a16z Global Market Insights: Trillion-Dollar Capital Expenditures, Evolution of AI-Native Architecture, and Full-Cycle Investment Landscape

The real major theme of this material is not "a16z is optimistic about AI," but a larger judgment: The past 15 years have been about "software eating the world," while the next 15 years are more likely to be about "AI driving capital back into the physical world." Chips, electricity, data centers, networks, manufacturing, robotics, defense, autonomous driving, and biomedicine will once again become the most important battlegrounds for technological capital. a16z has recently even established an $1.1 billion Machine Age Fund to invest in physical infrastructure such as chips, memory, networks, storage, data centers, robotics, electricity, and cooling. They explicitly state that AI is evolving from a software issue into a massive construction of physical infrastructure. Andreessen Horowitz But before diving into a deeper interpretation, let me first correct a few numbers in the material. First, the statement "the AI bubble has been debunked" is overstated. From the launch of ChatGPT on November 30, 2022, to September 2026, the S&P 500 indeed rose from about 4,080 points to about 7,600 points, nearly 90%. This number is basically valid. YCharts However, saying "the overall market P/E has fallen below 20 times" cannot be written like that. The current overall valuation of the S&P 500 is not universally below 20 times; some market indicators are still above 20 times. So a more accurate conclusion should be: Today, we cannot simply compare it to the 1999 internet bubble, but we also cannot prove that AI assets have no valuation risks. The real difference is: In 1999, many companies were: Valuations rose first, profits were discussed later. Today, leading companies like NVIDIA, Microsoft, Meta, Amazon, and Google at least have: Real revenue + real profits + real cash flow + real AI demand. This is a capital expenditure cycle supported by fundamentals. But: Having fundamentals ≠ any price is reasonable. These two concepts must be separated. ──────────────── 1. The most important judgment: AI is ending the golden age of "light asset technology" What has been Silicon Valley's most perfect business model in the past twenty years? Software. Why? Because the marginal cost of copying a software user is extremely low. Salesforce writes a set of software: The 10,000th customer and the 1 millionth customer do not need to build 1 million factories again. This is why software can have: 70% 80% or even 90% gross margins. So Marc Andreessen famously said in 2011: Software is eating the world. But AI has brought an interesting reversal. Traditional software: The more it is used, the closer the unit cost is to zero. Generative AI: The more it is used, the more GPUs, electricity, memory, and network must be consumed first. This means that the technological economy has regained: Physical constraints. Tokens are not truly free. A ChatGPT request is not an abstract "cloud." But a whole industrial chain: Power plants Transmission Transformers Data centers Cooling GPUs HBM Network switches Optical modules Servers Models Applications So AI has pulled the internet economy back from: Bits to: Atoms + Bits. This is why what is most worth understanding now is not "which AI app is popular." But: Who owns the indispensable means of production in the AI world? ──────────────── 2. AI is actually reinventing "capital-intensive technology" The last generation of entrepreneurial myths was: A few young people. A few computers. An office. With $2 million. They could possibly create a billion-dollar company. Facebook, Instagram, WhatsApp, Dropbox all belong to this era. AI infrastructure, however, is completely different. Today, training cutting-edge models may involve: Billions of dollars in computing power. Building large data centers: Billions of dollars. Advanced chip factories: Hundreds of billions of dollars. Electric power systems: Even decades of infrastructure assets. This means: Technology and capital markets are being deeply re-bound. This is very important. Because future technological competition is not just about: "Whose engineers are smarter?" It also includes: Who has the strongest financing ability? Who can get electricity? Who can buy GPUs? Who can secure HBM? Who can quickly build data centers? Who can obtain land and permits? Who can bear the upfront investment of billions of dollars? In other words: AI has turned technology back into industry. ──────────────── 3. This is the true meaning of "The Age of Atoms" Many people understand: AI = Chatbot. This is the shallowest layer. The real big trend is: AI will eventually enter: Factories Cars Military Logistics Energy Mines Agriculture Medical equipment Robots a16z is increasingly emphasizing the so-called electro-industrial stack: Electrification + materials + AI + industrial equipment + software control systems. It views minerals, power electronics, batteries, motors, sensors, and software as part of the same industrial stack. Andreessen Horowitz This is a huge paradigm shift in capital. The past internet solved: The information problem. For example: How to find a taxi? Uber. How to find a hotel? Booking. How to buy things? Amazon. How to disseminate information? Google / Facebook. The next stage will start solving: The physical execution problem. Not just: "Tell me what taxis are nearby." But: The car comes to pick me up by itself. Not just: "Tell me the warehouse inventory." But: The robot moves the goods by itself. Not just: "Help me design parts." But: AI controls the factory to produce by itself. This is the transition from: Software Economy to: Machine Economy. ──────────────── 4. Why does trillion-dollar CapEx not necessarily mean a bubble? Here we must learn a very important financial concept: The quality of capital expenditure cannot be judged by the amount; it must be judged by the capital return rate. Assuming Amazon builds a $10 billion data center. The question is not: "Is $10 billion too much?" But: How much can this $10 billion generate in the future: Revenue? EBITDA? Free Cash Flow? ROIC? If an investment of $10 billion generates $3 billion in free cash flow each year in the future, this could be a very good investment. If it can only generate $300 million, then it is a disaster. So what should really be observed in AI CapEx is: Incremental ROIC. Not absolute CapEx. a16z's latest material even states that the capital expenditure of just five hyperscalers in the U.S. may reach about $1 trillion next year. Andreessen Horowitz What does this mean? The global capital markets are engaged in a huge gamble: Can the revenue and productivity improvements brought by AI in the future support the computing assets built today? This is the core question of the entire AI bull market. ──────────────── 5. To understand this cycle, one must learn the "railroad analogy" AI infrastructure is very similar to the 19th-century American railroads. During the railroad construction period, the U.S. invested huge amounts of capital: Land Steel Bridges Stations Tracks Many railroad companies eventually went bankrupt. However: The railroads themselves changed the American economy. This is the most interesting rule in the history of technology investment: Technological revolutions can be real, but investors can still lose money. Railroads were real. Many railroad stocks died. The internet was real. Pets.com died. Fiber networks were real. Global Crossing went bankrupt. Smartphones were real. BlackBerry was eliminated. So today, when facing AI, we must separate two questions: Question 1: Is AI a huge technological revolution? More and more evidence points to: Yes. Question 2: Are the valuations of all AI companies today reasonable? That is a completely different question. This is the most important understanding for investors. Technology ≠ Security Price. Great technology does not necessarily mean all related stocks are great investments. ──────────────── 6. Why does "cheaper computing power lead to greater demand for computing power"? The most valuable economic concept in this material is: Jevons Paradox. 19th-century British economist William Stanley Jevons discovered: After the efficiency of steam engines improved, coal consumption did not decrease. Instead, it increased. Because: When coal became more economical, more industries began to use steam power. AI is the same. Many people think: As models become cheaper, the demand for AI infrastructure will decrease. Not necessarily. Assuming an AI task previously cost: $10. Now it drops to: $0.10. Companies will not just think: "Great, I saved $9.90." Companies will start asking: What tasks did I not let AI do before because they were too expensive? Then demand begins to explode. ──────────────── 7. A very simple example of AI economics Assuming a law firm previously had: 10,000 contracts. If the cost for AI to check one contract is: $20. The total cost is: $200,000. So the company might only check the most important: 1,000 contracts. If the model cost drops 100 times: $0.20. Then: Checking all 10,000 contracts: $2,000. Suddenly: Demand that did not exist before appears. Then the law firm might further think: Why check once a year? Can we check daily? Can we monitor in real-time? Can we continuously analyze: Contracts Emails Legal changes Court cases At this point, token usage might actually increase: 100 times. Even: 1,000 times. So what model companies most hope to see is actually: Decreasing unit reasoning costs + exploding total token consumption. Similar to historical: Bandwidth. Storage. Compute. Unit prices have been continuously declining. But global total consumption has been steadily increasing. ──────────────── 8. This is also why the "strongest model" may not necessarily be the biggest commercial winner. This is a point that entrepreneurs should learn from. Many people's first reaction when creating AI products is: "I want to use the best model." This is the wrong mindset. Real enterprise-level AI should ask: How strong of a model is needed for this task? For example: A customer asks: "When will my order arrive?" You don't need the smartest large model in the world. A small model is sufficient. But if a customer asks: "Please analyze the significant legal risks in this 200-page merger agreement." You might need the strongest model. Thus, we have: Model Routing. Simple tasks: Cheap models. Complex tasks: Expensive models. Extremely complex tasks: Multi-model validation. This is very similar to airline Revenue Management. Not every passenger is assigned to first class. What you optimize is: The overall economic efficiency of the system. So one of the biggest moats for AI applications in the future may not be: "Which model did I use?" But rather: How do I combine models? ──────────────── 9. The real profits from AI applications come from "Price–Cost Decoupling." This point is very important. Assume you are making an AI legal product. Customers are willing to pay: $100 for a complex legal analysis. Today your AI cost is: $30. Gross profit: $70. A year later, through: Caching Routing Fine-tuning Small models Distillation The cost becomes: $5. But customers are still willing to pay: $100. Then the gross profit increases from: $70 to: $95. This is: Price–Cost Decoupling. Businesses do not pay based on: "How many tokens did you use?" But rather based on: "How much value did you create?" This is where real huge profits may emerge in the AI application layer. ──────────────── 10. So what is the worst AI business model in the future? Very simple: Wrapper. Calling the OpenAI API. Putting on an interface. Then marking up the price. Because: The underlying models are getting stronger. Model companies can directly take over your functionality at any time. The strongest AI startups must have at least one of the following: Unique data. Proprietary workflow. Customer distribution channels. Industry integration. Regulatory capability. Brand. Network effects. High switching costs. Otherwise, as the model layer continues to improve: The value of your product may continue to be compressed. ──────────────── 11. Why is Vertical AI worth paying close attention to? Traditional SaaS sells: Software seats. Vertical AI starts selling: Labor outcomes. The difference is huge. Traditional software tells lawyers: "This is legal software, you operate it yourself." AI legal products in the future will tell law firms: "Give me the documents, and I will complete the work directly." Thus, the market boundary expands from: Software Budget to: Labor Budget. This is the real large TAM expansion in the AI application layer. Assuming in the U.S. a certain industry: Software spending is only: $10 billion. But labor spending is: $200 billion. The traditional SaaS TAM is: $10 billion. If AI starts to take on work, theoretically it enters: $210 billion economic pool. This is why the opportunity for AI software may far exceed that of the previous generation of SaaS. a16z has also publicly made similar judgments: AI applications are not simply competing for old software budgets, but may start to "eat labor." Andreessen Horowitz ──────────────── 12. The real situation of enterprise AI now is not whether there is adoption, but whether it has entered the core workflow. This is the key to distinguishing AI hype from AI value. When a company buys Microsoft Copilot for employees: It does not represent AI transformation. Employees occasionally: Summarize emails. Write PPTs. Polish text. This kind of value is very limited. The real second phase is: AI directly enters: Customer support Coding Sales Accounting Insurance claims Legal review Healthcare administration Procurement Once it enters these core workflows, AI begins to change: Headcount. Cost structure. Cycle time. Revenue. Error rate. At this point, CEOs and CFOs will truly start to care. So in the future, measuring enterprise AI should not ask: "How many employees use AI?" But rather ask: "How much economic activity is managed by AI?" This metric is much more advanced. ──────────────── 13. What should a boss really calculate? Not: How many users does AI have every day. But rather: Previously, a business process: 100 people × $100,000 cost = $10 million. After AI: 60 people + $1 million AI cost = $7 million. The company saves: $3 million per year. This is called: ROI. If the software company can take away: $1 million, The customer still nets: $2 million. Then both the software company and the customer win. This kind of AI will form a truly sustainable business model. ──────────────── 14. Why is public SaaS experiencing valuation reconstruction? Another very important data point in this material is basically established: a16z's latest statistics indicate that by 2026, about: 75% of public software companies will be profitable. But only about: 30% will maintain over 20% growth. a16z describes this as the result of "growth for profit" after the high-interest rate era. Andreessen Horowitz This is very interesting. Because SaaS is not dead. In fact: The financial quality is better than before. The real issue is: Growth Scarcity. The capital market is always willing to give high valuations for growth. But if you: Are profitable, But only grow: 8% 10% 12% The market will start to treat you as: A mature enterprise. And not as: A high-growth tech company. This is why many SaaS companies are not collapsing in business, But rather: Multiple Compression. ──────────────── 15. Why is real high growth increasingly staying in private markets? a16z has also observed: By 2026, among public software, internet, and FinTech companies, the market expects very few companies to have revenue growth exceeding 30%; meanwhile, there are still many high-growth companies in the private market. Andreessen Horowitz The underlying structure of the capital market has changed. Previously, companies needed to IPO: To raise money. To provide liquidity for employees. To allow VCs to exit. To build a brand. Now, the mature private market can provide: Billions of dollars in financing. Tender offers. Secondary liquidity. Private credit. Sovereign funds. Pensions. So companies can: Stay private longer. This is a significant structural change. ──────────────── 16. This is actually "privatizing the American growth economy." This is a very important trend in the financial market, but ordinary investors tend to overlook it. In the past, ordinary American investors could buy: Amazon very early. Microsoft very early. Google relatively early. Today, many companies may have already reached: $10 billion. $50 billion. $100 billion. Or even higher before going public. So a large amount of wealth creation has occurred in: The Private Market. This means that the future financial system may become increasingly divided: Ordinary residents: Public equities. Pensions, VCs, PE, sovereign funds, high net worth individuals: Private growth. If truly high-growth companies remain in the private market for a long time, Then: Who has access to the private market Will increasingly affect wealth distribution. This is a very significant institutional issue for the future of American capital markets. ──────────────── 17. Be very cautious about the conclusion regarding data centers and electricity prices. You wrote in your material: Data center capacity increases by 10%, residential electricity prices decrease by 0.4%. This figure indeed comes from a 2026 study. The study analyzed data from the U.S. from 2015 to 2024 and found that a 10% increase in data center capacity is associated with a decrease of about 0.4% in average residential retail electricity prices, and attempted to explain it using causal identification methods. Stack Futures The economic logic behind it is easy to understand: A large portion of the costs of the power grid are: Fixed Costs. For example: Transmission lines. Substations. Infrastructure maintenance. If large data centers continue to operate at high loads, These fixed costs can be spread over more kWh. Average costs may decrease. This is: Economies of Scale. However, it must not be written as: "Data centers do not push up electricity prices." Because the study itself also emphasizes: The historical results from 2015 to 2024 cannot guarantee that they will continue to hold in the future. If in the future: Electricity supply cannot keep up. Transformers are in short supply. Transmission congestion occurs. Gas prices rise. New power generation construction is too slow. Then data centers could very well push up marginal electricity costs. Emergent Mind a16z itself also acknowledges this risk in its energy policy article and advocates that large hyperscalers should bear the costs of grid upgrades they cause, rather than having residential users subsidize them. Andreessen Horowitz This is the complete answer. ──────────────── 18. The AI war may ultimately evolve into an "electricity war." Many people look at NVIDIA. But what they should really look at is one layer deeper: What do GPUs need? Electricity. So a very important bottleneck for AI in the end is not: Chip. But: Watt. a16z recently even wrote directly: "We are running out of watts." The scarce assets it listed include: powered land turbines transformers data centers GPUs advanced-node wafers HBM. Andreessen Horowitz This means that the energy industry may re-enter the core of the tech industry. Nuclear power. Natural gas. Geothermal. Energy storage. Grid software. Transformers. Transmission. In the past, many VCs thought these industries were: Too slow. Too heavy. CapEx too high. Now AI has turned these industries back into: technology bottlenecks. And investment history tells us: Bottlenecks are often profit pools. ──────────────── Nineteen, why will NVIDIA make money? Because it controls the bottleneck. Economic profits are usually not evenly distributed. Railroad era: Land and railroad nodes make money. Oil era: Oil fields and refining capacity make money. Internet era: Search portals make money. Mobile era: Operating systems and App Stores make money. Early AI: GPUs make money. Future bottlenecks may migrate. From: GPU to: HBM. Power. Networking. Cooling. Data centers. Robotics actuators. Even: Copper. So the smartest industry research always asks: What is the next constraint that limits the growth of the entire system? Instead of: "Who is the hottest company today?" ──────────────── Twenty, robots may be bigger than LLMs, this judgment is not crazy. Why? Because LLMs mainly enter: knowledge work. Robots enter: physical labor. The global labor economy is huge. Manufacturing. Warehousing. Logistics. Construction. Agriculture. Elder care. Catering. Household services. If AI can only help programmers write code, the market is already large. If AI can eventually: move. grasp. assemble. drive. clean. deliver. construct. Then the market directly enters: the global physical labor market. So the ultimate TAM of Physical AI could completely exceed that of pure language models. But the biggest difference is: LLM iterations: months. Robot iterations: involve mechanics, hardware, safety, supply chains, and are usually much slower. So: The potential market for robots is larger, but the commercialization speed may be much slower. This is the most important balance in analyzing robots. ──────────────── Twenty-one, the real value of autonomous driving is not "selling cars." This is where many people easily misunderstand Waymo and Tesla autonomy. The utilization rate of cars today is extremely low. The vast majority of private cars: 24 hours a day, may be parked for 22-23 hours. This is a huge capital waste. If autonomous driving truly matures: Vehicles can: operate continuously. Thus cars change from: Consumer Durable to: Productive Asset. Similar to: taxis. servers. airplanes. A car no longer just generates: "use value." But can generate: cash flow. Thus the industry structure may change from: OEM selling cars. to: Autonomous Fleet Network. This is a revolution at the business model level. ──────────────── Twenty-two, the American Dynamism behind it is actually "American Reindustrialization." a16z has increasingly focused on: Defense. Manufacturing. Aerospace. Energy. Robotics. Infrastructure. The reasons are very profound. The economic trend in the U.S. over the past few decades has been: Manufacturing outsourcing. Financialization. Softwareization. Asset lightening. China has accumulated a lot of: manufacturing supply chains. Industrial capabilities. Infrastructure capabilities. After the AI era, the U.S. suddenly rediscovered: You cannot only have: software. You also need: factories. chips. grid. drones. missiles. robots. manufacturing engineers. So the so-called: American Dynamism is essentially a: Technology-driven reindustrialization. This is also why companies like Anduril, SpaceX, and Palantir are so important. They represent a new type of American tech company: not: "making an app." but: rebuilding the country's industrial capabilities. ──────────────── Twenty-three, compress the entire a16z market judgment into an industrial chain. What you should really see is: AI model capabilities improve ↓ More AI tasks emerge ↓ Token demand explodes ↓ Inference growth ↓ GPU demand growth ↓ HBM, networking, server growth ↓ Data center growth ↓ Power demand growth ↓ Generation, transmission, cooling, transformer growth ↓ Manufacturing capital expenditure growth ↓ Robots and automation improve production capacity ↓ AI enters the real economy ↓ Labor productivity increases ↓ New applications and business models emerge This is not: "AI software industry." This is: a whole new capital formation cycle. ──────────────── Twenty-four, this is also why this round of AI cycles may affect U.S. GDP, not just Nasdaq. The biggest early impact of the internet in the 1990s: Information transmission. Today's AI infrastructure directly drives: Construction. Power. Manufacturing. Semiconductors. Real estate. Utilities. Engineering. So this investment cycle itself may contribute to: GDP. And this is an interesting macro change: In the past decade, U.S. growth has heavily relied on: consumption. Now AI may make: Private Fixed Investment become an important engine of growth again. If this trend continues long-term, The structure of the U.S. economy may gradually change from: Consumption-driven + digital services to: Consumption + digital services + large-scale technological industrial capital formation. ──────────────── Twenty-five, what should entrepreneurs really learn from this? Don’t just ask: "What AI products can still be made?" You should ask five more advanced questions: First, where is the biggest bottleneck in the entire industrial chain now? Electricity? Computing power? Data? Workflow? Talent? Regulation? Second, after costs drop by 100 times, what previously uneconomical behaviors will suddenly become economical? This is one of the best ways to find new markets. Third, am I selling software, or am I completing work? The latter's TAM is usually much larger. Fourth, does my value come from the model, or outside the model? If your product becomes useless after the model upgrades, then there is no moat. Fifth, what type of customers already have clear ROI, rather than just liking AI stories? The truly large enterprise software companies must ultimately enter: CFO's Excel. ──────────────── Finally, remember this entire logic in four sentences. The first sentence: The last round of technological revolution pulled capital out of the atomic world and invested in software; this round of AI revolution is reinjecting the intelligence generated by software back into the atomic world. The second sentence: The greatest economic value of AI may not be generating content, but transforming work that could only be done by humans into computable, replicable, and automatically executable production processes. The third sentence: The biggest winners in the future may not just be those who own the strongest models, but those who control bottlenecks, workflows, distribution, data, energy, and physical execution capabilities. The fourth sentence: If the core question of the last generation of entrepreneurship was "Can software eat this industry?" the real question for the next generation will be—"Can AI directly complete the work in this industry?" This is the core that this a16z material is most worth understanding for readers. It truly describes not an ordinary AI bull market, but a grander possibility: From the "digital world" of the internet era, entering the "automated world" of the machine age. And once this judgment is established, the biggest opportunities in the next decade will not only exist in interfaces like ChatGPT. It will spread down the entire industrial chain: Model → Chip → Data Center → Power → Manufacturing → Robotics → Automotive → Defense → Healthcare → the entire real economy. This is the true understanding of what is called the AI Supercycle.