Kimi K3 has arrived, will this become the "DeepSeek 2.0" moment for the US stock market?

Last Friday, the U.S. stock market's memory chip sector collectively declined again.

On the surface, this is just a continuation of the ongoing correction in chip stocks over the past two weeks. However, beneath the surface, a new variable from China is stirring the entire market: Kimi K3, released by Moonshot AI on July 17.

This open-source large model with 2.8 trillion parameters has topped the global code evaluation chart CodeArena, directly surpassing Anthropic's Claude Fable 5. What makes the market even more restless is its cost: the inference cost for a single task is only $0.94, less than half of Claude Opus 4.8, and basically on par with OpenAI's GPT-5.6 Sol.

Larger in scale, stronger in coding capability, and cheaper in inference cost—does this combination sound familiar?

In response, JPMorgan directly labeled it as the "DeepSeek 2.0 moment."

1. The Traumatic Memory of DeepSeek 1.0: Bloodshed in Chip Stocks and Bitcoin

To understand why the five words "DeepSeek 2.0" can keep Wall Street awake at night, we must first review what happened in March and April of 2025.

Last year, DeepSeek achieved performance close to GPT-4 levels with extremely low training costs. Once the news broke, the market fell into a collective panic over the overvaluation of AI chips:

• Nvidia's stock price plummeted from about $135 to around $85, a decline of nearly 40%.

• Bitcoin dropped from nearly $110,000 to around $75,000.

The logic is simple: if Chinese companies can train top models with fewer chips and lower costs, do American cloud providers and tech giants still need to continue frantically purchasing GPUs and HBM? Was the arms race for AI infrastructure from the beginning just an overpriced bubble?

At that time, the market gave a clear answer—plummeting.

Now, Kimi K3 has emerged with the label "the world's largest open-source model + lower inference cost," with a parameter scale even more exaggerated than DeepSeek's at that time. Once the fear psychology is activated, the selling pressure in the chip sector quickly spreads.

2. SemiAnalysis's Contrarian Interpretation: K3 is not eliminating GPU demand, but amplifying it

But there is another side to the story.

The semiconductor research firm SemiAnalysis recently provided a completely opposite framework of reasoning. Their core argument is that K3's massive parameter scale and inference architecture will not weaken the demand for high-end AI hardware; rather, it may become another long-term demand engine for Nvidia and its supply chain.

Breaking it down, SemiAnalysis's judgment is based on the following key facts:

First, the model is "too large," which actually requires more hardware.

K3's parameter scale exceeds 2.8 trillion, and the model weights alone require more than 1.5TB of HBM (high bandwidth memory) space. Currently, a single top Nvidia GPU's HBM capacity is far from sufficient to load the entire model. Even in relatively limited user concurrency scenarios, KV cache still needs to be heavily offloaded to CPU DDR5 memory and NVMe storage devices—HBM space will not only be insufficient but will also be tight.

Second, inference deployment places extremely high demands on hardware scale.

Moonshot AI previously revealed that K3 requires a large-scale expansion domain architecture composed of at least 64 high-end chips for efficient inference deployment. This level of cluster scale is highly consistent with the design concepts of Nvidia's GB200/GB300 NVL72 and other rack-level AI systems. In other words, K3 is not "replacing" high-end hardware but is "defining" the usage scenarios for the next generation of high-end hardware.

Third, the misreading of "linear attention reducing GPU demand."

There has been a popular view in the market: more efficient attention mechanisms (such as linear attention) mean less computation, which in turn means a decrease in GPU demand. SemiAnalysis believes this logic has fundamental flaws. The real impact may be exactly the opposite—more efficient model architectures lower the cost of single inference, significantly reducing the threshold for AI application deployment, thereby driving more enterprises, more scenarios, and larger-scale AI deployment demand.

From a macro perspective: improved model efficiency → lower unit costs → explosion of AI applications → total demand for computing power increases rather than decreases. This is a script that has been repeatedly played out in the semiconductor industry’s history—every "efficiency revolution" ultimately leads to a larger wave of hardware investment.

3. What is the market trading? A tug-of-war between fear and rationality

The current market is in a state of extreme division.

The logic of the bears is very intuitive: Kimi K3 has once again proven the breakthrough capabilities of Chinese AI companies in low-cost, high-efficiency paths. If this trend continues, AI manufacturers will eventually reassess their capital expenditure (capex) plans. The narrative of "supply not meeting demand" for AI chips may be replaced by the narrative of "just enough."

The bulls' rebuttal is equally strong: K3 is not a smaller version of DeepSeek but an enlarged version—it is so large that high-end hardware clusters cannot operate at all. SemiAnalysis's research indicates that HBM is not in surplus; rather, it may be even more scarce. Moreover, if AI applications become widely adopted due to lower costs, the total demand for computing power will show exponential growth.

Both sides have their merits. And this is precisely the most painful moment for the market—when two completely opposing logics can find solid data support, prices are not pricing fundamentals but are pricing emotions.

4. In the face of a divided market, you don't have to bet on a single direction

In this highly uncertain environment, the hardest part is not judging the direction but controlling the cost of exposure to the wrong direction.

Bulls say the pullback in chip stocks is a "wrong kill," while bears say the super cycle of AI chips has peaked. Both viewpoints have data support and endorsements from well-known institutions. For ordinary investors, rather than betting on a direction in this information fog, it is better to think differently—ensure that no matter which direction develops, you will not be completely knocked down.

The BIT platform's options feature will officially launch this week, providing several core tools for this "uncertainty trading":

① Hold chip stocks + buy put options (Protective Put)

If you are heavily invested in Nvidia, Micron, or SK Hynix but are worried that the K3 event will trigger further pullbacks, you can buy put options for the corresponding stocks. If the stock price continues to fall, the profits from the options can cover the losses from the stocks; if it rebounds, you can forgo exercising the options, with the maximum loss being only the premium of the options.

② Buy call options in one direction (Long Call)

If you agree with SemiAnalysis's "wrong kill" logic and believe that chip stocks will return to fundamentals after an emotional sell-off, you can directly buy call options. Bet on the rebound direction with funds far lower than the stocks, with the maximum loss locked in at the premium.

③ Buy put options in one direction (Long Put)

If you believe the "DeepSeek 2.0" storyline will fully replay and that chip stocks have further downside potential, you can buy put options to short directly. No need for margin calls or short selling, with the maximum loss being the option fee.

④ Buy both directions simultaneously (Long Straddle)

If you are confident that chip stocks will move but are uncertain whether it will be up or down, you can buy both call and put options simultaneously. As long as the stock price's volatility is large enough, profits in either direction may cover the costs of both options.

5. In Conclusion

The release of Kimi K3 has pushed AI chip investment to a critical crossroads. Is it the beginning of "DeepSeek 2.0," or the prelude to "demand being redefined"? The answer to this question may take time to fully reveal.

But before the answer is unveiled, the most valuable thing investors can do is not to guess who is ultimately right, but to prepare an exit strategy for their judgment errors.

BIT options cover core chip targets such as Nvidia, Micron, and SK Hynix. Whether you are bullish or bearish, there are corresponding tools to express your views—while keeping the worst-case losses locked within a range you can bear.

In an unclear market, those who have options hold the initiative.

Risk Warning: The market conditions, valuation calculations, and product descriptions mentioned in this article are for reference only and do not constitute investment advice. Trading U.S. stocks and their derivatives involves market volatility, leverage, and liquidity risks; short selling may face unlimited loss risks; options trading carries the possibility of total loss of premiums; past performance does not guarantee future returns. Investors should make prudent decisions based on their own risk tolerance and consult professional investment advisors if necessary.