嘉御资本卫哲：用“车、路、货”拆解AI商业化

Overall: Wei Zhe uses a "car, road, cargo" framework to break down AI commercialization: large models are the "car," transmission, storage, and computing power are the "road," and code generation, video generation, and conversational bots are the "cargo." Competition has already shifted from model racing to application-level "hauling cargo," which is the key framework for general readers to understand AI.

1. The model leadership window has shortened to about three months, and user switching costs are low. Users may switch "cars" if performance or price changes, so there is no need to blindly bet on any single strongest model.

2. Three types of applications have developed distinct logics: code generation has taken the lead in generating revenue, video generation builds barriers through a "content-feedback-retraining" loop, and conversational bots burn cash in the short term but are tied to traffic entry points.

3. For individuals or small teams, the most practical approach is to first clarify what "cargo" they want to haul and who will pay, then choose the "car" and "road," avoiding resource misallocation from the start.

Overall: Brand marketing and channel touchpoints are changing: AI search is becoming a new entry point, and businesses need to make their brands searchable, citable, and recommended by generative search; feedback data from content platforms is more valuable than the volume of public content.

1. Brands operate on Amazon, advertise on Google, YouTube, and other channels, and are starting to study generative engine optimization. Products and brands must adapt to the recommendation logic of AI search.

2. Backend metrics such as completion rate help determine which content works, making it easier for platforms to form a "content-feedback-retraining" loop. Brands should also feed user feedback back into product development and advertising.

Overall: Product R&D and price competition are also being reshaped by AI, and brands must use a "define the cargo first, then choose the car and road" mindset.

1. New AI hardware categories include AI glasses, AI toys, AI wearables, AI home devices, AI companions, and AI education hardware. Product definition must incorporate cost, power consumption, latency, privacy, compliance, OTA, return rate, and data feedback.

2. Computing costs directly affect product cost and market competitiveness. CEOs should participate in technology selection like a business decision; channel strategies and pricing logic will change accordingly.

Overall: The most valuable takeaway for cross-border e-commerce sellers is that platform algorithms are becoming AI-driven, ending the era of "manual operations." The cost of scaling up through headcount and reaction speed will continue to rise.

1. Risk alert: sellers have long relied on platform algorithms from Amazon, Google, YouTube, and others, but as platforms accelerate their AI adoption, scaling execution through human experience will become increasingly expensive; switching costs for large models are low, so don't pin hopes on a single model.

2. Opportunity alert: generative engine optimization is becoming a new growth area. Products and brands can be indexed, cited, and recommended by generative search, and traffic entry points are being redistributed.

Overall: The real accumulative competitive advantage is a data loop, not isolated AI tools.

1. AI tools for images, video, customer service, and document processing only replace repetitive work; they are point-based efficiency improvements.

2. Only when user judgments from product selection feed into content production, advertising feedback returns to product development, and customer service issues become input for the next round of operations and product improvement does a true data loop form.

3. The most realistic near-term task is to first figure out what "cargo" you want to haul, then decide which "car" and "road" to use.

Overall: Factories can see new product opportunities in the globalization of AI hardware: AI glasses, AI toys, AI wearables, AI home devices, AI companions, and AI education hardware are all emerging "cargo." Product definition should come before the "car" and "road."

1. Design and production must simultaneously consider cost, power consumption, latency, privacy, compliance, OTA, return rate, and data feedback; these variables determine whether AI hardware can be mass-produced and build user stickiness.

2. Different applications will in turn select computing power. For example, code generation requires sustained, stable, and cost-controlled computation, which may give rise to chips and system solutions beyond general-purpose GPUs, creating new orders for the supply chain.

Overall: AI is also changing factories' digitalization and e-commerce logic.

1. Enterprises need to connect user feedback, content data, and product development to form a "content-feedback-retraining" data loop, rather than using AI tools in isolation at each stage.

2. The adoption cycle for new technology is shortening, and CEOs need to be directly involved in technology selection. Factories making AI hardware must also speed up coordination among R&D, production, and channels, or they will fall behind in product cadence.

Overall: The core trend service providers can capture is that clients no longer only ask "which model is strongest," but need help sorting out "what cargo to haul, which car to choose, and what road to build."

1. Industry trends show the model leadership window has shortened to about three months, and users can switch models at low cost. Solutions must be portable and replaceable, not locked into a single model.

2. AI applications fall into three categories with different needs: code generation requires sustained, stable, low-cost computing; video generation requires storage and user-behavior data loops; conversational bots require long-term inference resources and entry-point defense design.

Overall: Client pain points and solutions are already clear.

1. Many enterprises still use "manual" tactics against platform algorithms, and using AI in isolation at each stage is only point-based efficiency. The pain point is that data across product selection, content, advertising, and customer service is not linked.

2. Providers can offer a "cargo-car-road" diagnosis, first helping clients identify applications that can generate revenue, data moats, or entry-point value, then designing model selection, infrastructure deployment, data feedback, and compliance solutions.

3. The globalization of AI hardware also creates service opportunities around latency, privacy, compliance, OTA, return rate, and data feedback. Each variable can be translated into a concrete product.

Overall: The most important signal for platforms is that AI will redistribute traffic entry points, and conversational bots may become the new entry point where users initiate searches, purchases, and service requests.

1. Search is hit more directly, social relationships are relatively stable, and e-commerce sits between search and social. Platforms should treat conversational bots as an entry-point defense investment, not just a source of short-term subscription revenue.

2. Model switching costs are low, and users can change "cars" at any time. Platforms should not hard-wire their core to one large model; they need to maintain flexibility at the application layer.

Overall: Platforms with content and user-behavior data have a unique advantage, but they also need to accelerate "road" construction.

1. Platforms hold backend data such as completion rates, enabling them to judge which content works and more easily form a "content-feedback-retraining" loop; data quality matters more than data quantity.

2. As users and "cargo" become more complex, transmission, storage, and computing power all need upgrading. Storage is like a parking lot, computing power is like dispatch and scheduling capacity. Building infrastructure early can create a long-term moat.

Overall: This article provides a reusable industry analysis framework: "car, road, cargo" separates large models, infrastructure, and applications, explaining the evolution from AI technical breakthroughs to commercial realization.

1. New dynamics: the model leadership window has shortened from six months to about three months, user migration costs are low, and betting on a single technical route is riskier. Training scale and rankings still matter, but the evaluation criterion is shifting to "whether people pay for applications."

2. Three business models: code generation corresponds to efficiency revenue, video generation to data loops, and conversational bots to entry-point defense. They should not be conflated.

Overall: New industry questions and governance implications also emerge.

1. Applications are starting to define vehicle types and roads in reverse. Different tasks impose different demands on training, inference, storage, and transmission, and the roles of storage and computing may need to be reassessed.

2. The adoption cycle for new technology is shortening, and CEOs are participating in technology selection, indicating that computing decisions have become business decisions. The globalization of AI hardware also involves cost, power consumption, latency, privacy, compliance, OTA, return rate, and data feedback. Professional services face new compliance, liability, and verification thresholds, warranting further research into relevant policies and standards.