Read in English ↓

42章经AI行业与基础设施：能力、交付、资本及价值归属

本文是来源限定的主题整理：保存42章经2023—2026年访谈中的框架、分歧与可继续验证的问题。这里的“当时”不能改成“现在”；没有独立核对的经营、融资、性能及市场数字不充当统计事实，也不据此提供证券买卖建议。公司工程案例另见相关硬件/前沿研究稿。

一、将一条宏大故事拆成五项命题

模型能力提高、应用有人用、客户愿付钱、公司有利润、资本给高估值是五件事。2024年3月莫傑麟把叙事拆成Scaling Law、竞争者位置和应用爆发：前两项升温并不自动完成第三项。2025年3月DeepSeek与Manus讨论又指出，模型贡献和应用交付不能用同一坐标评价；第三方模型不使应用没有价值，但把应用发布包装成基础模型突破会抬高错误预期。[S55][S26]

本文将因果链写为：能力边界变化→任务可以完成→交付足够可靠→持续使用与付费→收益在价值链中分配。资本投入既可提前支持每一环，也可在实际收益前把预期计入价格；因此股价或融资热不适合作为链条完成的验收。

2023年底莫傑麟用产业“明线”和利率、库存、需求等“暗线”区分可做的实践与外部条件；2026年朱宁进一步强调技术真实价值、企业盈利、估值合理相互独立。知道技术进步并不能解决价格、时间和承受后果的问题。其“新概念、流动性、政策支持、缺经验参与者”四条件明确为必要不充分，不能转成凑齐就能判断顶部的公式。[S29][S50]

2025年11月莫傑麟对谈进一步给出期限错配：数据中心与电力容量的长期建设、应用短期收入、市场季度预期并不同步。应画最终客户→应用→推理/训练采购→硬件/电力的资金链，标出付款、承诺、期限和毛利，避免把循环合作公告多次加总。该期说话人识别严重失准，Scaling Law“失效”与“未失效但评价困难”是对立意见；不能统合为预训练已结束。[S17]

二、阶段叙事有用，前提是不把阶段当日历

访谈时点	当期主要观察/预测	应保留的检验条件
2023年张津剑、陈昱	早期团队可能借资源完成学习；模型、中间层、应用有多层价值	技术可实现、团队积累、客户需求分别确认，不能只看融资 [S46][S48]
2024年初及年中曲凯	可控性不足先卖结果/限制任务；全力学与立即投入全部资金不同	人工返工、全链路交付与现金消耗；小闭环不等于大规模PMF [S3][S24]
2024年汪华	交互重构、复杂任务自动化、AGI三层	各层信心和证据不同，不能由第一层体验好推到全自治 [S8]
2025年初曲凯	Agent从信息服务走向执行结果，多模态走向参与	可计费结果、失败责任、采购信任和真实完成 [S28]
2025年8月莫傑麟	从模型智力叙事转向Token消耗和产业化	训练与推理需求、尝鲜收入与稳定PMF分开 [S52]
2026年张津剑与曲凯	大厂反应强、应用融资冷、模型势能增强	需求连续、实现迭代；融资故事不能冒充已做技术 [S11][S7]

这些不是彼此接续的确定产业年表，而是不同覆盖圈层的观察。2024年“生成打败推荐”是明确排除社区等因素的极限推演；2026年“应用没了”是反问，并非应用数量统计。比较它们应追问哪项约束改变、哪项需求仍在，而不是事后把每个新事件都收编为早已预测。[S3][S7]

2024年8月曲凯以Character.AI事件推演“模型商品化、模型战争近结束”，到2026年又观察模型势能更强，这正是应保存的观点演变，不能把前一期预测删掉后拼成连续正确叙事。其CAI分析还区分用户价值、产品经营、自研模型和退出结果：部分人员与授权安排不等于整家公司被股权收购；退出能否让部分人获利，也不证明产品盈利或陪伴需求失败。节目中的人物心理、监管动机与交易金额没有独立验证。[S39][S7]

三、模型训练、数据与成本：看目标函数，也看运行条件

2025年阶跃星辰朱亦博把Infra解释为计算、通信、存储、调度和框架的共同系统；同一模型可有不同首字延迟、输出速度、稳定性和每任务成本。高吞吐但首字等待很长，不一定适合聊天。小参数模型也不必然便宜，通信、内存访问、算子及硬件适配会改变账单。[S36]

训练与推理不应合并成“算力需求”。2025年8月莫傑麟认为预训练需求有周期、推理依赖应用使用，而RL大量生成又可能增加训练侧需求。这是当期解释，不代表二者以后必然同涨。朱亦博则用decoding说明目标重排：RL反馈也依赖大量输出，推理优化会影响训练迭代；优化MFU、训练速度与上线成本仍可能冲突。正确问题是“为哪一笔账优化”，不是哪个指标永远最大最好。[S52][S36]

2026年孟繁青进一步把数据价值放回闭环：格式正确、难度合适、无作弊只是初筛；训练后目标任务进步、旧能力不退化，才说明数据有效。模拟飞书/Notion等环境先要和真实环境行为一致，评分也须检查；否则越多轨迹越可能放大假进步。99条一分钟、一条100分钟的例子说明尾部任务拖慢整批，数据、算法和调度共同决定有效吞吐。[S33]

自我改进也分层：修结果、改Harness、产生训练数据、训练新模型并不是同一个成果。垂直任务可能由强模型建设流程、较小模型执行，但质量、隐私、成本和维护仍须实测。孟繁青不同意价格必然趋近免费；汪华2024年“GPT-4级模型每用户几分钱”、曲凯关于国内Token近零打开大DAU市场则是条件性预测。把这些分歧并列，比选最便宜的预测做预算更稳妥。[S33][S8][S52]

数据量也不是使用权。2023年Meshy的胡渊鸣已指出平台上能看到3D模型不等于可以训练；2023年底宏观讨论中的“国内有更多场景数据”不能绕过权属、质量和流程。真实工作轨迹、个人对话及外部教师模型输出同样要独立确认使用范围，不能由研究价值推导授权。[S32][S29][S33]

2026年8月逯雨鑫的低预算后训练案例把口径进一步拆开：Hugging Face Trending第一是热度，不是综合性能第一；数百美元覆盖个人项目部分数据与训练，不包含从零预训练、人力、失败试验和长期Serving。目标基准提高还可能伴随其他能力下降。其“任务未完却宣布完成”的数据病灶，要求对照实际工具状态，而不是只训练谦虚措辞；强教师轨迹也可能超出学生可吸收的容量。小模型路线需比较数据、部署、语言、安全和回归表现，不能只凭一项涨分。[S19]

Local AI同样有真实分歧：嘉宾看重敏感场景和创作消耗，主持人与制作人追问便利及SOTA差距。数据不发给模型厂商，只减少一条外流路径，插件、日志、同步和应用遥测仍需检查；本地不自动隐私无忧，自有显卡也不等于零运行成本。这个个案支持更多方案进入比较，不构成人人都应训练或本地部署的结论。[S19]

四、芯片和硬件价值来自系统位置与可采用性

2024年季宇将英伟达解释为芯片、软件和开发者生态的系统竞争：CUDA、框架、算子和习惯使替代不只是比单点性能。新部件可先补充既有CPU体系、渐进迁移，再扩大价值；真正的产品定义要说明服务何种任务、愿意放弃什么、对手为何受存量约束不能同样放弃。“更全能”也意味着更多平衡，不保证永远没有可拆分的矛盾。[S56]

这与朱亦博关于模型—硬件共设计相互支持，也存在边界：更深适配可创造效率，绑定特定架构又可能限制后续变化；热门开源模型既帮助生态采用，也可能固化投入方向。中立第三方若只靠短暂优化领先可能被上下游挤压，但两位都没有证明第三方必无机会。客户部署自由、迁移负担、可靠性与经营成本须一起比较。[S36][S56]

消费硬件不能直接套芯片生态叙事。2024年谢羽佳强调独立终端先回答相对手机补了什么场景/人群，再过能力、成本、使用意愿三关。12—18个月研发周期、模具与硬件不可随意回滚，使错误定义代价大于软件；OTA可更新内容和算法，却不能修复所有结构、供电和安全问题。合理毛利既要覆盖服务，也提供纠错空间，不是高售价的借口。[S9]

众筹还要区分测试兴趣、集中首发营销、融资执行展示。百万销售额可能亏损，早期支持者不等于大众市场；小规模ROI不能直接外推放量。渠道、CPA、退货、交付和后续复购，比单看首发金额更接近完整经营。这一组条件使“模型足够好即可做硬件”的推断站不住。[S9]

五、应用的价值归属，落在完整交付而非公司标签

2023年陈昱的Zilliz案例展示可迁移支点：团队从GPU加速数据库切向客户以图搜图暴露的向量需求，保留计算和数据库工程能力；原型容易，生产性能和全球商业化更难。他同时用PingCAP说明跨平台中立性可能成为独立公司价值。底层能力相同不代表部署范围、客户信任和交付责任相同。[S48]

2023年底Meshy又提供专业与UGC两种质量线。生成图看似3D，不代表Mesh拓扑、贴图、可编辑性、绑定和游戏预算都可用；普通用户还可能不知道怎样消费文件。中远景道具、游戏服装UGC或许先有惊喜，不能据此证明高规格角色生产已被替代。研究表示与既有生产管线的适配，是模型能力到收入之间不可省的一层。[S32]

企业交付也如此。Jenny在2025年底称95%准确的投行文件分析可能让人从头复核，净收益反而小；只服务两三家大客户、六七成收入集中也可能是定制业务，不能轻易宣布标准化PMF。2024年曲凯提出AI+人工卖结果是当时绕过不稳定性的路径，但必须计算返工和审阅，不可让“人工兜底”掩盖毛利。[S27][S3]

基础模型公司做应用，应用公司补RL或基础设施，说明边界随瓶颈移动。莫傑麟用ServiceNow的工作流交付、Figma的跨组织协作讨论Context；这些资产能否持续，取决于用户与交付任务，而非“应用层”标签。曲凯2026年提出持续服务同一类需求、随模型改变方案，则给经营者一个稳定轴：需求不必随资本热门名词频繁更换。[S52][S7]

六、Agent基础设施不只有算力与模型接口

张津剑2023年“Bot买酒”设想，已经包含需求理解、商家供给、通信追溯、支付权限、欺骗与争议责任；2026年又强调身份、代表权、交易和问责。两期共同说明：能生成交易建议，与能合法、可靠地代表用户完成交易不同。身份和责任仍是待建设问题，不能从“主体性”一词推出已有法律人格或无限支付权。[S46][S11]

对产品团队的编辑提炼是同时画两张图：技术依赖图（模型、执行环境、Context、工具与数据）和责任图（谁授权、谁持有数据、谁审核、失败谁恢复）。技术接口通了只验收前者的一部分。2025年结果付费愿景还需可定义结果、可核算失败和客户愿意迁移流程，不能用“卖劳动力”跳过信任成本。[S28]

七、用来源变化约束自己的判断

可复用的行业记录表应有：发表日期、当期角色与利益关系、观察样本、主张、相反解释、何种证据能推翻、下一次检查点。融资数量、ARR、估值、Token增速、成本、众筹ROI要各保留分母、周期和阶段；没有口径就不参与横比。2023年张津剑重看旧报纸恢复当时信息，片尾也警惕反复说服别人强化自身信念；朱宁则提醒懂偏差不意味着行动免疫。[S46][S50]

对产品团队而言，最可执行的行业研究结果不是“应用还是模型胜出”，而是固定真实任务集，检查质量、首字延迟、全程耗时、每成功任务成本和人工复核；将用户、供应商和团队预算放在同一场景。学习可持续投入，人员、算力与资金按已验证环节分段扩大。保留失败、预测落空及市场/技术两种时钟，才不会只收集支持既有方向的新闻。[S24][S36]

来源索引与核验边界

文中的S编号对应下列原节目。日期保留节目当时的时间；链接含官方嘉宾介绍与章节导航，口述和转写中的疑点不据猜测作定论。

• S3：24、25 年会是下一代浪潮最关键的两年 | AI 年终复盘（2024-01-13），节目原文。

• S7：AI 发展了 4 年，把应用发展没了？｜AI 年中复盘（2026-07-19），节目原文。

• S8：AI 的三个发展阶段、和移动互联网的异同、未来的演进路线｜对谈创新工场汪华（2024-08-10），节目原文。

• S9：AI 硬件的道、法、术 | 对谈三耳猫创始人谢羽佳（2024-04-20），节目原文。

• S11：All in AI 的第一个三年｜对谈绿洲资本合伙人张津剑（2026-01-10），节目原文。

• S24：上半年 AI 市场有多差？别刚起跑就想冲刺 | 我答莫傑麟（2024-07-13），节目原文。

• S26：世界怎么就「东升西落」了？聊聊二级市场与 DeepSeek+Manus 的热潮｜对谈莫傑麟（2025-03-22），节目原文。

• S27：中美 AI 创投的真实差异｜对谈 Leonis Capital 合伙人 Jenny（2025-12-27），节目原文。

• S28：为什么我们对 25 年 AI 极度乐观｜AI 年终复盘（2025-01-04），节目原文。

• S29：为什么我们开始乐观？关于经济发展的明线暗线、AI 与最佳实践 | 对谈莫傑麟（2023-12-02），节目原文。

• S32：从图形学到 AI 生成 3D：我们离 3D 版 Midjourney 还有多远？ | 对谈 Meshy.AI 创始人渊鸣（2023-12-16），节目原文。

• S33：从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青（2026-08-08），节目原文。

• S36：关于 AI Infra 的一切 | 对谈阶跃星辰联创朱亦博（2025-08-02），节目原文。

• S39：宿命论与万恶的 Scaling Law | 暴论 CAI 被收购（2024-08-17），节目原文。

• S46：投 AI 最猛的人 | 对谈绿洲资本合伙人张津剑（2023-07-29），节目原文。

• S48：揭秘 Minimax、Zilliz、PingCAP 背后的投资逻辑 | 对谈云启合伙人陈昱（2023-08-12），节目原文。

• S50：泡沫的四个必要不充分条件 | 对谈经济学者朱宁教授（2026-06-13），节目原文。

• S52：硅谷 AI 大转弯与二级市场的牛市｜对谈莫傑麟（2025-08-23），节目原文。

• S55：英伟达、Sora 与 AI 的三种核心叙事逻辑 | 对谈莫傑麟（2024-03-02），节目原文。

• S56：英伟达的阳谋、破绽与野心 | 对话行云创始人 Mackler 季宇（2024-06-10），节目原文。

• S17：「你觉得 AI 有泡沫吗？」——有！｜对谈莫傑麟（2025-11-22），节目原文。

• S19：一个人、两周、数百美元，如何训出登顶 Hugging Face 的模型 | 对谈研究员逯雨鑫（2026-08-29），节目原文。

来源中的人物心理、企业内部动机、市场数据、技术性能和预测均未自动升级为核验事实。来源介绍的官方身份也仅代表当期；本稿不向私人材料外延引出公开发布授权。

English version

AI Industry and Infrastructure Through 42 Zhangjing: Capabilities, Delivery, Capital, and Who Captures Value

This article is a synthesis bounded by its sources. It preserves frameworks, disagreements, and questions for further testing from 42 Zhangjing interviews published between 2023 and 2026. “At the time” must not become “today.” Business, financing, performance, and market figures that have not been independently checked are not treated as statistical facts or used to offer securities trading advice. Engineering examples from individual companies are discussed separately in related articles on hardware and frontier research.

1. Break a Grand Narrative into Five Propositions

Better model capabilities, applications attracting users, customers willing to pay, profitable companies, and high valuations from capital markets are five different things. In March 2024, Mo Jielin (莫傑麟) separated the narrative into scaling laws, competitors’ positions, and an application boom: growing momentum in the first two does not automatically deliver the third. A March 2025 discussion of DeepSeek and Manus similarly argued that model contributions and application delivery cannot be evaluated on the same axis. Using a third-party model does not make an application valueless, but presenting an application launch as a foundation-model breakthrough creates the wrong expectations. [S55][S26]

The editorial synthesis here expresses the causal chain as: a change in capability boundaries → a task becomes achievable → delivery becomes reliable enough → sustained use and payment → returns are distributed across the value chain. Capital can support each link in advance, while markets can also price in expectations before returns materialize. Share prices and enthusiasm for financing are therefore poor acceptance criteria for whether the chain is complete.

At the end of 2023, Mo distinguished the industry’s “visible threads” from “hidden threads” such as interest rates, inventories, and demand, separating practices a team could undertake from external conditions. In 2026, Zhu Ning (朱宁) further emphasized that technology’s real value, corporate profitability, and reasonable valuations are independent questions. Knowing that technology is advancing does not resolve price, timing, or the ability to bear the consequences. His four conditions—new concepts, liquidity, policy support, and inexperienced participants—were explicitly necessary but insufficient. They cannot be turned into a formula for identifying a market top whenever all four appear. [S29][S50]

The November 2025 conversation with Mo added a mismatch of time horizons: long-term construction of data centers and power capacity, short-term application revenue, and quarterly market expectations do not move in sync. A useful editorial proposal is to map the funding chain from end customers → applications → inference/training purchases → hardware/electricity, marking payments, commitments, terms, and gross margins rather than repeatedly adding together circular partnership announcements. Speaker identification in that episode is severely inaccurate. Claims that scaling laws had “failed” and that they had “not failed, but had become difficult to evaluate” are opposing views; they cannot be combined into a conclusion that pretraining was over. [S17]

2. Stage Narratives Help Only If Stages Are Not Treated as a Calendar

Interview date	Main observations or predictions at the time	Conditions that must remain part of the test
Zhang Jinjian and Chen Yu, 2023	Early teams might use resources to accelerate learning; models, middleware, and applications could all create value	Check technical feasibility, accumulated team capabilities, and customer demand separately rather than looking only at financing [S46][S48]
Qu Kai, early and mid-2024	When controllability is inadequate, sell outcomes or restrict tasks; learning intensively is different from immediately committing all available funds	Human rework, delivery across the full process, and cash burn; a small working loop is not large-scale PMF [S3][S24]
Wang Hua, 2024	Three layers: redesigning interaction, automating complex tasks, and AGI	Confidence and evidence differ by layer; a good first-layer experience does not establish full autonomy [S8]
Qu Kai, early 2025	Agents would move from information services toward executing outcomes; multimodal systems would move toward participation	Billable outcomes, responsibility for failure, procurement trust, and actual completion [S28]
Mo Jielin, August 2025	The narrative was shifting from model intelligence to token consumption and industrial adoption	Separate training demand from inference demand, and novelty-driven revenue from stable PMF [S52]
Zhang Jinjian and Qu Kai, 2026	Large companies were responding strongly, application financing was weak, and models were gaining momentum	Continuity of demand and implementation iterations; a financing story must not be presented as technology already built [S11][S7]

These are observations covering different circles, not a definitive industry chronology in which one stage necessarily follows another. The 2024 proposition that “generation beats recommendation” was an extreme thought experiment that explicitly excluded factors such as communities. The 2026 question “Have applications disappeared?” was rhetorical, not a statistical count of applications. Comparing these views requires asking which constraints changed and which needs persisted, instead of retroactively treating every new event as something already predicted. [S3][S7]

In August 2024, Qu Kai used the Character.AI event to project that “models are becoming commodities and the model wars are nearly over.” By 2026, he was observing stronger momentum behind models. This evolution in views is precisely what should be retained, rather than deleting the earlier prediction to assemble an uninterrupted story of being right. His CAI analysis also distinguished user value, product operations, proprietary models, and exit outcomes. Arrangements involving some employees and licensing do not mean the entire company was acquired through an equity purchase. An exit allowing some people to profit does not prove either product profitability or the failure of demand for companionship. Claims in the episode about individuals’ psychology, regulatory motives, and transaction amounts were not independently verified. [S39][S7]

3. Model Training, Data, and Cost: Examine Both Objectives and Operating Conditions

In 2025, StepFun’s Zhu Yibo (朱亦博) described infrastructure as a combined system of computation, communication, storage, scheduling, and frameworks. The same model can have different time to first token, generation speed, stability, and cost per task. High throughput with a long initial wait may be unsuitable for chat. A model with fewer parameters is not necessarily cheaper: communication, memory access, operators, and hardware adaptation can change the bill. [S36]

Training and inference should not be merged into a single category of “compute demand.” In August 2025, Mo argued that pretraining demand was cyclical and inference depended on application usage, while the large volume of generation required by RL might increase demand on the training side. This was an explanation offered at that time, not a claim that the two must always rise together thereafter. Zhu used decoding to illustrate changing priorities: RL feedback also requires large amounts of output, so inference optimization affects training iterations. Optimizing model FLOPs utilization (MFU), training speed, and production serving cost can still produce conflicts. The right question is “Which bill are we optimizing?” rather than which metric should always be maximized. [S52][S36]

In 2026, Meng Fanqing (孟繁青) placed data value back inside a feedback loop. Correct formatting, appropriate difficulty, and absence of cheating are only initial filters. Data is effective when training improves the target tasks without degrading previous capabilities. Simulated environments such as Feishu or Notion must first behave consistently with the real environments, and their scoring must also be checked. Otherwise, more trajectories may amplify false progress. The example of 99 tasks taking one minute each and one task taking 100 minutes illustrates how tail tasks slow down a whole batch. Data, algorithms, and scheduling jointly determine effective throughput. [S33]

Self-improvement also has distinct layers: correcting an output, changing the harness, producing training data, and training a new model are not the same achievement. A vertical task may use a strong model to build a process and a smaller one to execute it, but quality, privacy, cost, and maintenance still require measurement. Meng disagreed that prices must approach zero. Wang Hua’s 2024 prediction of “GPT-4-level models for a few fen per user,” and Qu Kai’s argument that near-zero domestic token prices could open markets with large daily active user bases, were conditional predictions. Keeping these disagreements side by side is more prudent than budgeting against whichever forecast is cheapest. [S33][S8][S52]

Data volume is not permission to use it. In 2023, Meshy’s Hu Yuanming (胡渊鸣) already noted that being able to view a 3D model on a platform does not mean it can be used for training. The late-2023 macro discussion that “China has more data from real-world scenarios” cannot bypass ownership, quality, and process requirements. Real work trajectories, personal conversations, and outputs from external teacher models likewise require separate confirmation of permitted use. Research value does not imply authorization. [S32][S29][S33]

Lu Yuxin’s (逯雨鑫) low-budget post-training case in August 2026 further separated the metrics. Ranking first on Hugging Face Trending indicates popularity, not first place in overall performance. The few hundred dollars covered part of one personal project’s data and training costs, not pretraining from scratch, labor, failed experiments, or long-term serving. A gain on the target benchmark might also accompany losses in other capabilities. The data problem he described—declaring completion when the task was unfinished—requires comparison against actual tool state, rather than merely training more modest wording. Trajectories from a strong teacher may also exceed what the student can absorb. A small-model approach needs comparisons across data, deployment, language, safety, and regression performance, not just one higher score. [S19]

Local AI involved real disagreement as well. The guest valued sensitive use cases and the compute consumed by creative work, while the host and producer questioned convenience and the gap from the state of the art (SOTA). Keeping data away from a model vendor removes only one possible leakage path; plugins, logs, synchronization, and application telemetry still need checking. Local deployment does not automatically eliminate privacy concerns, and owning a GPU does not mean operating costs are zero. This case supports including more options in comparisons, not a conclusion that everyone should train models or deploy locally. [S19]

4. Chip and Hardware Value Comes from System Position and Adoptability

In 2024, Ji Yu (季宇) explained Nvidia’s position as system-level competition across chips, software, and the developer ecosystem. CUDA, frameworks, operators, and established habits mean that replacing it requires more than comparing isolated performance. A new component can first complement the existing CPU system, migrate gradually, and then expand its value. A genuine product definition should explain which tasks it serves, what it is willing to give up, and why incumbents’ installed bases prevent them from making the same trade-off. Being “more general-purpose” also means balancing more requirements; it does not guarantee there will never be a conflict that a focused entrant can exploit. [S56]

This supports Zhu Yibo’s discussion of model–hardware co-design, while also exposing its limits. Deeper adaptation can create efficiency, but attachment to a particular architecture can constrain later changes. Popular open-source models can encourage ecosystem adoption while also locking in investment directions. A neutral third party relying only on a temporary optimization lead may be squeezed by upstream and downstream players, but neither speaker proved that third parties necessarily have no opportunities. Customer freedom in deployment, migration burden, reliability, and operating cost must be compared together. [S36][S56]

Consumer hardware cannot simply inherit the chip-ecosystem narrative. In 2024, Xie Yujia (谢羽佳) stressed that a standalone device must first explain which use case or audience it serves beyond what a phone provides, then pass three tests: capability, cost, and willingness to use it. A 12–18-month development cycle, tooling, and hardware that cannot be freely rolled back make a mistaken product definition more costly than in software. Over-the-air (OTA) updates can change content and algorithms, but cannot repair every structural, power, or safety issue. A reasonable gross margin must cover service and leave room for correction; it is not an excuse for a high selling price. [S9]

Crowdfunding also needs to be distinguished by purpose: testing interest, concentrated launch marketing, or demonstrating execution for fundraising. Sales of a million may still lose money, and early backers do not represent the mass market. Small-scale ROI cannot be directly extrapolated to higher volumes. Channels, cost per acquisition (CPA), returns, fulfillment, and later repeat purchases provide a fuller view of operations than launch proceeds alone. Together, these conditions undermine the inference that “a good enough model makes hardware viable.” [S9]

5. Applications Capture Value Through Complete Delivery, Not Company Labels

Chen Yu’s (陈昱) 2023 Zilliz example showed a transferable technical foothold. The team shifted from GPU-accelerated databases toward vector requirements revealed by customers’ image-search needs, retaining its capabilities in computation and database engineering. A prototype was easy; production performance and global commercialization were harder. He also used PingCAP to show how cross-platform neutrality could create value for an independent company. Sharing underlying capabilities does not mean sharing the same deployment scope, customer trust, or delivery responsibilities. [S48]

At the end of 2023, Meshy offered two further quality thresholds: professional production and user-generated content (UGC). An image that looks 3D does not establish that mesh topology, textures, editability, rigging, and game performance budgets are all usable. Ordinary users may not even know how to use the file. Mid-distance and background props, or UGC clothing for games, might produce early positive surprises; that does not prove that high-specification character production has been replaced. Adapting a research representation to existing production pipelines is an indispensable layer between model capability and revenue. [S32]

Enterprise delivery follows the same logic. At the end of 2025, Jenny said that investment-banking document analysis with 95% accuracy could make people review everything from the beginning, leaving little net benefit. Serving only two or three large customers, with 60–70% of revenue concentrated among them, might also indicate custom services rather than standardized PMF. In 2024, Qu Kai proposed selling outcomes through AI plus human labor as a way around instability at that time. Rework and review must still be counted; a “human backstop” must not obscure the gross margin. [S27][S3]

Foundation-model companies building applications, and application companies adding RL or infrastructure, suggest that boundaries move with the bottlenecks. Mo discussed context through ServiceNow’s workflow delivery and Figma’s cross-organization collaboration. Whether those assets endure depends on users and the tasks being delivered, not the “application layer” label. In 2026, Qu Kai’s suggestion to keep serving the same category of demand while changing the solution as models evolve gave operators a stable axis: needs need not change whenever a new term becomes popular with investors. [S52][S7]

6. Agent Infrastructure Extends Beyond Compute and Model APIs

Zhang Jinjian’s (张津剑) 2023 idea of a “bot buying wine” already included understanding demand, merchant supply, traceable communications, payment permissions, deception, and responsibility for disputes. In 2026, he again emphasized identity, authority to represent someone, transactions, and accountability. Together, these episodes show that generating a transaction recommendation is different from legally and reliably completing a transaction on a user’s behalf. Identity and responsibility remained problems to be built for; the word “agency” does not establish legal personhood or unlimited payment authority. [S46][S11]

The editorial proposal for product teams is to draw two maps together: a technical dependency map covering models, execution environments, context, tools, and data; and a responsibility map covering who authorizes actions, who holds data, who reviews, and who restores the system after failure. Connecting technical interfaces validates only part of the first map. The 2025 vision of outcome-based payment also requires definable outcomes, failures whose costs can be accounted for, and customers willing to migrate their processes. Calling it “selling labor” does not remove the cost of earning trust. [S28]

7. Use Changes in Sources to Constrain Your Own Judgments

A reusable industry record should contain publication date, the speaker’s role and interests at that time, observation sample, claims, alternative explanations, evidence that could disprove them, and the next checkpoint. Financing counts, annual recurring revenue (ARR), valuations, token growth, costs, and crowdfunding ROI each need their own denominator, period, and stage. Without a defined basis, they should not enter cross-case comparisons. In 2023, Zhang Jinjian revisited old newspapers to reconstruct the information available at the time; the episode’s closing remarks also warned that repeatedly persuading others can reinforce one’s own beliefs. Zhu Ning reminded listeners that understanding biases does not make one immune to them in action. [S46][S50]

For product teams, the most actionable result of industry research is not deciding whether “applications or models will win.” It is fixing a set of real tasks and checking quality, time to first token, total completion time, cost per successful task, and human review, placing user, supplier, and team budgets within the same scenario. Learning can receive sustained investment, while staffing, compute, and funding expand in stages around validated links. Keeping records of failures, missed predictions, and the different clocks of markets and technology helps avoid collecting only news that supports an existing direction. [S24][S36]

Source Index and Verification Boundaries

The S references identify the original episodes listed below. Dates retain the episodes' historical timing. The linked pages provide official guest introductions and chapter navigation; uncertainties in speech or transcription are not resolved through guesswork.

• S3: “2024 and 2025 Will Be the Two Most Critical Years of the Next Wave | AI Year-End Review” (2024-01-13), Original episode.

• S7: “After Four Years of AI Development, Have Applications Disappeared? | AI Midyear Review” (2026-07-19), Original episode.

• S8: “AI’s Three Stages of Development, Similarities and Differences with Mobile Internet, and Its Future Path | A Conversation with Sinovation Ventures’ Wang Hua” (2024-08-10), Original episode.

• S9: “The Principles, Methods, and Techniques of AI Hardware | A Conversation with Sanermao Founder Xie Yujia” (2024-04-20), Original episode.

• S11: “The First Three Years of Going All In on AI | A Conversation with Oasis Capital Partner Zhang Jinjian” (2026-01-10), Original episode.

• S24: “How Bad Was the AI Market in the First Half? Don’t Try to Sprint Just After Starting | Answering Mo Jielin” (2024-07-13), Original episode.

• S26: “How Did the World Become ‘East Rising, West Falling’? Public Markets and the DeepSeek + Manus Boom | A Conversation with Mo Jielin” (2025-03-22), Original episode.

• S27: “The Real Differences Between AI Venture Investing in China and the US | A Conversation with Leonis Capital Partner Jenny” (2025-12-27), Original episode.

• S28: “Why We Are Extremely Optimistic About AI in 2025 | AI Year-End Review” (2025-01-04), Original episode.

• S29: “Why Are We Becoming Optimistic? The Visible and Hidden Threads of Economic Development, AI, and Best Practices | A Conversation with Mo Jielin” (2023-12-02), Original episode.

• S32: “From Computer Graphics to AI-Generated 3D: How Far Are We from a 3D Midjourney? | A Conversation with Meshy.AI Founder Yuanming” (2023-12-16), Original episode.

• S33: “From Distillation to Synthetic Data to RSI: What Is the Next Focus of Model Competition? | A Conversation with Evolvent AI Co-Founder Meng Fanqing” (2026-08-08), Original episode.

• S36: “Everything About AI Infrastructure | A Conversation with StepFun Co-Founder Zhu Yibo” (2025-08-02), Original episode.

• S39: “Fatalism and the Much-Maligned Scaling Law | Provocative Takes on CAI’s Acquisition” (2024-08-17), Original episode.

• S46: “The Most Aggressive AI Investor | A Conversation with Oasis Capital Partner Zhang Jinjian” (2023-07-29), Original episode.

• S48: “Inside the Investment Logic Behind Minimax, Zilliz, and PingCAP | A Conversation with Yunqi Partners’ Chen Yu” (2023-08-12), Original episode.

• S50: “Four Necessary but Insufficient Conditions for a Bubble | A Conversation with Economist Professor Zhu Ning” (2026-06-13), Original episode.

• S52: “Silicon Valley’s Big AI Turn and the Public-Market Bull Run | A Conversation with Mo Jielin” (2025-08-23), Original episode.

• S55: “Nvidia, Sora, and Three Core Narratives of AI | A Conversation with Mo Jielin” (2024-03-02), Original episode.

• S56: “Nvidia’s Open Strategy, Vulnerabilities, and Ambitions | A Conversation with Xingyun Founder Mackler Ji Yu” (2024-06-10), Original episode.

• S17: “‘Do You Think AI Is in a Bubble?’—Yes! | A Conversation with Mo Jielin” (2025-11-22), Original episode.

• S19: “One Person, Two Weeks, a Few Hundred Dollars: Training a Model That Tops Hugging Face | A Conversation with Researcher Lu Yuxin” (2026-08-29), Original episode.

Statements in the sources about individuals’ psychology, internal corporate motives, market data, technical performance, and predictions do not automatically become verified facts. Official roles in source introductions apply only to the episode’s historical setting. This article does not extend publication authorization to private materials.