Kimi K3 is here. Will this become the "DeepSeek 2.0 moment" for U.S. stocks?

Last Friday, the US listed memory chip sector declined collectively once again.

On the surface, this appears to be merely a continuation of the chip stock pullback over the past two weeks. But beneath the surface, a new variable from China is roiling the entire market: Kimi K3, released by Moonshot AI on July 17.

This open-source model, boasting 2.8 trillion parameters, topped the global code evaluation leaderboard CodeArena, directly surpassing Anthropic's Claude Fable 5. What truly unsettled the market, however, was its cost: the inference cost per single task is only $0.94, less than half of Claude Opus 4.8 and roughly on par with OpenAI's GPT-5.6 Sol.

Larger scale, stronger coding capabilities, and cheaper inference costs—does this combination sound familiar?

In response, JPMorgan directly labeled it: a "DeepSeek 2.0 moment."

1. The Traumatic Memory of DeepSeek 1.0: Bloodbath for Chip Stocks and Bitcoin

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

Last year, DeepSeek achieved performance close to GPT-4 level with extremely low training costs. After the news broke, the market was instantly gripped by collective panic over the high valuation of AI chips:

• Nvidia's stock price plummeted from a high of around $135 to near $85, a decline of nearly 40%.
• Bitcoin fell from its all-time high near $110,000 to around $75,000.

The logic chain was simple: if a Chinese company can train top-tier models with fewer chips and at lower costs, will US cloud providers and tech giants still need to continue their frenzy of purchasing GPUs and HBM? Was the AI infrastructure arms race, from the very beginning, a bubble that was overpriced?

The market gave a clear answer at the time – a crash.

Now, Kimi K3 has emerged with the labels of "the world's largest open-source model + lower inference costs," its parameter scale being even more exaggerated than DeepSeek's was back then. Once the fear psychology is triggered, selling pressure in the chip sector spreads rapidly.

2. SemiAnalysis's Counter-Interpretation: K3 Isn't Eliminating GPU Demand, It's Amplifying It

But there's another side to the story.

Semiconductor research firm SemiAnalysis recently provided a completely opposite analytical framework. Their core argument is that K3's massive parameter scale and inference architecture, rather than weakening demand for high-end AI hardware, could instead become another long-term demand engine for Nvidia and its supply chain.

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

First, the model is "too large," which actually increases the need for hardware.

K3's parameter count exceeds 2.8 trillion, meaning just the model weights require over 1.5TB of HBM (High Bandwidth Memory) space. The current HBM capacity of a single top-tier Nvidia GPU is far from sufficient to load the entire model. Even in scenarios with relatively limited user concurrency, the KV cache still requires significant offloading to CPU DDR5 memory and NVMe storage devices – HBM space is not abundant; rather, it becomes a bottleneck.

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

Moonshot AI previously revealed that efficient inference deployment for K3 requires a large-scale scaling domain architecture consisting of at least 64 high-end chips. This level of cluster scale is highly consistent with the design philosophy of rack-scale AI systems like Nvidia's GB200/GB300 NVL72. In other words, K3 isn't "replacing" high-end hardware; it's "defining" the use case for next-generation high-end hardware.

Third, the misinterpretation that "linear attention reduces GPU demand" is flawed.

A popular market view previously held that more efficient attention mechanisms (such as linear attention) imply less computation, which in turn means lower demand for GPUs. SemiAnalysis believes this logic has a fundamental flaw. The real impact might be the opposite – more efficient model architectures lower the cost per inference, significantly reducing the barrier to entry for AI applications, thereby driving demand for AI deployment on a larger scale, across more enterprises and scenarios.

From a macro perspective: improved model efficiency → lower unit costs → explosion of AI applications → total computing power demand increases, not decreases. This is a script that has played out repeatedly in the history of the semiconductor industry – every "efficiency revolution" ultimately triggered a larger wave of hardware investment.

3. What is the Market Pricing? Fear vs. Rationality Tug-of-War

The current market is in a state of extreme fragmentation.

The bearish logic is intuitive: Kimi K3 once again proves the breakthrough capability of Chinese AI companies on the path of low cost and high efficiency. If this trend continues, AI companies will sooner or later reassess their capital expenditure (capex) plans. The narrative of "supply shortage" for AI chips may be replaced by the narrative of "good enough is sufficient."

The bullish rebuttal is equally powerful: K3 is not a scaled-down version of DeepSeek, but a scaled-up one – it's so large that it simply cannot function without high-end hardware clusters. SemiAnalysis's research indicates that HBM is not just in surplus; it might actually be in even shorter supply. Moreover, if AI applications become widely adopted due to lower costs, total computing power demand will grow exponentially.

Both sides have merit. And this is precisely the market's most painful moment – when two completely opposite logics can both find solid data support, price is not pricing fundamentals, but rather sentiment.

4. Facing a Fragmented Market, You Don't Have to Bet on a Single Direction

In this highly uncertain environment, the hardest thing isn't judging the direction, but controlling the cost of being wrong.

Bulls say the chip stock pullback is a "mistaken sell-off," while bears say the super-cycle for AI chips has already peaked. Both views have data backing and well-known institutional endorsements. For ordinary investors, rather than betting on a direction amidst this fog of information, a different approach might be better – ensuring that no matter which direction the market takes, you won't be completely knocked out.

The options functionality on the BIT platform precisely provides several core tools for this kind of "uncertainty trading":

① Holding Underlying Chip Stocks + Buying Put Options (Protective Put)

If you are heavily invested in Nvidia, Micron, or SK Hynix but are worried about further pullbacks triggered by the K3 event, you can buy put options on the corresponding stocks. If the stock price continues to fall, the profit from the options can offset the loss on the underlying shares; if it rebounds, you let the options expire, and your maximum loss is only the option premium paid.

② Buying Call Options Directionally (Long Call)

If you agree with SemiAnalysis's "mistaken sell-off" logic and believe chip stocks will revert to fundamentals after emotional selling, you can directly buy call options. Use far less capital than buying the underlying stock to bet on a rebound, with the maximum loss capped at the premium paid.

③ Buying Put Options Directionally (Long Put)

If you believe the "DeepSeek 2.0" script will play out fully and chip stocks have further room to fall, you can buy put options to directly short the market. No need for borrowing shares or margin calls; the maximum loss is the option premium.

④ Buying Both Directions Simultaneously (Long Straddle)

If you are confident that chip stocks will move, but unsure whether up or down, you can buy both call and put options simultaneously. As long as the stock price fluctuation is large enough, the profit from either direction can potentially cover the cost of both options.

5. Final Thoughts

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 itself.

But before the answer is revealed, the most valuable thing an investor can do is not to guess who will ultimately be right, but to prepare an exit strategy for their own potential misjudgment.

The BIT options feature goes live this week, covering core chip stocks such as Nvidia, Micron, and SK Hynix. Whether you are bullish or bearish, there are corresponding tools available to express your view – while locking your worst-case loss within a range you can tolerate.

In a market with unclear direction, those with choices hold the initiative.

Risk Disclaimer: The market conditions, valuation estimates, and product descriptions mentioned in this article are for reference only and do not constitute investment advice. Trading in US stocks and their derivatives involves market volatility, leverage, and liquidity risks. Short selling stocks may carry the risk of unlimited losses. Options trading carries the risk of total loss of premium. Historical performance does not guarantee future results. Investors should make prudent decisions based on their own risk tolerance and consult professional investment advisors when necessary.