● HBM Demand Run Could Outlast GPU Peak
Will AI Data Centers and HBM Memory Demand Really Peak Out?
The core point of this issue is not the news that “Meta and SpaceX are renting out leftover data center resources.”
The truly important point is that AI data centers are changing from “server warehouses” into “factories that produce tokens,” and in this structure, HBM memory, DRAM, LPDDR, and SRAM demand follows more persistently than GPU demand.
Recently, the stock market has seen renewed debate over semiconductor outlooks, the memory cycle, and the semiconductor peak-out narrative.
But Vice President Kim Ji-hyun of the SK AI Committee has a fairly clear perspective.
Short-term stock prices can be swayed by desire and sentiment, but the structural direction of AI infrastructure investment and memory demand is still difficult to say has turned downward.
In particular, when the era of AI agents opens, token usage will surge, and demand for data centers, GPUs, NPUs, ASICs, and HBM memory to process those tokens is likely to grow even further.
1. The semiconductor peak-out debate should separate stock prices from industrial demand
Recently, as stock prices of memory semiconductor companies such as Samsung Electronics, SK Hynix, and Micron have swung sharply, many have begun asking whether this is already the peak.
In particular, because expectations related to HBM memory are being reflected quickly in stock prices, even slightly negative news is fueling concerns about a semiconductor peak-out.
Vice President Kim makes an important distinction here.
Corporate value reflects current revenue, future growth potential, and market participants’ desires all at once.
Current revenue and future growth potential can be analyzed to some extent through technology and demand.
But expectations, fear, and desire attached to stock prices are difficult to predict.
So it is dangerous to simply conclude that “if the stock price falls, industrial demand has also weakened.”
Conversely, it is also overly simplistic to think that “if industrial demand is strong, stock prices will keep rising.”
This debate requires separating investor sentiment swings from the structural changes in AI infrastructure demand.
- Stock prices can swing in the short term due to desire and fear.
- AI data center investment should be viewed as medium- to long-term infrastructure demand.
- HBM memory demand is connected not only to GPU sales volume but also to changes in how AI is used.
- The memory cycle should be interpreted differently from the old DRAM-centered cycle.
2. Why the Meta and SpaceX “surplus data center” controversy is close to noise
The most widely discussed issue in the market recently was the idea that companies like Meta and SpaceX could lease out unused data center resources to other firms.
At first glance, this news can make people think, “Have we built too many AI data centers?” or “Has GPU demand now peaked?”
But Vice President Kim interprets this issue not as a signal of demand collapse, but as a signal that market opportunities have expanded.
Meta is not, in the first place, a traditional cloud provider like AWS, Microsoft Azure, or Google Cloud.
Meta is closer to a company that built large-scale in-house data centers to operate Facebook, Instagram, and AI services.
But if external data center supply is scarce and AI infrastructure prices rise, Meta would naturally want to sell off excess resources.
This is less about “resources going unused because there is no demand” and more about “a market has opened where they can be sold at a premium.”
SpaceX or xAI-related resources can be viewed similarly.
If a particular AI service does not grow as quickly as expected, GPUs or data center resources already secured may temporarily sit idle.
But if AI infrastructure supply is tight, leasing those resources externally is a natural business decision.
In other words, it is difficult to conclude that overall AI data center demand has weakened just because a few companies have surplus resources.
| Issue | Surface-level interpretation | Vice President Kim’s reinterpretation |
|---|---|---|
| Potential leasing of Meta’s data centers | AI infrastructure is sitting idle | It is considering external sales because the cloud market is profitable |
| Use of SpaceX and xAI resources | AI demand is weaker than expected | Short-term mismatches in some resources are being monetized in a high-priced market |
| GPU rental price fluctuations | GPU demand has peaked out | The supply structure and demand channels are being reorganized |
3. Data center demand is not shrinking; it is growing at the national level
To understand AI data center demand, you need to look at national power-based data center expansion plans rather than the surplus resources of individual companies.
Based on the approximate figures presented in the original, the United States, China, and South Korea are all moving to significantly expand data center scale.
The United States is said to be aiming to expand current data center power demand of about 54 GW to around 130 GW by 2030.
China is mentioned as moving from about 32 GW to around 60 GW.
South Korea is discussing a target of raising its data center scale, currently estimated at less than 1 GW, to around 10 GW by 2030.
The exact numbers can vary depending on policy, grid capacity, and investment speed.
But the direction is clear.
AI data centers are not a shrinking industry; they are infrastructure expanding at the level of national competitiveness.
As data centers increase, demand for GPUs, NPUs, ASICs, servers, power equipment, cooling systems, and HBM memory rises together.
For South Korea’s economy in particular, memory semiconductor demand inside those data centers is far more important than the data centers themselves.
- Data center expansion is a core indicator of AI infrastructure investment.
- A 1 GW data center is not just a building, but a combination of GPUs, memory, and power grids.
- As AI data center demand grows, HBM memory and high-performance DRAM demand grow together.
- South Korea can hold a dual position as both a data center operating country and a memory supply country.
4. Why AI needs more memory as it advances
When many people talk about AI semiconductors, they only look at GPUs.
But in the actual AI computation structure, HBM memory attached next to the GPU is extremely important.
As AI models become larger, conversations get longer, files are uploaded, and multi-step reasoning is performed, more information must be retrieved quickly.
That is where high-bandwidth memory, HBM, comes in.
At first, only a small number of HBM stacks were attached around a single GPU, but as AI models grow, the direction is toward needing more memory configurations, such as 4, 8, or 16 stacks.
In simple terms, as AI gets smarter, not only does the “calculator” need to improve, but the “workbench” must become larger too.
If the GPU is a calculator, HBM is the ultra-fast workbench right next to it.
If the workbench is too small, no matter how fast the calculator is, bottlenecks occur because data must be constantly reloaded.
That is why, when looking at AI semiconductor prospects, you need to examine not only GPU shipments but also the increase in HBM capacity per GPU.
5. Memory is not just HBM
The most important part of this discussion is that memory demand does not end with HBM alone.
Past memory cycles were mainly driven by DRAM used in PCs, servers, and smartphones.
But memory demand in the AI era is becoming far more diversified.
AI data centers need HBM.
Smartphones, PCs, cars, and robots with on-device AI need LPDDR.
SRAM becomes important for ultra-fast inference and cache processing.
For conversation history, training data, user files, and log storage, SSD and NAND demand are also connected.
| Memory type | Main use cases | Meaning in the AI era |
|---|---|---|
| HBM | GPUs, AI servers, data centers | High-bandwidth memory that resolves the key bottleneck of large-scale AI computation |
| DRAM | Servers, PCs, general IT devices | Receives both baseline computing demand and demand around AI servers |
| LPDDR | Smartphones, laptops, on-device AI | Low-power, high-performance memory for running AI inside devices |
| SRAM | On-chip cache, high-speed inference | A key element for improving AI inference speed and power efficiency |
| SSD·NAND | Data storage, cloud, user records | Expanded data storage demand as AI usage rises |
In this structure, memory demand does not move simply as the cycle of one product group.
If HBM is insufficient, DRAM production capacity shrinks, DRAM prices rise, and the impact can spread to PC, console, and server markets.
In fact, recent cases in which some game console or hardware companies adjusted production plans due to rising memory prices can be viewed in connection with this trend.
6. In the era of AI agents, token demand surges
One of the points Vice President Kim emphasized most was AI agents.
Until now, generative AI has mainly been used in a way where users input questions and receive answers.
But AI agents perform multi-step tasks on behalf of users.
They search, read, compare, write code, test, revise, create reports, and connect all the way to execution.
In this process, they use far more tokens than ordinary chatbots.
The original explains that AI agents can use as many as 1,000 times more tokens than general generative AI.
Tokens are the minimum processing unit and billing unit used when AI understands and generates sentences.
So an increase in token usage means an increase in AI computation, data center usage, GPU usage, and memory usage.
This is why data centers are increasingly being called “token factories.”
If companies begin introducing one AI agent per person, department-specific AI agents, and task-specific AI agents, token demand could become much larger than it is now.
- AI agents go through far more computational steps than simple question-answering.
- As computational steps increase, token usage rises.
- Rising token usage leads to higher AI data center demand.
- Greater AI data center demand leads to higher GPU and HBM memory demand.
7. If efficiency improves, will demand fall?
Some argue that if AI model optimization and data center efficiency improve rapidly, data center demand may decline.
Even for a 1 GW data center, the amount of tokens produced can vary depending on hardware configuration, model optimization, and software efficiency improvements.
For example, some 1 GW data centers may produce 100 tokens, while others may produce 200 or 300.
With such improvements, a logic emerges that says, “Can’t we handle the same demand with fewer data centers?”
But Vice President Kim mentions Jevons paradox here.
Jevons paradox refers to the phenomenon where, when the efficiency of using a resource improves, usage does not decrease; instead, it becomes cheaper and more convenient, and total usage increases.
A classic example is coal: when coal efficiency improved, coal consumption did not fall but instead spread more widely across industry.
AI may be similar.
As token prices fall and AI response speeds increase, people do not use AI less; they use it more.
It will not be limited to services like ChatGPT or Claude, but will also be embedded in PowerPoint, Excel, business SaaS, cars, smartphones, robots, and manufacturing equipment.
Ultimately, efficiency improvements are more likely to expand use cases than reduce demand.
8. Why the spread of B2B AI is even more significant
Individual users open and use services like ChatGPT, Claude, and Gemini directly.
But a large share of actual AI revenue comes from the enterprise market.
The original explains that a substantial portion of Anthropic’s revenue comes from B2B.
This matters because even if users do not consciously recognize they are using AI, AI can already be invoked inside many business systems.
For example, it can be used to plan slides inside PowerPoint, analyze expenses inside accounting systems, and draft responses inside customer service solutions.
Even if users think, “I do not use Claude,” Claude or another AI model may actually be operating behind enterprise software.
As this structure spreads, AI usage grows much faster than visible app usage.
The moment AI becomes not a separate service but a basic function embedded in all software, data center and memory demand can grow once again.
9. Why South Korea’s three major mega projects matter
The original mentions data centers, semiconductor fabs, and physical AI as South Korea’s three major mega projects.
These are not industries that move separately.
Data centers are the infrastructure that produces AI tokens.
Semiconductor fabs are the foundation that increases memory supply capacity for HBM, DRAM, LPDDR, and more.
Physical AI refers to the area where AI operates in the real world, such as robots, manufacturing, automobiles, and smart factories.
South Korea has a strong manufacturing base and global memory semiconductor companies like Samsung Electronics and SK Hynix.
So as AI data centers and physical AI grow, South Korea can gain a linkage that benefits its economy.
However, this project is not simply a matter of building factories.
Power, transmission networks, cooling water, land, permits, talent, and securing global customers must all come together.
Because AI data centers consume enormous amounts of electricity, power strategies such as nuclear power, SMRs, solar, wind, ESS, and local self-generation are also needed.
In the end, the key is not direction but speed and execution.
- Semiconductor fabs are responsible for the memory supply chain used in global AI infrastructure.
- Data centers can become the core of a Korea-style token factory strategy.
- Physical AI is a card that can raise manufacturing competitiveness again.
- The power grid and transmission infrastructure can become bottlenecks for AI industry growth.
10. Whether it is GPU, ASIC, or TPU, memory is still needed
Recently, major tech companies have been building their own AI chips to reduce dependence on Nvidia GPUs.
Google uses TPUs, Amazon is developing its own AI chips, and Meta and Microsoft are also strengthening their own accelerator strategies.
Tesla is also directly building AI chips and supercomputing infrastructure.
Looking only at this trend, one might think that if Nvidia GPU demand falls, HBM demand will also fall.
But Vice President Kim’s perspective is different.
Whether it is GPU, ASIC, or TPU, high-performance AI computation still requires memory in the end.
AI chips can be designed in-house, but the number of companies that can stably supply high-performance memory in large volumes is limited.
In the current global memory market, the key players are Samsung Electronics, SK Hynix, and Micron.
So from South Korea’s perspective, it can view the entire global AI chip ecosystem as its customer base, not just Nvidia.
Whether OpenAI, Anthropic, Google, Amazon, Meta, Microsoft, Tesla, AMD, or Intel builds AI infrastructure, memory is required.
This is the most important structural strength in the outlook for South Korea’s semiconductor industry.
11. The AI infrastructure market is becoming vertically integrated
The AI market is increasingly becoming vertically integrated, moving beyond simply selling GPUs.
The first stage is a structure that bundles GPUs and HBM together.
The second stage is a structure that leases data centers.
The third stage is a structure that produces and sells tokens from the data center.
The fourth stage is a token production structure optimized for a specific AI model.
The fifth stage is a structure that bundles AI agents and applications and sells them together.
As the market becomes more integrated up and down the stack, competition to reduce intermediate infrastructure costs intensifies.
The reason foundation model companies such as OpenAI and Anthropic are directly interested in data center investment is also because of infrastructure costs.
In a structure where a substantial portion of costs is paid to cloud providers, profitability can be limited.
But whether they build their own data centers, rent from cloud providers, or use in-house AI chips, memory demand does not ultimately disappear.
12. The China variable is a risk, but South Korea still has time
China is building an independent AI ecosystem in response to U.S. AI semiconductor regulations.
Centered around Huawei, its own AI chips, servers, and software ecosystem are growing.
In the long term, China is also likely to try to advance its memory technology.
This is clearly a risk for South Korea’s semiconductor industry.
But as of now, Korean companies still have strengths in the performance, quality, yield, and reliability of high-performance HBM and advanced memory.
So what matters most is not a debate that drags things down, but rapid execution.
When demand is open, production capacity and the technology gap must be widened further.
Rather than pausing out of concern for the memory cycle, South Korea needs a strategy to preempt high-value-added memory supply chains in line with the spread of AI data centers and on-device AI.
The most important point not often stated in other news
The most important point is not “AI data center demand,” but “the increase in memory capacity per GPU.”
Most news focuses only on how many data centers are being built or how many Nvidia GPUs are being sold.
But what is truly important to South Korea’s economy is that the amount of HBM memory attached to each GPU can keep increasing.
As AI models get larger, context windows get longer, and agents perform multiple tasks simultaneously, the memory bottleneck around GPUs becomes more severe.
Then even if GPU sales remain at the same level, HBM demand can grow further.
In this structure, you should not judge memory demand based only on news that “GPU demand may slow.”
The second most important point is that memory demand does not end with HBM.
If on-device AI enters smartphones, PCs, cars, robots, and smart glasses, demand for LPDDR and SRAM will also rise.
In other words, the memory cycle in the AI era is likely to become much more complex and longer than the past DRAM cycle centered on PCs and smartphones.
The third most important point is that efficiency improvements do not directly translate into lower demand.
As AI becomes cheaper and faster, it can enter more applications and industries.
This is Jevons paradox, and it is the most underestimated variable in the AI infrastructure market.
Checkpoints for investors and industry observers
- HBM supply capacity: The speed at which Samsung Electronics and SK Hynix expand HBM production is key.
- HBM capacity per GPU: This may become a more important variable than GPU shipments.
- Spread of AI agents: You need to check whether token usage is surging.
- Power infrastructure: Data center expansion is impossible without a power grid.
- On-device AI: It can increase demand for new memory types such as LPDDR and SRAM.
- Big Tech’s own AI chips: Dependence on Nvidia may fall, but memory demand is likely to remain.
- China’s technology catch-up: A long-term risk, but in the short term South Korea’s technology gap is an opportunity.
< Summary >
The controversy over surplus data centers at Meta and SpaceX is closer to a monetization issue for some resources than to a signal of collapsing AI demand.
AI data centers are continuing to expand at the national level, and data center expansion is connected to demand for GPUs, NPUs, ASICs, and HBM memory.
Because AI agents use far more tokens than ordinary generative AI, they are turning data centers into “token factories.”
When efficiency improves, demand may not fall; instead, prices may drop and usage may rise, as in Jevons paradox.
Memory demand is diversifying not only into HBM but also DRAM, LPDDR, SRAM, and SSD.
Whether it is GPU, ASIC, or TPU, high-performance AI computation ultimately requires memory.
South Korea can hold a key position in the AI semiconductor supply chain through Samsung Electronics and SK Hynix.
In conclusion, the semiconductor peak-out debate may be valid from a short-term stock price perspective, but it is still difficult to say that the structural demand for AI infrastructure and HBM memory has turned downward.
[Related Articles…]
- HBM Memory and the Outlook for the AI Semiconductor Supercycle
- AI Data Center Investment and Changes in Global Power Infrastructure
*Source: [ 티타임즈TV ]
– “스페이스X와 메타 남아돈다는 건 소음일 뿐” (김지현 SK AI위원회 부사장)
