Last week I attended an event for public pension funds. I was there to talk about AI and crypto. Having given a similar presentation to endowments and foundations in late January, what struck me most was not simply how much the AI narrative had changed in three months. It was how much my own life with AI had changed in just three months.
On the JetBlue flight to the event, I spent nearly the entire five hours working on my laptop, going back and forth between large language models and my AI assistant, OpenClaw, back in Brooklyn. At the endowment and foundation event earlier this year, OpenClaw was just gaining global attention. Software stocks and other industries were under attack from fears of obsolescence driven by Claude. It was on that trip that, because of OpenClaw, I ordered the first of many Apple hardware purchases since then. By the time I stood in front of the pension audience last week, the shift was obvious. AI progress is a locomotive, and three months of change is almost impossible to comprehend. AI was no longer just the operating system of how I work. I now had digital employees working for me all day.
That reflective experience felt like a microcosm of what the entire world is going through right now. Standing on stage in front of CIOs and allocators responsible for decisions that influence tens of trillions of dollars, I could see the gap between the speed of change and the speed of institutional response. Watching the survey results and listening to panel discussions, I did not feel that investors fully grasped the structural shift that has already occurred. Even I, someone immersed in this every day, only recognized the magnitude of the last three months after stepping back and reflecting on it.
The world has shifted faster than portfolios and most investors can adjust. AI adoption itself is taking longer than the expansion of AI capabilities, and many institutional investors are not set up to make dramatic shifts quickly. That is the key point. Benchmarks are still weighted for the world that won the last decade, not necessarily the world that will win the next one. This will be a decade of benchmark arbitrage because investors will adjust more slowly than AI and crypto are moving. In my view, 2026 will be remembered as the beginning of the rise of AI agents, and with that rise, the investment opportunity has moved from the software world into the physical world.
The End of the Margin Era
For the last fifteen years, the dominant investment phrase was Jeff Bezos’s famous line: “Your margin is my opportunity.” It was the perfect description of the software era. Code scaled globally. Distribution costs collapsed. Network effects created winner take all markets. The largest technology companies used software, platforms, data, and cloud infrastructure to attack profit pools across media, retail, advertising, enterprise software, transportation, and finance.
The result was a historic period of margin capture and market concentration. A small number of companies became the dominant drivers of equity returns, index performance, and investor imagination. They were the winners of the software age, and because they won so decisively, they now dominate the S&P 500, the MSCI World, and the way most investors think about growth.
That era is not over because software no longer matters. Software still matters enormously. But the opportunity has shifted. The next decade will not be defined only by who writes the best code. It will be defined by who can build, power, cool, connect, manufacture, and deploy the physical infrastructure required for intelligence to enter everything. The threat to the code moat built by humans is AI’s ability to convert ideas into monetization in minutes.
The new phrase is no longer “your margin is my opportunity.”
The new phrase is “your CapEx is my opportunity.”
This is one of the most important investment changes of our lifetime. Artificial intelligence has moved from the digital world into the physical world. It is no longer only a story about models, applications, and software productivity. It is now a story about the heavy infrastructure layer required to scale intelligence: chips, power, cooling, chemicals, optical networks, datacenters, advanced packaging, memory, batteries, automation, robotics, and the reindustrialization of the global economy.
The world spent the last decade optimizing for asset light software businesses. The next decade will require an enormous asset heavy buildout.
The Five Layer AI Economy
Jensen Huang has described the coming transition as a roughly $90 trillion physical world upgrade. Whether one focuses on that exact number or the broader direction, the message is clear: AI is not just a data center CapEx cycle. It is not just hyperscalers buying GPUs. It is a full stack rebuild of the global economy so intelligence can be embedded into every company, every factory, every device, every car, every phone, every robot, and eventually every workflow.
The data center is only the beginning. The larger opportunity is the conversion of the physical world into an AI native operating system.
That is why the five layer AI cake is such a useful framework. At the top are applications and workflows. Below that are models and AI platforms. Beneath that is data infrastructure and management. Then come chips, compute, storage, and networking. At the base are energy, hardware, manufacturing, and commodities.
Investors naturally gravitate toward the top of the stack because the top looks like the old world. It looks like software. It looks like margins. It looks like scalability. At the top, there is now abundance. But the constraint is increasingly at the bottom, where scarcity and bottlenecks lie. AI demand is no longer limited by imagination. It is limited by the physical stack: power availability, heat, land, permitting, substations, memory, networking, and materials.
That means the most important investment question is changing. In the software era, the question was: which company can take someone else’s margin? In the AI infrastructure era, the question is: who receives the CapEx dollars required to make intelligence ubiquitous?
This is where benchmark arbitrage begins.
Why the Benchmarks Are Wrong
The major global equity benchmarks still reflect the winners of the last era. The S&P 500 and MSCI World are heavily weighted toward the companies that dominated the software, internet, platform, and cloud age. That made sense. Those companies generated enormous returns, expanded margins, and built deep competitive moats. But benchmarks are backward looking by design. They tell you who won the last cycle, not necessarily who will receive the marginal dollar in the next one.
There is also a momentum element inside the construction of these indexes. The biggest weights become the biggest weights because they were the winners. Their market capitalizations rise, the indexes allocate more capital to them, passive flows reinforce that dominance, and the process continues. In the case of the Mag 7, that dominance took most of the 2010s to build. It was a long compounding process. The world gradually moved toward software, cloud, mobile, digital advertising, e commerce, and platforms, and the benchmarks slowly came to reflect that reality.
As someone who grew up being trained to handicap horse races, I always go back to what Charlie Munger said: “The model I like, to sort of simplify the notion of what goes on in a market for common stocks, is the pari mutuel system at the racetrack. If you stop to think about it, a pari mutuel system is a market. Everybody goes there and bets, and the odds change based on what is bet. That is what happens in the stock market.”
The current market weightings reflect the bets people have made about who they think the winners of the future will be. Historically, when change was more linear, you had time to adapt your views. AI is different because AI is moving like a locomotive. The speed and power of this transition are already creating visible strain across the physical economy. Datacenter demand is running into the limits of the grid, the construction cycle, the permitting process, the semiconductor supply chain, and the materials needed to build it all. The bottlenecks are not theoretical. They are the evidence that the physical world is being forced to respond to a digital intelligence wave moving far faster than the capital stock was built to handle.
That is why I use the phrase benchmark arbitrage, even though this is not benchmark arbitrage in the traditional sense. Most arbitrage situations are thought of as short term events. An index addition. An index deletion. A rebalance. A forced buyer. A forced seller. A gap that closes over days, weeks, or months.
This is different. This is structural. It may take years for the benchmarks to fully reflect the new AI economy. But the size and speed of the change make it feel like an event happening right now. The arbitrage is not that an index committee is about to make one adjustment. The arbitrage is that the real economy is already changing faster than the benchmark can evolve.
If the next decade is defined by AI Cap Ex, then today’s benchmarks are likely underweight the areas that matter most. They are underweight the physical inputs required to scale intelligence. This includes power, electrical infrastructure, advanced manufacturing, chemicals, optical connectivity, semiconductor equipment, packaging, and the fragmented industrial supply chains now sitting directly in the path of the AI spending wave.
This creates a rare moment. Investors can look at the world not as it is currently represented in the benchmarks, but as it may need to be represented ten years from now. That is benchmark arbitrage. It is a structural mismatch between where capital is currently allocated and where the physical economy must go.
The Cost of Underinvestment
The irony is that the prior software era helped create this opportunity. For years, capital flowed toward asset light businesses and away from asset heavy industries. Investors rewarded recurring revenue, high gross margins, buybacks, low capital intensity, and terminal value stories built on long duration cash flows. At the same time, many parts of the physical economy were neglected. Commodity capacity was underbuilt. Grid infrastructure aged. Industrial supply chains became optimized for cost, not resilience. Manufacturing was pushed offshore. Hardware became less fashionable than software.
Now AI is exposing the cost of that underinvestment.
The same investors who spent years rewarding companies for needing little capital now have to recognize that AI requires enormous capital. The winners are not only the companies deploying AI. They are also the companies selling the inputs needed for everyone else to deploy AI.
If every Fortune 500 company needs its own AI infrastructure, if every country wants sovereign AI, if every factory needs automation, if every car becomes an AI computer on wheels, if humanoids move from concept to production, and if every device becomes intelligent, then the bottlenecks will define the profits.
The receivers of the CapEx dollars become the new margin takers.
The Terminal Value Problem
This also explains why software has become more difficult to value. AI is disrupting the terminal value philosophy that supported many long duration assets. In the old model, investors could look three, five, or ten years out and assume that dominant software franchises would continue compounding with limited disruption. But AI changes the speed of competition. Coding is becoming cheaper. Intelligence is becoming more widely available. The cost of building software is collapsing.
That does not mean every software company fails. It means the durability of future margins is harder to underwrite. When the pace of change becomes exponential, the confidence interval around terminal value widens. A business that looked unassailable three years ago can suddenly face new competition from AI native workflows, agents, open source models, cheaper code generation, or a customer deciding to build internally rather than buy another software seat.
The market is beginning to understand that software may still be valuable, but the old assumptions about duration, pricing power, and defensibility need to be reexamined.
This is the other side of benchmark arbitrage. The old winners are not disappearing, but their dominance was built for a slower world. Their index weights reflect years of compounding in an era when software had scarcity value, code created durable moats, and terminal values could be modeled with more confidence. AI compresses that timeline. It questions the durability of some software margins at the same time it creates urgent demand for the physical inputs required to scale intelligence.
That is what makes the current moment so unusual. The market is not waiting ten years to recognize the strain. It is seeing it now across the physical supply chain. The indexes still carry the weight of the last era, but the bottlenecks are already pointing to the next one.
The Speed of Code Meets the Speed of Steel
Hardware and commodities face the opposite dynamic. They were ignored because they were messy, cyclical, capital intensive, and fragmented. But those are precisely the characteristics that can create opportunity when demand shocks arrive. If supply is slow to respond and demand accelerates, pricing power can emerge in unexpected places.
The world cannot prompt its way into more electricity. It cannot instantly create more transformers. It cannot magically permit new data centers, manufacture more high bandwidth memory, expand advanced packaging capacity, or produce the specialty chemicals required for leading edge semiconductors overnight.
The digital world moves at the speed of code. The physical world moves at the speed of steel, copper, silicon, chemistry, energy, and regulation.
That gap is the investment opportunity.
AI is forcing the fastest moving technology in history to collide with some of the slowest moving supply chains in the economy. That collision creates bottlenecks. Bottlenecks create pricing power. Pricing power creates earnings revisions. Earnings revisions eventually force benchmark weight changes. The investor’s job is to identify those changes before the benchmark does.
Follow the Bottlenecks
The old world was about concentration. The new world is about fragmentation. The Mag 7 captured the margin in the software age because software rewards scale, distribution, and network effects. The AI physical buildout rewards a much broader ecosystem. The winners may include semiconductor companies, packaging suppliers, memory producers, power equipment manufacturers, grid operators, industrial automation companies, thermal management firms, optical networking providers, specialty chemical companies, construction firms, and commodity producers.
The opportunity spreads across countries, sectors, and supply chains.
That does not mean every hardware or commodity company is a winner. It does not mean investors should blindly buy anything related to AI infrastructure. The point is more precise. The world is moving from a software dominated investment regime to a full stack AI buildout regime. In that world, the most attractive opportunities may appear in places that benchmarks still treat as secondary.
Investors need to stop viewing AI only through the lens of applications and start viewing it through the lens of constraints. Where is the shortage? Where is the bottleneck? Where is the underinvestment? Where does permitting take years? Where does supply require physical capacity? Where are the receivers of the CapEx dollars?
That is the new map.
This is also why I am currently building my thematic portfolio around the five-layer AI cake. If the opportunity is moving from the software layer into the physical infrastructure layer, then active management has to evolve with it. The goal is not simply to own a static basket of AI winners. The goal is to understand where we are in the cycle, where the bottleneck sare forming, where capital is flowing, and how the weights should shift across the stack as the 90 trillion dollar buildout progresses. Early in the cycle, that may favor semiconductors, advanced packaging, power equipment, and optical connectivity. As the cycle broadens and commodity shortages and power constraints take move, the focus may move toward data centers, energy, chemicals, cooling, automation, and eventually applications, agents, and humanoids. Benchmark arbitrage only becomes actionable if it can be translated into portfolio construction, active reweighting, and a disciplined process for moving capital along the cake as the AI economy evolves.
The software era taught investors to follow margin capture. The AI era will teach investors to follow capital expenditure. The previous winners made money by using code to compress costs and attack incumbents. The next winners may make money because everyone else must spend capital to survive.
The world has changed. The opportunity has shifted. The benchmarks still reflect the last era, but the physical world is already being rebuilt for the next one.
Your margin was my opportunity.
Your CapEx is my opportunity.
By Jordi Visser, head of AI macro nexus research for 22V.
