● DeepSeek Shock 2-0, AI Stocks And Semiconductors Slammed
Is This DeepSeek Shock 2.0? The Real Drivers of AI Tech Stock and Semiconductor Correction
The core of this issue is not simply that “Chinese AI beat U.S. AI.”
The key point is that improvements in Chinese open-source AI models, Nvidia GPU demand, the memory semiconductor cycle, big tech AI capex, and the correction in technology stocks are all linked.
Although the market has labeled Kimi K3 as DeepSeek Shock 2.0, a closer look shows that some of the fear is overstated, while some developments may actually support higher semiconductor demand.
This report summarizes the Kimi K3 debate, the real competitiveness of Chinese AI, Nvidia and HBM demand, the outlook for memory semiconductors including SK hynix, Samsung Electronics, and Micron, and whether the AI investment cycle is truly weakening.
1. The Kimi K3 debate: “DeepSeek Shock 2.0” may be overstated
The latest market-moving event is Moonshot AI’s Kimi K3 model.
Following reports that Kimi K3 outperformed Claude Fable 5 and GPT 5.6-level models on some benchmarks, technology stocks and semiconductor sentiment weakened sharply.
However, one important distinction should be made first.
The area in which Kimi K3 was said to outperform U.S. AI models is not overall AI capability, but a specific frontier coding use case closer to web UI and UX implementation.
On aggregate benchmarks, Kimi K3 is still ranked among the top tier, but it is not widely regarded as clearly superior to Claude and GPT across all dimensions.
According to major AI benchmark institutions, Kimi K3 is generally viewed as a top-three model.
In other words, headlines suggesting that Chinese AI has fully surpassed U.S. AI are materially more aggressive than the underlying evidence.
- Kimi K3 showed strong performance in specific coding tasks.
- Its aggregate performance still appears to trail GPT and Claude in several assessments.
- Chinese AI capability is improving rapidly.
- However, the DeepSeek-style “much cheaper and nearly as good” narrative is less clear in this case.
2. The key difference between DeepSeek and Kimi K3: the cost advantage is no longer as large
DeepSeek shocked the market because it was perceived to be significantly cheaper than U.S. models while delivering roughly 80% to 90% of their performance.
Kimi K3 is different.
Its token pricing may appear lower than GPT or Claude.
The issue is that it often uses more tokens to complete the same task.
In other words, a lower unit price does not necessarily translate into lower total job cost.
For example, even if GPT is more expensive on a per-1 million output-token basis, Kimi K3 may consume more tokens to achieve the same output, resulting in similar total cost.
This is one of the most overlooked points in recent coverage.
“Lower token price” and “lower real-world job cost” are not the same thing.
- GPT 5.6-class models are estimated to cost about $1 per task.
- Kimi K3 is also estimated to cost around $0.9 per task in practice.
- Claude Fable-class models are relatively expensive, but remain strong in user experience.
- Kimi K3 is better understood as a model that impressed through capability, not simply through low cost.
3. Chinese AI is no longer just about cost efficiency
The most important shift in the Kimi K3 debate is that the positioning of Chinese AI is changing.
In the past, Chinese models were generally seen as weaker but cheaper alternatives.
More recently, they have advanced to the point where they can challenge U.S. frontier models on capability alone.
In addition to Kimi K3, Alibaba’s Qwen series, GLM, MiniMax, and DeepSeek have all improved rapidly.
Kimi K3 is reportedly a model in the 3 trillion-parameter range.
Alibaba’s next-generation Qwen model is also reportedly being discussed at more than 2 trillion parameters.
The repeated emergence of models at this scale indicates that China is no longer just a follower, but a major participant in the AI competition.
China’s AI competitiveness also reflects its talent pipeline.
Elite institutions such as Tsinghua University continue to produce top-tier AI and computer science talent, feeding startups and core research teams.
There is also a notable flow of Chinese engineers returning from institutions such as Stanford and MIT into China’s AI ecosystem.
For Korea, the more urgent issue is AI talent outflow and the treatment of engineering talent.
4. Kimi K3’s acknowledged weakness: user experience still favors U.S. models
Moonshot AI has also acknowledged that Kimi K3 still trails Claude Fable 5 and GPT 5.6 in user experience.
This is important.
In AI, benchmark scores alone do not make a model the best choice in real-world work.
Developer tools such as Claude Code and OpenAI Codex matter as much as, or more than, raw model performance.
In practice, the model is only one part of the workflow.
Agent reliability, code-editing flow, file handling, and context retention are often more important than the model’s isolated benchmark score.
This is where Anthropic and OpenAI still appear to have a meaningful moat.
Even if Chinese models are catching up on benchmarks, there remains a gap in production productivity.
- Chinese models are competitive in basic chat use cases.
- Claude Code and Codex remain strong for complex development workflows.
- In enterprise AI agents, workflow integration matters as much as model quality.
- For that reason, it is too early to conclude that U.S. AI companies have lost their advantage.
5. Why Kimi K3 may increase demand for Nvidia and memory semiconductors
Markets initially worried that better Chinese AI models would reduce demand for Nvidia.
In reality, the opposite may be more plausible.
Training and inference for a model like Kimi K3 require massive accelerator infrastructure.
Models in the 2 trillion to 3 trillion parameter range are difficult to run efficiently without large-scale GPU clusters.
Analysts suggest that Kimi K3 requires a scale-up domain with at least 64 accelerators.
The most likely platform for this kind of workload remains Nvidia.
If Kimi K3 had been trained primarily on Huawei chips, that would likely have been highlighted more prominently.
Instead, many analysts believe that Kimi K3, Qwen, and GLM were likely trained largely on Nvidia hardware.
Memory is another critical point.
Large models require such extensive weights and KV cache capacity that HBM alone is insufficient.
As a result, DDR5 DRAM and NAND flash are needed for offloading.
In effect, the growth of large models may increase demand not only for HBM, but also for standard DRAM and NAND.
- Improvement in Chinese AI models may increase GPU cluster demand rather than reduce it.
- Large models raise demand for HBM, DDR5, and NAND flash.
- Kimi K3 may reinforce the long-term demand case for memory semiconductors.
- This is one of the most important points in current semiconductor cycle analysis.
6. The Jevons effect: lower AI costs may increase, not reduce, infrastructure demand
The market made a similar mistake during the DeepSeek shock.
It assumed that cheaper AI models would reduce data center investment.
In practice, lower costs expanded usage more quickly than expected, increasing total infrastructure demand.
This is the Jevons effect.
As token prices fall, enterprises and individuals use AI more frequently.
Higher usage raises inference demand.
Higher inference demand increases demand for GPUs, HBM, DRAM, NAND, networking equipment, and optical components.
Therefore, AI cost efficiency may create short-term fear, but over time it can expand the overall AI infrastructure market.
7. Google Gemini delay suggests friction in the big tech AI race
Google is another company that should be monitored closely in this context.
The delay in the launch of the next Gemini Pro model has heightened concerns.
Bloomberg reported that internal coding performance may have fallen short of targets.
This does not mean Google has lost the AI race, but it does suggest that in coding-focused AI, it is facing pressure from OpenAI, Anthropic, and some Chinese models.
Another issue for Google is talent retention.
Key personnel from Google DeepMind and other AI groups have reportedly moved to Anthropic, OpenAI, and startups.
Anthropic and OpenAI have stronger equity-based retention tools because of their IPO prospects.
This is intensifying the internal talent war across the sector.
There is also ongoing debate over whether TPU architectures are fully aligned with the current Sparse MoE model trend.
Nvidia’s GPU ecosystem remains strong not only because of chip performance, but because of CUDA, networking, software, and cluster operations expertise bundled into one platform.
8. AI investment is difficult to stop: it has become a strategic asset
The real message from Kimi K3 is not that AI investment is ending.
It is more likely that the U.S. will continue to expand AI investment.
AI is now a strategic asset tied to national security, defense, cybersecurity, and industrial competitiveness.
If Chinese AI is approaching U.S. frontier models in some areas, it is unlikely that the U.S. government or major technology companies will scale back AI spending.
Meta, Google, xAI, OpenAI, Anthropic, Microsoft, and Amazon are all central participants in the AI arms race.
Even if concerns over efficiency continue, intensifying competition with China is likely to keep data center capex elevated.
This is a key variable for the global macro outlook.
Rising AI investment affects semiconductors, power infrastructure, cooling, optical communications, data centers, and cloud infrastructure broadly.
9. Are OpenAI and Anthropic actually generating profits?
The most important question in the AI bubble debate is whether OpenAI and Anthropic are actually monetizing their platforms.
At present, Anthropic appears to be the fastest-growing in revenue terms.
Its enterprise API business and the rise in Claude Code usage are driving rapid growth.
The key driver is B2B API revenue.
Although consumer subscriptions may cost only $20 or $200 per month, heavy users’ token consumption can translate into thousands or even tens of thousands of dollars at API pricing.
As a result, consumer subscriptions can be low-margin or loss-making, while enterprise API usage can generate high margins.
Anthropic’s gross margin is estimated to have reached around 60%.
OpenAI is also estimated to be operating with gross margins in the 40% range.
These figures do not account for training costs, personnel, marketing, or additional infrastructure spending, so they are not equivalent to net profit.
Still, the claim that AI generates no real revenue is not supported by current data.
- OpenAI’s large base of free users weighs on short-term profitability.
- Anthropic is monetizing more effectively through enterprise APIs and developer tools.
- AI agent usage still represents a very small share of total users.
- As adoption expands, revenue upside remains significant.
10. The semiconductor correction appears more related to positioning than fundamentals
Semiconductor indices have corrected sharply from recent highs.
The Nasdaq has been relatively resilient, but memory semiconductors and AI hardware-related stocks have shown much greater volatility.
SK hynix, Samsung Electronics, Micron, SanDisk, and optical communication names have all seen substantial pullbacks.
However, a stock decline does not necessarily imply a deterioration in fundamentals.
Much of the current weakness appears to reflect profit-taking, fund rebalancing, crowded long positioning, and unwinding of leveraged products.
As semiconductor longs became a crowded trade, institutional investors likely had to reduce exposure.
This came alongside concerns about Chinese AI models and broader AI capex fatigue.
That said, memory pricing, data center investment, HBM demand, DRAM shortages, and NAND demand remain firm.
For technology stock analysis, it is important to separate short-term price action from medium- to long-term industrial demand.
11. The memory semiconductor cycle: 2028 supply overhang may be priced in too early
The market’s main concern is a potential memory oversupply in 2028.
The argument is that new fabs from Samsung Electronics, SK hynix, and Micron could begin affecting supply in late 2027 and push the market into oversupply in 2028.
This risk cannot be dismissed.
Memory semiconductors remain a cyclical industry.
However, semiconductor stocks typically discount conditions two to three quarters ahead.
The market currently appears to be pricing in 2028 concerns as early as 2026.
That effectively means fears are being pulled forward by more than six quarters.
As a result, the recent correction may reflect more of a discounting of future risk than a genuine deterioration in current demand.
Several global brokerages still see DRAM shortages persisting into 2027.
Analyses from UBS, Morgan Stanley, JPMorgan, and Nomura suggest that data center demand may absorb much of the incremental supply.
AI data centers require significantly more DRAM and NAND than legacy PC or smartphone demand.
12. DRAM and NAND pricing remain firm
DRAM prices have risen sharply over the past year.
Server DDR5 pricing has strengthened due to persistent supply tightness.
Some institutions expect third-quarter DRAM prices to post double-digit percentage gains.
Even if fourth-quarter increases slow, that implies slower price growth, not price declines.
NAND flash is also improving.
Demand is strengthening as AI data centers require larger storage and caching layers.
Next-generation AI server architectures may require significantly more NAND flash than before.
AI agents need long-term memory, user history, and context storage.
- DRAM shortages continue to center on server DDR5.
- HBM expansion may further tighten supply for standard DRAM.
- NAND demand may increase as AI server storage layers expand.
- The memory cycle is evolving beyond the legacy PC and smartphone model.
13. HBM is changing the moat structure of memory suppliers
In the past, DRAM was a commodity product.
Customers could relatively easily switch among Samsung Electronics, SK hynix, and Micron depending on price.
This left memory suppliers vulnerable when the cycle turned.
HBM is different.
HBM must be matched with specific AI chips and packaging architectures from Nvidia, AMD, or Broadcom.
It is not a simple plug-and-play memory module.
It requires customized packaging and validation.
That makes supplier switching more difficult.
This is why SK hynix, Samsung Electronics, and Micron may be able to establish stronger lock-in in HBM and next-generation high-bandwidth memory.
As HBM’s share rises, earnings volatility at memory suppliers may decline relative to the past.
This is a major structural shift in the semiconductor cycle.
14. Optical communications and AI hardware also matter
As AI data center investment rises, GPUs and memory are not the only requirements.
Networking equipment, optical modules, optical cables, switches, power systems, and cooling systems are also needed to connect large GPU clusters.
Optical communication stocks have already seen sharp gains and subsequent corrections.
However, as AI cluster size increases, data movement requirements rise dramatically.
In next-generation AI servers, GPU-to-GPU, rack-to-rack, and data center interconnection become increasingly important.
As a result, optical communications could become a major theme in the 2027-2028 AI infrastructure cycle.
Short-term volatility remains high, but the structural direction still warrants attention.
15. Macro risk: oil and inflation still need monitoring
Even if AI and semiconductor fundamentals remain strong, macro risks cannot be ignored.
Rising tensions in the Middle East and issues involving the Strait of Hormuz have pushed oil prices higher again.
If WTI and Brent rise further, inflation expectations could move higher as well.
That would weaken expectations for rate cuts and pressure growth stock valuations.
The market is currently balancing a strong structural growth story in AI and semiconductors against macro risks such as oil, inflation, and geopolitics.
For that reason, technology stock analysis should also consider rates, the dollar, and global liquidity conditions, not just earnings.
16. What the news flow is not emphasizing enough
First, the real significance of Kimi K3 is not that China can build cheaper models, but that China has entered the high-performance model competition.
Chinese AI is no longer only a low-cost alternative.
It is increasingly challenging U.S. models on performance.
Second, as Chinese AI models grow larger, they may increase, rather than reduce, demand for Nvidia and memory semiconductors.
Larger models require more GPUs, HBM, DRAM, NAND, and networking infrastructure.
This is the opposite of the fear-driven narrative.
Third, the AI capex debate should not focus only on spending figures.
OpenAI and Anthropic are building real monetization paths through enterprise APIs, developer tool lock-in, and expanding AI agent usage.
Fourth, memory semiconductors are no longer a simple cyclical industry.
HBM is a customized product, and AI data centers structurally require more DRAM and NAND.
This shift supports a re-rating case for Samsung Electronics, SK hynix, and Micron.
Fifth, the recent semiconductor decline appears more closely tied to positioning and sentiment than to a deterioration in fundamentals.
Leveraged products, institutional rebalancing, and crowded momentum exposure likely amplified the selloff beyond the underlying industry reality.
17. Key investment checkpoints
- Monitor whether big tech earnings calls raise AI capex guidance.
- Track whether Google, Meta, Microsoft, and Amazon are increasing or reducing data center spending.
- Watch for expansion in long-term supply agreements at SK hynix and Samsung Electronics.
- Check whether DRAM and NAND pricing remains stronger than market expectations.
- Assess whether HBM4 transition drives both price and volume growth.
- Monitor how much memory capacity is added in Nvidia’s next GPU platform.
- Track whether Middle East tensions and oil prices continue to pressure inflation.
- Maintain risk discipline in leveraged ETFs and high-volatility names.
18. Conclusion: DeepSeek Shock 2.0 is more about structural change than fear
Kimi K3 is an important development.
It shows that Chinese AI has advanced enough to challenge U.S. frontier models in some areas.
However, it is premature to interpret this as the end of AI investment, a collapse in Nvidia demand, or a peak in the semiconductor cycle.
Rather, Kimi K3 suggests that AI competition is intensifying and that the U.S.-China AI arms race is likely to expand further.
Large model competition requires more GPUs, more memory, and more data center infrastructure.
Although AI investment and the semiconductor cycle are undergoing a correction, the medium- to long-term industry direction remains intact.
Stock prices can remain volatile.
This is especially true for technology and semiconductors.
But investors should focus on actual demand, pricing, capex, and monetization metrics rather than fear-driven headlines.
On that basis, the structural growth case for AI and memory semiconductors does not appear to be over.
< Summary >
Kimi K3 is an important signal that Chinese AI capability is improving rapidly.
However, it does not appear to fully outperform U.S. AI models across all dimensions.
Its unit token pricing may be lower, but total task cost may not differ materially from GPT.
As Chinese AI models scale up, demand for Nvidia GPUs, HBM, DRAM, and NAND may increase rather than decrease.
Despite concerns over AI excess, OpenAI and Anthropic are building monetization paths through enterprise APIs and developer tools.
The semiconductor correction appears driven more by positioning, profit-taking, and deleveraging than by a fundamental break.
The memory semiconductor cycle is becoming structurally different because of HBM and AI data center demand.
In conclusion, DeepSeek Shock 2.0 looks more like a signal of intensified AI competition than a signal that the cycle is ending.
[Related Articles…]
- AI Investment Cycle and Global Technology Stock Outlook
- Semiconductor Cycle and Memory Market Analysis
*Source: [ 월텍남 – 월스트리트 테크남 ]
– 딥시크쇼크 2.0 왔다? 기술주 및 반도체 논란 종결 짓습니다.
