中文精读
热度、交付和价格怎样错位：莫傑麟谈DeepSeek与Manus之后的预期
42章经单期精读 · 产品定位与资本叙事
嘉宾：莫傑麟，家族办公室资深从业者（当期介绍）｜主持：曲凯
发布：2025年3月22日｜时长：56分18秒｜整理：2026年10月18日
官方单集与节目介绍
“东升西落”先描述人怎样改变预期
曲凯到美国后，发现大家谈论中国与美国资产的语气迅速变化。他的疑问不只是看好哪边，而是为什么两年悲观可以突然逆转，新的乐观又能持续多久。
莫傑麟把这解释为预期、风险偏好与实际趋势的交错。DeepSeek和Manus成为当时的标志事件，但技术进步、应用可用、企业收入与证券价格不是同一种证据。本篇保留2025年3月的讨论，不把节目中的政策、行情或公司比较当作当前结论，也不提出交易选择。
1. 一个事件改变的是预期差，不一定是全部现实
线索：01:00–09:30。 莫傑麟认为，美国此前的高预期遇到宏观不确定性，中国较低的预期则遇到新的技术成果，资产反应因此可能相反。投资者担心衰退，与当时数据尚未明确指向衰退，可以同时存在；预期不是已经发生的经济结果。
他进一步把美国叙事概括为智能与扩展规模，中国叙事概括为应用和投入产出。DeepSeek以较低成本取得可比效果，在他看来挑战了“只有不断堆资源”的故事。这是其覆盖圈子的理解，不能给两国全部团队贴统一标签；具体性能、训练成本与比较方法未在节目中完整展示。
编辑推论：遇到突发热潮，先问原本预期是什么、事件提供了哪些新证据、哪些问题仍没回答。成本下降可能使某任务可用，也可能刺激更多调用；只按模型单价推算企业利润，仍缺用户量、延迟和总任务成本。
2. Manus被放到模型坐标上，应用价值就容易失焦
线索：10:15–21:12。 曲凯观察到，专业个人用户既会分享，又有付费能力，可用AI产品在供给稀缺时容易出现“脉冲式”传播。但Manus当时尚未大规模开放，热度先由媒体与创始团队已有内容影响力推动，广泛用户体验并未同步出现。
他认为反噬的一部分来自坐标错位：DeepSeek贡献在模型层，Manus则把已有技术组织成应用。把后者抬到同一种AGI突破的位置，就会让演示承担不匹配的期待。AGI在这里是通用人工智能的未来目标，不能由一次任务演示证明。
使用第三方模型不自动否定应用价值；自研模型也不能自动保证交付。讨论海外是否关注，同样不能成为唯一标准：传播圈子、受众任务和产品阶段都不同。双方希望媒体给团队更多时间，这是一种态度，并不取消产品接受实际检验的责任。
编辑推论：发布时应说明模型、工具与产品各自完成什么，给目标任务、适用用户和失败边界。传播越快，越需要接住真实访问和使用；访问量、邀请热度与长期付费之间仍隔着完整体验。
3. Capex买来能力和时间，股价并不等于利润
线索：22:05–33:24。 Capex指资本性支出，例如购买计算卡、建设数据中心，是前置投入。曲凯问，为什么公司宣布更多成本，市场反而可能上涨？莫傑麟的解释是，投资者将其理解为未来产业机会，改变了对公司的估价。
这不能推出“投入越多越好”。资金花出去后，还要变成可用能力、用户价值与经济回报。节目讨论腾讯借助开源模型跳过部分追赶阶段，甚至称其最大受益者；这是当时商业判断，并无对比数据证明收益排名，也不是公司内部路线已被核实。
开源在嘉宾看来不仅提供可调用或可部署的成果，还给其他团队路线与工程参考，减少探索弯路。但借鉴之后仍需接入产品、运营服务、承担成本，并根据真实任务检验。
编辑推论：记录投资计划、实际部署、用户使用和收入变化四个时点。市场可以先重估未来，经营必须逐步完成；不能从一段股价走势倒推投入有效，也不能把开源可获得当成商业效果已到手。
4. 技术部署有具体约束，宏观与产业也有不同节奏
线索：33:48–43:38。 嘉宾区分推理应用与大训练集群：前者偏部署和服务，后者对大规模互联有高要求。他用这个差异讨论硬件限制，但“推理不用互联”等口述过度概括，不能直接用作架构选择；模型、并行方式和规模都会改变部署需求。
他还沿用明线、暗线比喻：产业趋势是明线，地产、债务、消费和就业是更慢的宏观背景。某类科技资产的乐观，不等于全部经济问题同时解决。节目关于房价、政策成效和特定行业的描述没有独立数据审计，本篇只保留分析尺度的差别。
另一个节奏差是价格与产业。市场可能在很短时间把多年预期计入价格，企业却要长期研发、交付。价格又可能经并购能力、融资成本和员工期权反过来影响公司，形成反馈，而不是一面完全被动的镜子。
编辑推论：研究中同时画出技术、经营和预期的时间轴。故事跑得快，不代表业务进度也快；资本反馈带来的资源，应由后续交付检验，不能让讲故事永久替代结果。
5. 一线认知有价值，仍须转换成另一套能力
线索：44:36–56:07。 莫傑麟用基础设施从业者和Agent调用量作例子：亲手做产品的人可能更早感知需求，而与投资者交流又能知道市场期待的水位。Agent即能使用工具推进任务的系统，它可能增加每次任务的调用次数；这种体感值得研究，却不是特定资产必涨的证明。
他承认行业理解转成价格、策略和风险管理并不简单。高波动也不意味着人人应交易，长期持有和不同策略可以并存。离开一线全职投资，反而可能失去原来信息优势，所以二级市场是否合适仍取决于个人能力和选择。
片尾列出三组观察：模型与基础设施怎样分配价值，原生应用怎样发展，垂直领域怎样形成产品。他还保留一个张力：闭源模型定价权可能受压，同时有资源持续训练的公司也可能减少。竞争压力与供给集中可以并存，不能只摘一个方向当终局。
编辑推论：把观察问题写成可更新清单，比提前选唯一赢家更有用。公司、产业与价格之间有多次转换，每一步都需要证据，本期没有提供适用于个人的投资决策程序。
用五层记录复盘一次产品发布
以下为编辑练习。选一次AI产品发布，分别留下一份当期证据：
| 层次 | 应记录什么 | 不能替代什么 |
|---|---|---|
| 技术 | 任务、版本、性能与成本条件 | 真实用户结果 |
| 产品 | 开放范围、完成率与失败案例 | 持续需求 |
| 使用 | 真实体验、复访与任务频次 | 收入与利润 |
| 经营 | 付费、成本、退款及交付责任 | 证券回报 |
| 预期 | 宣传承诺、媒体传播与价格反应 | 以上各层实际进度 |
隔一段时间重新填写，保留最初页面和承诺。若传播先于体验，就先扩大验证；若产品有效但经营不成立，就继续查价值与成本。五层可能相互影响，却不能互相冒充。
回听导航与关联阅读
- 01:00：风险偏好与预期差。
- 07:27：DeepSeek改变了哪种叙事。
- 11:58：传播脉冲与真实体验的距离。
- 14:35：模型和应用的评价坐标。
- 24:39：资本投入为什么影响预期。
- 41:13：价格与产业时间不同步。
- 54:56：一线认知能否长期迁移。
相关主题：发布沟通、交付证据、市场叙事。可对照莫傑麟在2024年谈三种AI叙事和2025年初的应用展望，保留观点随事件变化的时点。
资料边界：完整阅读所供转写、旧精读和官方介绍，未逐句复听、实测产品或核验行情。宏观、成本、公司内部策略、性能、受益排名和市场预测均为当期观点；未据其形成当前投资、技术部署或政策结论。
English edition
Hype, Delivery and Prices: 莫傑麟 on Expectations After DeepSeek and Manus
42 Zhangjing episode analysis · Product positioning and capital narratives
Guest: 莫傑麟, an experienced family-office practitioner, as introduced at the time | Host: Qu Kai
Published: March 22, 2025 | Duration: 56 minutes 18 seconds | Prepared: October 18, 2026
Official episode and introduction
“East rising, West falling” first describes changing expectations
After reaching the United States, Qu noticed rapid changes in how people discussed Chinese and American assets. He asks why two years of pessimism could reverse suddenly and how long new optimism might last.
The guest explains intersecting expectations, risk appetite and actual trends. DeepSeek and Manus were contemporary landmarks, but technical progress, usable applications, corporate revenue and securities prices are different kinds of evidence. This analysis preserves the March 2025 discussion without treating policy, market movements or company comparisons as current conclusions or proposing trades.
1. An event can change the expectation gap without changing all reality
Episode reference: 01:00–09:30. The guest argues that previously high American expectations encountered macroeconomic uncertainty, while lower Chinese expectations met new technical achievements. Asset reactions could therefore diverge. Investors anticipating recession and contemporary data not yet clearly showing recession can coexist. Expectations are not completed economic outcomes.
He characterizes American narratives as intelligence and scaling, and Chinese narratives as applications and returns on investment. In his view, DeepSeek’s comparable performance at lower cost challenged a story of adding ever more resources. This reflects his circles rather than a classification of every team in both countries. Specific performance, training costs and comparison methods are not fully presented.
Editorial interpretation: During a sudden surge, identify the previous expectation, the new evidence and the unanswered questions. Lower costs may make tasks usable or stimulate more calls. Model prices alone cannot establish corporate profits without volumes, latency and whole-task costs.
2. Putting Manus on a model benchmark obscures application value
Episode reference: 10:15–21:12. Qu observes that professional individual users both share products and can pay. Usable AI products in limited supply may experience bursts of distribution. However, Manus was not broadly available then: media and the team’s existing content influence generated attention before widespread direct experience.
He attributes some backlash to mismatched coordinates. DeepSeek contributes at the model layer; Manus organizes existing technology into an application. Promoting both as the same kind of AGI breakthrough makes demonstrations carry inappropriate expectations. AGI, or artificial general intelligence, is a future objective that one task demonstration cannot establish.
Using third-party models does not automatically invalidate an application, while developing a model does not guarantee delivery. Overseas attention is not a sufficient standard either: distribution circles, audience tasks and product stages differ. The speakers ask media to allow teams time. That position does not remove responsibility for practical evaluation.
Editorial interpretation: A launch should explain what models, tools and the product each accomplish, along with target tasks, suitable users and failure boundaries. Faster distribution increases the need to handle real access and use. Visits, invitation demand and lasting payment remain separated by an entire experience.
3. Capital expenditure buys capability and time; share prices are not profit
Episode reference: 22:05–33:24. Capex is capital expenditure, such as computing cards and data centers: spending ahead of returns. Qu asks why announcing additional costs might raise prices. The guest interprets this as investors reassessing a company’s future industrial opportunity.
It does not follow that more spending is always better. Expenditure must become usable capability, customer value and economic returns. The discussion suggests Tencent could skip part of a catch-up phase through open models, even calling it the largest beneficiary. This is a contemporary business judgment without comparative evidence ranking benefits or verifying internal strategy.
The guest sees open source as more than an available model: it also supplies routes and engineering references, reducing exploratory mistakes. Borrowing those lessons still requires product integration, service operation, cost ownership and task-level evaluation.
Editorial interpretation: Record investment plans, actual deployment, usage and revenue changes at separate dates. Markets can reassess the future before businesses complete it. A price chart cannot prove spending effective, and accessible open technology does not mean commercial value has already arrived.
4. Deployment has specific constraints; macroeconomics and industries move at different speeds
Episode reference: 33:48–43:38. The guest distinguishes inference services from large training clusters, emphasizing interconnection in the latter. He uses this distinction to discuss hardware constraints. Broad statements that inference needs no interconnection cannot guide architecture directly: model, parallelization and scale change deployment requirements.
He also distinguishes visible industrial trends from slower background issues such as property, debt, consumption and employment. Optimism about selected technology assets cannot establish resolution of every economic problem. Property prices, policy effects and individual industries lack independent data checks here; the analysis retains the difference in analytical scale.
Prices and industries create another timing gap. Markets may price years of expectations quickly while companies research and deliver over time. Prices can then affect acquisitions, financing costs and employee options, creating feedback rather than acting solely as a passive reflection.
Editorial interpretation: Maintain separate timelines for technology, operations and expectations. A fast story is not equally fast business progress. Resources obtained through capital feedback require subsequent delivery; storytelling cannot permanently substitute for results.
5. Frontline understanding still needs conversion into different skills
Episode reference: 44:36–56:07. Infrastructure practitioners and agent usage illustrate the guest’s argument: builders may feel demand earlier, while talking with investors reveals prevailing expectations. Agents—systems using tools to advance tasks—may increase calls per task. That experience merits investigation without proving any particular asset will rise.
He acknowledges that translating industry understanding into prices, strategy and risk management is difficult. Volatility does not require everyone to trade, and long-term holding can coexist with other approaches. Leaving frontline work to invest full time may weaken the original information advantage. Suitability therefore depends on individual capabilities and choices.
The ending presents three questions: value allocation between models and infrastructure, development of native applications and products in specialized domains. It preserves a tension: closed-model pricing power might weaken while fewer companies retain resources for continued training. Competitive pressure and concentrated supply can coexist, so neither observation establishes a final industry outcome alone.
Editorial interpretation: An updateable question list is more useful than selecting one inevitable winner. Several conversions connect a company, its industry and its price. Each needs evidence, and this episode offers no personalized investment decision procedure.
Review a launch through five layers
For an editorial exercise, choose one AI launch and retain contemporary evidence separately:
| Layer | Record | Cannot substitute for |
|---|---|---|
| Technology | Tasks, versions, performance and cost conditions | Actual user outcomes |
| Product | Availability, completion rates and failures | Continuing demand |
| Use | Direct experience, repeat visits and task frequency | Revenue and profit |
| Business | Payment, cost, refunds and delivery responsibility | Securities returns |
| Expectations | Promises, media distribution and price reaction | Actual progress in every preceding layer |
Update after an interval while preserving initial pages and promises. If attention precedes experience, expand validation first. If the product works but its economics do not, examine value and costs. The layers can influence one another without impersonating one another.
Listening guide and related reading
- 01:00: Risk appetite and expectation gaps.
- 07:27: Which narrative DeepSeek changed.
- 11:58: Distribution bursts versus direct experience.
- 14:35: Different coordinates for models and applications.
- 24:39: Capital spending and expectations.
- 41:13: Prices and industrial time.
- 54:56: Whether frontline understanding remains transferable.
Related themes: launch communication, delivery evidence and market narratives. Compare the guest’s 2024 discussion of three AI narratives and the early-2025 application outlook, preserving when events changed views.
Source boundary: The complete supplied transcript, earlier note and official introduction were read, without sentence-by-sentence audio checking, product testing or market-data verification. Macroeconomics, costs, internal strategies, performance, beneficiary rankings and forecasts remain contemporary views. They establish no current investment, deployment or policy conclusion.