Silicon Valley’s Great AI Turn and the Public-Market Bull Market | A Conversation with 莫傑麟

Summary

• 莫傑麟’s view is that the central axis of US AI in 2025 has shifted from Scaling Law to Token consumption: July grew more than 20% from June, resembling the rapid increase in usage time during the mobile-internet boom. The incremental demand is coming first from the labs’ own products—OpenAI, Anthropic, xAI and others—not APIs; ChatGPT’s daily active users and usage time are also rising. AI is moving from an “alien species” that must constantly prove AGI into Search, office software and enterprise Workflow, and the industry’s overarching sensation is that “everything keeps speeding up.”

• Even if existing models stop getting smarter for now, they are already capable of unlocking far more demand than most people expect. Enterprises want productivity gains, partial labor substitution or greater reliability; consumers want simpler ways to search and work; applications and Infra make existing intelligence cheaper, faster and more reliable. GPT-5’s significance is therefore not a leap in IQ, but hiding model selection, integrating the Stack and interface, and shifting competition from Benchmarks toward Full Stack industrialization and efficiency.

• The boundaries between models, applications and Infra are blurring, and the new competitive rule is that companies “cannot have a weak link.” OpenAI and Gemini are accelerating product development, while application companies such as Cursor and Manus are training models and building Sandbox, Browser and Agentic layers; Anthropic’s former Coding focus has also been diluted by Gemini and GPT-5’s catch-up. 曲凯 cautions that PMF remains scarce and model companies will readily absorb validated use cases; 莫傑麟 believes end-to-end integration is an inevitable organizational ambition, while Context is more likely to become the part that applications truly retain.

• AI Infra investment logic still builds upward from the most foundational GPU layer, but falling Token costs will also create demand for ASICs, AMD, inference providers, and China’s optical-module and PCB supply chains. Nvidia’s immediate upside driver is that Token consumption and Inference have not declined, while its software, supply chain and “always-on” support make customers willing to buy the full stack from one vendor. The logic can hold over the long term, but hardware is not software: “The market expects demand from you at this level, but you cannot deliver,” so individual stocks ultimately depend on delivery.

• The US AI bull market has no clear short-term break point yet, but the core test has shifted from Scaling Law to whether training demand and inference demand can hand off positively to each other. Nvidia shares fell briefly when DeepSeek emerged, but subsequent RL progress, Meta’s team expansion and Token growth pushed both demand pools higher, taking Nvidia from roughly $3T toward $4T and closer to $5T. The next questions are whether Meta’s new team can deliver a model in 6—12 months, and whether “vibe revenue” can harden into stable application consumption with demonstrable ROI.

• America’s AI “bubble” is not a scam, but a state in which real productivity, resource allocation and high valuations coexist. 莫傑麟 roughly estimates that AI or technological development may account for two-thirds or more of US economic momentum; as long as industrialization keeps accelerating, the broad direction remains upward, but replacing people will not happen overnight, and regulation, enterprise decision-making, Infra costs and even a 3—6 month stagnation could trigger major volatility. China is driven by more variables—debt resolution, household deposits, risk appetite and the manufacturing cycle—so the surface narrative may be “East up, West down, then East up and West up,” but the underlying dynamics are different.

• Whether A-shares can form a durable bull market depends not only on the industrial narrative, but on whether quality companies, risk appetite and the wealth effect for retail investors can form a closed loop. Cases including DeepSeek, Miners, JazzMark, Pop Mart and Sino Biopharmaceutical are turning the previously “not lying flat” positive expectations into actual delivery; Tencent has also become a listed-market AI proxy through its investment, hiring and strategic moves. 莫傑麟 places particular emphasis on the new generation of retail investors’ information tools and Conviction: with high-turnover institutions trading against one another and long-term capital trading very little, individual investors who hold long-term views on Tesla, Palantir and others may become an important marginal force in trend markets.

Deep dive

1. Token Consumption Replaces Benchmarks as the Industry Thermometer for 2025

• 莫傑麟’s overall judgment is that the industry completed a “great turn” after the pessimism of “pre-training hitting a wall” and re-entered an acceleration phase in 2025; the central evidence is not the leaderboard but Token consumption.
• The data that stayed with him most was July growing more than 20% from June, with the pace continuing to accelerate quarter over quarter; the curve is “very much like the fastest-growth stretch of the mobile-internet era.”
• Pressed by 曲凯, he made clear that the main incremental demand is first coming from OpenAI, Anthropic and xAI’s own products, not APIs; ChatGPT in the US, 豆包 in China and DeepSeek are all expanding daily active users and usage time.

2. Even If Models Pause, Existing Intelligence Can Release Massive Demand

• Users once treated AI as an “alien species,” waiting for it to announce AGI and prove that it had surpassed ordinary people; now it is quietly embedding itself in Search, office software and enterprise Workflow, with usage time rising alongside adoption.
• 莫傑麟’s biggest update in thinking is that even if models remain at their current level, the scenarios they can unlock are already extensive; demand is “actually much stronger than people think.”
• Enterprises want productivity gains, partial labor substitution or greater stability, while consumers want simpler ways to complete tasks; application and Infra companies are responsible for turning those needs into actual demand.

3. GPT-5’s Significance Is Not Showmanship but Making Intelligence Full Stack

• GPT-5 did not primarily prove that it was dramatically smarter than other models; it hid the complexity of model selection and integrated the Stack, Infra and front-end interface into a unified product.
• The stronger developer feedback was that it was “very useful”; commercially, it also distinguishes “thinking without thinking,” while pricing RL Tokens and Coding Tokens separately and assigning them to different teams for sales and execution.
• 曲凯 asked whether this had moved into the comfort zone of Chinese entrepreneurs, and 莫傑麟 broadly agreed: apart from Thinking Machines Lab, whose Demo is still due in the third quarter, the market has accepted the current architecture, with intelligence continuing to be “squeezed out” of every layer.

4. Agents Are Becoming a New Interface Like Apps, with Context Determining the Closed Loop

• 莫傑麟’s analogy is: “In the past, every company and every product needed an App; now every use case or product needs an Agent.”
• Agents have entered the industrialization phase of the application ecosystem, with the bottlenecks shifting to stability, speed and cost; Manus burning through its Credits quickly is a vivid example of Tokens not yet being finely optimized.
• 曲凯 added the capability progression: CAI-style companion products were built on GPT-3.5’s chat capability, Agents on o1’s reasoning capability, and the next wave of “wrapper” opportunities may come from multimodality.

5. The Model-Application Boundary Is Blurring; “No Weak Links” Is the New Rule

• 曲凯’s objection is worth preserving: AI Coding has become as competitive as China’s early internet, while genuinely PMF-backed scenarios remain scarce; once a model company sees a wrapper product validate demand, it may simply build the feature itself.
• 莫傑麟 believes new organizations are actively combining technology and product: OpenAI has recruited multiple startup Founders, while Manus, Jasper and others are building Sandbox, Tool Use, Browser and Agentic Infra.
• Application companies such as Cursor are beginning to train models, while model labs are developing their own chips and expanding products; all are testing whether they can bypass the division of labor beyond GPUs and build end-to-end systems.
• Anthropic won favor through its focus on text Scaling Law and Coding, but Gemini and GPT-5 have rapidly closed the Coding gap; a key advantage in this era may simply be that a company “cannot have a weak link.”

6. Vertical PMF Lives in Workflows That Cannot Be Validated Purely Through RL

• Coding and mathematics can produce precise Rewards, but many industry tasks cannot rely solely on I/O Validation; people, compliance and specialized language environments must remain in the workflow.
• 莫傑麟 still cites Harvey as the vertical application company with the highest ARR; several healthcare startups are also growing ARR rapidly, and healthcare institutions interviewed in the field have rated their problem-solving ability highly.
• Finance and insurance likewise cannot be won through technical Demos alone; they require Documentation, compliance and industry Workflows to be embedded into Context. Chinese entrepreneurs, meanwhile, have been among the first to make consumers feel intelligence through products such as Mantis, Jasper and Fellow.

7. o1’s Underestimated Upgrade Was “Understanding You Better”; GPT-5 Rewrites Evaluation

• o1 received a limited response at launch, but in retrospect its longer IO time made the model better at understanding users and seem to have “enough emotional intelligence,” while lowering the difficulty of Prompting.
• OpenAI and Google models winning gold and silver medals in the MO also shows that traditional tools have become poor at distinguishing model quality; 莫傑麟 admits that in many scenarios he can no longer tell whether a model has actually improved.
• In the industrialization phase, the more important metric is cost at equivalent performance. GPT-5 therefore opened a new stage of efficiency competition after a major reset in AGI expectations.

8. AI Infra Has at Least Four Layers, with the Upper Layers Closer to User Value

• The bottom layer is card-utilization efficiency between models and GPUs; DeepSeek accumulated substantial engineering experience through quantization and can make good use of previous-generation cards, while 梁云峰 is viewed by both speakers as a strong Infra-style Founder.
• The second layer sits between models and applications: Together AI, Fireworks and others optimize inference and service stability for customers while serving as Backup Vendors in multi-model deployments.
• The third layer is Agentic Infra: models do not merely answer questions but book tickets, operate Browsers and execute real tasks; E2B and Browserbase are examples cited by 曲凯.
• The fourth layer is Context. The Context Engineering discussed by Minus CTO Peak asks what materials allow a model to understand the user and how those materials can be turned into action.

9. The Middle Layer May Become a Commodity, While Context Is More Likely to Be an Application Company’s Core Asset

• After Anthropic stopped providing services for certain models, application companies realized they could not rely on a single lab and began seeking Backup Vendors such as Together AI and Fireworks.
• The debate is whether inference optimization is merely a Commodity: much of the technology is open source, Nvidia wants to reduce the friction of Agent deployment, and OpenAI and Anthropic have incentives to bring the capability back in-house; the question is whether third parties can sustain their gross margins.
• On Context, 曲凯 believes it will be difficult to create a unified third-party Infra layer; 莫傑麟 agrees, because much of the content cannot be validated through RL and may ultimately become the core through which application companies understand users and industries.

10. US Enterprise Stickiness Supports To B, While Chinese Teams Are Pushed Toward To C

• 莫傑麟 puts it bluntly: US Enterprise is “not really a VC business model”; it lacks disruptive substitution, but benefits from a stable business environment and relatively low cancellation rates.
• The implicit premise of the industry is that existing people or vendors are difficult to replace, while new people or companies also struggle to produce disruptive innovation.
• 曲凯’s practical constraint is that China has few successful To B cases, while Chinese teams face even greater difficulty selling to enterprises in the US, so many Chinese Founders are almost “forced” into overseas To Consumer markets.

11. The Underlying Explanation for Nvidia’s Continued Rise Is Still Token Consumption

• As the inference technology paradigm develops, inference consumption has not declined; overall Token consumption, including Inference, is rising, so Nvidia continues to benefit.
• The rise in AMD and ASICs shows that the market is already betting on efficiency: enterprises want to use proprietary chips or alternative cards for inference, adding competition for Nvidia.
• Nvidia’s software, supply chain and after-sales support create powerful stickiness; when Silicon Valley customers encounter execution problems, Nvidia is almost “always on call,” which is why AMD has not yet converted its cost advantage into sufficient share.

12. The Cost-Reduction Thesis Can Hold for Years, but Hardware Stocks Ultimately Need Delivery

• China has long faced restrictions including H-card and the current B-card export ban, and has tried using more alternative cards for inference; Cambricon, 新易盛, 中际旭创 and PCB manufacturers have also become public-market proxies for the thesis.
• 莫傑麟 emphasizes that hardware cannot keep getting faster the way software can: “The market expects demand from you at this level, but you cannot deliver.” A durable thesis does not mean every company can fulfill it.
• 阿瓦国 has earned strong recognition for its ability to deliver to major customers; the longer-term question is whether inference, training and Agent Infra will independently produce a “new Snowflake” or be absorbed into Palantir-style platforms.

13. FOMO Has Shifted from Buying Chips to Hiring People, with Meta the Most Visible Example

• Google’s technical reserves span hardware, text models, VIVO Three, video and robotics; what remains to be tested is not capability but whether a very large organization can change how it collaborates and deliver what it has accumulated.
• Meta’s new team has produced two competing interpretations: whether a temporary assembly of stars lacks a shared Vision, or whether people from different labs with frontline engineering and launch experience are enough to compete on the next generation of models.
• In 2023, FOMO meant “you must have a model, buy cards and pile on compute”; in 2025, hiring and acquisitions by Meta, Microsoft and ServiceNow show that FOMO has spread to Post-training, products and core talent.
• 莫傑麟’s explanation is that application companies have felt firsthand that AI is moving closer to commercial deployment: “It is not just some AGI that remains out of reach.”

14. America’s “AI Bubble” Has Productivity Behind It, but Valuations Will Still Whipsaw

• 莫傑麟 maintains his earlier view: AI is the “only bubble” among all assets, but “bubble is in quotation marks”; it is not a scam, but a state in which an important technology receives high valuations, incentives and resources.
• He roughly estimates that AI or technological development may contribute two-thirds or more of US economic momentum; the Chinese market still depends on multiple variables, including debt resolution, household deposits and risk appetite, and the manufacturing cycle.
• A 3—6 month stagnation need not break the long-term trend; the real friction comes from regulation, enterprise decision-making, technical stability and Infra costs, alongside market volatility caused by Trump’s interventions in technology companies and geopolitics.

15. Nvidia’s Key Test Is Whether Training and Inference Can Hand Off Positively

• Nvidia’s demand divides into training and inference: Pre-training is more like cyclical capital expenditure, while Inference depends on whether real users keep consuming Tokens over the long term.
• The bearish case is that training peaks while application demand created by short-term enthusiasm fails to persist; the bullish case is that inference becomes large enough for Nvidia to retain share through its architecture and services.
• When DeepSeek emerged, the stock first plunged; subsequent RL progress and Meta’s hiring drove training demand, while applications drove inference, turning the expected negative Trade-off into simultaneous growth on both sides.
• This explains Nvidia’s move from roughly $3T toward $4T and closer to $5T; the next validation points are whether Meta’s team can deliver within 6—12 months and whether application consumption persists.

16. “Vibe Revenue” Must Become Stable Token Consumption for the Application Bull Market to Get a Second Leg

• The US market calls some AI revenue “vibe revenue”: part of the usage comes from excessive expectations or the amplification of broader behavior, and users may not yet clearly know what they will ultimately use AI for.
• 曲凯 believes the To B math remains straightforward: if the original market is $100B and AI improves efficiency by just 5%, that represents $5B in value; enterprises’ desire to reduce headcount management will be even more durable.
• 曲凯 says Cursor’s annual ARR should be around $500M; Chinese AI companies have moved from a handful surpassing $10M last year to some challenging $50M or even $100M, though the overall scale remains small.

17. China’s Listed AI Proxies Remain Concentrated in Hardware and Large Platforms

• DeepSeek and ByteDance, the most native AI players, are not listed, so the market first traded Cambricon, Nvidia’s supply chain and optical modules; Alibaba also drew attention through model performance, especially in open-source models, before Tencent’s strategic investment received greater focus.
• Tencent has become the new core proxy because its level of investment and strategic commitment to domestic AI has risen markedly this year; the market will also watch whether it is making progress and can deliver.
• The investment process has moved earlier as well: during the mobile-internet era, investors waited for PMF, Market Share and Retention; now the market first tests whether key talent has real engineering experience and a track record, and whether management is pursuing the right path.

18. A-Shares Need Quality Supply, Risk Appetite and the Wealth Effect to Close the Bull-Market Loop

• Looking back to January 2024, many bearish expectations had already been priced into Chinese assets, but the Founders and company heads around 曲凯 and 莫傑麟 had not lain flat.
• Two years later, Chinese teams have begun proving that they can achieve global leadership in certain areas; in AI, the examples include DeepSeek, Miners and JazzMark, while Pop Mart and Sino Biopharmaceutical provide real-economy evidence.
• The first condition for a bull market is a group of quality companies that keep delivering results, something already visible among the Xiaomi cohort; the second is rising risk appetite, which comes from the positive feedback of “I made money in the market.”
• 莫傑麟 is therefore relatively optimistic on A-shares, but makes no definitive prediction: the core variable is how retail investors understand new technology assets and whether they can transfer lessons observed in US equities back into China.

19. The New Generation of Retail Investors May Have More Conviction Than High-Turnover Institutions

• 曲凯 offered the conventional counterargument: the US has a higher institutional share and more long-term capital, so in theory it should be more stable than retail-dominated A-shares. 莫傑麟 distinguishes between “money that is held” and “money that actually trades.”
• His rough breakdown is that more than half of US equity capital comes from quantitative and platform-style institutions, while long-term capital, though large, trades infrequently; high-frequency capital trades against itself and follows machines, making it difficult to create a durable bull market on its own.
• The remaining 20%—30% may be retail investors, who have stronger tools through Robinhood, Futu, professional media and KOLs, yet may hold more firmly to long-term views on Tesla, Palantir and others than institutions do.
• A-shares historically had short bull markets and long bear markets, making it difficult for individuals to build a stable feedback system; US equities, with long bulls and short bears, allow investors to turn their understanding of companies into strategies, profits and the next trade.

20. Software Candidates Have Emerged, but Each Company Is Betting on a Different Context

• Palantir first entered view through enterprise AI contracts, then saw its consulting implementation, organizational model and AI Stack increasingly emulated by enterprises; even with retail investors pushing it toward nearly 100x or even 200x PE, it still has fundamental support.
• Reddit’s value lies in high-quality community data that has not been over-commercialized, and ChatGPT uses its data heavily; 莫傑麟 believes Chinese investors can study it through experience with Zhihu and Xiaohongshu.
• ServiceNow began with difficult Workflow consulting and software delivery, gradually occupying management mindshare before expanding into data, CRM and other add-ons; it is a useful case for Chinese teams going to the US to pursue Enterprise.
• Figma’s moat lies not only in generative design but in the Context of cross-company collaboration; its roughly $40B market cap also provides recruiting resources. Duolingo, meanwhile, is again viewed as an AI Loser because ChatGPT’s overall usage and voice interface have strengthened while family-plan data underperformed expectations.

21. The Primary Market Rewards Only the “Special 1%” and Accelerates Polarization

• Thinking Machines Lab secured a valuation of more than $10B with an all-star team, while its product Demo is not due until the third quarter; 莫傑麟 sees it as evidence that the US market is becoming even more Founder-driven.
• His internal formulation is that “the 1% is not necessarily excellent; it may be special”: differences in training scale and model experience are creating a valuation chasm between the first and second tiers unlike anything seen in more than a decade in the industry.
• 曲凯 cautions that headlines about Meta “paying $100M to hire people” omit complex terms, deal assemblers and ancillary transactions; secondary stakes and SPV transactions may also contain layered structural risks.
• China’s primary market is still contracting and looks more like a “functional department”; AI enthusiasm may be 10x last year’s level, but transaction volume is nowhere near 10x, while financing for genuinely top-tier companies may be 10x larger.

22. One-Person Companies Are Not the End Point; Organizational Efficiency and Commercial Ability Are the Variables

• Chinese AI Founders increasingly believe that technology will continue advancing, but must still manage the coopetition with model companies: use models to gain capabilities without being eaten by them while standing on the main road of the model stack.
• Overseas payment capacity, willingness to pay and financing conditions are better, so going abroad remains the preferred option; 曲凯 nevertheless believes China is not impossible if Token costs eventually fall close to zero, reopening the path of large DAU, advertising and social products.
• On Vibe Coding and one-person companies, 莫傑麟 accepts that efficiency can be higher and fewer people may be needed, but rejects treating headcount as the objective: “If 2 people can do this better, why must it be 1?”
• US independent developers can earn several million dollars a year not only because Coding is cheap, but because the ecosystem is mature; typical Chinese independent developers often have strong engineering ability but lack commercial experience.

23. Multimodality May Be the Next Wave After Agents, While Investment Research Is Also Being Rewritten by AI

• Video Token growth has already surpassed text, but current AI video is concentrated mainly in special effects, editing and image-to-video; 曲凯 believes these are more like consumer AI in the multimodal phase than truly Native products.
• What they are waiting for is a Manus-like multimodal product after models such as VIVO Three continue to improve quality and reduce costs: it could be a content community, multimodal reasoning product or something closer to a world model, potentially starting a new cycle lasting 6 months or even longer.
• 莫傑麟 plans to further AI-enable investment research, a workflow built around productivity and Documentation, through an internal project tentatively called What If; it will first serve his own research, then organize researchers, engineers and investors around discussions with a cap of roughly 100 people.