Chipmaker shares stabilized on July 21 after a sharp week of selling, as fears of a new shock to semiconductor demand eased and traders reassessed the impact of a fast-growing artificial intelligence model that is proving expensive to run despite its efficiency.
SK Hynix ended the session down 1.86%, but several other semiconductor-related names, including Western Digital and Micron, showed modest recoveries. The mixed performance suggested the sector had reached a temporary balance after heavy volatility driven by concerns that a new generation of low-cost AI models could reduce demand for advanced chips.
Those fears centered on Kimi K3, a large language model with 2.8 trillion parameters and unusually low inference costs. Its launch initially revived comparisons with last year’s DeepSeek-driven selloff, when traders worried that more efficient AI systems could weaken demand for expensive graphics processors and memory chips.
But the market’s interpretation has shifted. Rather than showing that AI can run with less hardware, Kimi K3’s rapid adoption appears to be putting more pressure on computing infrastructure. Its popularity has strained available capacity, pushing the debate back toward whether major technology firms will need even more servers, memory, storage and specialized processors to meet demand.
That change in perception helped calm the semiconductor sector. DeepSeek 1.0 had been viewed by many traders as a symbol of “efficiency replacing scale.” Kimi K3 is now being interpreted differently: as an example of “efficiency driving demand.” In simple terms, cheaper AI usage can lead to more usage, and more usage can still require more chips.
The market’s attention also turned to Alphabet, which announced development of a new AI server chip known as Frozen v2. The processor is designed to embed parts of Alphabet’s Gemini model directly into silicon, reducing the amount of data that must move between computing units and memory. That approach could lower power consumption, reduce latency and improve performance for specific AI workloads.
Alphabet’s announcement added to the view that the AI hardware race is entering a more specialized phase. Large technology companies are no longer relying only on standard graphics processing units. They are increasingly designing custom chips for their own models, data centers and software ecosystems.
Shares tied to AI hardware briefly moved higher after the Alphabet news, with one closely watched chip-linked name rising from about $350 to $359 during trading. The reaction reflected trader interest in the chip’s planned deployment before 2028 and in what it signals about the long-term shift toward custom AI infrastructure.
Why Kimi K3 changed the narrative
Kimi K3 arrived at a sensitive moment for semiconductor stocks. After several quarters of strong gains tied to AI spending, traders had become alert to any sign that software breakthroughs could reduce demand for hardware.
The initial concern was straightforward. If a model can deliver strong results at much lower inference costs, companies might need fewer high-end chips to run AI services. That was the fear that followed DeepSeek 1.0, which became closely associated with the argument that better model design could reduce the need for expensive infrastructure.
Kimi K3 complicated that argument. While the model is efficient, its popularity has increased pressure on the systems needed to serve users. Lower inference costs can make AI services cheaper to offer, which can expand use across businesses, developers and consumers. If usage grows faster than efficiency improves, total demand for computing power can still rise.
This is a key reason the semiconductor selloff slowed. Traders began to distinguish between efficiency that reduces hardware needs and efficiency that unlocks more consumption. The second outcome can be supportive for the chip supply chain, especially for companies that provide high-bandwidth memory, storage systems, networking equipment and advanced system components.
The market is now trying to determine which effect will dominate. If future AI models keep becoming cheaper to run but also attract far more usage, chip demand may remain strong. If efficiency gains eventually outpace growth in workloads, the pressure on hardware spending could increase.
For now, Kimi K3 appears to have eased the worst fears of a sudden AI infrastructure slowdown. But it has not ended the debate. The sector remains highly sensitive to any evidence that demand for servers, memory and accelerators is either accelerating or cooling.
Alphabet pushes deeper into custom AI silicon
Alphabet’s Frozen v2 chip is important because it points to a deeper change in the structure of AI computing. The company is developing a chip that places parts of the Gemini model closer to the hardware itself. By reducing data movement between memory and compute units, the design aims to improve energy efficiency and response speed.
That matters because data movement is one of the biggest sources of power use in large-scale AI systems. Running advanced models requires vast amounts of data to move continuously through servers, memory chips and networking systems. If a chip can reduce that movement, it can lower electricity use and improve the economics of AI deployment.
Frozen v2 is expected to deliver six to ten times the energy efficiency of Ironwood, Alphabet’s seventh-generation tensor processing unit. If achieved at scale, that improvement would be significant for data centers, where power availability has become a major constraint.
The chip also shows how major technology firms are trying to control more of their AI stack. Instead of buying general-purpose chips and adapting software around them, companies such as Alphabet are building processors tailored to their own models. This can reduce dependence on outside suppliers, improve performance for internal workloads and support more predictable long-term infrastructure planning.
For the semiconductor industry, this trend cuts both ways. Custom chips can reduce the appeal of standard graphics units for large corporate data centers. At the same time, they create new demand for advanced manufacturing, packaging, memory, interconnects, testing equipment and design services.
The result is not a simple decline in chip demand. It is a shift toward narrower and more technical demand segments. Suppliers that can adapt to custom architectures may benefit, while those tied mainly to generic components could face more pressure.
Earnings will test the rebound
The calm in chip shares may prove temporary. Several major technology companies are scheduled to report earnings in the coming week, and traders are focused on whether AI-related capital spending remains strong.
Alphabet, Meta and Microsoft are expected to face close scrutiny over their data center budgets, chip purchases and long-term infrastructure plans. These companies have been among the largest drivers of AI hardware demand. Any sign that they are slowing server expansion or becoming more cautious on spending could quickly weigh on semiconductor shares.
SK Hynix will also be closely watched because of its role in high-bandwidth memory, a critical component for AI accelerators. Traders will be looking for evidence that memory pricing, demand and margins remain strong. A softer outlook could raise doubts about whether the recent strength in AI-linked memory can continue.
The challenge for the sector is that expectations are already high. After multiple quarters of stronger-than-expected AI spending, simply maintaining growth may not be enough to trigger a large positive market reaction. Companies may need to show both rising demand and confidence in future spending to satisfy traders.
That raises the risk of renewed volatility. If large technology firms signal that infrastructure demand remains strong, chip shares could extend their stabilization. If they suggest that spending growth is slowing, the recent recovery could reverse quickly.
Options activity has become more important ahead of these results. With share prices moving sharply on small changes in guidance, traders are increasingly using options to manage risk around earnings announcements. That reflects uncertainty not only about company results, but also about how the market will interpret them.
Custom chips reshape the server market
The move toward custom microchips is changing the economics of large-scale computing. Standard graphics processing units remain central to many AI workloads, but large cloud and internet companies are increasingly exploring chips built for their own needs.
This shift makes sense at scale. The largest technology firms operate massive data center networks. Even small efficiency gains can translate into large savings in electricity, cooling, hardware replacement and operating costs. A custom chip that performs well on a company’s most important workloads can be worth billions of dollars over time.
The change also reflects the growing specialization of AI models. As companies build models for search, advertising, coding, video, enterprise software and consumer assistants, the hardware needed to run those models may become more specialized as well. A general-purpose chip offers flexibility, but a custom processor can be more efficient for a narrower set of tasks.
For chipmakers, this means the competitive landscape is becoming more complex. Demand no longer depends only on who can build the fastest processor. It also depends on who can support custom designs, advanced packaging, memory integration and power-efficient systems.
Memory and storage suppliers remain important in this environment. Even highly efficient AI chips need large amounts of fast memory and reliable storage. If model usage continues to grow, pressure on these parts of the supply chain may remain high.
However, the rise of custom chips could reduce pricing power in some parts of the market. If large technology companies develop more of their own hardware, they may gain leverage over suppliers. They may also spread orders across a wider group of manufacturers, reducing dependence on any single provider.
Digital-asset traders watch the hardware cycle
The semiconductor cycle also matters for digital-asset markets. Tokens tied to shared computing networks, decentralized infrastructure and peer-to-peer processing capacity often respond to the broader demand for machine power.
When large technology companies increase spending on servers and data centers, the market tends to pay more attention to alternative computing networks. These networks attempt to rent unused processing power, storage or bandwidth to software developers and smaller companies that cannot easily access large cloud contracts.
Sustained hardware spending by major corporations can support the argument that compute remains scarce. If demand for AI processing continues to outpace the supply of physical servers, shared server networks may have an opportunity to prove their practical value.
At the same time, digital-asset traders remain exposed to sharp swings. If major technology firms cut or delay server budgets, tokens linked to compute demand could fall quickly. In such periods, some traders may shift toward stablecoins or other less volatile digital assets to reduce exposure to sudden market declines.
The link between chip spending and token prices is not always direct. Many digital assets move for reasons unrelated to semiconductor demand, including liquidity conditions, regulation, software updates and broader risk appetite. Still, the hardware cycle has become more important as AI and decentralized infrastructure narratives overlap.
Utility-based digital coins tied to low-cost processing power may attract more attention if centralized cloud capacity remains expensive or limited. Smaller software builders, artificial intelligence startups and independent developers often need affordable computing access. Shared networks that can provide reliable service at lower cost could benefit if they solve performance, availability and trust challenges.
Compute demand remains the central question
The broader issue is whether global computing capacity can keep up with demand. AI systems, cloud software, video platforms, enterprise automation and digital services are all increasing the need for power, land, cooling systems, chips and network infrastructure.
Morgan Stanley analyst Brian Nowak has projected that global compute capacity could quadruple to 120 gigawatts by 2028. That estimate highlights the scale of the buildout now being debated across equity, credit and digital-asset markets.
A fourfold increase would require enormous spending on data centers and related infrastructure. It would also require reliable access to electricity, advanced chips, memory, storage and cooling technology. In many regions, power supply has already become a bottleneck for new data center projects.
Server budgets among the largest technology companies are expected to remain a central market signal. Some estimates place combined spending by the four biggest technology giants at about $725 billion this year, with much of that money flowing into data centers, AI hardware and custom computing systems.
That flow of capital has supported semiconductor shares, but it has also raised the stakes. If spending continues, the market may view recent weakness as a pause in a longer AI infrastructure cycle. If spending slows, traders may question whether chip valuations have moved too far ahead of actual demand.
For now, chip shares have found a measure of stability after fears of a “DeepSeek 2.0 moment” eased. Kimi K3 has not removed concerns about efficiency, but it has shown that cheaper AI can also create more usage. Alphabet’s Frozen v2 has added another signal that the future of AI hardware will be more specialized, more power-conscious and more closely tied to the design of individual models.
The next test will come from earnings and spending guidance. The semiconductor sector has regained its footing, but the balance remains fragile. Traders are waiting to see whether the world’s largest technology companies will keep expanding their AI infrastructure at the pace markets have come to expect.
Explore how AI volatility shapes chip and crypto markets in Toobit’s latest insights on AI market moves.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
