The most interesting scene in the market from July 7 to 12, 2026, was the performance competition among AI models.

OpenAI unveiled GPT-5.6 and ChatGPT Work, while in China, open-weight models including GLM-5.2 are rapidly closing the performance gap with US frontier models. With Grok and Meta also releasing models emphasizing price-performance, the cost of using frontier-level AI is now likely to continue to decrease.

However, the exact opposite is happening in the semiconductor and data center markets.

Memory is scarce, advanced packaging orders are overflowing, and there isn't enough power or water to build data centers. Demand is spreading to metals like copper and aluminum, cooling systems, and optical communication equipment.

Models are getting cheaper, but the physical world required to actually run those models is becoming more expensive.

This is what I see as the core insight of this week.

The price reduction of AI models may not be a factor that reduces AI infrastructure investment, but rather a factor that enables more agents and longer tasks, leading to an explosion in computing consumption.

1. AI's Next Competition Isn't Model Performance, But Workload

How smart GPT-5.6 is, or which model, Claude or Grok, leads in benchmarks, is certainly important.

However, from an investment perspective, there's a more significant change than model rankings: the length and complexity of tasks people are entrusting to AI are rapidly increasing.

OpenAI recently revealed that its internal agentic token usage has increased approximately 22-fold over the past six months, and the proportion of computing used for internal coding inference has grown 100-fold. ChatGPT Work has integrated Codex, and Codex's weekly users already exceed 5 million.

In the past, AI generated one answer for one question.

Today's agents search for information, read documents, write code, correct errors, review results, and then rework. A single user command can internally trigger dozens to hundreds of model calls.

A new input factor has entered the production function of knowledge work.

Previously, it was roughly as follows:

Human Time × Skill Level

In the age of agents, one more factor is added:

Human Judgment × Supervised Agent Labor × Computing

Even if the number of users no longer significantly increases, the tokens and computing consumed per person can continue to grow. AI infrastructure demand is no longer solely dependent on 'user growth' but gains a second engine: increased consumption per user.

However, a large number of tokens generated does not mean that value increases at the same rate.

More time might be spent reviewing and correcting AI-generated output. As agents take over execution, the enterprise bottleneck shifts from execution itself to defining what to ask for and judging if the results are correct.

AI may not eliminate domain expertise; rather, it could increase the value of verifiable expertise.

2. Model Prices Go Down, But Infrastructure Prices Go Up

In the AI model market, price-performance is rapidly improving.

If cheaper models catch up to the performance of frontier models, the pricing power of model companies may weaken. However, this does not necessarily mean a decrease in demand for data centers and semiconductors.

Just as lower electricity prices don't necessarily reduce electricity consumption, when token prices fall, companies begin to apply AI to tasks that were previously uneconomical.

Cheap inference creates more AI agents, in-house tools, personalized services, and simulations. The cost of running a single model decreases, but the total amount of computing consumed by the entire industry can actually increase.

Goldman Sachs projects that global AI capital expenditure could increase from approximately $765 billion in 2026 to $1.6 trillion annually by 2031. The cumulative investment from 2026 to 2031 is expected to reach approximately $7.6 trillion.

At the heart of this investment are still semiconductors.

Recently, $K000660 SK Hynix's US ADR issuance attracted orders more than 7 times the planned volume. The final fundraising amounted to $26.5 billion, making it the largest stock issuance by a foreign company in the US market ever. This demonstrates how highly the market values the structural growth potential of HBM and AI memory.

The problem is that HBM is not the only thing in short supply.

AI chips include not only GPUs but also CPUs, HBM, network chips, power semiconductors, substrates, and numerous passive components. To integrate these into a single system requires advanced packaging and testing.

While TSMC is rapidly expanding its CoWoS production capacity, it is still insufficient to keep up with increasing demand. The reason orders are being diversified to other packaging companies like ASE, Amkor, and Intel, not just TSMC, is not because TSMC's competitiveness is weakening, but because the overall market's supply capacity is insufficient.

A good company and a good price are different.

However, it is difficult to view the current memory and packaging shortages as merely short-term scarcity. This is because AI infrastructure demand is expanding from just GPUs to the entire computing system.

3. The Data Center Bottleneck Is Now Power Grids and Water

The perspective is also changing in the data center market.

Previously, it was important which cloud company bought more servers. Now, it's more important whether a data center, once built, can actually receive power.

Gartner estimates that global data center power consumption will increase by 26% from 447 TWh in 2025 to 565 TWh in 2026. Specifically, power consumption by AI-optimized servers is projected to increase by 84% in 2026 alone and surpass that of traditional servers by 2027.

The International Energy Agency projects that global data center power consumption will increase to approximately 945 TWh by 2030. This is more than Japan's entire current electricity consumption. In the US, data centers could account for nearly half of the increased electricity demand by 2030.

Now, the data center bottleneck is not merely a shortage of power generation.

What matters is whether power can be delivered to the necessary locations, at the necessary times, without interruption. In regions where power grid connections take a long time, data centers may not be able to operate even if land is secured.

Cooling methods are also changing.

As the power density of AI racks increases, there is a shift from traditional air cooling to direct-to-chip and closed-loop liquid cooling. In Korea, $K066570 LG Electronics is developing a prototype AI server rack compliant with NVIDIA Vera Rubin specifications and preparing for mass production. This is an approach to integrate HVAC, batteries, and data center construction capabilities.

LG Electronics is not a company that creates AI models.

However, as AI data centers increase, server racks and cooling systems are absolutely necessary. This is why we should look at 'AI infrastructure that must be passed through' rather than 'apps AI will replace.'

4. Sovereign AI Is a Supply Chain Issue, Not a Model Issue

This week, observing the Chinese AI industry, I re-confirmed something:

Sovereign AI should not be understood as merely developing a proprietary language model.

Even if a country possesses frontier models, if it relies on foreign countries for GPUs, HBM, foundries, packaging, data centers, and power grids, it is difficult to say that it has secured complete AI sovereignty.

Recently, open-weight models like China's GLM-5.2 are rapidly closing the gap with US frontier models in some coding and agentic tasks. Their prices are only a fraction of US models, attracting growing interest from global startups and SMEs.

At the same time, DeepSeek is reportedly embarking on the development of its own AI chips for inference. This move aims to reduce reliance on NVIDIA and Huawei by collaborating with chip design, foundry, and memory companies.

This is important because China's AI competition does not end with models.

Alibaba and Tencent have cloud and models, SMIC handles foundries. CXMT and YMTC are building memory supply chains, and Huawei is simultaneously expanding AI accelerators, networks, and computing clusters.

This does not mean that China can immediately and completely replace NVIDIA, TSMC, and SK Hynix. There are still gaps in advanced process yields, HBM, and equipment performance.

However, the value of 'the best performing system' and 'a system that can be operated independently in one's own country' are different.

As US export controls tighten, Chinese companies will prioritize supply stability and self-reliance over mere technical performance. In this trend, not only frontier model companies but also Chinese foundries, memory, and semiconductor equipment companies should be considered.

However, when looking at Chinese semiconductors, one should also be wary of exaggerations like 'they will soon completely catch up with the West.'

The investment point for the Chinese supply chain is not necessarily that it will achieve the highest performance, but that it is becoming a sufficiently good alternative that can be used internally within China.

5. AI's Economic Impact Spreads Beyond Tech Companies

The AI industry cannot be explained solely by data center investments.

As the cost of using AI decreases, people can access services they previously paid for, such as tutoring, translation, travel planning, software, and document creation, for free or at a very low cost.

Stanford Digital Economy Lab estimates that the consumer surplus gained by US generative AI users increased from approximately $116 billion in mid-2025 to $172 billion in early 2026. This analysis suggests that the time and cost savings for consumers are greater than the revenue companies earn from AI services.

What is important here is that AI is not simply creating new revenue, but can reallocate existing consumer spending.

If AI replaces tutoring, translation, and simple software subscriptions, that much spending power can shift to areas that AI cannot directly replace, such as dining out, travel, and entertainment.

Conversely, companies that sold simple information provision or repetitive execution are likely to face price pressure.

Beneficiaries in the AI era do not necessarily have to be companies that create AI models.

Companies that possess unique data and customer relationships and use AI to lower costs or improve decision-making quality can also be potential beneficiaries. The case of insurer $TRV Travelers adopting an AI-based claims processing system is a prime example.

Summary from an Investment Perspective

The trends observed this week can be summarized into four points:

The most important thing is that model companies and infrastructure companies should not be evaluated with the same logic.

In the model market, price-performance competition can intensify. Today's top model might be caught up by a low-cost model in a few months.

In contrast, memory, advanced packaging, power grids, and cooling infrastructure are difficult to scale up quickly. It takes several years to build factories, install equipment, and pass customer verification.

Therefore, while the competitive advantage of models changes rapidly, bottlenecks in the physical supply chain are likely to persist longer.

Conclusion

Ultimately, what the market showed this week is one thing:

AI is becoming increasingly cheaper, but the real-world resources needed to use this cheaper AI at scale are becoming scarcer.

When model prices fall, AI investment does not end. Instead, as more people entrust AI with longer tasks, token and computing consumption can increase.

In that process, bottlenecks continue to shift from GPUs to HBM and general DRAM, foundries, advanced packaging, power grids, cooling, water, and metals.

Therefore, going forward, it may be more important to identify the supply chain that AI must pass through to perform its work, rather than simply observing which AI model ranks first in benchmarks.

From this perspective, the stock groups to continue watching include $K000660 SK Hynix, $K005930 Samsung Electronics, $MU Micron, $TSM TSMC, $AMAT, $LRCX, $VRT, $ETN, $K066570 LG Electronics, $H09988 Alibaba, and $H00981 SMIC.

However, structural growth and the short-term direction of stock prices are separate issues.

As expectations for AI and semiconductors have grown, leveraged funds and theme-based products have also rapidly increased. Even if the industry grows long-term, if prices and supply-demand run ahead of fundamentals, corrections can occur at any time.

A good company and a good price are different.

Now is not the time to worry about whether AI is ending, but rather to more precisely distinguish which industries' supply shortages and pricing power are connected to AI's growth.