● AI Shock Fears, Memory Demand Surge
Is This DeepSeek Shock 2.0, or an Exaggerated Fear? A One-Stop Summary of Kimi K3, AI Semiconductors, and the Memory Semiconductor Cycle
The core point of this issue is not simply that “Chinese AI beat U.S. AI.”
The truly important point is that the performance gains of Chinese open-source AI models have shaken sentiment toward tech stocks, while also potentially increasing demand for Nvidia GPUs, HBM, DRAM, and NAND.
Added to this are Google Gemini delays, the monetization of Anthropic and OpenAI, memory semiconductor supply shortages, and inflation risks from rising oil prices.
So in this article, we will organize the DeepSeek Shock 2.0 debate in a news format from the perspectives of AI model performance, cost structure, semiconductor demand, Big Tech investment, and the global economic outlook.
1. The core point of the Kimi K3 controversy: did Chinese AI really beat the U.S.?
The recent market mover was Kimi K3, released by China’s Moonshot AI.
As some benchmarks showed Kimi K3 surpassing Claude and GPT-family models, fears spread that “DeepSeek Shock 2.0 has arrived.”
-
The symbolism is huge.
The fact that a Chinese open-source model outperformed top U.S. models in certain coding benchmarks is clearly a market-shaking event.
-
However, interpreting it as the overall No. 1 model is exaggerated.
Based on the original source, the areas where Kimi K3 appears ahead are mainly specific domains such as frontend coding, UI·UX generation, and web page composition.
In comprehensive benchmarks, it is still reasonable to interpret Claude and GPT-family models as remaining at the top.
-
The score gap is not something that can be called overwhelming.
Although the graph makes Kimi K3 look far ahead, the actual score difference is said to be around 3%.
That degree of difference carries strong symbolic meaning, but it is hard to say it is a gap large enough to overturn the entire real-world user experience.
In conclusion, Kimi K3 is evidence that “Chinese AI has caught up frighteningly fast.”
But interpreting it as “U.S. AI is finished” or “AI semiconductor investment is collapsing” is closer to fear that has gone too far.
2. What makes it different from DeepSeek Shock: Kimi K3 is not as cheap as it seems
The reason DeepSeek Shock shook the market so strongly was price-to-performance.
The key message at the time was that “GPT-level performance was achieved at a much lower cost.”
But Kimi K3 has a slightly different structure.
-
The token unit price may look low.
On the surface, Kimi K3’s output token pricing appears cheaper than GPT or Claude.
-
But the total task cost can be similar.
The issue is that Kimi K3 may use more tokens to complete the same task.
Even if the unit price is low, heavy token usage can make the actual cost of finishing a job comparable to GPT.
-
Real-world speed and user experience also matter.
A model is not good just because it is cheaper.
In agentic development environments that developers actually use, such as Claude Code and OpenAI Codex, model performance matters, but so do the tool ecosystem, response stability, and task continuity.
So it is risky to interpret the Kimi K3 issue as “China has completely taken over the AI market through cost efficiency.”
Rather, it is more accurate to say that Chinese AI is now moving from cost-efficiency competition to performance competition.
3. The real change in Chinese AI: from China of cost-efficiency to China of performance
What must not be missed in this issue is that Chinese AI companies’ technological level has risen rapidly.
Moonshot AI, Alibaba Qwen, Zhipu AI, MiniMax, DeepSeek, and other Chinese AI firms are no longer just making low-cost models.
-
They have entered the large-parameter model competition.
Kimi K3 is mentioned as a model on the scale of about 3 trillion parameters.
There is also a trend indicating that Alibaba’s Qwen line is preparing models above 2 trillion parameters.
-
China also has a strong AI talent pool.
Through elite AI education systems such as Tsinghua’s Yao Class, top engineering talent continues to be produced.
As this talent gains research experience in the U.S. and returns to China’s AI ecosystem, technological competitiveness is rising quickly.
-
Manufacturing competitiveness and AI are being combined.
If you look at Roborock, DJI, Chinese EVs, robots, and smart manufacturing equipment, China is already building a structure that pushes both hardware and software together.
Chinese AI should not be underestimated.
That said, the fact that Chinese AI has become stronger and the conclusion that U.S. AI investment, Nvidia demand, and memory semiconductor demand are collapsing are completely different matters.
4. The market misses the core point: Chinese AI still needs Nvidia and memory
The most important part of this controversy is what kind of infrastructure a large model like Kimi K3 actually runs on.
This is a key point that is relatively less covered in other news reports or YouTube videos.
-
A Kimi K3-scale model requires massive compute infrastructure.
A model on the scale of 2 trillion to 3 trillion parameters is difficult to handle with a single GPU or a small server setup.
Large GPU clusters, high-speed networks, and large-capacity memory infrastructure are essential.
-
Training infrastructure is likely to depend heavily on Nvidia.
If a Chinese company trained on Huawei chips, it would likely have promoted that fact actively.
Without such mention, the market naturally gives a high probability to Nvidia GPU usage.
-
As models get larger, HBM alone is not enough.
As weights and KV cache grow, it becomes difficult to fit all the data into HBM next to the GPU.
Ultimately, DRAM, NAND, SSDs, and high-speed storage are all needed as well.
This is important.
As Chinese AI models advance, Nvidia GPU demand does not fall; rather, demand for high-performance GPUs and memory semiconductors can increase.
In particular, once AI agents begin handling long-term memory, user context, codebases, and corporate databases, NAND and SSD demand is also likely to rise.
5. The rebound effect: as AI gets cheaper, infrastructure demand does not decline — it rises
Even during the DeepSeek Shock, the market initially reacted incorrectly.
That was because it assumed cheaper AI models would reduce GPU demand.
But the actual trend was close to the opposite.
-
When token costs fall, usage rises.
If AI becomes cheaper to use, companies apply it to more tasks.
-
When usage rises, total infrastructure demand grows.
Even if the cost per individual task falls, if total workload explodes, demand for GPUs, data centers, power, and memory actually increases.
-
This is the rebound effect seen in the AI industry.
As efficiency improves, total consumption does not decline — it becomes larger.
So instead of interpreting the Kimi K3 issue as a signal to cut AI investment, it is also necessary to view it as a sign of expanding AI usage.
6. Google’s crisis: Gemini delays and AI talent outflow
Google is also an important variable in this trend.
The original source mentions reports that the release of Gemini’s next-generation model is delayed, and that its coding performance in particular fell short of internal targets.
-
Google is a pioneering AI company, but it has recently been losing ground in the coding model competition.
While Claude Code and OpenAI Codex are rapidly taking over the developer ecosystem, Gemini is getting somewhat disappointing reviews in real-world usage evaluations.
-
Key talent outflow is also a burden.
There is a flow of Google AI talent moving to Anthropic, OpenAI, and startups.
The stock compensation available at companies like Anthropic or OpenAI, which are preparing for IPOs, is a strong incentive for Big Tech engineers.
-
There is also debate over the limits of the TPU structure.
As AI model architectures have recently evolved toward directions like Sparse MoE, some argue that the Nvidia GPU platform may be even more advantageous.
Still, Google should not be seen as finished.
That is because it has search, cloud, YouTube, Android, TPU, and data assets all in one company.
However, from a tech-stock investment perspective, it is worth checking that Google is not the overwhelming leader in the current AI coding model competition.
7. Are AI companies really making money? Anthropic and OpenAI monetization
In the AI bubble debate, the most important question is one.
“Is AI actually profitable?”
The original source emphasizes the revenue structures of Anthropic and OpenAI.
-
Anthropic has strong enterprise API revenue.
Its number of consumer subscribers may be smaller than OpenAI’s, but enterprise customers use the API in large volumes, driving rapid revenue growth.
-
OpenAI bears a heavy free-user burden.
ChatGPT’s user count is very large, but the conversion rate to paid users is limited.
A large number of free users is good for brand and ecosystem expansion, but it is a burden on short-term profitability.
-
AI heavy users dominate total token usage.
Developers, automation users, and enterprise engineers consume enormous amounts of tokens.
Some heavy users may consume compute equivalent to API costs far exceeding monthly subscription fees.
-
B2B API margins are likely high.
Subscription-based B2C may lose money, but enterprise usage-based APIs can produce high margins.
In the end, OpenAI and Anthropic are not companies that are already delivering perfect profits across every area right now; they are closer to platform companies that can turn the monetization switch much more strongly at any time.
If this structure holds, it will be difficult for AI semiconductors and cloud infrastructure investment to slow down easily.
8. The AI arms race: the U.S. cannot easily reduce investment
AI is no longer simply a productivity tool.
It is a national strategic asset.
From the U.S. government’s perspective, Chinese AI reaching a level that threatens top U.S. models is a highly sensitive issue.
-
AI is being treated like a nuclear-level strategic technology.
AI is connected to defense, cybersecurity, intelligence analysis, unmanned weapons, scientific research, and semiconductor design.
-
If China catches up, the U.S. will find it hard to reduce investment.
Rather, it is highly likely that the government and private sector together will increase investment in AI data centers, AI semiconductors, and power infrastructure.
-
Meta, Google, xAI, OpenAI, and Anthropic all can’t easily step out of the competition.
That is because once they fall behind in AI competition, they may lose platform dominance.
From this perspective, Kimi K3 may be less a signal of reduced AI investment and more a signal of an expanded AI arms race.
9. Sharp semiconductor stock decline: fundamental problem or supply-demand problem?
Recently, semiconductor indices and AI hardware stocks have undergone a major correction.
In particular, many memory semiconductor, optical communications, and AI server component stocks have fallen significantly from their highs.
However, the original source sees the core of this decline as a supply-demand issue rather than a collapse in fundamentals.
-
Semiconductor stocks rose too much in too short a time.
From the perspective of fund managers, portfolio rebalancing and profit-taking became necessary.
-
The selling pressure in high-beta momentum stocks was excessive.
In some indicators, the selling pressure is interpreted as having reached levels similar to those seen during financial crisis periods.
-
Leverage products tied to Korean semiconductors also increased volatility.
It is argued that leveraged flows related to Samsung Electronics and SK hynix distorted the underlying stock flows and made short-term trading by foreigners and institutions easier.
-
Many research reports still say fundamentals remain strong.
Forecasts continue to suggest that DRAM, NAND, and HBM shortages may continue through 2026 and 2027.
Just because stock prices have fallen does not mean the semiconductor industry has already turned downward.
Stock prices reflect supply-demand conditions and sentiment in advance, but they do not always reflect fundamentals accurately.
10. Core point of memory semiconductors: the debate over 2028 oversupply
In the market, the scenario that “DRAM oversupply will arrive in 2028” is circulating.
This logic has increased concerns that the memory semiconductor cycle may be peaking out.
But there are many points here that need to be examined carefully.
-
Semiconductor cycles are usually priced in 2 to 3 quarters ahead of time.
Yet the market is already pricing in 2028 concerns strongly starting from 2026.
This may be an excessive discounting of a future that is too far away.
-
There are many forecasts that shortages will continue through 2027.
Major institutions such as UBS, Morgan Stanley, Goldman Sachs, and Nomura, while differing in degree, suggest memory supply may remain tight through 2027.
-
AI data centers are a new demand pillar.
In the past, DRAM demand was centered on PCs and smartphones.
Now AI servers, GPU clusters, and AI agent infrastructure are creating new demand.
If you apply the old memory cycle directly, you may conclude that “the time when PER is low is the peak.”
But if AI data centers are creating structural demand, we need to rethink whether the same extreme loss cycle as in the past will repeat.
11. DRAM, NAND, HBM: why all of them may be in short supply
When looking at memory semiconductors, focusing only on HBM is not enough.
AI infrastructure moves HBM, DRAM, and NAND together.
-
HBM is a high-value-added memory attached next to AI GPUs.
HBM is essential for Nvidia, AMD, and Broadcom AI chips.
As the industry moves to HBM4, unit prices and capacity are likely to increase.
-
DRAM handles the memory demand of the entire server.
As AI models grow and data centers expand, general server DRAM demand also increases.
-
NAND becomes more important in the AI agent era.
If AI needs to remember, search, store, and reuse more data, demand for high-performance SSDs and NAND will inevitably rise.
In particular, large AI models increasingly require functions such as KV cache offloading, long context, and user-specific memory storage.
This shift can elevate NAND and SSD demand from mere storage devices to key parts of AI infrastructure.
12. Structural changes brought by HBM: the moat of memory companies grows
In the past, DRAM was relatively close to a commodity product.
Customers could easily switch between Samsung Electronics, SK hynix, and Micron depending on who offered the best price.
But HBM is different.
-
HBM is highly customized.
It requires packaging and validation tailored to Nvidia GPUs, AMD GPUs, and Broadcom ASICs.
-
Once adopted, it is hard to replace.
HBM is not a simple plug-in component; it is a component designed together at the chip packaging stage.
-
Long-term supply contracts become important.
Customers want stable supply, and memory companies can reduce earnings volatility.
As this structure strengthens, memory semiconductor companies may deserve higher valuations than in the past.
Of course, the limitation of being a cyclical industry still remains, but if you look only in the exact same way as before, you will miss some things.
13. Global economic variables: oil prices and inflation risk
You cannot look only at AI and semiconductors.
Macroeconomic variables are still important.
The original source mentions U.S.-Iran tensions, risks around the Strait of Hormuz, and rising oil prices.
-
If oil rises again, price pressure increases.
When WTI and Brent crude surge, transportation costs, energy costs, and production costs all rise together.
-
Expected inflation may rise again.
If inflation risks re-emerge, expectations of rate cuts weaken and that can put pressure on tech stock valuations.
-
The Trump factor and geopolitical risks are difficult to predict.
The market is likely to repeatedly price in expectations of war ending and fears of renewed escalation.
Therefore, when looking at the global economic outlook, the AI investment cycle remains strong, but oil and interest-rate variables must continue to be monitored.
14. Tech stock valuation: this is not just an expensive market
After the recent correction, tech stock valuations have become less burdensome in some ranges.
The original source mentions periods in which the forward P/E ratios of the M7 and major tech stocks have fallen below historical averages.
-
Tech stock revenue growth remains strong.
Thanks to AI cloud, semiconductors, and data center investment, revenue growth in the tech sector is stronger than in other sectors.
-
EPS growth is also superior for tech stocks.
If AI infrastructure and software monetization happen simultaneously, earnings growth can become even stronger.
-
Earnings season matters.
Stocks that have fallen a lot recently can rebound sharply if earnings come in strong.
Conversely, stocks that have already risen a lot can correct if earnings fail to meet expectations, even if results are good.
In the end, the core of this earnings season is not just beating earnings.
What matters most is whether hyperscalers say they will continue increasing AI capex.
15. The most important point that other reports talk about less
-
First, the interpretation that Nvidia and memory demand will fall as Chinese AI models strengthen is too simplistic.
Large models require larger GPU clusters, more HBM, more DRAM, and more NAND.
-
Second, even if token unit prices fall, total costs may not fall.
If the model uses more tokens or tasks take longer, the actual job cost can remain similar.
-
Third, AI model competition is both a software competition and a semiconductor arms race.
The U.S. is more likely to increase AI investment rather than reduce it because of China’s catch-up.
-
Fourth, memory semiconductors have entered a demand structure different from the old PC and smartphone cycles.
AI data centers are creating structural demand for DRAM, NAND, and HBM.
-
Fifth, stock price declines and industry deterioration must be distinguished.
The recent semiconductor correction can reasonably be interpreted as being driven more by supply-demand, leverage, profit-taking, and momentum unwinding than by a collapse in fundamentals.
16. Investment perspective summary: what to watch now
-
AI model competition
You need to look at the coding performance and real user experience of Kimi K3, Qwen, DeepSeek, Claude, GPT, and Gemini together.
-
Big Tech AI capex
It is important to see whether Microsoft, Google, Meta, Amazon, and xAI are reducing or increasing data center investment.
-
Memory prices
You should check the rate of increase in DRAM, NAND, and HBM prices, as well as whether long-term supply contracts extend into 2027.
-
Semiconductor earnings season
Guidance from TSMC, ASML, Nvidia, AMD, Micron, Samsung Electronics, and SK hynix is important.
-
Macroeconomic variables
Oil prices, expected inflation, interest rates, and geopolitical risk can affect sentiment toward tech stocks.
-
Risk management
Leveraged products can recover quickly in a rebound, but in a correction they greatly increase account volatility.
In particular, the strategy of increasing leverage at the top can damage the investment habit itself.
< Summary >
Kimi K3 is a strong signal that Chinese AI technology is rising rapidly.
But the interpretation that “DeepSeek Shock 2.0 means AI semiconductors are finished” is close to exaggeration.
Kimi K3 is strong in certain coding areas, but Claude and GPT remain strong in overall user experience and comprehensive performance.
In addition, Kimi K3 may not actually be that cheap in terms of total task cost.
As Chinese AI advances, demand for large GPU clusters, HBM, DRAM, and NAND may actually grow.
The rebound effect, in which usage rises as AI becomes cheaper, is also important.
Memory semiconductors may remain in shortage through 2027, and concerns about 2028 oversupply may have been priced in too early.
Google is being shaken by Gemini delays and talent outflows, while OpenAI and Anthropic are showing monetization potential centered on enterprise APIs.
For tech stock investing, AI capex, memory prices, earnings season, and oil and inflation risks must all be considered together.
In conclusion, this correction is more reasonably viewed as a process of cleaning up overheated supply-demand conditions rather than the end of the AI cycle.
[Related Articles…]
*Source: [ 월텍남 – 월스트리트 테크남 ]
– 딥시크쇼크 2.0 왔다? 기술주 및 반도체 논란 종결 짓습니다.
