ME News

wrote a column ·

Navigating the AI Compute Panic: Five Key Themes for AI Technology in the Second Half of 2026

Author and source: Wall Street News

Recently, Meta's announcement of selling computing power has triggered significant market turbulence, leading to notable volatility in the tech sector. There is widespread concern about potential oversupply of computing capacity, raising doubts over whether the investment thesis for AI hardware has already broken down. At this moment of market divergence, Wall Street Journal invited Kong Rong, Deputy General Manager of Guolian Minsheng Securities and Head Overseas Tech Analyst, to share her latest outlook on AI and technology for the second half of 2026, following extensive on-the-ground research and in-depth discussions across the industry chain.

The biggest recent shock in the global tech sector came from Meta’s announcement to lease out its computing power externally. The market’s two biggest fears—‘Is computing capacity oversupplied?’ and ‘Is Meta abandoning large models?’—are, in our view based on field research, both overreactions.

To understand this issue, let’s first clarify the facts.

Meta has consistently increased its capital expenditures (capex) significantly. However, unlike Google, Amazon, or Microsoft, it lacks an established cloud business foundation. Given its continued capex growth and relatively tight cash flow, Meta’s decision to lease out a portion of its computing capacity is entirely understandable.

Here’s an important reference point:SpaceX (and xAI prior to their merger) also built up substantial computing capacity earlier, and later, during its IPO process, leased part of its computing cluster to Anthropic.This was because Anthropic faced severe computing shortages after its Coding (AI-assisted programming) use case took off this year, forcing it to seek computing resources globally. This arrangement generated positive and stable cash flow for SpaceX and provided Meta with a key strategic insight.

Even more noteworthy data includes:Currently, the payback period in North America’s computing power leasing market is around two years or slightly more, making it still a highly profitable business.Even in an environment where acquiring GPUs is relatively easy, the profit potential of computing power leasing remains substantial. This is one of the key reasons Meta is considering entering the space.

Looking at the broader context: NVIDIA is actively promoting its 'New Cloud' initiative. As global AI demand surges, computing capacity remains insufficient—existing cloud service providers (CSPs) alone cannot meet market needs. NVIDIA has rolled out an integrated offering combining computing power and financial services to serve as foundational infrastructure for global AI development. More companies are expected to enter this space. This is what we learned during GTC this year—From NVIDIA’s and GPU manufacturers’ perspective, they too recognize this market opportunity and are pushing computing power leasing and the New Cloud business.。

Returning to the question of whether Meta might abandon large models: based on our understanding, Meta’s large model team continues hiring and releasing models as scheduled, with no disruptions observed. Although Meta’s models are not currently among the absolute top tier (which today is dominated by Anthropic and OpenAI—both of which are considering IPOs this year or in the near term), Meta’s core rationale for building models remains unchanged: integrating them with internal products to drive commercialization and maintain competitiveness.

From a competitive landscape perspective, the model race in the U.S. is already entering a consolidation phase, gradually narrowing down to four or five major players. But why are companies like Meta and Google still committed to developing their own models?Because model capability translates into control over future traffic gateways.If Meta were to abandon model development and cede AI capabilities, companies like OpenAI could build similar AI products, effectively nullifying Meta’s value as a traffic gateway. From this standpoint, Meta will not casually give up on models and AI.

Earlier, when markets were trading at elevated levels with crowded positioning, reactions to Meta-related news were particularly intense. However, based on the facts we’ve gathered and the current state of the competitive landscape,the market has overreacted with excessive panic regarding Meta’s move to sell computing power.。

Finally, from an investment perspective,a payback period of two and a half years means that the more you invest—and the more computing power you own—the stronger and more sustained the future profitability of your business will be.Cloud providers have consistently increased their capital expenditures (capex) to attract more customers and sustain business growth. This week, we also saw Meta announce yet another multi-billion-dollar data center project in Canada, demonstrating its ongoing commitment through tangible investments amid market fears.

Since March this year, the market’s core narrative has begun shifting from 'big tech continuing to ramp up capex' to 'strengthening AI monetization capabilities.'

Over the past three-plus years, capex has been the market’s most closely watched metric—so long as capex kept rising, the investment thesis for AI hardware remained intact. But now, entering the fourth year of the AI boom,relying solely on 'faith' to justify capex spending is actually unsustainable.。

We started seeing signs of Anthropic and OpenAI’s revenue-generating ability as early as Q1 this year. By mid-year, we learned that their combined annualized commercial revenue was approximatelyover $100 billion,already approaching half the annual revenue of major tech giants. For example, Meta generates roughly $200 billion in annual revenue; together, these two large model companies are nearing half that scale.

The continued validation of profit-generating effects and commercialization capabilities forms the most critical foundation of the investment thesis for AI hardware.。

Although there was一度 concern that commercialization momentum might be slowing, the growth rate we currently observe remains quite healthy. Additionally, we’ve seen the release of new models (such as Anthropic’s latest Fbale model), which—based on our own evaluations—demonstrate strong capabilities and are indeed relatively expensive. This implies they will drive sustained growth in annual recurring revenue (ARR), as revenue and commercialization closely track model and product capabilities.

Looking ahead to the next three quarters, we believe overall model-related commercial revenue remains in a healthy state. This will serve as a key indicator for investing in AI hardware and the broader AI sector.It is also a leading indicator for forecasting major tech firms’ capital expenditures (capex).—because these large tech companies allocate capex primarily based on clear evidence that these two model providers are indeed generating profits, and they are willing to continue investing capex as long as this 'profit-generating effect' persists.

Based on our industry engagements in Silicon Valley and South Korea, aside from the rapid iteration of AI model capabilities, one of the next major scenarios worth close attention isrobotics as a key focus area.。

Robotic opportunities and investments have already gone through multiple rounds of thematic trading. However, at this point in the year, the maturity of AI capabilities will bring substantive changes to robotics.

Tesla’s Optimus production line retrofit: a clear signal of mass production

Robots represented by Tesla's Optimus have gradually entered mass production. At Tesla’s Fremont factory in California, Elon Musk recently shared an image—showing that the original vehicle production line at the Fremont facility is nowbeing converted into a robot production line.This is a significant signal: mass production has reached a substantive stage. Additionally, at its Texas factory, we can also see new robot production lines under construction.

From Tesla and Optimus’s progress so far, although there has been prolonged discussion about the timeline for mass production, based on current overseas developments, we understand that:overall capacity is indeed ramping up, and production lines have already begun conversion—preparations are underway for significantly larger-scale mass production next year.。

Supply chain inventory buildup begins: mass production timeline can be viewed more optimistically

Another critical observation comes from the supply chain. During Micron’s earnings call last week, the company highlighted future applications of memory and HBM (High Bandwidth Memory) in robotics. Investors have been closely watching memory stocks, which rose sharply earlier this year, primarily driven by demand from AI servers. But beyond AI servers, what other scenarios could drive substantial memory demand?Micron mentioned robotics during its earnings call.。

Our research in South Korea also indicates that, against the backdrop of rising AI hardware demand, Tesla is making preparations across its broader supply chain—primarily gearing up for future component inventory buildup.

This suggests we can take a more optimistic view on the pace of robot mass production next year and beyond. Previously, many believed progress on robots had stalled due to a lack of recent news, but the current reality tells a different story:Some production lines and capacity are undergoing upgrades, with the expectation that output will increase in the future—particularly next year.This is driving the global robotics industry from the initial conceptual phase (0 to 1) into a phase of mass production and real-world deployment (1 to 10)!

The investment thesis is shifting: from pure thematic speculation to actual mass production and commercialization.

Over the past few years, robotics has been largely conceptual, with limited scale in mass production. However, starting in the second half of this year and going forward, it is gradually entering a substantive phase of mass production. This opportunity couldshift from pure thematic hype to investments grounded in real-world mass production and deployment.In terms of direction, we believe this is a particularly important theme for the second half of this year.

Of course, ramping up production will take time, as it differs from vehicle manufacturing—Elon Musk also noted on X that robotic production lines are fundamentally different from those for cars. Nevertheless, tangible progress is being made.

Aside from robotics, another key area identified through our discussions isMLCC (Multilayer Ceramic Capacitors)。

which has also attracted significant market attention. Some view it as 'the next memory sector,' while others argue its barriers to entry and technological thresholds aren't as high as those in memory. We believeBeyond the steadily growing demand for MLCCs in AI servers, overall demand will also emerge in the medium to long term across areas such as robotics and satellites.。

4.1 The key difference this cycle: AI-driven demand sustainability could last longer than in previous cycles.

Recently, the memory sector has experienced sharp market swings. Early trading activity centered around a view we began highlighting last year—that investor attention toward memory is gradually increasing.

At that time, when discussing with many professionals across the electronics industry and broader tech sector, most believed this cycle wouldn’t differ significantly from prior ones—specifically, that once memory makers started expanding capacity, the overall opportunity in memory would end.

However, we now recognize the biggest distinction this cycle lies in how one interprets the AI opportunity. From our current perspective,AI, as a foundational productivity tool, is generating opportunities and continuity that appear larger than those driven by the internet or general-purpose computing, and its duration is likely to be substantially longer.。

This will determine that hardware opportunities in memory will last longer than in prior cycles. We’re not suggesting memory stocks are no longer cyclical or fundamentally different; rather,we need to define how long this cycle will actually last. Is this a supercycle, and if so, how extended will its duration be?

Looking at current memory demand—whether it’s HBM demand in servers, ongoing acceleration in model capability iterations (as noted earlier, model capabilities continue advancing rapidly with no visible ceiling yet), or future multimodal and AI agent requirements—we observe continuously expanding data volumes. With a solid foundational view on AI, one can reasonably estimate the duration and sustainability of future memory demand.

4.2 Assessment on memory price increases: Balancing interests across stakeholders is required; it cannot be decided by a single company alone

Regarding pricing, we observe differing market views—some believe price hikes will be rapid, while others think the increase won’t be that pronounced. Indeed, demand-side dynamics have accelerated the pace of price increases, but current upstream price hikes shouldn’t be judged solely based on individual company statements; a holistic assessment is needed.

For example, HBM price increases significantly impact AI server shipments and the total amount cloud service providers (CSPs) can afford to spend. We’ve already observed mounting pressure on cloud vendors’ capital expenditures and cash flows.Overall HBM price increases will indeed follow market demand, but must also take into account the need to balance interests across all parties involved.。

Demand remains robust overall at this point. The only potential exception would be if commercialization of AI drives further non-linear growth—as mentioned earlier, Anthropic and OpenAI combined are now generating just over USD 100 billion in annualized revenue. If new models or AI products from these two companies trigger another wave of non-linear, jump-like growth in AI adoption, upstream hardware prices could see renewed or even steeper upward pressure.

Otherwise, with capex already at elevated levels and companies increasingly reliant on fundraising to sustain investments, any significant additional price hikes would place immense strain on all players. A key consideration is whether this is a short-term trade or a long-term business—a judgment that stems from our broader view on AI itself. Regarding overall price increases, especially for the most critical component, HBM, we believe pricing will dynamically adjust in line with market developments while balancing stakeholder interests.It’s not as simple as one company unilaterally deciding on its own.。

4.3 Assessment on memory capacity expansion: Significant near-term supply additions unlikely; supply-demand gap persists

On capacity expansion, last week we saw the South Korean government convene the two memory giants—Samsung and SK Hynix—to unveil a five-year plan outlining future investment strategies. This has raised concerns among market participants about the potential onset of a major industry-wide capacity ramp-up.

However, our view is:Judging from the current pace of capacity expansion across the board, it generally takes about two years or more, so significantly increasing capacity in the short term is not easy to achieve.。

Secondly, from the perspective of actual implementation by these companies going forward—which is also a critical point to watch—their real pace of future expansion will be subject to detailed internal planning next year, as we understand it. Therefore, the extent of any increase can be observed more closely later on.

Additionally, market participants are also highly focused on which companies in the memory sector will truly shape the market’s future—primarily domestic players.CXMT and YMTC (Yangtze Memory Technologies)Their entry into public markets and subsequent participation in competition will significantly impact the overall market. This warrants continued close attention—particularly their upcoming new products and shifts in customer dynamics. However, among the few memory companies already listed, the current pace of capacity expansion is proceeding at a measured rate, with due consideration given to their actual execution capabilities.

From a supply-demand perspective for the overall market,the supply-demand imbalance is likely to persist for a relatively extended period.。

4.4 Potential Shift in Memory Valuation Logic: From PB to PE

In terms of memory investments overall, investors often worry that once capacity expansion begins, the opportunity may already be over. We emphasize that:the supply-demand imbalance will continue to exist.On one hand, this reflects the sustained and rapidly accelerating demand for AI; on the other hand, supply struggles to keep pace with the speed of demand. Under these circumstances, the overall opportunity in memory/storage has not ended.

Regarding the potential shift in valuation logic going forward: previously, the market may have primarily traded memory/storage based on price increases. As mentioned earlier, pricing itself involves numerous considerations—particularly for HBM. If the opportunity in memory/storage continues, what will be the underlying logic?

We believe it has evolved intoa longer-term and more sustainable business. If the cycle extends beyond our initial expectations or lasts significantly longer than anticipated, could its valuationshift from a price-to-book (PB) basis to a price-to-earnings (PE) basis? From recent earnings calls, we can already see that major global institutional investors—particularly long-term, overseas, top-tier funds—have started allocating to memory/storage assets. Why? Because within the AI hardware segment, memory/storage remains attractively valued on a PE basis.

Of course, this requires a clear distinction from the previous valuation framework. The core issue remains whether profitability can sustainably grow. Therefore, as we look ahead to the second phase, the key focus will continue to be on the sustainability of memory/storage’s overall profitability and demand.

Looking even further out, the critical factor will be capacity—who can consistently expand production, which may signal enhanced profitability and drive sustained growth across the sector.

4.5 Memory/Storage Remains Within a Super Cycle

The overall conclusion is that the memory sector remains within a supercycle. Although we cannot say it is no longer a cyclical industry, the AI-driven cycle and opportunities in this round may extend its duration beyond previous cycles. This will likely sustain memory demand and investment opportunities for a longer period than in the past.

Under these circumstances, short-term market volatility will be relatively high due to news such as capacity expansions. The core focus remains on demand—so long as demand persists, the overall memory opportunity will continue. This trend has been reinforced by the significant post-listing rally in SK Hynix's US-listed ADRs andthe upcoming listing of ChangXin Memory Technologies (CXMT), which continues to sustain investor interest and momentum in the broader memory industry.We remain highly optimistic about the global competitiveness and capabilities of Chinese memory companies.

5.1 Impact of Chinese models on the U.S. market: Open-source models capturing market share

From the perspective of global model competition, although U.S. models are indeed powerful, they are also very expensive. This leaves a segment of demand unmet—such as small and medium-sized enterprises, startups, and industrial firms—that lack sufficient and sustained budgets to continuously cover the high costs of these models.

This inevitably creates room for alternative solutions, and open-source models are indeed capturing market share in this space. We have observedZhipucapabilities that are highly regarded not only in domestic discussions but also globally, including strong endorsements from industry professionals. Therefore, we believe open-source models are currently gaining market share and will continue to secure additional share going forward. This presents a significant opportunity for the development of Chinese models.

Looking at this dimension of global model competition, beyond Zhipu AI, other ChineseKimi, DeepSeek, Qwen, all of which have also attracted considerable attention. In fact, based on our conversations with companies in the industry, it’s not just about reputation—these models are indeed being used in practice. As we’ve previously seen, some overseas companies’ CEOs have mentioned in public forums that, due to the high cost of U.S. models in the early stages, they built their internal products and drove product iterations primarily using open-source models (likely dominated by Chinese ones).

5.2 Domestic Large AI Model Market Landscape: Intensifying Shakeout, Continued Global Competitiveness Expected

Currently, China’s large AI model landscape is largely dominated by three major tech giants: Alibaba, Tencent, and ByteDance; alongside roughly three or four leading startups, including Zhipu AI, Kimi, and DeepSeek mentioned earlier. The competitive landscape will inevitably evolve going forward, likely consolidating similar to what we’ve seen among U.S. companies. A preliminary round of market consolidation has already occurred over the past two years, and the upcoming phase of this shakeout is expected to be even more intense.

We remain highly optimistic about the global competitiveness of Chinese large models. Strong foundational capabilities—rooted in open-source development and bolstered by China’s robust AI talent pool—provide a solid basis for continuous model iteration and product enhancement. Coupled with our relative advantages in pricing and cost-effectiveness, Chinese models are well-positioned to gain broader global market recognition.

Viewed over a longer-term horizon, the primary driver of AI adoption in the first half of this year has largely come from select overseas tech giants, which initially encouraged unlimited internal usage by employees. However, as companies begin to prioritize cost efficiency and sustainable returns, they will inevitably conduct rigorous cost-benefit analyses. In this context, competitive Chinese open-source models will certainly become a natural and necessary option for consideration going forward.

63K Views