DeepSeek 2.0 timing never came: chip stocks stabilize, with the US stock market waiting for the earnings season

On July 21, after the close of U.S. stocks, the chip sector that had been thrown into turmoil earlier—because of Kimi K3—finally managed to catch its breath.

The “DeepSeek 2.0 moment” that the market had feared would spread further did not take off. Although SK Hynix still fell 1.86%, flagship names like SanDisk and Micron—whose shares had previously been leading—turned upward, showing signs that the selloff had stabilized and a rebound could be underway. After more than a week of brutal selling, the entire semiconductor sector appears to have found a temporary equilibrium.

Why this time didn’t Kimi K3 replicate the panic script from DeepSeek 1.0?

I. K3’s narrative changed: not “enough compute,” but “compute scarcity”

The reason DeepSeek 1.0 triggered a frenzy back in March and April last year was the core narrative: “low cost, high efficiency, and no shortage of compute.” It trained models close to top-tier performance using a very small amount of GPU resources, directly challenging the investment logic that “AI performance must come from endlessly stacking chips.”

But Kimi K3 tells a completely different story.

Although K3 shocked the industry with 28 trillion parameters and extremely low per-inference costs, another side effect after its release was equally notable: compute scarcity. The sheer popularity of K3 exceeded expectations, quickly putting pressure on the inference infrastructure in the “Moon’s Dark Side,” making compute capacity expansion an urgent priority.

What does this mean? It means K3 is not proving “you don’t need that many chips,” but proving “even if model efficiency is high, the pace of demand growth will always stay ahead of supply.” This “compute scarcity” narrative is precisely the most beneficial support for the chip sector—it helps investors believe again that demand for AI chips won’t disappear as model efficiency improves; instead, it may keep expanding because application scenarios can explode.

DeepSeek 1.0’s logic was “efficiency replaces scale,” while K3’s logic is “efficiency releases demand.” These two narratives have opposite effects on chip stocks. This is also the most fundamental reason why “DeepSeek 2.0” didn’t unfold with the same kind of panic.

II. Google’s ace card: “engraving” Gemini into the chips

At a key moment when the chip sector is looking for direction, Google throws out a potential game-changer.

Alphabet is developing a brand-new AI server chip codenamed Frozen v2. The design approach is extremely aggressive: directly writing part of the Gemini model’s architecture into the silicon itself.

This is not the traditional sense of “optimizing software to fit hardware.” It’s more like “baking the model blueprint into the chip”—by reducing the movement of data between compute units and memory, sharply cutting power consumption and latency per inference.

Google engineers expect Frozen v2’s efficiency ratio to reach an astonishing level: the number of tokens that can be processed per unit of power will be 6 to 10 times that of the currently most advanced Ironwood TPU. For comparison, Ironwood is Google’s seventh-generation TPU, and even it only doubled performance per watt versus the prior generation. The generational leap of Frozen v2 far exceeds any earlier chip upgrade from Google.

After the news broke, Alphabet’s stock price briefly rose from around $350 to around $359. The market is clearly re-evaluating Google’s long-term competitiveness in AI infrastructure. If Frozen v2 can truly be commercialized and deployed before 2028, Google will have one of the most efficient large-model inference infrastructures globally, making its advantages in AI service costs difficult for competitors to replicate.

But for the entire chip sector, Frozen v2’s significance goes far beyond “this is just Google’s own thing.”

It sends a key signal: AI giants are not slowing their investment in hardware because model efficiency is improving. On the contrary, they are pushing the competition down to a more bottom-layer, more customized dimension—application-specific chips (ASICs). From Nvidia’s general-purpose GPUs to Google’s dedicated TPUs, and then to Frozen v2 that directly “carves” the model architecture into silicon—the competition in AI hardware is shifting from “who has more cards” to “whose cards are smarter, more efficient, and more specialized.”

This shift implies that demand for AI chips won’t shrink because “models get smaller and cheaper.” On the contrary, the trend toward specialization and customization will create more diverse and more granular chip demand—which is a long-term positive for the entire semiconductor industry chain.

III. The big test of earnings season: next week will determine the direction

Although the chip sector has temporarily stabilized, the challenge is not over.

Over the next week or so, key players across the AI industry chain will begin to release their quarterly results. The significance of this earnings season is unusual—because the market isn’t just looking at whether the numbers are good; it’s looking at several critical questions that determine whether AI narratives can continue:

For Google, Meta, and Microsoft, will capital expenditures keep burning? Can revenue generated from cloud business and AI services cover the increasingly heavy depreciation, leasing, and electricity costs? If not, the “AI monetization” narrative will show cracks, and the turning point in capex growth may arrive earlier than we think.

For SK Hynix, will the money cloud vendors spend ultimately translate into pricing power for storage chips, market share expansion, and profit growth? Hynix needs to prove in its earnings that it is truly an element of the AI industry chain that makes money, not merely a “middleman that earns the spread.”

What this earnings round is really about isn’t just whether capex is “high.” It’s whether results can “keep beating expectations.” After multiple quarters of “surprises,” the market’s threshold for beating expectations has been pushed to a very high level. Any marginal slowdown—whether guidance is lowered, wording is softened, or key data misses expectations—could become the trigger for another round of selling.

IV. Written in closing: in the storm eye of earnings season, options are an important anchor

The chip sector is going through a highly uncertain window. Bulls say K3’s compute scarcity and Google’s Frozen v2 prove that AI hardware demand has not yet peaked; bears say capex growth is about to top out, valuations have already been drained, and earnings won’t be able to keep beating expectations indefinitely.

In an environment where information is torn in different directions, the risk of betting on a single direction is far greater than the potential upside.

BIT’s brokerage department believes options functions are precisely the tools that fit this kind of market environment—“no clear direction, but volatility is extreme”:

• Hold the chip stocks + buy put options: insure the position before earnings, lock down downside risk

• Buy call/put options in a single direction: use capital far lower than the stock to bet on the post-earnings direction; the maximum loss is only the premium

• Buy in both directions at the same time: not sure if earnings are a surprise or a scare? Bet on both sides—profit if volatility is large enough

The gale force of earnings season is approaching. In a market with no clear direction, only those who have options are entitled to talk calmly.