormously in the current autonomous driving and physical AI competition.

Large corporations cannot move in isolation, universities cannot limit themselves to papers, research institutes cannot focus only on project execution, and the government cannot remain confined to regulatory review.
Industry, academia, research institutions, and government must define roles clearly while experimenting, failing, and improving rapidly toward a shared objective.

The M·AX Alliance can serve as such a collaboration platform.
If manufacturing AX, autonomous driving, robotics, AI semiconductors, and data platforms are integrated, Korea can move beyond second-generation catch-up and create an opportunity in third-generation physical AI-based future mobility competition.

16. Key indicators to watch

How quickly Tesla improves commercialization in the Level 2+ market
How far Waymo expands its operating area for robotaxi services
How rapidly Chinese autonomous driving firms complete vertical integration across data and semiconductors
Whether Hyundai Motor and Kia establish a large-scale data collection system using production vehicles
Whether the Korean government creates a public autonomous driving data platform and standardization framework
Whether domestic AI semiconductor firms and automakers jointly develop customized vehicle AI chips
Whether a simulation-based edge-case data platform becomes institutionalized

If even one of these indicators lags, Korea is likely to remain in a follower position.
Conversely, if data, AI models, semiconductors, and manufacturing AX move together, Korea’s automotive industry could secure a strong position in the next cycle.

< Summary >

Autonomous driving competition has shifted from automotive engineering to competition in AI, data, and AI semiconductors.
First-generation rule-based autonomous driving reached its limits, and Tesla and China are currently leading second-generation end-to-end AI autonomous driving.
Waymo leads in the Level 4 robotaxi segment, while Chinese firms are pursuing it.
Korea must quickly close the second-generation gap while preparing for third-generation physical AI and world-model-based autonomous driving.
The main battlegrounds are data acquisition, data sharing, simulation learning, customized AI semiconductors, and industry-academia-government collaboration.
Because future mobility is central to Korea’s economy and export competitiveness, the entire country must move quickly and cohesively.

[Related Articles…]

*Source: [ 경제 읽어주는 남자(김광석TV) ]

– [M·AX 4화] 테슬라도, 중국도 뛰고 있다… 한국 자율주행의 마지막 승부수 | 경읽남과 토론합시다 | 전병욱 소장

● Shockwave-Rebound-AI-Chips

Samsung Electronics and SK Hynix Are Not Over: Why the AI Semiconductor Cycle, HBM, General-Purpose DRAM, and the CPU Battle Are Still in the Early Stages

The key drivers of the second-half market are threefold.

First, geopolitical risk from the Trump administration, along with oil and interest-rate variables.

Second, AI data center investment and semiconductor supply tightness.

Third, a structural shift from an Nvidia GPU-centered market toward CPUs and AI agents.

At first glance, volatility in Samsung Electronics and SK Hynix shares may suggest that the AI semiconductor cycle has peaked.

However, a closer look indicates that the AI semiconductor ecosystem may still be at an early stage of transitioning from GPU training toward inference, on-device AI, and physical AI.

Investors focusing only on HBM may miss a significant part of the story.

Going forward, the key variables will also include general-purpose DRAM shortages, rising CPU demand, hyperscaler CAPEX pressure, competition from Chinese memory makers, and re-rating potential for Korean semiconductor equities in ADR form.

1. The Starting Point of Second-Half Market Volatility: Trump and Middle East Risk

The main reason for recent market volatility is geopolitical risk.

In particular, rising tensions in the Middle East can destabilize oil prices, and oil price volatility can affect U.S. interest rates and the global growth outlook.

If conflict in the Middle East persists, risks around the Strait of Hormuz may rise.
The Strait of Hormuz is a critical bottleneck in global oil transportation.
Instability in this region can drive oil prices sharply higher in the short term.
If oil rises above $100 per barrel, inflationary pressure may reaccelerate.

If inflation rises again, central banks will find it difficult to cut rates.

They may instead maintain restrictive policy or even keep the possibility of further tightening in play.

The market complication is that President Trump favors rate cuts, while tariffs and war-related risk can create conditions that push rates higher.

2. How Trump’s Negotiation Style Moves Markets

Trump’s negotiation style generally follows five steps.

He initially makes demands that are difficult for the other side to accept.
If rejected, he applies pressure through tariffs, military signaling, or political statements.
He then lowers the demand partially to create the appearance of a concession.
If negotiations stall, he repeats strong rhetoric and pressure.
Finally, he presents the outcome as a major personal achievement.

This approach can work in negotiations, but it creates substantial uncertainty for financial markets.

The market can move from expecting a near-term agreement to pricing in renewed escalation within hours.

As a result, the second-half market will be driven by when conflict de-escalates, how far oil prices rise, and whether expectations for U.S. rate cuts remain intact.

3. Why Oil at $100 Is So Damaging for Equities

Higher oil prices raise logistics costs, utility costs, manufacturing costs, and consumer prices.

When inflation rises, consumers can purchase less with the same income.

Corporate cost burdens increase, while demand may weaken as consumption slows.

Higher inflation can lead to weaker corporate earnings.
Higher rates raise interest expenses for leveraged companies and households.
When rates rise, cash and bonds become more attractive relative to equities.
Risk appetite weakens, and volatility in the KOSPI and Nasdaq can increase.

For this reason, the market is not moving solely on Samsung Electronics and SK Hynix earnings.

Even when semiconductor earnings are strong, share prices can fall because macro uncertainty suppresses valuations.

4. Why AI Semiconductor Earnings Are Strong but Share Prices Can Fall

Recent earnings from U.S. AI semiconductor-related companies have generally been solid.

Storage companies such as SanDisk and Western Digital have also reported favorable operating trends.

Still, share prices have often fallen sharply.

There are three main reasons.

Market expectations have already become too elevated.
Hyperscaler CAPEX keeps rising, increasing pressure on cash flow.
Chinese semiconductor makers are advancing faster than expected.

Hyperscalers are companies that operate large-scale data centers.

Key examples include Microsoft, Meta, Alphabet, Oracle, and Amazon.

These firms are making massive capital expenditures to build AI data centers.

The issue is that CAPEX continues to rise while FCF, or free cash flow, is weakening.

As cash becomes tighter, firms may need to issue debt or sell corporate bonds.

That leads the market to question whether AI investment can continue at the same pace.

5. Why AI Data Center Investment Is Still Difficult to Stop

Hyperscalers are spending heavily not simply because AI is fashionable.

The reason is that securing an early lead in AI platforms can expand into finance, insurance, telecom, healthcare, retail, and workflow automation.

AI can substitute for certain banking functions.
It can automate insurance consulting and contract management.
It can handle telecom services and customer support.
It can also support medical diagnostics and health management.

For big tech firms, failing to invest now could mean falling behind in future platform competition.

That is why they continue to build data centers even before full profitability has been proven.

According to some foreign broker estimates, AI-related capital spending could expand from about $293 billion in 2024 to more than $1 trillion by 2027.

The exact number is less important than the direction.

AI infrastructure investment is still in an expansion phase rather than a contraction phase.

6. Why Samsung Electronics and SK Hynix Matter in Nvidia’s AI Server Architecture

Nvidia is not simply selling graphics cards.

It is closer to selling a complete AI server system.

An AI server contains multiple components.

GPUs handle large-scale computation.
HBM acts as high-speed temporary memory adjacent to the GPU.
CPUs serve as the system’s control unit.
General-purpose DRAM stores frequently accessed data for the CPU.
NAND and SSDs handle large-capacity storage.
Networking equipment connects servers across the system.
Software ecosystems such as CUDA also play a critical role.

Samsung Electronics and SK Hynix matter because of HBM and DRAM.

As AI servers proliferate, demand is not limited to GPUs.

HBM next to the GPU, general-purpose DRAM next to the CPU, and storage devices all rise together.

7. The Quietly More Important Variable Than HBM: General-Purpose DRAM Tightness

The market tends to focus on HBM alone.

However, a quieter but potentially stronger variable ahead is general-purpose DRAM tightness.

In Nvidia’s next platform, Vera Rubin, HBM stacking may rise to as many as 16 layers.

If HBM moves from 4-layer and 8-layer products to 12-layer and 16-layer configurations, more high-performance DRAM wafer capacity may shift toward HBM.

That could tighten supply for general-purpose DRAM used in PCs, mobile devices, and standard servers.

Rising HBM demand shifts more advanced DRAM production capacity toward HBM.
That reduces available supply for general-purpose DRAM.
AI servers also increase CPU-related DRAM demand.
As a result, general-purpose DRAM pricing may strengthen.

This structure may improve the pricing power of Samsung Electronics, SK Hynix, and Micron.

In other words, the core AI semiconductor cycle may broaden beyond HBM into a wider DRAM supply shortage.

8. Chinese Semiconductor Catch-Up Should Not Be Ignored

Another source of market concern is the progress of Chinese semiconductor companies.

In the past, Chinese semiconductor technology was often dismissed, but sentiment has changed recently.

Two key issues stand out.

China’s push to localize lithography equipment.
The growth of memory companies such as CXMT and YMTC.

In semiconductor manufacturing, lithography equipment is essential for drawing circuit patterns on wafers.

At the leading edge, ASML effectively dominates the global market.

Because of U.S. export controls, China cannot freely obtain ASML’s most advanced equipment.

China’s announcement of domestic lithography development has raised concerns that a major bottleneck may be narrowing.

It remains difficult for Chinese equipment to catch up with ASML in the short term.

However, semiconductor valuations can weaken quickly once the technology gap begins to narrow.

9. The Pressure Created by CXMT and YMTC

CXMT is China’s leading DRAM company.

Its market share is estimated to have risen from roughly 3% in the past to around 8%, increasing its visibility.

With a recent IPO that raised substantial capital and strong policy support from the Chinese government, its strategic importance has grown.

Chinese authorities also supported the listing process in ways that increased investor attention.

As a result, the company has become a symbol of China’s memory self-sufficiency efforts.

YMTC is attracting attention in NAND.

Together, CXMT in DRAM and YMTC in NAND are shaping China’s memory localization strategy.

The market is particularly concerned that global firms such as Apple could increase use of Chinese memory if supply tightness persists.

If CXMT’s share rises beyond 10% and toward 15%, pressure on the existing three-player memory structure could intensify.

10. Why Samsung Electronics and SK Hynix Still Have the Advantage

Although China-related risk is real, Korean memory makers still hold a strong position in the AI supply chain.

In particular, high-performance HBM and server DRAM are not markets where production alone is sufficient.

Quality certification from customers such as Nvidia, AMD, and major cloud providers is required.
Yield management is critical for high-performance products.
Packaging, thermal performance, power efficiency, and reliability must all be validated.
Scale and long-term supply credibility are also essential.

Chinese companies may increase their share in general-purpose DRAM.

However, replacing Samsung Electronics and SK Hynix in premium AI server memory is unlikely in the near term.

That is why short-term volatility and long-term industry direction should be analyzed separately.

11. Is AI the Same as the Dot-Com Bubble?

Bubble concerns around AI continue.

However, the current AI cycle differs materially from the dot-com bubble of the early 2000s.

During the dot-com era, any company with “.com” in its name could rally sharply.

Companies such as Pets.com drew attention despite weak fundamentals and unsustainable business models.

Heavy marketing spending, low barriers to entry, and weak monetization eventually led to collapse.

By contrast, today’s AI ecosystem includes companies with real earnings.

Nvidia, TSMC, SK Hynix, Samsung Electronics, Broadcom, AMD, and Micron are all benefiting from actual AI infrastructure demand.

Prices can still become overheated, but the industry itself is not merely an empty theme.

12. The Market’s Understated Reality: Hyperscalers Are Under Pressure, Suppliers Are Benefiting

Many reports focus only on the expansion of AI data center investment.

More important is the structure of profit distribution.

At present, hyperscalers are bearing the heaviest burden.

Microsoft, Meta, Oracle, and Alphabet are spending heavily to build data centers.

As GPU, HBM, DRAM, SSD, and networking equipment costs rise, the financial burden increases further.

By contrast, suppliers of constrained components gain pricing power.

TSMC can raise foundry prices.
SK Hynix and Samsung Electronics can benefit from higher HBM and DRAM prices.
Micron can also benefit from the improving memory cycle.
SanDisk and Western Digital can benefit from stronger NAND and storage demand.

In AI, the most stable profits may accrue to bottleneck suppliers rather than end users.

This is a key point that is often underemphasized in market commentary.

13. The Next Battleground May Be CPUs, Not GPUs

From 2023 through 2025, GPUs were the leading focus of the AI market.

Nvidia captured the largest gains, and AMD and Broadcom also benefited from related expectations.

Going forward, CPUs may become more important.

The difference between GPUs and CPUs can be described simply.

A CPU is like one professor solving difficult problems sequentially.
A GPU is like tens of thousands of elementary students solving simple problems simultaneously.

GPUs are stronger for AI training.

Training requires parallel processing of large datasets.

That is why HBM demand is also so strong.

Inference, however, may place greater importance on CPUs.

Inference is the process of delivering answers based on previously trained models.

Because user requests differ and contextual judgment matters, control and orchestration become more important.

14. CPU Demand May Rise as the AI Agent Era Develops

An AI agent is essentially a personalized AI assistant.

For example, an insurance salesperson could ask AI to do the following.

Organize the customers to contact today.
Identify the 10 customers most likely to renew.
Draft messages for each customer.
Summarize previous consultation records.

These tasks are not solved only through parallel computation.

They require personalization, conditional logic, sequencing, and contextual understanding.

That is why CPUs, memory, and on-device AI chips may gain importance in the inference era.

Healthcare is similar.

AI can compare blood glucose data from yesterday, today, and the past several years, while also organizing medication records.

As these functions move into smartphones, wearables, and robots, opportunities in on-device AI and physical AI may expand.

15. Why Nvidia, AMD, and ARM Are Competing in the CPU Market

Traditionally, the CPU market has been dominated by Intel and AMD.

Intel has long been a major force in the x86-based CPU market, while AMD has grown into a strong competitor.

However, ARM-based CPUs are emerging as a new challenger.

ARM’s advantage is lower power consumption.

That is a meaningful strength in AI data centers, where power usage is extremely high.

Nvidia should not be viewed as only a GPU company.

Its next-generation platform is moving toward an integrated CPU-GPU approach designed to capture more of the AI server market.

AMD is also leveraging its CPU base to challenge in GPUs and AI accelerators.

As a result, AI semiconductor competition is likely to become a platform contest spanning CPUs, GPUs, HBM, DRAM, networking, and software.

16. Why SK Hynix ADR and Samsung Electronics ADR Matter

SK Hynix has issued ADRs in the U.S. market.

ADRs are depositary receipts that allow foreign shares to trade in the U.S.

This matters because it may support valuation re-rating.

While Micron is often valued at around 11x forward 12-month earnings, SK Hynix often trades at a lower multiple.

If Korean companies receive a discount simply because they are Korean, broader U.S. market access could reduce that discount.

ASML and TSMC both saw improved access to global investors through ADRs and later moved closer to U.S.-market valuation standards over time.

If Samsung Electronics also moves toward ADR consideration, that could matter for foreign inflows and valuation.

17. July and August Volatility Is Also a Flow-Driven Issue

Second-half volatility is not driven only by fundamentals.

July and August are common periods for rebalancing by global pension funds and asset managers.

Stocks that rose sharply in the first half may face trimming pressure.
Some European pension funds have limits on single-stock weightings.
If a stock rises too much, it may need to be reduced mechanically.
In Korea, the National Pension Service may also face pressure to adjust positions in Samsung Electronics and SK Hynix.

In such conditions, share prices can weaken even if earnings remain strong.

When combined with short selling, long-short strategies, and hedge fund flows, the result can be repeated sidecar and circuit-breaker events.

18. Why ETFs May Be Preferable to Individual Stocks

In technology, winners and losers can change quickly.

Today’s leader is not guaranteed to remain the leader tomorrow.

For that reason, excessive concentration in a single stock can create significant losses when competitive dynamics shift.

ETFs, by contrast, reduce exposure to firms that lose momentum and naturally increase exposure to those that survive and scale.

Investors can use a domestic semiconductor ETF that includes both Samsung Electronics and SK Hynix.
An ETF tracking the Philadelphia Semiconductor Index is another option.
An AI ETF that includes U.S. AI semiconductors and big tech can also provide diversification.
Combining broad equity ETFs with technology ETFs can help reduce volatility.

Even if a stock such as Oracle or Meta moves sharply, an ETF can be balanced by other holdings.

For investors seeking long-term AI exposure but concerned about single-company risk, semiconductor and AI ETFs may be a practical alternative.

19. Key Checklist for Investors

When evaluating Samsung Electronics, SK Hynix, and AI semiconductor exposure in the second half, the following variables should be monitored together.

Whether Middle East risk eases and oil prices stabilize.
Whether expectations for U.S. rate cuts remain intact.
Whether hyperscaler CAPEX continues rather than declines.
Whether weaker FCF leads to actual investment slowdown.
Whether HBM and general-purpose DRAM prices continue to rise.
Whether CXMT and YMTC continue to gain share.
How the CPU competition between Nvidia, AMD, and ARM evolves.
Whether SK Hynix ADR and Samsung Electronics ADR developments support valuation re-rating.

20. Conclusion: Samsung Electronics and SK Hynix Are Not Over, but Await Reassessment Amid Volatility

It is difficult to argue that Samsung Electronics and SK Hynix will rise steadily in the second half without interruption.

Trump-related risk, oil prices, U.S. rates, Chinese semiconductor competition, and global portfolio rebalancing all increase volatility.

However, the AI semiconductor industry itself does not appear to be weakening.

AI data center investment continues, HBM demand remains strong, and the possibility of broader DRAM shortages is increasing.

As inference AI, AI agents, on-device AI, and physical AI develop, CPU and memory demand may enter a new phase.

Ultimately, the key is not to react only to short-term price swings, but to identify where bottlenecks lie in the industry.

At present, the main bottlenecks in the AI ecosystem remain high-performance semiconductors, HBM, DRAM, foundry capacity, power, and data center infrastructure.

Samsung Electronics and SK Hynix remain at the center of those bottlenecks.

For investors who lack confidence in individual names, a diversified approach through semiconductor and AI ETFs may be more practical.

The AI semiconductor cycle is not over; it is moving from GPU-centered training toward the next phase led by CPU inference and AI agents.

< Summary >

The main variables for the second half are Middle East risk, oil prices, U.S. rates, and AI data center investment.

AI semiconductor earnings remain solid, but share prices are volatile due to high expectations and CAPEX concerns.

Samsung Electronics and SK Hynix may benefit not only from HBM but also from broader DRAM tightness.

Chinese competitors such as CXMT and YMTC are a risk, but Korean firms retain strong competitiveness in premium AI memory.

The AI market may continue to expand from GPU training toward CPU inference, AI agents, on-device AI, and physical AI.

For investors concerned about single-stock volatility, semiconductor and AI ETFs may offer a more balanced approach.

[Related Articles…]

*Source: [ Jun’s economy lab ]

– 삼전닉스 아직 안 끝났습니다. 이제 시작입니다(ft.최고민수 작가 2부)

NextGenInsight.Net

Like this: