The Real Question Behind the Technology Boom
Artificial intelligence is no longer merely a technology sector; it is a rewrite of the production function. [Fact] In 2025, global corporate AI investment reached 344.7 billion, with generative AI as the central destination of capital. These numbers show that capital markets no longer view AI as a laboratory technology, but as a core variable shaping enterprise efficiency, software architecture, and industrial structure over the next decade.
Yet AI has not entered maturity. [Inference] It is closer to the early stage of hypergrowth: technical capability has been validated, user demand has been validated, and capital investment is scaling, but commercial closure, stable margins, and organizational transformation are not yet fully proven. In other words, ChatGPT proved that people are willing to use AI; enterprise purchasing proved that companies are willing to pay for AI; but the industry has not yet fully proven that AI can reliably create measurable profit across most sectors.
That is exactly why AI is worth studying. It is not a finished story; it has just moved from model demonstration into workflow transformation. Over the next five to ten years, the biggest change will not be that models answer more like humans, but that AI begins replacing, compressing, and augmenting large parts of white-collar and engineering workflows. Software companies used to sell tools; in the future, they may sell outcomes. Enterprises used to buy SaaS seats; in the future, they may buy agents that execute tasks.
Industry Stage: Hypergrowth, but Not Risk-Free Prosperity
[Fact] Menlo Ventures data shows that enterprise generative AI spending reached 11.5 billion in 2024, representing roughly 3.2x year-over-year growth. More than half of that spending went to the application layer, showing that enterprises are not merely buying foundation models but are searching for software that directly improves productivity. McKinsey’s 2025 survey also shows that AI use has become widespread, but many companies remain stuck in pilots, partial deployment, and workflow adaptation.
[Inference] This means the AI industry has moved beyond the question of whether demand exists and into the question of whether value can be delivered at scale. Early internet companies had to prove that users would go online. Cloud computing companies had to prove that enterprises would migrate IT systems. AI companies now need to prove that they can embed themselves into enterprise workflows. This stage is characterized by rapid revenue growth, while cost, organizational friction, and product reliability remain key constraints.
[Prediction] Over the next five to ten years, AI’s core drivers will come from four directions: stronger reasoning capabilities, lower inference costs, enterprise workflow redesign, and the combination of multimodality with robotics. This forecast depends on continued model progress, no severe disruption in compute supply, and no full regulatory blockage of data and model usage. If these assumptions fail, AI will still grow, but growth will likely be much slower than current market expectations.
Business Model: Strong, but Value Will Not Be Evenly Distributed
The AI business model looks simple on the surface: enterprises, developers, and consumers pay for model capability, software features, or automated outcomes. The harder question is value distribution. The current value chain includes AI chips, cloud platforms, foundation models, AI applications, and vertical agents. Every layer can make money, but margins, moats, and capital intensity differ dramatically.
[Fact] Chip companies such as NVIDIA benefit from training and inference demand, with strong supply constraints and ecosystem advantages. Cloud providers such as Azure, AWS, and Google Cloud absorb compute demand. Model companies such as OpenAI, Anthropic, and Google DeepMind sell APIs, enterprise subscriptions, and platform capabilities. Application companies wrap AI into concrete use cases such as coding, legal work, customer support, design, financial analysis, and office collaboration.
[Inference] Long-term profits are most likely to come from two types of companies: those that control critical infrastructure such as GPUs, cloud, and data centers, and those that deeply embed AI into enterprise workflows. Simply having a “smarter model” may not create a permanent moat, because model capabilities may converge, prices may fall, and open-source models will continue to compress the premium on base intelligence. The harder-to-copy advantages are distribution, enterprise data, workflow integration, reliability, security, compliance, and user habits.
The biggest debate is whether AI will strengthen SaaS or destroy it. One view argues that AI will replace traditional software because users will no longer need to open many tools; they will simply ask agents to complete tasks. Another view argues that AI will increase software value because incumbents already control customers, data, and workflow entry points. My judgment is that both will happen: weak SaaS will be compressed, strong SaaS will become AI-native, and new AI-native software companies will emerge in vertical markets.
Market Space: Not an Industry, but a New Infrastructure Layer
AI’s market space cannot be measured only by AI software revenue. Looking only at model subscriptions and API revenue underestimates its impact; counting every industry that AI may transform overestimates short-term revenue. [Fact] Stanford AI Index 2026 shows that global corporate AI investment exceeded $580 billion in 2025, indicating that AI-related capex, software procurement, hiring, and infrastructure investment have become a systemic trend.
[Inference] AI’s TAM should be viewed in three layers. The first layer is infrastructure, including GPUs, data centers, electricity, cloud services, and networking. The second layer is software, including AI office tools, AI coding, AI customer support, AI marketing, AI legal work, and AI financial analysis. The third layer is the physical economy, including robotics, autonomous driving, manufacturing, healthcare, energy, and logistics. The first layer monetizes first, the second layer determines margins, and the third layer determines whether AI can become a multi-trillion-dollar long-term market.
[Prediction] Over the next decade, the biggest incremental opportunity will not be chatbots, but enterprise agents and Physical AI. Enterprise agents will rewrite white-collar workflows, while Physical AI will move intelligence from screens into factories, warehouses, hospitals, and homes. This judgment assumes AI system reliability continues to improve, robotics hardware costs decline, and enterprises are willing to redesign workflows. If agents cannot reliably execute complex tasks, or robotics hardware progresses more slowly than expected, AI’s market will still be large, but the path of monetization will remain more software-heavy than physical-economy-heavy.
Competitive Landscape: Models Are Not the Only Battlefield
The current global AI competitive landscape can be divided into four groups. The first group is infrastructure winners, represented by NVIDIA, AMD, Broadcom, and data center and cloud infrastructure companies. The second group is cloud and platform companies, represented by Microsoft, Amazon, and Google. The third group is foundation model companies, represented by OpenAI, Anthropic, Google DeepMind, Meta, and xAI. The fourth group is application-layer companies, including Cursor, Perplexity, Harvey, Glean, ServiceNow, and Salesforce.
[Inference] The most stable winners are still at the infrastructure layer, because all models and applications require compute. NVIDIA’s moat is not only its GPUs, but CUDA, developer habits, supply chain capability, and customer lock-in. Cloud companies benefit from enterprise customer access and compute resources. Model companies benefit from technology, brand, and developer ecosystems, but competition at the model layer will be brutal because training costs are high, prices fall quickly, and open-source alternatives keep improving.
The most uncertain layer is applications. The application layer may produce the largest number of new companies because it is closest to customer budgets and concrete ROI. But it is also the layer most easily copied by large platforms. If an AI application is merely a chatbot added to old software, its moat is weak. If it rewrites a high-value workflow such as contract review, code generation, sales follow-up, insurance claims, medical documentation, or financial research, it has a real chance to become a durable company.
Fewer Than Ten Key Metrics
- [Fact] Global corporate AI investment reached about $581.7 billion in 2025, up 130% year over year.
- [Fact] Private AI investment reached about $344.7 billion in 2025, up 127.5% year over year.
- [Fact] Enterprise generative AI spending reached about $37 billion in 2025, roughly 3.2x higher than 2024.
- [Fact] More than half of enterprise AI spending went to the application layer, showing demand shifting from models to productivity tools.
- [Fact] The number of newly funded AI companies grew by about 71% in 2025.
- [Fact] The United States remains the center of AI investment, with private AI investment far ahead of China.
- [Fact] McKinsey shows that AI usage has broadly diffused, but scaled deployment and financial returns remain uneven.
- [Inference] Compute, electricity, and data center capex are becoming hard constraints on AI growth.
- [Inference] Enterprise AI procurement is moving from “let’s try it” to “can it produce measurable ROI?”
- [Prediction] The most important new demand over the next decade will come from agents, AI-native software, and robotics.
The Five Rules
First, AI is close to an explosion, but not every AI company is. [Fact] Capital, enterprise procurement, and user adoption are rising together. [Inference] The real explosion is not chat, but AI entering enterprise workflows. The test is not model launch events, but whether enterprises are willing to keep paying to save labor, improve efficiency, and shorten delivery cycles.
Second, AI has the potential to grow for more than ten years. [Inference] The reason is not one popular application, but its ability to enter nearly all knowledge work and part of physical work. Office work, programming, customer support, education, law, healthcare, manufacturing, and finance will all be affected. As long as model capabilities improve and costs decline, this growth cycle is unlikely to end quickly.
Third, AI is easy to distribute, but high-quality delivery is hard to copy. [Inference] Software can be distributed globally, while APIs and cloud services lower the barrier to entrepreneurship, which means low-end AI applications will become highly commoditized. The hard part is embedding AI into complex workflows and making it stable, controllable, compliant, and auditable. The future moat is not “I connected to a large model,” but “I can continuously produce results in real business operations.”
Fourth, AI meets real long-term demand. [Inference] Enterprises always pursue two things: lowering costs and increasing revenue. AI hits both. It can reduce repetitive labor and improve sales, R&D, customer service, and operations. This is not short-term entertainment demand; it is long-term productivity demand.
Fifth, AI is suitable for long-term commitment, but not blind commitment. [Inference] Someone who only learns prompt tricks may quickly be replaced by tool iteration; a company that only wraps a model API can be copied easily. The areas worth committing to are AI system design, agent workflows, industry know-how, data engineering, AI infrastructure, robotics, and AI safety governance.
The Biggest Opportunities and Risks
The first major opportunity is enterprise agents. Their value is not answering questions, but completing tasks. Customer support, sales, finance, legal, HR, operations, and data analysis all contain many standardized workflows that still require judgment. If agents can execute these workflows reliably, enterprise AI spending will move from innovation budgets into core operating budgets.
The second opportunity is AI-native software. Traditional software is human-centered, while AI-native software is task-centered. In the future, users may no longer click through buttons one by one; instead, they will describe a goal and let the system mobilize data, tools, and workflows to complete the task. This change will rewrite the product structure and pricing logic of many software companies.
The third opportunity is robotics and AI in the physical economy. Digital AI solves problems in information flows, while Physical AI solves labor, movement, and manipulation problems in the real world. If AI can enter warehouses, manufacturing, healthcare, home services, and energy systems, its market space will become much larger than pure software.
The risks are equally clear. The first is cost risk: compute, electricity, data centers, and inference costs may compress profits. The second is competitive risk: once model capabilities converge, price wars may make it difficult for many model companies to earn excess returns. The third is regulatory and safety risk: copyright, data privacy, hallucinations, accountability for automated decisions, and national security reviews may all slow down AI deployment.
Institutional Investment Judgment
From an institutional investor’s perspective, artificial intelligence deserves long-term attention, but the investment thesis cannot stop at “which model is the strongest.” Model capability matters, but long-term returns usually come from the most irreplaceable positions in the value chain. The current AI industry resembles a combination of early cloud computing and early mobile internet: infrastructure monetizes first, platforms redistribute traffic, and the application layer eventually produces new giants.
The companies I would pay the most attention to fall into three groups. The first group consists of infrastructure companies that control compute, cloud, chips, and data center resources, because they are the toll roads of AI growth. The second group consists of software companies that can embed AI into core enterprise workflows, because they are closest to budgets and measurable ROI. The third group consists of robotics and industrial AI companies, because they determine whether AI can move from the digital world into the real economy. By contrast, I would be cautious about applications with no distribution, no proprietary data, no workflow depth, and no real differentiation beyond wrapping a general-purpose model.
For entrepreneurship, the best direction is not to build another general chatbot, but to choose a vertical scenario that is high-frequency, high-value, workflow-heavy, and outcome-verifiable. Examples include legal contracts, insurance claims, medical documentation, sales operations, financial auditing, code migration, game content production, and industrial maintenance. Startups should avoid entry points that large platforms can easily copy and instead enter areas that require domain knowledge, accumulated data, and long-term customer trust.
For career choice, the most valuable roles in the future will not be simple Prompt Engineer roles, but AI product manager, agent system designer, AI application engineer, data infrastructure engineer, AI safety and governance specialist, robotics software engineer, and industry AI solution architect. These roles share one thing: they are not just about “using AI,” but about understanding business, systems, data, reliability, and commercial outcomes. The people who can connect AI capability with real business value will capture more durable opportunities than those who only follow tool-level trends.
If I could choose only one direction to commit to for the next ten years, I would choose artificial intelligence, but not vague “AI” in general. I would choose the intersection of enterprise agents, AI-native software, and Physical AI. The reason is that this path connects technological progress, enterprise budgets, and real productivity improvement at the same time. The biggest change brought by AI is not that machines can chat, but that software is beginning to turn from a tool into labor. Once this becomes true, the cost structure, organizational structure, and job structure of the entire business world will be repriced.