Introduction

Over the past decade, artificial intelligence has made the leap from technological breakthroughs and product explosions to large-scale expansion of infrastructure. Today, the industry is moving beyond the phase of "just increasing investment to receive valuation rewards" and entering more serious operational tests: Can models consistently generate revenue? Can data centers improve utilization? When will the massive capital expenditures be recouped? And as GPU supplies gradually improve, how will electricity, grid connections, cooling, and engineering capacity become the new scarce resources?

This article is based on a course sharing about the investment logic in the AI industry and insights from Silicon Valley, attempting to organize the most noteworthy changes in the second half of AI from the perspectives of industry cycles, value chains, storage prices, competition among models, energy constraints, and Physical AI.

Core Summary

• The AI industry can be summarized as three stages of evolution: technological breakthroughs and the advent of visual AI from 2016 to 2021, valuation clearing and the productization inflection point of GPT-3.5 in 2022, and the AI infrastructure construction cycle since the end of 2022.
• The end of 2025 will become an important watershed: the market shifts from rewarding "bold capital expenditure" to requiring companies to prove revenue, cash flow, utilization, and return cycles.
• The AI industry chain can be divided into downstream models and applications, midstream cloud and computing power services, and upstream semiconductors and energy. The pricing power at the model level determines whether costs and profits can be transmitted upstream.
• The storage price increase cycle indicates that a rise in prices does not automatically equal long-term opportunities; when the terminal's bearing capacity reaches its limit, demand contraction, alternative solutions, and new supplies will dampen prices and valuations.
• The computing power bottleneck is shifting from chip supply to energy infrastructure. The future expansion of data centers will depend not only on GPUs but also on power access, transmission and distribution, cooling, grid connection, and construction capacity.
• The competition for large models is diverging into three routes: extreme performance, extreme cost-effectiveness, and vertical scenarios; the real decisive factor is whether a sustainable unit economic model can be formed.
• Humanoid robots and Physical AI still face issues of insufficient three-dimensional world data and limited generalization capabilities, making them more suitable for standardized, controllable scenarios in the short term; autonomous driving is closer to scaled commercial operations.

From AlphaGo to Large Models: Three Discontinuities in AI

Looking back at the development of the AI industry over the past decade, it can roughly be divided into three stages.

First Stage: Technological Breakthroughs and the Fall of Visual Intelligence

In 2016, AlphaGo's victory over Lee Sedol made the public intuitively aware of the possibility of artificial intelligence surpassing humans in complex tasks. Subsequently, interest in the AI industry rapidly increased, but for a considerable time, commercialization mainly focused on computer vision.

During this period, companies like SenseTime and Megvii became typical representatives of visual AI. The market recognized the technological value of AI but was still looking for a wider range of product forms and commercial scenarios. It can be said that AI at that time was "smart enough," but there was still some distance from serving ordinary users on a large scale and with low barriers.

By around 2020, models like GPT-3 had made researchers aware of the potential of generative AI, but ordinary users had not yet truly perceived its changes. The reason is not complicated: the technical capabilities were in rudimentary form, yet product interactions had not matured.

Second Stage: Valuation Clearing and the GPT Moment

In 2022, global liquidity tightened, and the valuations of risk assets fell, leading the market through a broad "de-bubble" process. For the AI industry, this round of adjustments was not just a cooling of the capital markets, but also created space for the next industrial cycle.

The real turning point occurred at the end of 2022. GPT-3.5 entered the public eye with a more user-friendly chat interaction mode, and generative AI transitioned for the first time from research results to products that could be experienced at scale. The combination of technological breakthroughs, improved product maturity, and macroeconomic changes constituted an important starting point for a new round of AI cycles at this time.

Third Stage: Expansion of AI Infrastructure and ROI Testing

Starting from the end of 2022, AI entered a phase of infrastructure construction. Data centers, GPUs, CPUs, storage, network connectivity, cabinets, and supporting power systems all became key investment directions in the industry chain.

In the first half of this stage, the market focused more on whether companies were sufficiently proactive in expanding capital expenditures: building more data centers, purchasing more chips, and deploying more computing power were often seen as necessary actions to seize the future.

However, by the end of 2025, the market's evaluation criteria began to change. Companies were no longer recognized just for "spending more;" investors began to question more specific issues:

• Can capital expenditures be converted into revenue?
• Is the utilization rate of data centers sufficiently high?
• Can equipment depreciation, financing costs, and electricity costs be covered?
• Will the prices of computing power leasing be maintained?
• Do model services have the ability to raise prices continuously?

Thus, AI entered the second half of "accounting." Growth is still important, but the quality of growth, return cycles, and cash flow capabilities have become even more critical.

Re-understanding the AI Industry Chain: Models, Cloud, and Energy

Understanding AI investment logic requires a unified perspective on the industry chain.

Within this framework, model vendors are the end that directly faces users, including general large models, industry models, and AI application service providers. They determine whether AI capabilities can be converted into real income.

Cloud vendors act more like "commercial real estate developers in the digital age": first investing funds to build data centers and then providing computing power as cloud services or rentals to customers. Their business model is not mysterious; the core lies in balancing investments, utilization rates, leasing prices, depreciation, and payback cycles. A healthy cloud business needs to be able to recover capital investments within a relatively clear period.

Meanwhile, semiconductors and energy form the upstream foundation of the industry chain. Over the past few years, GPUs, storage, servers, and network devices have become the most focused components; however, as data center scales rapidly expand, new constraints are increasingly evident: chips may not be the only scarce resource; stable, accessible, and scalable power is becoming a more significant bottleneck.

"Full Industry Chain Inflation" is Being Broken

The reason the AI infrastructure cycle can bring about significant industrial opportunities fundamentally relies on a mechanism similar to "full industry chain inflation."

Taking traditional real estate cycles as an example: when downstream housing prices and land prices continue to rise, developers can bear higher costs for cement, steel, equipment, and labor. When downstream prices can increase, profits and price spaces gradually transmit to midstream and upstream, ultimately forming profit expansion across the entire industry chain.

The AI industry is no different.

If model vendors can continuously raise service prices or expand income through higher value products and greater usage, then cloud vendors can endure higher computing power construction costs, and upstream GPU, storage, server, and energy suppliers will find it easier to achieve higher prices and profits.

However, the problem is: do model services really possess long-term pricing power?

As Chinese model vendors accelerate their catch-up, open-source model capabilities improve, and global competition among large models intensifies, the price war at the model level is weakening this premise. If model capabilities gradually converge and service prices continuously decline, the industry chain may shift from "both volume and price rising" to "volume increases without price increases."

This will directly change investment judgments: demand growth still exists, but valuation expansion and profitability enhancement may not occur simultaneously.

Insights from the Storage Cycle: A Price Increase Does Not Equal a Long-Term Opportunity

The storage industry is a typical example for understanding cyclical fluctuations.

At the bottom of the cycle, the market often displays languid prices, bleak trading, and lack of interest from investors; however, when demand improves, inventories decline, and supply expansion is insufficient, prices can surge rapidly. In the early stages of a price increase, the market usually provides higher expectations and valuations.

But prices cannot rise indefinitely.

The course used consumer-grade memory as an example to illustrate that when the prices of end products rise to levels that users find difficult to bear, demand will conspicuously contract. For consumers, excessively high hardware prices do not automatically lead to more purchases, but may instead cause users to delay upgrades, reduce configurations, or even forgo related consumption.

This reflects a simple economic principle: price increases on the supply side may ultimately trigger a collapse on the demand side.

For the enterprise market, the situation may be somewhat different. Large cloud vendors and data center clients have a stronger rigid demand for high-performance storage and may still accept higher prices in the short term. However, once prices become excessively high, they will also incentivize customers to look for alternative suppliers, optimize software architecture, adjust procurement rhythms, or drive new capacity into the market.

Therefore, assessing cyclical goods cannot be limited to "whether prices are rising," but must also consider three questions:

• Is the price increase driven by real demand growth, or short-term supply disruptions?
• Does the downstream have the ability to continue passing on costs?
• Will high prices harm terminal demand, or attract new supply into the market?

From this perspective, the storage industry is likely to enter a phase of stable demand, rational pricing, where returns are more reliant on dividends and buybacks, rather than continually depending on valuation increases for achieving high elastic growth.

Computing Power Bottlenecks are Shifting from "Chip" to "Electricity"

Over the past two years, when discussing AI, the most common question has been "Are there enough GPUs?" However, observations from the frontlines of Silicon Valley show that real constraints are shifting: increasingly, the question is no longer whether chips are available, but whether there is sufficient electricity, transformation facilities, and grid connection capabilities to enable these chips to actually operate.

The construction of a large AI data center goes far beyond merely purchasing servers. It also requires:

• A stable and sufficiently large power supply.
• Transmission, transformation, and distribution infrastructure.
• Cooling systems and engineering capacity for data centers.
• Land, approvals, grid interconnections, and construction timelines.
• Long-term predictable energy costs.

This means that even if GPUs are in place, if the data center cannot timely connect to the grid, the computing power assets cannot be converted into revenue.

This also explains why the focus of AI infrastructure investment is shifting from purely semiconductors to energy and electrical equipment: power generation capacity, grid upgrades, transformers, transmission and distribution equipment, backup power, cooling systems, and the engineering capacity of data centers may all become important variables that affect the speed of industry expansion.

For investors and industry participants, this signifies an important change: the future excess returns on the AI industry chain may not just come from "selling computing power," but also from "actually powering up the computing power."

Competition Among Large Models: Differentiation in Performance, Price, and Engineering Capability

The large model industry is experiencing clear differentiation.

One type of company persists in pursuing the strongest model capabilities, maintaining a leading edge through larger training scales, higher inference costs, and more aggressive technical routes. This path requires continuous massive capital investments and depends on stronger performance to build brands, users, and pricing power.

Another type of company places greater emphasis on engineering efficiency and cost-effectiveness: not necessarily pursuing parameter scale or leaderboard supremacy every time, but focusing on whether models are sufficiently user-friendly, inference costs are low enough, developers are willing to integrate, and enterprises are willing to purchase.

There is no absolute superiority or inferiority between these two routes; the key is whether the business model can close the loop:

• The extreme performance route needs to prove that users are willing to pay a premium for leading capabilities.
• The extreme cost-effectiveness route needs to demonstrate that cost advantages can be converted into scale and profits.
• The vertical application route needs to show that industry scenarios can create higher switching costs and willingness to pay.

What deserves attention is not the frenzy when models are released but whether model capabilities can generate sustainable revenue, retention, and unit economic benefits.

Physical AI is Still Early Stage, Autonomous Driving is Closer to Commercialization

"Physical AI" and humanoid robots are currently among the most closely watched cutting-edge directions, but their challenges are far more complex than training a language model.

Language models primarily handle internet data such as text, images, code, and video; but robots need to complete perception, understanding, planning, and execution in the real three-dimensional world. This means they must not only "see" objects but also understand distance, direction, force, friction, gravity, touch, spatial relations, and the consequences of actions.

For example, just because a robot learns to open a door does not mean it can close it successfully; learning to pick up a cup does not guarantee it can operate stably in different materials, weights, and tabletop environments. The core challenge lies in insufficient generalization capabilities and the lack of high-quality three-dimensional, multimodal, embodied interaction data.

Therefore, robots are more likely in the short term to land in highly standardized, repetitive, and relatively controllable scenarios, such as factories, warehousing, logistics, cleaning, and specific service processes, rather than immediately becoming universal assistants that adapt to any home environment.

In contrast, autonomous driving has come closer to scaled commercial deployment under specific regions, weather conditions, and operational requirements. Its essence is also Physical AI, but it has clearer task boundaries, denser data sources, and more direct paths to commercialization. In the future, the development of autonomous driving will continue to depend on safety, costs, regulations, and operational efficiency rather than a single technology indicator.

Conclusion: The Opportunities in AI are Not Over, Just the Evaluation System has Changed

The construction of AI infrastructure is not yet complete, but the phase of "expansion means value" is coming to an end.

The upcoming industrial opportunities are more likely to focus on segments that can solve actual constraints and create real cash flow: more efficient models, more controllable inference costs, higher utilization rates of data centers, more stable energy supplies, and intelligent systems capable of long-term operation in the real physical world.

For industry participants, the most important capability is no longer just telling technical stories, but answering a more realistic question: how does this technology ultimately create revenue, reduce costs, and maintain returns amidst ongoing competition?

As AI enters the "accounting era," the real watershed is not who has more concepts, but who can turn concepts into replicable, deliverable, and sustainable business results.

Note: This article is based on course recordings and summaries; the market judgments, industry views, and cases mentioned only represent the analysis framework of the speaker and do not constitute investment advice.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。