Global AI and compute-linked equities have gone through a sharp swing in recent weeks, with Korean memory names leading the decline and U.S. AI hardware stocks correcting as a group. Morgan Stanley China chief economist Xing Ziqiang framed the moment as a “halftime break” for AI investing, arguing that the next phase may shift away from pure compute “pick-and-shovel” suppliers toward real AI adopters that can use the technology to cut costs, raise efficiency, and lift revenue.
PANews held a Twitter Space on the same theme on the evening of Aug. 11, bringing together Investment TALK, BIT brokerage head Elio Cui, Virtuals ecosystem lead Harry, and The Mu co-founder Sun. The discussion focused on six areas: what drove the latest drawdown, whether the market’s main line is moving from hardware suppliers to AI adopters, whether the semiconductor and compute-hardware trade is over, how to think about HALO assets, where China’s digital infrastructure opportunity may sit, and how capital expenditure and interest rates are feeding into the trade.
Was the sell-off about fundamentals, or a clearing of leverage?
Investment TALK opened with a full market recap. In his account, March was dominated by concern over geopolitical conflict. In April, retail investors largely missed the move, a sharp contrast with the aggressive dip buying seen during last year’s tariff-driven decline. In May, money rushed into the space and chip stocks at one point rose 100% in two months, creating an overcrowded trade with substantial leverage. June turned into a sideways phase, and the Federal Reserve meeting became the starting point for the pullback. By July, the market had entered a leverage-clearing stage.
His conclusion was direct: the correction had little to do with fundamentals and was, at its core, a washout of sentiment premium and leverage. He pointed to the fact that capital expenditure at Microsoft, Amazon, and Google was still rising rather than falling, while SpaceX had also made a major commitment to data center investment.
Elio Cui agreed and laid out three measurable signals he would watch for a genuine turn in fundamentals. First, capex guidance would need to be revised lower. Second, cloud and AI orders would have to fall behind the pace of investment. Third, credit spreads would need to widen materially. From BIT’s vantage point as a brokerage directly connected to U.S. exchanges, he said funds were still in a net inflow state, there had been no large-scale outflow, and trading remained active. In his words, “it’s completely not at that stage yet.”
Harry added a view from crypto markets. Retail sentiment, he said, had clearly been rattled, but larger users were still buying into weakness. He also said capital could be seen continuing to move out of crypto and into AI-related U.S. equities.
Sun approached the question from the startup ecosystem. His wording was blunt: “AI adoption is actually getting bigger and bigger. A lot of founders now bring AI into almost any project they build. I think the problem is very simple right now — the math doesn’t work.” In his view, the earlier phase saw too much heat and too much speculation. The market now needs a cooling period before shifting into a stage where investors examine how much monetization AI companies can really deliver.
From compute suppliers to AI adopters: who is actually turning AI into revenue?
On Morgan Stanley’s call that the market’s main line is changing, Investment TALK pointed to specific names from a portfolio perspective: large-cap tech such as Amazon and Microsoft, along with selected software companies including Adobe.
He highlighted a change the market had underestimated for much of the past year and a half. Investors had broadly assumed software companies would be “disrupted” by AI. Yet earnings from companies such as Microsoft showed that if software firms can integrate AI with their existing businesses, revenue growth can accelerate while margins avoid obvious damage. He expects AI deployment to show up more on the enterprise side, monetized through existing software companies’ installed user bases, with business models shifting from one-time licensing toward token-based pay-as-you-go pricing.
At the same time, he warned against treating all hardware names the same. “The names that were cut in July need to be looked at one by one. Some are optical modules, some are chips. As long as the fundamentals hold up, I still think they can recover over the long run.”
Elio Cui added a B2C2C ecosystem perspective. He said “Xiaolu,” a Bitcoin mining company under his group, has fully transitioned into AI data centers, or AIDC, making it a small-scale reflection of the broader move from Bitcoin compute power to AI compute power. He also laid out a supply-chain transmission path for semiconductors: the memory price increase cycle has not finished, with capacity already booked beyond 2027; optical modules are still moving through an upgrade path from 800G to 1.6T and then 3.2T; and CPO and silicon photonics are longer-dated incremental themes.
He then offered a standard for calling the AI infrastructure story “done.” “When is the whole AI infrastructure main line finished? It has to be when overbuilding causes a collapse in supply-side pricing. I don’t think the buildout phase is finished.”
Harry named two segments he believes capital may rotate toward in the second half of the year: optics and photonics, and robotics. His emphasis was simple — follow the money.
Sun added two field observations from startup circles. First, many software engineers are beginning to move toward hardware, while aerospace companies are also using AI to speed up R&D. Second, the concern that large models might swallow all AI applications is being tested and, in his view, disproved; large models cannot absorb every narrow and specific use case. Still, he raised a sharp question. Outside of geeks and users in the tech world, willingness to pay for AI on the consumer and enterprise side remains limited. “All this power and all these things are being built, but everyone’s capital chain is still the same, so the math doesn’t clear. It feels a bit strange.”
Has the semiconductor and compute-hardware trade already run its course?
Investment TALK said clearly that he does not think the logic line is over. He used Intel, one of his holdings, as an example, arguing that the company’s internal transition in areas such as optical interconnect can help it through this cycle.
He also urged investors to keep expectations grounded. “This vehicle is too heavy. Some people have to get off before it becomes easier to pull it up again. Any overheated asset will go through a process like this.”
He pointed to another incremental theme as well. Pressure from open-source models in mainland China on two U.S. closed-source models could gradually shift AI value creation away from the closed-model layer and toward the hyperscaler and application layers. He noted that the coding track represents a market of about $2 trillion, while the broader U.S. software market is around $6 trillion. In his view, the room for real landing at the application layer is much larger for software companies than for coding alone.
Elio Cui explained the gap between strong fundamentals and weak share prices from a broader angle. The issue, he said, is not the industry outlook itself but liquidity and credit. “This round of investment is too large, and a big part of it is being filled with debt issuance. If rates or credit spreads widen, financing costs rise, and the market becomes less patient about waiting for returns. In the end, the leverage may not be coming from retail investors. It may be the listed companies themselves taking on a lot of leverage — and that leverage is easier to break than retail leverage.”
Harry added a valuation-cycle comparison. SK Hynix and Samsung have historically traded below 10 times earnings, while companies with more software-like characteristics such as Nvidia have traded above 20 times earnings. The market is still trying to decide whether this memory cycle is a “supercycle,” and he said that judgment is unlikely to change much over the next three to four months.
Why HALO assets, power, and energy are being discussed as repricing candidates
This was the densest part of the conversation. Investment TALK started from the capex tempo of Meta and SpaceX and argued that the bottleneck in compute has already moved from GPU shipments to power supply. He then described a three-layer power hierarchy from cheaper to more expensive options: direct supply from the grid; self-provided power, or BTM, behind the meter, with examples including turbine-linked solutions associated with Siemens, Caterpillar, and GE; and then higher-cost fuel-cell approaches such as Bloom Energy.
His view was that the power-demand thesis will only grow stronger over the next year to year and a half.
Elio Cui then offered what may have been the most actionable framework of the session: a tiered way to think about HALO assets. “Power infrastructure is hard to obsolete and hard to eliminate. That’s the biggest difference between it and GPUs. GPUs are outdated in three to five years, but power plants, grids, and transformers can be used for many years.”
His ranking put power infrastructure and electrical equipment in the first tier; nuclear power and natural-gas generation companies in the second; cooling and data-center-related assets in the third; and raw materials such as copper and aluminum only in the fourth tier.
He returned several times to what he called a first-principles approach. “You only get higher investment returns by staying closer to the center and closer to first principles. Once you move outward to traditional industrial engineering and construction, those are distant relatives. They’re not as core.”
He also warned that not every AI infrastructure name still deserves to be bought. The key question is whether the pace of earnings revisions can outrun the pace of valuation expansion. If valuation has already risen too quickly and forward earnings do not have room to move up with it, then the asset may already be overpriced.
Sun added a forward-looking signal from the startup side: energy is becoming a new hotspot for entrepreneurship, ranging from geothermal generation to household-level distributed solar. His phrasing was simple: electricity is permanent.
China’s digital infrastructure opportunity: open source and lower cost
Investment TALK said openly that he does not cover A-shares, but from a U.S. equity perspective he sees a clear line of logic. If Chinese open-source models do not materially damage acceptance of U.S. closed-source models, then U.S. cloud vendors may actually benefit. API cost is part of software companies’ cost base, and the more tokens they consume — even at a lower unit price — the more the cloud providers carrying those workloads stand to gain.
Building on that, Elio Cui framed both a bear case and a bull case for HALO assets. On the bear side, better model efficiency, lower inference cost, and overbuilding in infrastructure could, in theory, depress the price of compute. He called that the biggest risk to HALO assets. Still, he leaned more constructive because AI demand is growing exponentially. Even if supply-side efficiency keeps improving, demand growth may still outpace those gains. He compared this path with historical patterns in IT and the internet, where performance improved, infrastructure at times looked excessive, and penetration still kept climbing.
Sun delivered one of the most information-rich sections of the discussion, drawing from observations inside founder communities. In his view, the core advantage of Chinese AI models is not just lower cost but open source. “Different countries and even different universities are asking whether they can set up joint research institutes with Chinese AI model companies. The reason is that many countries have realized the United States is not a stable long-term partner, and there are concerns around data security and embedded values in closed-source models.”
He cited a specific example: a London institution focused on AI safety and privacy research had initially believed Chinese AI performance was still insufficient, but after Kimi K2 gained traction, it began reassessing that view and started seeking cooperation with Chinese open-source models in order to build a domestic “sovereign AI.”
He also mentioned regions such as Africa and Latin America, where AI penetration remains low but the base of founders and users is large. There, the monthly subscription price of $100 to $200 for closed-source products such as OpenAI’s creates a high threshold. A combination of Chinese open-source models and lower-cost infrastructure built with local photovoltaic deployment could have a better chance of gaining traction.
Sun’s five-year view was that if OpenAI and Anthropic remain closed source — and he argued they probably have to, because their business models would not work otherwise — more and more countries will build their own sovereign AI systems. A large share of them, he said, would likely choose Chinese open-source models because they are cheaper, more efficient, and easier to customize.
Capex, the Fed, and when strong fundamentals might reconnect with weaker stock prices
Investment TALK used a recent market contradiction to illustrate the current dislocation. About a month and a half ago, investors were debating whether Micron had become “de-cyclical” and deserved a valuation of 15 to 20 times earnings. At the same time, the market did not want to reward the large tech companies effectively “paying” Micron through their spending. He argued that this made little sense. Whether memory suppliers can escape cyclicality ultimately depends on whether downstream big-tech capex generates real returns.
In his view, the latest earnings from Microsoft, Amazon, and Google have already begun to validate that point, and the turn in market sentiment started during the earnings season two weeks ago.
He also laid out a key scenario. “If capex in 2027 can stay in the range of $1.6 trillion to $1.8 trillion, that growth rate is acceptable, as long as the money really comes back. Once capex falls off a cliff, the memory story no longer holds — at its core, it’s still a cyclical sector.”
Elio Cui explained the sector’s sensitivity to liquidity through discount-rate mechanics. “AI is a classic high-valuation, long-duration growth sector. A big part of the pricing comes from discounting future cash flows. If rates move at all, the elasticity on valuation will be much larger than the change on the earnings side.”
When the discussion turned to the possibility of a 25-basis-point rate hike in the fourth quarter, Investment TALK said that would be more “market noise” than the main line unless the Federal Reserve truly reopened a full hiking cycle. On that basis, he sees limited long-term impact on AI fundamentals. Elio Cui stressed that the effect of rates on corporate debt issuance costs — and then on capex pacing — still deserves attention. Sun brought the focus back to applications: Fed rates, he said, primarily affect financing conditions for AI companies and new ventures. The stage of pure narrative and pure storytelling has most likely passed. What matters next is who can actually build businesses with cash flow.