AI Industry Digger: Wait for the Bubble to Burst, Then Pick Through the Wreckage! 2027 Will Be a Brutal Year

Summary

• The Digger has sharply cut his exposure since May–June and is now almost entirely in cash, calling a definitive top in Korea, Japan and A-shares, declaring 2026’s ferocious bull market over, and saying 2027 “could be a year of very large index declines.” Technically, the semiconductor index is far above its 5-year moving average; even if it closes this year flat, it could retest that average next year, implying another 40% to 50% decline. The market “could still stage one more small rally” ahead of U.S. and Taiwan elections, but everything outside U.S. equities has already been “beaten to a pulp”—the cycle is in its final stretch.
• The core bearish case is that, outside coding, “there is no other fast, easy market capable of absorbing a large market,” while open-source models have crossed the threshold of “whether they work, or even whether they work well,” turning competition into a price war. Image models are already deeply commoditized: he is seeing Google Cloud’s Nano Banana usage decline, alongside lower costs and prices. He expects LLMs to follow the same path. The ARR of leading AI Labs is “quite a large short-term bubble”; once the leading applications weaken, the pressure will flow through midstream CSPs to downstream semiconductors.
• Gross margin is the pressure point: the Digger believes “intelligence itself is commoditized … it is a commodity,” making extreme margins upstream and downstream hard to sustain, and questions whether NVIDIA’s 75% to 85% margin is justified. Mr. Z cited SK Hynix’s roughly 86% gross margin and the SanDisk CEO’s view on prices or margins over the next 4 years; under his own assumptions, Mr. Z roughly values SanDisk at $160B. Storage remains cyclical—it is simply showing the upswing today.
• This cycle looks more like real estate than the 2000 telecom bubble: companies are generating real profits, but the entire leverage stack rests on the assumption that leading AI Labs will grow rapidly and sustain their margins, while data centers are packaged into bonds and refinanced in a structure not fundamentally different from real-estate bond refinancing. Copper and steel in China peaked around 2007 while property remained hot for more than a decade, making them an analogy for storage today. The financial cycle leads the real economy; by the time results reflect reality 6 months later, the related stocks may already have fallen sharply.
• The Digger sees Jensen’s moves as business: he cited bringing in Wall Street capital and financing on the order of RMB500B; Mr. Z described NVIDIA using its balance sheet to backstop Neocloud, with the Digger stressing that it would backstop only part of the exposure. Jensen’s support for open-source models, in the Digger’s view, is aimed at preventing a 2-company closed-source duopoly from developing in-house ASICs and damaging demand for NVIDIA GPUs. Data centers will eventually be used, like fiber-optic cables laid ahead of demand, but whether current 75% to 85% margins are reasonable remains an open question. If Anthropic IPOs in October, Mr. Z believes “the IPO itself could be the top.”
• Robotics is a sentiment trade, not a fundamentals trade: Mr. Z said Unitree Technology opened roughly 600% higher, peaking above RMB360B, or more than $53B in market value, versus Figure AI’s roughly $40B private valuation; the Digger sees no mass deployment in robotics within 5 years. Data collection is far harder than in autonomous driving: manual assistance and tactile data are both difficult. Mr. Z is not overly concerned about China’s foundation models or manufacturing capabilities and thinks the long-term winner could be a Chinese player; the Digger only agrees that robotics becomes attractive once the technology crosses its threshold.
• Front-line penetration data show coding user counts and usage intensity are already maxed out, with subscription prices falling from $200 to $100 and still declining: “per-user spend can only go down from here, not up.” The Digger’s team switched from Claude Code to Codex around May–June, so he expects Anthropic’s ARR to slow first, followed by OpenAI. To change his view, AI Labs must explain what the market is, how they will compete and how they will preserve pricing.
• Playbook: the Digger advises retail investors to stay flat and wait to pick through the wreckage after the market blows up; those who are invested and remain bullish should use a small amount of capital to buy puts for protection. Mr. Z is most negative on high leverage and Neocloud, viewing companies such as CoreWeave as risk-transfer vehicles for major CSPs. Storage companies have ample cash flow and are unlikely to fail, but their valuations remain high relative to a normal cycle. He is highly cautious on SOX and sees software, cloud and large CSPs as relatively safe. His overall call: U.S. equities may not fall much, but semiconductors could fall a lot.

Deep dive

1. Opening Stance: Fully in Cash; Korea, Japan and A-Shares Have Already Peaked

• The Digger said that after his last exchange in May–June, he cut exposure across U.S. stocks, A-shares, Hong Kong stocks and Japanese stocks: “At this point I should have only a little exposure to U.S. stocks, but it’s not much … basically I’m sitting on a full cash position.”
• His market call was unequivocal: Korea, Japan and A-shares have “definitely already peaked,” cannot make new highs and are unlikely even to approach their previous highs.

2. No Second Fast Market Outside Coding; Open Source Has Turned Competition Into a Price War

• His core framework is that the strongest leg of the semiconductor rally over the past year-plus was built around the coding narrative and surging ARR at 2 leading AI Labs. There are now 2 problems: outside coding, “there is no other fast, easy market capable of absorbing a large market,” while open-source models—including mainland Chinese models, Grok and Meta’s catch-up models—have advanced extremely quickly over the past 3-4 months.
• The essential difference from the DeepSeek moment is that models then still had substantial room to improve. Open-source models today have “already moved past the stage of whether they work, or even whether they work well,” leaving users more focused on cost and price.
• He is seeing first-hand that Imagen, GPT Image and Qwen-Image have entered intense commoditization. Google Cloud’s Nano Banana appears to be using less compute, while costs and prices have also fallen. “At least from where I sit today, I can see LLMs falling into a similar situation over the next period of time,” which he views as a highly unfavorable trend.

3. Leading AI Labs’ ARR Has a Short-Term Bubble; the Gross-Margin Chain Is Fragile

• In a prior post on X, the Digger argued that leading AI Labs’ ARR “has a bubble in the short term, and it’s quite a large bubble.” Over the long term, he expects ARR to rise again as other business lines emerge across industries.
• This is the top layer of the entire AI narrative. Once it weakens, midstream CSPs, downstream AI applications and semiconductors all become vulnerable: downstream gross margins will not hold, and ultra-high upstream margins will not hold either. Once margins reach extreme levels, both the logic and the narrative become fragile.
• He viewed the rebound after the last selloff as reasonable: the immediate fundamentals still work, earnings remain very strong and some buybacks are supporting prices. But at least for now, the long-term logic does not hold up.
• The CSP risk is that current data still looks excellent, even as the market shifts from closed-source to open-source models. “Closed-source models have extremely high gross margins; open-source models do not.” Token consumption may continue to rise, but whether pricing can hold is debatable.
• The anchor is downstream application scenarios. SaaS can create value, but its main advantages are know-how and distribution, which must be developed industry by industry. The process will not be as fast, easy or replicable as coding. Without a rapid breakthrough in a new application scenario, neither RLVR nor any other narrative matters.

4. Unitree’s Surge Is Sentiment, Not Fundamentals; No Mass Robotics Deployment Within 5 Years

• Mr. Z described the market action: Unitree Technology reportedly opened roughly 600% higher, at one point reaching more than RMB360B in market value, or above $53B, versus Figure AI’s roughly $40B private valuation. “This is not trading fundamentals at all; it is trading sentiment.” He also mentioned 王兴兴 appearing in a suit without a smile and 沈南鹏 of Sequoia at the lunch.
• The Digger has consistently been negative on robotics because data collection is difficult. In autonomous driving, every driver helps collect data; robotics companies often have to pay people to manually guide the machines, making data collection slow. Tactile data, beyond vision, is even harder to gather.
• His view is that the robotics industry—including Tesla and Figure—will not see mass deployment within 5 years. He would become constructive once the technology crosses its threshold, but sees no sign of that breakthrough yet.
• Mr. Z offered the opposing view: China’s “brain” does not need to be feared, pointing to DeepSeek, 月之暗面’s Kimi and Qwen as powerful models built at less than one-tenth the U.S. cost. China also has a manufacturing advantage, much as it does in autos, so the eventual winners may still be Chinese companies. The Digger agrees only that robotics becomes investable after the technology crosses its threshold; until then, he remains negative.

5. Storage: Extreme Margins; SanDisk Is Worth Only $160B Under Mr. Z’s Assumptions

• Mr. Z recalled SK Hynix’s earnings report late last month showing a gross margin of “about” 86%, and mentioned Apple saying it would raise prices and pass costs on to consumers. The Digger believes margins at that level are unsustainable and is concerned by the SanDisk CEO’s claim that related prices or gross margins can hold for the next 4 years.
• Mr. Z challenged the idea that memory “is no longer cyclical.” It remains cyclical; it is simply in an upcycle now. When supply becomes excessive or demand weakens, it will turn down again. Samsung and SK Hynix reporting strong earnings while their stocks fall is also a danger signal.
• His rough math: if SanDisk generates $30B in annual after-tax cash flow as the CEO suggests, that is $120B over 4 years. Add a $40B premium at the cycle trough, and his DCF assumptions imply a value of roughly $160B. If future Token prices and GPU rental rates decline, the assumption that prices can hold for 4 years collapses and the valuation must be reset.

6. This Cycle Looks More Like Real Estate Than the 2000 Telecom Bubble

• The Digger’s characterization is that “this AI cycle is fundamentally a bet on infrastructure.” The U.S. is building a new generation of infrastructure, including data centers and power capacity; when construction is funded with heavy leverage, upstream supply is bound to tighten.
• His analogy is Chinese real estate. The sector began developing around 2000 and remained hot for more than a decade after 2007 and 2008, while copper and steel in China had already peaked around 2007. He sees no fundamental difference between that cycle and storage today.
• There is a gap between the financial cycle and the real-economy cycle. In early 2023, CSP earnings were still weak and some companies were cutting capex, but semiconductors had already started rising. Earnings were still poor when semiconductors first rallied and only came through 6 months or more later.
• If there is no new narrative or change in industry logic 6 months from now, today’s strong earnings will gradually be reflected in the numbers. By then, however, the related stocks will probably already have fallen sharply.
• He rejects a simple comparison with the 2000 telecom bubble: many companies then did not generate such high profits, whereas companies today are producing real profits. The current leverage rests on the assumption that leading AI Labs will continue to grow rapidly and sustain their margins. Data centers are packaged into bonds, sold and refinanced, which is fundamentally similar to refinancing real-estate bonds; once prices start falling, the highly leveraged structure could collapse rapidly, as seen in Korea.

7. Jensen’s Game: Backstops, Wall Street Capital and Open-Source Support Are All Business

• The Digger believes AI Labs generate almost no cash flow yet, with funding coming mainly from CSPs and Middle Eastern sovereign wealth funds. Jensen bringing investment banks in through Wall Street, with Blackstone and others reportedly raising roughly RMB500B, is fundamentally about continuing to bring in money to sustain the buildout.
• Mr. Z described NVIDIA using its own balance sheet to backstop smaller providers such as Neocloud. The Digger added that Jensen would backstop only part of the exposure, not all of it. For Jensen, rapidly selling GPUs while maintaining maximum margins is obviously beneficial.
• The Digger sees a commercial motive behind Jensen’s support for open-source models. If models end up controlled by 2 closed-source companies, those companies will understand their own models best and could build ASICs in-house, hurting demand for NVIDIA GPUs. Open-source models, by contrast, are used separately by different players and are not tied to a single fixed model, which is better for NVIDIA. “These are all business decisions”; public statements are not enough—the underlying interests matter too.
• Long-term direction and current pricing have to be separated. Data centers will not be permanently overbuilt, just as fiber-optic cables laid in advance were eventually used. But whether NVIDIA’s 75% or even 85% margin is normal remains a valid question. If downstream margins are not nearly as high, current prices are the product of extreme demand compression and cannot hold over the long term.

8. RLVR Deconstructed: Transformer Has Essentially Reached the End of the Road

• The Digger’s technical history runs as follows: Transformer emerged around 2016 or 2017, after which models progressed from pretraining to reinforcement learning and then online learning. He sees RLVR as essentially an upgraded form of online learning, allowing a smarter model to select its own training samples—potentially training every few minutes.
• His conclusion is blunt: at the online-learning stage, the Transformer path has “basically reached the end of the road.” The next technical iteration is unlikely to rely on Transformer alone; a new model architecture will be needed to move intelligence up another level. Transformer alone cannot reach AGI.

9. Coding Penetration Is Maxed Out; ARPU Can Only Move Lower

• The penetration curve is already saturated. After Anthropic released Claude 4.6 last year, roughly 2 or 3 people out of 10 around him were using it. By March–April and April–May this year, essentially all 10 were using it. Each person may be using Codex, Claude Code and Cursor simultaneously, with usage intensity also maxed out.
• Prices have fallen from $200 to $100 and are still declining. His view is that each user’s spending can only move lower from here, not higher.
• His team switched from Claude Code to Codex around May–June, leading him to expect Anthropic’s ARR to show a slowdown first, followed by OpenAI. OpenAI has taken substantial share, but it is giving value back to users through frequent quota resets and lower Token prices; the relevant data may show up over the next 1-2 months.
• The API side is also shifting toward open source. Many enterprise technology companies, both overseas and in mainland China, have already switched their LLMs to open-source models such as GLM and Kimi. There are 2 reasons: cost and data security. For SaaS and business companies, data is business know-how; once AI Labs obtain that data, they may acquire the business capability through post-training.
• Palantir, Cohere and Base are also gradually switching to open-source models. The Digger expects most software companies to become increasingly wary of this data risk. Once OpenAI and Anthropic hit a ceiling in coding, they will inevitably pursue ARR and compete for the business of SaaS companies.

10. OpenAI’s Announcement and What Would Falsify the ARR Case

• Mr. Z relayed OpenAI’s announcement that it had paused reinforcement-learning training for 2 weeks on the latest model scheduled for deployment, citing the model’s achievement of a critical cybersecurity capability threshold defined by OpenAI. He was unsure whether this reflected a genuine security reason or factors such as insufficient GPU capacity. The Digger said it sounded “pretty nonsensical,” while making clear that he did not know the facts.
• The Digger wants AI Labs to answer several questions: Can ARR continue to grow? What exactly is the market? How will they compete? And how will they preserve pricing? “You can’t just tell me there’s no problem with competition and that the market is still huge.”
• Even if Token volumes grow exponentially, that does not mean ARR—or Anthropic’s or OpenAI’s ARR—will grow exponentially at the same rate. Jevons Paradox may hold for the industry as a whole, but it does not necessarily hold for any specific company.

11. Rates: Corporate Debt Crowds Out Government Debt; “Collapse, Recession, Then Cuts”

• The Digger believes high 10-year and 30-year Treasury yields partly reflect strong AI demand, but also reflect corporate debt “crowding out” government debt. Several major CSPs are absorbing market capital, making the price of money more expensive.
• Mr. Z said that if AI growth is slower than expected, the rate pressure becomes even more awkward. He cited the Druckenmiller fund’s 13F positions in rate-sensitive sectors such as housing and autos, and called Kevin Warsh the current Fed chair. Because Warsh previously worked at Morgan Stanley, Mr. Z inferred that Morgan Stanley might know something others do not. His expectation is that rates will neither continue rising sharply nor fall aggressively.
• The Digger says he is not particularly strong on macro, but sees little chance of continued rate hikes in a high-rate environment. Another possible path is that some link in the chain collapses first under high rates, triggering a recession and then rate cuts; assets that had been priced at elevated levels would return to fair value before gradually recovering.

12. First-Hand Bubble Signals: Deifying Experts, Redundant Agent Teams and “Believing Only What You See”

• The Digger believes that outside a small group of top-tier people, most model-training work is now easier than before: AI helps with data cleaning and model architecture. Yet the industry still treats these people as if they were gods.
• He currently works on LLM post-training, previously worked on an enterprise system analytics engine, and earlier worked in search advertising; he studied computer vision. Seeing the industry and financial markets from both angles gives him a different perspective on what is happening.
• Agent systems are another example of the bubble. Most companies are building them, many teams are doing essentially the same thing, and results still have not arrived. Once enterprises begin auditing ROI, redundant teams will contract quickly, and excess compute may ultimately become redundant capex.
• The cycle usually starts with industry insiders being pessimistic. By the middle and late stages, almost every hardware company tells you it is “doing extremely well.” Industry people believe only what they can see, but when these developments are everywhere around you, the cycle is often already near its top. Once capital floods into the industry, overcapacity follows.
• Mr. Z added that many Taiwanese semiconductor companies cannot use these LLMs directly because privacy and confidentiality are obstacles. The Digger believes enterprise know-how resides in workflows, databases and modules; expecting Codex or Claude Code to directly absorb an entire industry is essentially a SaaS model that requires industry-by-industry penetration and will face resistance from incumbents.

13. Private-Market Endgame: 2 Fundraises in 1 Month; “The IPO Itself Is the Top”

• Mr. Z cited Jane Street’s lead investment in AI inference-chip startup Etched: $700M at a $21B valuation. Just 1 month earlier, Etched had completed a $300M Series C at a valuation above $10B, with Sequoia, a16z and others participating. “They did 2 financings within 1 month.” He sees this as a warning that the AI frenzy may be fading and private companies are rushing to raise money in the private market.
• Mr. Z believes OpenAI and Anthropic may both be under pressure, and speculated that if Anthropic IPOs in October, the IPO itself could be the top. The Digger was more cautious: it could at least be a short-term top, while the longer-term upside remains uncertain.
• The Digger believes Jensen and Wall Street capital are tied to the same boat, so a successful IPO should not be difficult. But once capital comes in and investors cash out, the structure could also come to an end.
• Mr. Z said Jensen put everyone on the same boat in 2022 and 2023, and now everyone is trapped. The Digger added that long-term comprehensive intelligence built on data centers is not the problem; the questions are whether GPU prices are too high and whether it is reasonable for Jensen to capture the profits.

14. Playbook: Stay Flat and Wait; 2027 Could Be a Highly Damaging Year

• The Digger’s advice to retail investors is to stay flat, wait for these trades and structures to blow up one day, then pick up assets after prices fall. If AI is viewed as infrastructure construction, investors could later consider companies the government would certainly rescue, such as Intel.
• Mr. Z said institutions with existing positions should rebalance. Retail investors who remain extremely bullish could use a small amount of capital to buy puts as a hedge. Both stressed that investors should not trade randomly.
• On timing and magnitude, the Digger believes 2026’s ferocious bull market is over. If no new application scenario emerges, 2027 could be a highly damaging year, with a very large index decline. The semiconductor index has deviated substantially from its 5-year moving average; even if it closes this year around current levels, it could retest that average next year, implying a 40% to 50% decline. Taiwan passive-component maker Yageo is already down roughly 40% to 50%.
• There “could still be one more small rally” before elections in the U.S. and Taiwan, but he believes the cycle is in its final stretch. Outside U.S. equities, global markets have already fallen sharply; U.S. liquidity is the best, and Bessent and Trump have continued to support the market.
• Mr. Z is most negative on high leverage and Neocloud. Companies such as CoreWeave are fundamentally risk-transfer vehicles for major CSPs such as Microsoft: they generate no cash flow, have low profits and face high borrowing costs. Storage companies have ample cash flow and are unlikely to fail, but even after substantial declines their P/B ratios remain high relative to a normal cycle. He is highly cautious on SOX and sees software, cloud and large CSPs as relatively safe; sectors with sufficiently low expectations that are already trading at low levels can be monitored. Mr. Z’s overall view is that U.S. equities will not fall much, but semiconductors could fall a lot.

15. Aside: A Look at Crypto; In the AI Era, Your Attention Has to Be Your Own

• Mr. Z observed that if semiconductor capital takes profits or stops out, it may turn to Bitcoin, while stressing that this is not investment advice. When AI and semiconductors were hottest in February, he believed liquidity had already begun to deteriorate. During the correction from late June and mid-July through today, Bitcoin fell sharply but continued to attract buyers absorbing supply.
• Mr. Z reviewed the performance of MSTR, Coinbase and Solana-related DATs. He described Forward Industries as the largest Solana DAT—or the second-largest DAT among Solana-related companies—and noted that some DATs still rose modestly even as most assets fell.
• The Digger says he does not understand Crypto, but after looking at the performance of related companies, he believes there are indeed signs of something developing. He also sees the Crypto industry separating the wheat from the chaff, with few good projects in the Crypto primary market. Crypto-focused VCs such as Paradigm have also started setting up funds to invest in AI.
• The Digger cited Etched completing 2 financings in 1 month as a possible warning that the AI frenzy is fading. Mr. Z separately warned that Bitcoin could continue falling to $50,000, while emphasizing that this is his view, not investment advice.
• The methodological conclusion: AI should be treated only as a tool; your attention has to remain your own. Form an initial view first, then use AI to organize the evidence and test whether the view is solid. Without your own thesis, asking AI questions continuously will not help you observe the industry properly.
• Mr. Z agreed that tools are 100x more capable in the AI era, but judgment and core logic will become even more differentiated. Choice matters more than effort—and choice is actually harder.