Crossing the Computing Power Panic: Five Main Threads in AI Technology in the Second Half of 2026

In recent times, news that Meta is selling computing power has caused severe turbulence in the market, with significant volatility in the technology sector. The market generally worries about an oversupply of computing power, and thereby questions whether the logic behind AI hardware investment has already broken down. At this point of market divergence, Wall Street Insights invited Kong Rong, Deputy General Manager and Overseas Technology Chief Analyst at Guolian Minsheng Securities. After conducting on-site research and in-depth exchanges with the industry chain, she shared the latest outlook for AI technology in the second half of 2026.

1. Meta selling compute power ≠ compute power oversupply → Meta continues to build new data centers: opportunities found amid excessive panic

The biggest shock to the global technology sector recently comes from Meta’s announcement that it will rent out computing power to external parties. The two issues the market worries most about—“Is compute power oversupplied?” and “Is Meta going to give up on large models?”—in our research, are both overreactions.

To understand this, first let’s sort out the facts.

Meta’s prior capex (capital expenditures) has been increasing substantially, but unlike Google, Amazon, and Microsoft, it doesn’t have an existing cloud business foundation. Against the backdrop of continued capex increases and tight cash flow, considering renting out part of its computing power externally is completely understandable in itself.

Here’s an important reference point: SpaceX (and xAI before the merger) also made large computing power reserves early on. Later, during the listing process, it rented out part of its computing clusters to Anthropic, because once Anthropic’s Coding (AI programming assistance) scenarios took off this year, it experienced a severe shortage of computing power—meaning the whole world was scrambling for compute. This brought SpaceX positive and stable cash flow, and it also provided Meta with an important insight.

More worth关注 data is: the payback period in North America’s compute leasing market is currently about more than two years, and it’s still a fairly profitable business. Even in an environment where GPUs are relatively easier to obtain, the profitability of compute leasing remains significant. This is one of the core reasons Meta is considering entering the field.

Look at the broader backdrop again: Nvidia is pushing its “new cloud” business. After global AI demand picks up, compute power is still not enough, and relying only on the existing CSPs (cloud service providers) makes it difficult to meet market demand. Nvidia has launched support solutions that include compute power plus financial services, as fundamental infrastructure for global AI development. In the future, more companies will enter this space. This is what we learned during this year’s GTC—from Nvidia and GPU companies’ perspective, they also see this market opportunity and are promoting compute leasing and the new cloud business.

Back to whether Meta might abandon large models. What we learned is that Meta’s large model team hiring and model releases have not been affected, and the large model push is moving forward according to the original timeline. Although Meta’s models are not currently the very top tier among the “frontline” (the main frontline players are currently Anthropic and OpenAI; both are considering IPOs this year or in the near future), Meta’s core logic for building models is to integrate them with internal products, realize commercialization, and maintain competitiveness. That strategic goal has not changed.

From the logic of the competitive landscape: model competition in the US has already entered a convergence process, gradually narrowing down to four or five companies. But why do Meta and companies like Google insist on building models? Because model capability means the battle for future traffic entry points. If Meta gave up on building models and on AI capabilities, then companies like OpenAI could build similar AI products—meaning Meta’s traffic entry value would go to zero. From this perspective, Meta is unlikely to give up on building models and AI casually.

In the early period, with the market at high levels and trades relatively crowded, reactions to the Meta event were intense. But based on the facts we learned and the overall situation across different parties in the competitive landscape, the market is too panicked about Meta selling compute power.

Finally, from an investment perspective: a payback cycle of two and a half years means that the more you invest and the more computing resources you have, the more the future business profitability effect will be sustained. Cloud vendors have also been increasing capex to gain more customers and keep business growth continuing. And we also see that Meta has announced this week that it will build a one-hundred-billion data center project in Canada, using real ongoing investment to respond to market panic.

2. AI commercialization revenue inflection point: from “belief in capex” to “seeing real revenue”

Since March this year, the core logic in market trading has started to shift from “big companies continue to increase capex” toward “AI monetization and cash-generating ability strengthening.”

In the past more than three years, the market’s most watched indicator has been capex—if capex keeps increasing, it implies the overall AI hardware investment logic is still intact. But this year marks the fourth year of the AI wave, and investing in capex purely based on “belief” is actually difficult to sustain.

We could already see this in the first quarter this year: the profitability of Anthropic and OpenAI began to become visible. By mid-year, we learned that their combined annualized commercial revenue is roughly over $100 billion, already approaching about half the revenue scale of major enterprises. For example, Meta’s annual revenue is around $200 billion. Put the two model companies together, and it may be close to about half of those enterprise revenues.

Sustained validation of monetization profitability and commercialization capability is the most important foundation for the AI hardware investment logic.

Although people were once concerned that monetization might start slowing down, what we currently understand is that the growth rate remains fairly healthy. In addition, we have also seen the latest model releases (such as Anthropic’s new Fbale model, etc.). Based on our own evaluations, the capabilities are strong, and it’s also indeed relatively expensive—meaning it will drive continued growth in overall revenue (ARR). That’s because revenue and commercialization basically move along with model capability and product capability.

Looking further into the next three quarters, we believe that the overall model commercialization revenue will still remain in a relatively healthy state. This will be a very important core indicator for investing in AI hardware and the entire AI sector—and it is also an important forward-looking indicator for assessing big companies’ capex. Because when big companies invest in capex, they mainly see that these two model companies are actually making money, and they are willing to keep investing in capex while continuing to see the “profitability effect.”

3. Robot mass production entering a substantive stage: from theme speculation to mass production on the ground

Based on our industrial exchanges in Silicon Valley and South Korea, besides the rapid iteration of AI model capabilities, the next major scenario that deserves close attention is robots.

Opportunities and investments in robotics have already gone through many rounds as a theme investment. But at this point in time this year, as AI capabilities mature, robots will undergo tangible changes.

Tesla Optimus production line renovation: the mass production signal is clear

Robots represented by Tesla Optimus have already gradually entered the mass production stage. At the Fremont factory in California, Musk recently released an image—basically, the production line that used to manufacture cars at the Fremont factory is now being renovated to produce robot production lines. This is an important signal: mass production has moved into a more substantive phase. In addition, at the plant in Texas, we can also see that it is building new robot production lines.

From the progress of Tesla and Optimus, we can infer that although they previously talked for a particularly long time about the mass production schedule, based on current information from overseas, overall production capacity has started to be ramped up, and production lines have begun to be converted—this is to prepare for much larger scale mass production next year.

Supply chain begins inventory stocking: mass production timeline can be more optimistic

Another important observation dimension comes from the supply chain. At Micron’s earnings call last week, it also mentioned future applications of storage and HBM in robots. People are paying attention to storage—this has risen quite a lot earlier this year, mainly because of trading opportunities tied to AI servers. But beyond AI servers, in what other scenarios will storage be used more? Micron’s earnings call mentioned robots.

From our research in South Korea, we can also see that amid the current rise in AI hardware, from the overall supply chain perspective, Tesla is making some preparations mainly to do supply chain inventory stocking for later stages.

This means that for the mass production schedule of robots next year and beyond, we can be more optimistic. Because although people previously believed that robot progress had not shown more news for a long time, the current substantive situation is: some production lines and capacity are undergoing renovation, and capacity is expected to be ramped up in the future (especially next year). This will help the global robotics industry move from the former concept stage of 0–1 into the 1–10 mass production and on-the-ground deployment stage!

Investment logic is changing: from pure theme to mass production deployment

In the past few years, robotics have been concept-driven, with overall mass production scale remaining relatively small. In the second half of this year and looking forward, it is gradually entering a substantive mass production stage. This opportunity may shift from the former pure theme speculation to an investment thesis centered on real mass production on the ground. In terms of direction, we think this is an important direction in the second half of this year.

Of course, it will take some time to ramp up, because it’s different from building cars—Musk also mentioned on X that the robot production lines are completely different from the existing car lines. But it is still making real substantive progress.

Apart from robots, another direction that stands out from the exchanges is MLCC (multilayer ceramic capacitors).

Market attention is also high, but some views think it is the “next storage,” while others believe its barriers and thresholds are not as high as storage. We believe that in addition to steadily growing demand for AI servers, overall demand for MLCC will also manifest in the mid-to-long term across areas including robots and satellites.

4. Storage super cycle: AI reshapes demand durability

4.1****The biggest difference in this round: the demand durability brought by AI may be longer than in previous rounds

During this time, the storage sector may see sharp surges and sharp drops. In the early phase, overall trading has been driven by a judgment we started sharing since last year: storage is gradually drawing more attention.

At that time, when we discussed with people in the electronics industry and across the industry, people believed this cycle would be no different from previous cycles. That includes the view that once storage companies start capacity expansion, the overall storage opportunity would end.

But what we’ve seen in this round’s biggest difference is how you view this AI opportunity. Looking at it now, because AI is a bottom-layer efficiency tool, the opportunities and their continuity it creates—at least for now—are bigger than the internet, and bigger than gene computing (should be “cloud computing” or “general computing”); and its duration is also likely to be longer.

This will determine that the hardware opportunity for storage will last longer than previous rounds. We can’t say it’s “different,” or that it’s no longer a cyclical stock type. Rather, we need to define: how long is this cycle supposed to be? Is it a super cycle—and how long will it last?

If you look at current storage demand, whether it’s HBM demand in servers, future model capability iterations (we also mentioned at the beginning that model capability iteration is still accelerating, with capabilities getting stronger and still no ceiling in sight), as well as future multimodal demand, Agent demand, and so on—you can also see that the data volume keeps growing. With a baseline judgment about AI, you can roughly estimate how long future storage demand will last and how long it will remain durable.

4.2****On storage price increases: need to balance interests among all parties; not determined by a single company

Regarding pricing, we see different market views: some think price increases happen quickly, others think the increases aren’t that obvious. It’s true that demand drives the pace of price increases, but the current price increases at the upstream level can’t be judged only by each company’s narrative—it needs to be assessed overall by combining factors.

For example, regarding HBM price increases, the core impact is also on AI server shipments: the overall amount of money that CSP companies can pay. We actually have also seen the current pressures on cloud vendors—capex and cash flow. So we believe HBM’s overall price increases will indeed move with market demand, but it also needs to consider and balance the interests of all parties.

From the current overall demand side, it’s indeed fairly strong. Unless one potential scenario happens: commercial monetization and AI further experience non-linear growth. We mentioned earlier that Anthropic and OpenAI combined have annual revenue around just over $100 billion. If later these two companies achieve AI non-linear growth and big jumps due to new models and AI products, then upstream hardware could again pull prices up or even more sharply.

Otherwise, at the current level, capex is already at a high point, and if people need to rely on financing to invest further, then a more substantial price increase would create significant pressure for all companies. The key is to balance: is it a short-term trade or a long-term business? This is also judged based on the overall perspective of AI itself. For overall price increases—especially for the most core HBM price increases—we believe it will dynamically adjust based on changes in the overall market, and still needs to balance the interests of all parties. Therefore, it’s not that simple for one company to decide unilaterally on a single dimension.

4.3****On storage capacity expansion: short-term expansion can’t add a lot of capacity; supply-demand gaps will persist

Regarding capacity expansion, we may have seen it last week as well: the South Korean government convened two storage giants—Samsung and SK hynix—to make a five-year plan, including how total investment will be carried out over the future. The concern for everyone here is whether a major capacity expansion is coming.

But our view is: based on the overall capacity-expansion pace today, it basically takes two years to more than two years for it to materialize, so it’s not easy to expand a lot of capacity in the short term.

Second, from the perspective of these companies’ actual future implementation, this is also a key observation: they actually plan in detail for how to carry out real expansion next year. So the magnitude of ramp-ups can be observed later and assessed how it will actually work out.

In addition, what many people are also paying attention to is that among storage companies, the ones that truly will affect the future market are domestic players CXMT and Yangtze Memory (YMTC, Yangtze Memory Technologies Co.). Once they enter the market through their listing and competition participation, what impact will that have on the overall market? This may be a continuous focus going forward—through their later new products and changes in overall market customers. But judging from the current pace of the existing listed companies, capacity expansion also follows a certain cadence, and it’s still necessary to consider the capability to actually execute and materialize.

For the market overall, from supply-demand, the imbalance between supply and demand will continue for a relatively long period.

4.4****Potential shift in storage valuation logic: from PB to PE

For overall storage investing, people worry that if the opportunity is only about capacity expansion, it might end. We want to emphasize: the supply-demand imbalance will persist. On the one hand, it is determined by the fact that AI demand is continuous and rapidly and continuously pulling demand. On the other hand, supply and upstream supply are hard to keep up with the pace of demand. In this situation, the overall storage opportunity does not end.

About changes in valuation logic going forward: previously, the market mainly traded storage based on price increases. As said, price increases themselves require considering multiple factors, especially for HBM. So if the storage opportunity continues, what will its logic be?

We think it will turn into a longer, more continuous business. If it goes beyond or lasts longer than we imagined overall time cycle, can its valuation switch from PB valuation to PE valuation? In this round’s earnings reports, you can also see that some big overseas, global, long-term first-tier institutions have started to allocate to storage. Why? If you look at the storage industry within the AI hardware segment, using PE valuation, it is still cheap.

Of course, this needs to be distinguished from the old valuation framework. The core here is still that its profitability can continue to grow. So we believe that in the second stage, the core lies in the sustainability of storage’s overall profitability and the sustainability of overall demand.

Looking even longer term, the key is to see capacity—who can continuously expand, which could also mean improved profitability and in turn drives overall sustained growth.

4.5****Storage still remains within a super cycle

The overall conclusion is: storage in this round is still within a super cycle. Although we can’t say it’s no longer the so-called “cyclical commodity category,” because the AI cycle and this round of AI opportunities may make us see it lasting even longer than previous rounds, it also will pull overall storage demand and storage opportunities for a longer time than in the past.

In this situation, in the short term the market will be relatively volatile due to some capacity-expansion news and the like. The core still depends on the demand side: as long as demand exists, the overall storage opportunity will continue. After SK hynix’s (Hynix) listing on US stock ADR and a big jump, and as CXMT (CXMT Storage) is about to be listed, investment and enthusiasm for the overall storage industry will continue. We still strongly believe in Chinese storage companies’ competitive strength and capabilities in global competition.

5. Domestic large models capturing global market share: cost-performance advantages + open-source path

5.1****China models’ impact on the US market: open-source models抢占份额

Returning to the global model competition perspective. US models are indeed strong, but they are also very expensive. This means that some demand cannot be satisfied—some smaller companies, startups, and industrial companies may not have sufficient, sustained budgets and costs to continuously cover such expensive models.

So it inevitably leaves some demand, and open-source models are indeed抢占市场份额. Including what we’ve seen about Zhipu’s capabilities—this is not limited to domestic discussions. Globally, including industry evaluations, it has received high praise. So we believe that open-source models are currently taking market share, and will continue to obtain some share in the future. This creates important opportunities for the development of China’s models.

From the perspective of global model competition, besides Zhipu, attention is also high for China’s Kimi, DeepSeek, and Qwen. In fact, we learned from some companies in the industry that people are not only talking about it reputationally—actually, they are using it (transcription/recognition: “in actual”—the original says “在面,” but it should be “实际”). Including earlier, we also saw some overseas companies; their CEOs mentioned in some public meetings that because US models were expensive in the early stage, they also based their internal product development and product iteration on open-source models (mainly likely Chinese models).

5.2****Market landscape for domestic large models: elimination rounds intensify; global competitiveness remains promising

For domestic models, we can basically see three major players: Alibaba, Tencent, and ByteDance. Startups may have around three or four companies, including the ones mentioned earlier such as Zhipu, Kimi, DeepSeek, etc. Looking forward, the competitive landscape will inevitably change, and it will appear to “shrink” like US companies. The past two years have already entered a certain elimination stage. Next, the elimination competition may become even more intense.

As for China’s model competitiveness in global markets, we remain very optimistic. Based on open-source and our relatively strong existing AI talent teams, it will strengthen our model capability and ongoing iteration, as well as product capability, creating a solid foundation. Combined with our relatively lower price and cost-performance advantage, this will help China’s models gain more global market recognition.

Especially from a longer-term dimension: in the first half of this year, the main drivers for AI might have been from certain overseas big companies. Earlier, they endlessly encouraged employees and internal use. If it’s a company considering costs and sustained effective returns, it will definitely calculate this economic equation. For open-source Chinese models with competitiveness, it will be a natural option they consider in the future.

Risk disclosure and disclaimer

        The market involves risks; investment involves caution. This article does not constitute personal investment advice, and it does not consider any specific users’ special investment objectives, financial conditions, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article align with their specific circumstances. Investment decisions are made at your own risk.