PANews’ 168X program brought back an AI industry participant known as "QihongF44102" on Aug. 19 for a long discussion on where the AI and semiconductor rally now stands. The timing mattered. Semiconductor names had just bounced after a violent deleveraging phase, while Unitree’s stock debut on the same day briefly surged 600%, pushing its market value above 360 billion yuan, or more than $53 billion. That was higher than Figure AI’s roughly $39 billion private-market valuation cited in the show. The question running through the episode was blunt: is this a reset before another move higher, or a top being confirmed?
Still mostly in cash after cutting exposure in May and June
The guest said he reduced positions after his previous appearance and has kept exposure low ever since. He said he had cut U.S., A-share, Hong Kong and some Japanese equity holdings, and now only keeps a small amount of U.S. stock exposure. Other markets, he said, are no longer in his portfolio.
His market view is not uniform across regions. For Korea, Japan and mainland China’s A-share market, he said the AI and semiconductor trade has likely already topped out. In his judgment, those markets are unlikely to make new highs and may not even revisit prior peaks.
Beyond Coding, he does not see a second fast-scaling AI market
On the industry side, he kept returning to Coding, which he described as the core narrative behind the rapid annual recurring revenue growth at two leading AI labs. That growth, in turn, fed the strongest stretch of the semiconductor rally.
Now he sees two problems. First, outside Coding, he does not see another AI use case that can quickly scale into a very large market. Second, open-source models have improved sharply over the past three to four months, both in China and the U.S. He argued that the market is no longer at the stage of asking whether open-source models are usable, or even whether they are good enough. In his view, they have already crossed that threshold. Once that happens, users start caring more about price and cost.
He cited image models as one sign of where competition is heading, mentioning Google’s Nano Banana, GPT image and Qwen image. In that segment, he said, products are already moving into severe homogenized competition. From his frontline perspective, competition has reduced compute utilization and pushed cost and pricing lower. He expects large language models could head in the same direction. That is why he had previously posted on X that leading AI labs face a short-term bubble, even if they may rise again over the long term as more industries adopt AI.
Open-source switching is showing up in real companies
The guest said many agentic AI companies, both overseas and in mainland China, are already replacing closed-source large language model stacks with open-source ones. He gave two reasons.
- Cost: open-source models are now strong enough for many tasks.
- Data security: SaaS companies treat their data as core knowhow and do not want top AI labs learning their business too easily.
He named Perplexity and Coinbase as companies gradually moving toward open-source models and said similar shifts are taking place in China. His argument was that once Coding starts hitting limits, OpenAI and Anthropic will inevitably move harder into SaaS territory, which gives software companies another reason to protect their own data and business logic.
The gross-margin question runs from apps to cloud to chips
The conversation then shifted to profitability across the AI stack. Host Mr. Z brought up SK Hynix’s earnings at the end of the previous month, noting 86% gross margin. He also mentioned that Apple had responded with significant price increases, passing costs on to consumers. The guest said he does not think that kind of margin is sustainable and added that he is also worried about claims from SanDisk’s CEO about maintaining margin over the next four years.
His framework is simple. Upstream margins only hold if downstream economics remain strong. If application-layer profit starts weakening, expecting upstream suppliers to maintain extremely high margins becomes difficult. And once margins hit an extreme, the logic behind them gets fragile.
He said the recent rebound in semiconductor names makes sense in the short run because earnings are still strong and some buybacks have helped. But the longer-term case, in his words, is not well supported at the moment.
He applied the same thinking to the cloud service provider layer. CSP data still looks very good because demand remains there. But as usage shifts from closed models to open-source models, he said investors need to watch a critical difference: closed models can support ultra-high margin, while open-source models do not necessarily do that. Token consumption may continue rising, but whether pricing can hold is another matter.
That is why he said investors should focus on application demand. If no new AI use case breaks out quickly, the rest of the story gets weaker. So far, he said, there are very few scenes outside Coding. SaaS can create some demand, but it depends on knowhow and distribution and has to be pushed industry by industry, which is much slower.
Mr. Z pointed to another signal: Samsung and SK Hynix both reported strong numbers, yet their shares fell after earnings. The host called that dangerous. They also discussed whether memory names are still cyclical stocks. The guest said he largely agreed that they remain cyclical; the cycle just happens to be showing its strong side right now, and that could reverse if oversupply or weaker demand returns.
He says the setup looks more like real estate than 2000
The guest described the current AI boom as a buildout of infrastructure. The U.S., he said, is constructing a new layer of basic infrastructure, including data centers and power systems. When that buildout is heavily leveraged, upstream suppliers naturally tighten.
Still, he sees warning signs. The first is where the money comes from. CSPs are the main source because AI labs themselves generate little cash flow, he said, with additional funding coming from sovereign money in the Middle East. He added that Jensen is now going to Wall Street and bringing in investment banks and firms such as Blackstone, with talk of $500 billion. In his reading, that means pulling more money into the system to keep it running.
The key question is whether all that capital can earn excess returns. If Anthropic can maintain 70% to 80% gross margin over the next year, then the assumption works. But he openly questioned whether that assumption is credible now.
His deeper view is that intelligence itself becomes commoditized. If the products become similar, preserving very high margin gets hard. If that is true, then upstream players should also see margins fall over time, even if demand continues to spread through applications.
Asked whether the comparison should be the 2000 telecom bubble, he said he has never focused much on that analogy and instead sees real estate as the better match. His reason was that the 2000 bubble included many companies with no profits at all, while the current cycle has real profits. The parallel, in his telling, is a financing structure built on the shared assumption that leading AI labs can keep growing very fast and preserve high margins, allowing debt-backed expansion of data-center assets. That resembles a real-estate model where leverage keeps building because everyone assumes prices will continue rising.
He also drew a historical comparison with China’s property cycle. The housing boom that began around 2000 kept running for years, but copper and steel peaked in 2007. He argued that memory today is not fundamentally different. Financial markets and the real economy move on different clocks. In early 2023, the real cycle was still poor, CSP earnings were weak and capex was being cut, yet semiconductor stocks had already started rising. Now he thinks the reverse may be happening: earnings still look strong, but if no new narrative or new industrial logic appears over the next half year, the peak in earnings will begin to show up in stock prices, and stocks are likely to move first.
The host also cited total capex of 725 billion this year among several CSPs and brought up what he called an NVIDIA backstop, where NVIDIA uses its balance sheet to support Neocloud-driven AI data-center expansion and partially absorb default risk. The guest said the long-term case for data centers still stands, since intelligence will eventually rely on that infrastructure, much like fiber laid early was eventually used. His issue is price. He questioned whether Micron earning 75% to 85% margin and NVIDIA earning 75% margin should be treated as normal.
He added that Jensen is only backstopping part of the Neocloud risk, not all of it. Selling GPUs faster is good business for NVIDIA. He also argued that Jensen’s support for open-source models is commercially rational. If a couple of closed-model giants dominate the market, they could build their own ASICs. A more open and fragmented model ecosystem is better for NVIDIA’s GPU business.
A SanDisk valuation exercise points to about $160 billion
The guest said companies that have pushed price increases the hardest are not names he wants to own here. Long-term demand may still grow quickly, but when price has already moved too far, future earnings per share have to be filled in by volume. By the time volume becomes the main support for EPS, he said, those stocks are often already near a major high.
He used SanDisk as an example. In the exercise he described on the show, after-tax cash flow is roughly $30 billion a year. If pricing stays unchanged for four years, as he said the CEO suggested, that would imply about $120 billion in profit over four years. Add another $40 billion as a cyclical-bottom premium for residual value, and the company would be worth around $160 billion on a DCF basis. If token prices fall and GPU orders weaken, he said, that four-year pricing assumption would have to be revisited.
Unitree’s valuation surge did not change his view on robotics
The show also moved into robotics after Unitree’s debut. Mr. Z said Unitree surged about 600% at the open and reached more than 360 billion yuan in market value, or over $53 billion, above Figure AI’s roughly $39 billion to $40 billion private-market valuation. He called it a sentiment trade rather than a fundamentals trade.
The guest said he has consistently been negative on robotics, including during earlier periods of heavy speculation. In his view, mass deployment is still far away. Some of today’s trading in the robotics theme is focused on supply-chain parts such as screws and dexterous hands, but key bottlenecks remain unsolved.
He highlighted data collection as the central problem. Robotics data is difficult to gather at scale, unlike autonomous driving, where regular drivers provide data continuously. He said many robotics companies are still collecting training data by having people physically guide robots through tasks, which is slow. Tactile data, not just visual data, is also hard to obtain. Based on that, his judgment is that robotics will not see large-scale commercialization within five years, including projects at Tesla and Figure.
He did not dismiss the long-term direction. He said he would turn positive once the technology crosses a clear threshold. But at this stage, he does not see any sign of that inflection point.
Mr. Z then argued that China may have an edge because domestic open-source models such as DeepSeek, Kimi and Qwen are already strong and reportedly built at less than 10% of U.S. cost. He also said, based on his recollection, that DeepSeek had subscribed to nearly 5% of Unitree’s equity. The guest replied that China does have advantages in manufacturing-heavy industries, including the embodied and industrial side of robotics, but he remains unconvinced until the threshold really arrives.
RSI may be the last step on the current Transformer path
On recursive self-improvement, or RSI, the guest split the issue in two: technology and penetration.
On technology, he said AI has gone through several rounds since Transformer emerged in 2017, moving through pre-training, post-training, reinforcement learning and then online learning. In his view, today’s RSI is basically an upgraded form of online learning, where smart enough models select their own samples and retrain themselves, potentially every few minutes. Once the field reaches that layer, he said, the Transformer path is close to exhausted. A meaningful next jump will likely require a new model architecture.
On penetration, he said most of the people around him who use AI deeply, or were likely to use it deeply, are already covered. He identified the inflection in Coding around the release of Claude Sonnet 4.6 last year. At that time, maybe two or three out of ten people around him used such tools. By March, April and May this year, he said, it was effectively ten out of ten, with users often running several products at once, including Codex, Claude Code and Cursor.
At the same time, open-source competition has pushed subscription pricing lower. He said fees have gone from 200 to 100 and are still falling as costs come down. Colleagues and enterprises around him are also switching from closed models to options such as GLM and Kimi. His conclusion is that ARPU for users is more likely to decline than rise.
He tied that to revenue expectations. In the environment around him, the switch from Claude Code to Codex happened around May and June, so he expects Anthropic’s ARR slowdown to show up first, followed by OpenAI. OpenAI may have gained market share, he said, but it has also been resetting quotas, discounting tokens and cutting price to compete. He thinks that effect could start showing up within one or two months.
Mr. Z also brought up OpenAI’s notice that it was pausing reinforcement learning training for a newly deployable model for two weeks after hitting what it defined as a threshold for “critical cybersecurity capability,” and asked whether another explanation might exist. The guest said he does not know the internal story. What matters to him is whether these companies can still grow ARR, identify where the market is, explain how they compete and show how they keep pricing intact. So far, he said, he has not seen signs that would overturn his view.
He added that even if token usage keeps rising exponentially, that does not mean ARR will do the same, and it does not automatically imply the same outcome for Anthropic or OpenAI valuations. Jevons paradox may apply at the industry level, he said, but not necessarily to every individual company. He also pointed to high 10-year and 30-year Treasury yields as a sign that the market has placed high expectations on ARR growth, since corporate debt is competing with government debt and leading CSPs are pulling in a lot of capital. If those growth assumptions are wrong, the setup gets more uncomfortable.
Bubble signs inside the industry
One of his more pointed comments was about internal industry behavior. He said that, outside a very small group of top people, much of model training work has actually become easier than before because AI helps with data cleaning and model setup. Yet the industry has elevated many of those workers to near-mythical status. To him, that is a bubble signal in itself.
He also stressed the gap between industry and market perspectives. At the start of a cycle, insiders tend to be pessimistic. By the middle or later stages, they often become highly optimistic because they see demand directly around them. Today, he said, nearly every hardware company one speaks to says business is very good. In early 2023, the message was the opposite. The problem is that insiders may not realize that when everything around them looks booming, the cycle is often already near the top.
He gave one more example. Large companies are all building agentic systems, and many teams are doing very similar work without clear outcomes yet. If the industry cools and companies start measuring return on investment more strictly, he said, those duplicate efforts could be cut quickly. Much of that redundancy has already been turned into capex, and nobody knows exactly where the excess compute will surface first.
Fast private-market fundraising and an IPO window
The primary market came up late in the show. Mr. Z cited news that Jane Street led a $700 million round in AI inference chip startup Etched at a $21 billion valuation. Only a month earlier, he said, the company had raised a $300 million Series C at a valuation above $10 billion, with Sequoia and a16z participating. Raising two rounds in one month, he argued, is a warning sign that private companies are trying to grab capital while the window is still open.
The guest agreed and said OpenAI and Anthropic are surely in a hurry as well. The show also mentioned market talk that Anthropic could pursue an IPO in October. Mr. Z wondered whether the market still has enough resilience to absorb that kind of deal. The guest’s answer was direct: the listing itself could mark the high, at least a short-term one. What happens later depends on the business. But because Jensen and Wall Street capital are now tied to the same ship, he thinks getting the deal out should not be difficult. After that, early investors can cash out, and the next phase begins.
Crypto also entered the conversation
Mr. Z said his team has crypto roots and noticed in February and March that fewer people wanted to hear crypto-focused content, which pushed them to study AI and semiconductors instead. But during the correction from late June and mid-July through the date of the show, he said Bitcoin had not fallen much. He believes someone has been accumulating. He also cited what he described as the largest Solana DAT, Forward Industries, saying it was up 2% to 3% during a week when much of the market was falling. His personal takeaway was that if money from semiconductors starts taking profit or cutting losses, some of it could slowly turn back toward Bitcoin. He made clear that this was not investment advice.
The guest said he does not know crypto well, but he had looked at the charts for MSTR and Coinbase and thought the idea might have something to it.
His preferred positioning: stay in cash, use puts if needed
At the end, Mr. Z asked what investors should do with exposure to memory, Neocloud and GPU names.
The guest said retail investors are best off staying out and waiting. If the whole move is treated as an infrastructure buildout, then after the break investors can buy the companies governments would almost certainly support. He used Intel as an example. For people already stuck in chip and memory positions but still bullish long term, he suggested buying put options as insurance. Institutions that must remain invested should keep portfolios more balanced.
Sector by sector, he is most cautious on highly leveraged companies. If the market breaks, leverage tends to snap fast. He described many Neoclouds as vehicles taking on risk transferred from large CSPs such as Microsoft: they build data centers under contract, but do not generate their own cash flow and often carry high debt loads. He remains wary there. Memory companies at least still have strong cash flow, he said, so they probably will not die, but even after a drawdown they remain elevated relative to normal cycle valuations, including on price-to-book. He said he is highly alert on semiconductors and on SOX. Software looks safer to him, and so do the large cloud and CSP players. Other sectors can be fine if expectations are already low and prices are already low. His broader view is that the U.S. stock market may not fall that much overall, but he remains quite bearish on semiconductors.
If no new application arrives, 2027 could be a very damaging year
When asked whether the violent 2026 bull market is basically over, he answered yes. His more pessimistic scenario is that if no new application use case emerges, 2027 could become a very damaging year, with large index declines.
He said investors can look at the annual semiconductor index chart and its five-year moving average. The deviation is already large. Even if the index finishes this year flat, he thinks it could still revisit the five-year average next year, which in his estimate would imply a 40% to 50% decline. He also acknowledged that elections in the U.S. and Taiwan could produce one more short-lived upward push before that. But his own view is that the rally has reached its late stage.
He added that outside the U.S. market, many global markets have already fallen hard. U.S. equities have held up better partly because of superior liquidity, and he said both Bessent and Trump have been defending that market. Mr. Z noted that Taiwan stocks have already shown some advance reaction, citing Yageo’s 40% to 50% decline.
On rates, Mr. Z mentioned Stanley Druckenmiller’s 13F and positions in rate-sensitive sectors such as housing and autos, then asked whether the right reading is that rates probably do not have much room to rise, but deep cuts are also unlikely unless growth breaks. The guest said macro is not his main field, but he does not think more rate hikes are very likely at these levels. The bigger risk, in his view, is that something at the top of the market cracks first, producing recession and then rate cuts. The high-flying parts would reset toward more normal levels, and the market could then rebuild from there.
He ended by making clear that he is not bearish on AI as an industry. Once expectations across the stack are reset and companies return to more reasonable levels, he believes the industry can keep developing in a healthier way. The issue, in his words, is not the direction of travel. It is the current price.