Summary

AI does have a bubble, but it is concentrated in prices and expectations, not in the essence of technological progress. 莫傑麟 defines a bubble as “expectations higher than reality” and stresses that having a bubble “is not necessarily a bad thing, nor is it necessarily about to burst”; he believes AI is already extremely intelligent and often surpasses humans. Primary-market valuations and public-market capitalizations need to be assessed by market and sector.
The common denominator in this debate is not whether models work, but whether model companies’ ROI can cover their rapidly expanding investment. Meta is poaching talent; OpenAI, xAI and others are planning data centers exceeding 20-30 GW; Nvidia is investing in model companies, rapidly lifting investment while perceived returns are flattening. 曲凯 summarized the shift: “We didn’t look at ROI before; now we’re looking at ROI.”
The recent pullback in U.S. equities and Nvidia cannot be attributed to an AI bubble alone; macro conditions, liquidity and risk appetite matter just as much. When DeepSeek emerged in January, Nvidia fell below $90, but the current decline has not reached the same magnitude. 莫傑麟 therefore argues that “if the market fully believed the bubble story, it should have fallen more”; for now, multiple factors appear to be eroding the optimism that prevailed earlier.
The real regime change is from waiting for the next AGI-style breakthrough to demanding industrialization, cost efficiencies and delivered applications. People no longer broadly expect the next generation of models to produce a decisive aha moment. Model companies are competing on applications and revenue, while application companies are building models in return; AI is moving from the scaling-law “dream multiple” to an actual P/E, but still lacks a new “faith recharge.”
The two disagree sharply on whether the scaling law has failed, but agree that new investment must be justified by capability or returns. 莫傑麟 sees an effective scaling law as the prerequisite for larger investment, comparing the situation to a chain restaurant expanding only to be told it must “start by making the concrete.” 曲凯 rejects the idea that the scaling law has failed, arguing that it has merely become “impossible to evaluate.” Many founders believe the current bottlenecks lie more in cost, infra, context and the agent layer.
The bubble is highly structured: China’s primary market is relatively healthy, the U.S. primary market is more frothy, and public markets have yet to complete the pricing shift from hardware to software. Comparable companies in the U.S. can command valuations at least 10x those in China; Cursor can still reach a $10B valuation despite persistent losses. In public markets, data-center names such as Oracle have fallen further, while Nvidia has held up relatively well because “the factual result is that Nvidia is still the most profitable.”
For investors, more important than deciding whether there is a bubble is identifying the new winners and losers after the cycle turns. In 2023, the two worried that Chinese models were falling behind; two years later, 曲凯 says open-source models are now entirely Chinese. Cambricon also shows how sharply industry views can reverse within a year. “There is no ceiling on the winner’s upside,” while To C, cloud and foundation models could all reinforce concentration at the top.
For founders, the right response to a bubble is not to stop, but to manage the cadence of fundraising, valuation and cash flow. Bubbles provide growth capital but can turn large numbers of participants into “fuel” when they burst. The PayPal experience through the dot-com bubble points to the same bottom line: “you need to generate your own cash.” Meanwhile, Nvidia has announced a $100B investment in OpenAI, which has also signed a procurement agreement with AMD and a new agreement with Broadcom, suggesting that the upstream compute landscape is being rearranged ahead of financial results.

Deep dive

1. Separate AI’s value from its price

莫傑麟 opens with a direct “yes,” then immediately dismantles the assumption that a bubble must inevitably collapse: if a bubble is “an expectation higher than reality,” rapidly developing industries almost inevitably come with one. A bubble can bring capital and room to grow; “having a bubble is not necessarily a bad thing, nor is it necessarily about to burst.”
莫傑麟 separates the question into value and price. People actually building AI rarely discuss bubbles and may even find the question laughable, because AI’s intelligence is already very strong—“it surpasses humans in many cases.” There is therefore no bubble in the essence of AI’s development; primary-market valuations and public-market capitalizations are a separate matter.
曲凯 notes that when people discussed a bubble last year, the main evidence was sustained growth in tech companies’ revenue and EPS, with forward P/E at roughly 30x. On P/E alone, some of the prices were defensible. 莫傑麟 adds that expectations are not fully captured by valuation multiples: two companies may buy the same Nvidia cards, but one may be betting on AGI while the other is betting on cost efficiencies, producing very different expectations for data-center ROI.

2. The latest bubble narrative was triggered by an investment surge

Earlier doubts tended to focus on a single point: during DeepSeek, the debate was about building a model “with very little money”; in the previous round, it centered on GPT-5 not being trained and hardware constraints preventing training from finishing. This time, the arguments span raw intelligence, RL, post-training, applications, business models, inference costs and Chinese open-source models, ultimately converging on one question: ROI.
莫傑麟 traces the trigger sequence from Meta’s talent poaching, to OpenAI, xAI and others planning more than 20-30 GW of data centers, and then to Nvidia investing in model companies. Capital behavior has continuously expanded investment, while perceived returns have at least failed to strengthen in tandem.
曲凯 asks why these issues, which had been raised for at least a year, have suddenly converged. 莫傑麟’s answer is that the magnitude has changed: the amount U.S. tech companies are investing in AI is now larger than the combined amount invested over the preceding several years, and it has coincided with macro volatility. 曲凯 sums it up: “We didn’t look at ROI before; now we’re looking at ROI.”

3. A stock-market pullback is not standalone evidence of a bubble bursting

On the causes of the U.S. equity decline, the two cite a pullback after the earlier rally, rate-cut expectations, a U.S. government shutdown, geopolitics, liquidity and AI valuations. 莫傑麟 refuses to force a ranking, but is certain that “it is not falling solely because of this AI bubble”; the common result is a broader decline in risk appetite.
When DeepSeek emerged in January, Nvidia fell below $90; the current decline has not reached that level. 莫傑麟 reasons from this that “if the market fully believed the bubble story, it should have fallen more.” For now, the market appears to be taking back part of its earlier optimism rather than fully pricing in a bubble.
Industry sentiment has not broadly turned cold. The decline in risk appetite is structural and limited, while many people around them are still buying Google stock. At the time of recording, Gemini 3 was due the next day and Nvidia’s earnings the day after; expectations for Gemini 3 were high, while Nvidia’s earnings expectations were relatively controllable but extremely important. The real risk is that market “sentiment is somewhat fragile” and prone to using events to find direction.
曲凯 cites an intriguing market consensus: AI bulls want the market to fall first so they can buy the dip, while bears believe “it should have fallen anyway.” Both sides may therefore want a short-term pullback, meaning price action alone cannot prove which long-term narrative is correct.

4. The AGI faith cycle is giving way to an execution cycle

Over the past two or three years, the industry tracked how the next model would solve math and coding, what data and how many cards it would use, and which people it would poach, because all of these pointed to what stronger models might unlock. 曲凯’s instinct has changed: very few people now expect the next GPT to deliver another stunning leap.
Starting with o1, the narrative gradually shifted from pre-training toward post-training and RL; DeepSeek then pushed reasoning models to a peak of attention. 曲凯 asks whether RL now appears to be running out of steam, and whether DeepSeek’s failure so far to deliver the new version the market expected is consuming the old expectations.
莫傑麟 defines this as a cycle change, not merely a bubble. Standardized benchmarks can no longer reliably measure model quality, and founders around them generally believe capability is “already OK.” Over the next three years, pre-training may no longer be the sole focus; attention will shift to extracting value from existing models, vertical To B, business models, and cost efficiencies.
曲凯 infers management teams’ views from their actions: model companies are visibly competing on applications, engineering and revenue, while application companies are building models in return, and Cursor has begun launching its own models. If Sam Altman still saw a high-certainty path to a pure intelligence leap, he might not be directing so much effort toward the application layer. 曲凯 stresses, however, that outsiders can only infer from actions and cannot access the internal information.

5. Is the scaling law broken, or simply unmeasurable?

莫傑麟 first argues that larger investment only makes sense if the scaling law works. If a 10x investment could produce AGI, the market might accept “putting an entire country into it.” But if investors provide the money to expand a restaurant chain and are ultimately told it must first “start by making the concrete,” they will naturally ask where the value is.
曲凯 disagrees that the scaling law has failed, arguing that his research suggests it has merely become “impossible to evaluate.” Models already surpass humans in many settings; the remaining problems center on RL, generalization and the final step of capability. The absence of a unified benchmark is not enough to declare the scaling law dead.
Many founders say that now “it’s not a model problem,” but a problem of cost, infra, context and the agent layer. 曲凯 compares the phase change to moving from having no infrastructure at all to having the infrastructure in place and needing to furnish the interior. Industry participants are therefore more likely to see this as the next phase of the sector than as deterioration.

6. China and the U.S., primary and secondary markets, are at different bubble levels

莫傑麟 sees no obvious bubble in China’s primary market and considers it relatively healthy overall. Comparable companies in the U.S. can readily raise valuations more than 10x those in China; leading Chinese projects may be valued at only one-tenth or less of their U.S. peers. The U.S. primary market is frothy: some companies’ data are not solid, yet Cursor has reached a $10B valuation while continuing to lose money.
莫傑麟 is more cautious on China’s public market, because many names are extensions of the U.S. public market and tied to Nvidia’s supply chain. But optical modules and Cambricon, among others, at least have earnings support and cannot be treated the same way as companies based purely on expectations.
The U.S. public market response is more anomalous. If the training cycle is stabilizing, some companies clearly have bubbles, yet Nvidia has not fallen sharply this time; Oracle and other OpenAI data-center-related names have fallen more instead. The market still appears to believe that building data centers requires buying Nvidia cards, making Nvidia a beneficiary.
The market has long expected a shift “from hardware to software,” but the migration has not been completed at the price level. The fact remains that Nvidia is the most profitable company; software has yet to produce success and profitability on the same scale. History has also repeatedly shown that shorting or underestimating Nvidia was wrong, so investors’ expectations have not fully converted into prices.

7. The bubble measures the water temperature; the cycle is the real question

曲凯 mentions a podcast by an a16z partner; 莫傑麟 adds that the person experienced the 2000 dot-com bubble. At the time, almost any company could list and then surge, revenue was not a prerequisite, and even taxi drivers discussed stocks. Yet at that point, relatively few people publicly said, “This is a bubble.”
曲凯 compares this with China’s A-share market around last year’s National Day holiday. During the rally, you could hear security guards downstairs discussing stocks, everyone was opening brokerage accounts, and the market fell immediately after the holiday. AI has not yet reached that kind of mass participation; if everyone knows there is a bubble, prices should theoretically reflect it.
AI’s rhythm looks more “pulse-like” or “stepwise.” It heated up in the first half of 2023 on models such as GPT-3.5; Chinese primary-market investment nearly froze from the second half of 2023 through the first half of 2024; then o1, DeepSeek and coding capabilities reignited the boom from the second half of 2024 through the first half of 2025.
What is missing now is a larger “faith recharge.” At the time of recording, the market expected Gemini 3 to bring multimodality, coding and a strong product combination, but it was not broadly betting on a quantum leap toward AGI. If another DeepSeek-style moment appears, a new concept could still take over and restart the short cycle.

8. A cycle shift will reshuffle winners rather than cool everything evenly

曲凯 believes AI’s To C businesses naturally exhibit strong winner-take-most effects. 莫傑麟 adds that the same applies if models are defined as To C products. Cloud computing has high fixed-cost barriers, so there are usually only two or three major players. Foundation models are not only expensive; they also depend on whether the leader truly understands AI, potentially producing an even stronger winner effect than mobile internet.
In 2023, the two still worried that Chinese models were falling behind and that China lacked enough cards. Two years later, 曲凯 says, “open-source models are all Chinese models now.” 莫傑麟 particularly warns that Silicon Valley discussions of Chinese models often omit ByteDance, whose talent reserves may be an important input when assessing its long-term position.
Cambricon offers a case study in rapidly changing sentiment. In 2024, many industry experts were still voicing concentrated doubts. According to the program, its first earnings report after its inference chip appeared in October 2024; by the time of recording, it was November 2025—only roughly a year later.
In 2025, the company attracting concentrated criticism was instead OpenAI, partly because of Sam Altman’s series of aggressive moves. 莫傑麟 preserves the reverse possibility: even if OpenAI is now different from its original form when Ilya was still there, that may not prevent it from becoming a powerful commercial company. Short-term sentiment and long-term conclusions may point in opposite directions.

9. Public markets are replacing long-term absolute views with sentiment and expectations

In preparing for the discussion, the two read a large number of bubble articles. 曲凯 notes that some articles even misidentified the years corresponding to data-center investment, then used those figures to conclude that OpenAI’s cash flow was about to collapse. 莫傑麟 adds that public markets may not scrutinize these underlying calculations now; the quality of the inputs behind emotional conclusions may be completely inadequate.
Public markets have shifted sharply from value investing toward financial engineering. 莫傑麟 believes professional investors care about recent sentiment, narratives, alpha and expectations that will affect buying over the next three months, and are unwilling to hold excessively long-term, absolute views.
曲凯 says he once told a professional public-market investor with a strong institutional track record that everyone around him was buying Google. 莫傑麟 adds that the investor showed no reaction because they had not researched Google.
莫傑麟 divides information into four layers: sentiment, expectations, trends, and “this thing is absolutely right.” Public-market investors are usually more sensitive to the first two, while founders are more sensitive to the latter two. On this bubble debate, 曲凯 believes the most that can be said is that “expectations are not low, while sentiment is deeply depressed,” which could mean the decline is not over—or that sentiment is so weak that a rebound is possible.

10. The market is shifting from dream multiples to P/E

莫傑麟 describes the current position as the intersection of two cycles and two standards. People inside the industry believe applications will inevitably land and the future will be better, but the market has not yet seen sufficiently clear delivery. Valuations are moving from the AGI and scaling-law “dream multiple” toward an actual P/E constrained by industrialization results.
A bubble can give startups more capital to grow, but when it bursts it turns many participants into “fuel.” Great companies also often emerge from the same bubble. The key is therefore not accurately calling the top, but staying clear about which phase one is in, what one is doing, and what one wants to exchange for.
莫傑麟 believes founders must also become good investors. Beyond management, technology and product, the cadence of fundraising and valuation are both determinants of company success. “Everyone knows it’s a bubble—should I stop? I definitely don’t think so,” but whether and when to raise can determine whether a company makes it through the cycle.
曲凯 uses PayPal’s survival through the dot-com bubble to extract the bottom line: “you need to generate your own cash,” and maintain healthy cash flow. Without cash flow, a company dies as soon as the bubble bursts, leaving no chance to survive into the next phase or wait for faith to return.

11. Interests and structure are being reshaped upstream in compute

Looking back at the booms in SaaS, new consumer brands, education, gaming and offline chains, 莫傑麟 believes the channel usually profits first. In education, gaming and new consumer, the channel captured the money; in chains, shopping malls and other location channels benefited. Further up the chain, that money came from investors and ultimately LPs.
莫傑麟 believes the more important event in the third quarter of 2025 was Nvidia announcing a $100B investment in OpenAI. In the same week, OpenAI announced a procurement agreement with AMD, followed two days later by a new agreement with Broadcom. These arrangements reflect Sam’s investment and ROI mindset as an entrepreneur “with an investor’s background.”
Google’s TPU offers another compute route worth watching, and TPU’s commercial value to Google deserves a separate discussion. The arrangements involving Nvidia, AMD and Broadcom all point toward semiconductors; meanwhile, Oracle and other OpenAI data-center partners came under greater pressure as the market worried about data-center profits, construction feasibility and power shortages.
莫傑麟 believes the upstream compute structure has already shifted at the level of logic and expectations, even though the change has not yet been fully reflected in financial results. OpenAI’s data centers and related issues require further unpacking. The bubble debate measures the emotional temperature; the more important task is identifying the structural changes and the winners and losers they will create. Future discussions will also cover pre-training, post-training, video-model cards, token optimization and PMF.

AI Update: NVIDIA’s Record Revenue, Elon’s Data Centers in Space & Gemini 3’s Insane Performance Epstein Files Fallout, Nvidia Risks, Burry's Bad Bet, Google's Breakthrough, Tether's Boom