🚪 Model opens, supply chain deepens

지난 week, we discussed how as AI models become cheaper, their usage will increase, eventually requiring more GPUs, memory, power, and data centers.

This week, the next scene has become a little clearer.

In China, the open-weight model competition is accelerating again. Companies are moving towards selecting and using multiple models based on task difficulty and cost, rather than using the most expensive frontier models for all tasks.

In the real world, however, the exact opposite is happening.

Memory is locked into long-term contracts, and advanced packaging production capacity is limiting customer growth. Power grids cannot keep up with data center demand, and even test equipment, connectors, optical communication, and cooling components—all necessary to complete AI servers—are emerging as new bottlenecks.

Models are becoming increasingly open, but the supply chain that transforms AI into real value is becoming more complex and deeper.

Here's what I see as the core point this week:

Scarcity in the AI era is shifting from the models themselves to the data, workflow, packaging, testing, power, and cooling supply chains that turn models into actual services.

1. The economics of the task are more important than model performance

In the AI model market, performance gaps are rapidly narrowing.

Not only US frontier models, but also Chinese models like DeepSeek, Qwen, Kimi, and GLM series are increasing their price-performance ratio, expanding the range of choices.

This change doesn't simply mean that Chinese models are catching up to US models.

The important point is that the companies' questions themselves are changing.

In the past, they asked:

Which model is the smartest?

In the future, they are likely to ask:

Which model should I use to complete this task most economically?

High-performance models can be used for complex coding and legal reviews, while cheaper models can be used for simple document classification or search. Within a single service, multiple models will be connected, and a structure that automatically selects based on cost, speed, and accuracy will become widespread.

This could reduce the monopolistic power of any single model.

Instead, the importance of platforms through which corporate data and requests pass, regardless of which model is chosen, will grow.

What's important here is not just looking at companies that directly create new AI models.

An approach that focuses on where data and usage pass through, no matter which model wins, can be more stable.

2. Companies pay for results, not tokens

As model options increase, companies will prioritize return on investment over simple performance.

Traditional SaaS was billed based on the number of seats.

If 100 employees use it, 100 seats are paid for.

However, as AI begins to replace actual work, it becomes difficult to explain the product's value solely by the number of seats. One employee might operate multiple AI agents, or a system used by hundreds of people might not generate actual work performance.

Therefore, the following billing methods are likely to increase in the future:

• Number of customer inquiries processed
• Volume of documents and code tasks completed
• Hours of work saved
• Revenue generated
• Insurance claims resolved
• Security analyses completed

The competitiveness of AI SaaS will be built not on the number of features, but on how effectively it completes specific tasks for customers.

In this trend, companies that already dominate internal corporate workflows may have an advantage.

Models can be purchased.

However, connecting to an organization's data, permissions, approval processes, and exception handling is not simply a matter of purchase.

Therefore, the bottleneck in enterprise AI is shifting from model performance to workflow design and system integration.

3. Data and workflows are stronger moats than models

As AI becomes more commoditized, the value of proprietary data may increase.

Take medical AI as an example: models that excel at solving medical exam questions will continue to grow in number.

However, what questions doctors actually ask during consultations, what evidence they trust, and what answers they reconfirm are difficult to obtain simply by training a model.

This is because such data can only be acquired by being embedded within the actual workflow of medical professionals.

The same applies to robots.

In the future, companies that collect skilled workers' hand movements, gaze, work sequences, and exception handling may be as important as humanoid robot manufacturers.

The reason for filming workers' movements in a factory is not simply to obtain video footage.

It is to transform how humans make judgments, move in what sequence, and respond to unexpected situations into data that robots can learn from.

Once collected, work data becomes an asset that can repeatedly train multiple robots.

The true asset in the age of humanoids may not be a single robot, but the work data that enables robots to learn tasks.

However, simply having a lot of data is not enough.

Data quality, usage rights, update frequency, and connectivity to actual work are crucial.

A company that acquires new, high-quality data daily through its work processes may have an advantage over a company that holds large amounts of old data.

4. AI servers are not completed by GPUs alone

When discussing AI semiconductors, most people immediately think of GPUs and HBM.

These remain central.

However, the market is already well aware of this fact. Therefore, it is necessary to look more deeply into the supply chain behind GPUs, rather than just the leading stocks.

As AI chips become more complex, the difficulty of processes such as wafer cutting, polishing, bonding, inspection, and testing also increases.

Designing a good chip is not enough.

It is necessary to secure yield during mass production and verify that the packaged chips actually function normally.

Backend Bottlenecks in Advanced Packaging

These companies are interesting because their demand can increase as process complexity rises, regardless of who wins the AI chip race.

Whether $NVDA NVIDIA wins, or proprietary ASICs increase, mass production of new AI chips absolutely requires inspection, testing, substrates, and packaging.

However, a good industry and a good price are different.

Some Japanese and European equipment companies have already largely priced in AI packaging expectations, so rather than chasing them immediately, it's necessary to consider both the pace of earnings growth and valuation.

5. Taiwan's supply chain is not just TSMC

When discussing the Taiwan AI supply chain, $TSM TSMC is often the only company mentioned.

However, an actual AI server is not made by a foundry alone.

An entire ecosystem is needed, including ODMs that design and assemble servers, data center switches, CPU sockets, power components, and cooling systems.

An interesting company in this trend is $TW:3533 Lotes.

As CPU and GPU performance increases, the amount of power and signals that sockets and connectors must transmit also rises. As servers become more advanced, the technical difficulty of connectors, which once seemed like simple components, also increases.

$TW:3017 Asia Vital Components is also noteworthy.

As the power density of AI racks increases, it becomes difficult to manage heat with air cooling alone. As the transition from internal server fans and heatsinks to liquid cooling systems progresses, the role of thermal management companies grows.

These are not AI leading stocks, but rather an approach that looks at specific bottlenecks in the supply chain that actually complete a server.

6. Power must travel from the power plant to the server rack

The increasing power demand of AI data centers is already a familiar topic.

However, a data center doesn't immediately become operational just because there's a lot of power generation.

The electricity generated at the power plant must be delivered to the server rack, passing through transmission lines, substations, cables, distribution panels, and power supply units.

The equipment required for this process is surprisingly extensive.

In particular, $J5803 Fujikura is an interesting candidate.

Fujikura is known as a wire and cable company, but it is being re-evaluated due to the increasing demand for optical fiber and high-density optical wiring for AI data centers. This is because as AI clusters grow, the importance of optical communication networks connecting servers increases.

Europe's Prysmian and Legrand are in the same trend.

Not all power companies benefit equally from the increase in data centers. Companies that generate power and companies that deliver power to actual facilities have different roles.

In this cycle, bottlenecks may appear in components closer to the server rack, such as cables, distribution, power conversion, and thermal management, rather than in power plants.

7. Memory is shifting from a commodity component to a strategic asset

While openness is accelerating in the model domain, national barriers are rising in the semiconductor supply chain.

The US views China's memory and semiconductor supply chain as a national security issue, and China is expanding its own memory production capacity, centered around CXMT and YMTC.

In the past, memory was considered a commodity component, where suppliers could be chosen based on price.

But now, it has become a strategic asset that powers AI servers, smartphones, automobiles, and defense systems.

Models can be open-sourced, but memory production capacity, yield, and customer validation are difficult to replicate in a short period.

However, this article doesn't need to repeatedly explain only $K000660 SK Hynix and $K005930 Samsung Electronics.

The more interesting area to watch this time is the equipment and materials needed for memory processes.

Japan's $J5332 TOTO is mostly known as a sanitary ware company, but it also operates in the area of ceramic components and electrostatic chucks for semiconductor equipment.

As AI semiconductor production increases, the demand for ceramics, vacuum components, and materials repeatedly used inside equipment can also increase.

This is an approach not to chase leading stocks at high prices, but to find components that are necessarily replaced and consumed as AI grows.

8. Even if the industry is good, stock prices can collapse

The fact that AI infrastructure demand is strong and that related stocks always rise are not the same thing.

When looking at an industry, at least three things must be distinguished:

1. Is industry demand actually increasing?
2. Are the company's sales and profits increasing?
3. Are the current stock price and investor positioning safe?

Even if the first and second points are true, if the third collapses, stock prices can undergo significant correction.

Especially in areas with high market expectations like AI and semiconductors, leveraged and thematic products, options, and passive funds can push prices up faster than fundamentals.

Even if the long-term direction of the industry is correct, buying at too high a price can lead to not making a profit for a long time.

Therefore, more important than repeatedly mentioning the same stocks is finding the next stage of the supply chain within the same logic that the market has not yet fully recognized.

A good company and a good price are different. A good industry and a good stock are also different.

Summary from an Investment Perspective

This week's trends can be summarized by the following groups of stocks:

This does not mean that existing core stocks are unimportant.

$K000660 SK Hynix and $MU Micron are still at the center of HBM, $TSM TSMC in advanced foundries, and $E:ASML in lithography equipment.

However, these are now closer to benchmarks that the market is well aware of.

Future investment ideas are likely to come not from repeatedly explaining these leading stocks, but from tracing how the growth of these leaders translates into the performance of equipment, component, material, and data companies.

Conclusion

This week's trends show one thing:

AI models are rapidly becoming commoditized.

Companies are starting to choose suitable models for each task rather than being tied to a single model, and they are prioritizing data and business outcomes over the models themselves.

However, the real world that actually runs AI is becoming even more complex.

Equipment for cutting, bonding, and inspecting chips, sockets and switches for connecting servers, cooling systems for dissipating heat, and cables and distribution equipment for delivering power are all necessary.

Therefore, in the future, the following questions may become more important than simply saying "AI is growing":

Is it needed no matter which model wins? Is supply difficult to increase quickly? Is it deeply embedded in the customer's process or workflow? Is it used repeatedly rather than being a one-time sale?

From this perspective, interesting candidates to watch include $E:BESI, $J6857 Advantest, $J6146 DISCO, $CAMT Camtek, $TW:3533 Lotes, $TW:3017 Asia Vital Components, $J5803 Fujikura, $E:PRY Prysmian, $CFLT Confluent, $ESTC Elastic, $E:VACN VAT Group, and $J5332 TOTO.

This does not mean that all these stocks are at a good price right now.

Some already have high expectations, and for others, AI demand still accounts for a small portion of their overall performance.

However, it is clear that the AI bottleneck is shifting from GPUs and HBM to packaging, testing, connectors, optical communication, cooling, power delivery, and industrial data.

Now is the time to look for the next, less-known bottlenecks in the AI supply chain, rather than repeatedly following the same leading stocks.