I increasingly believe the long term is easier to predict than the short term.
I have no idea which AI model will lead in five years, which robotics architecture will win, or what exact mix of generation will power the grid in 2035. I have much higher conviction in a few broader outcomes: intelligence will get dramatically cheaper and more abundant, more of it will move into the physical world, we are going to need a lot more electricity, and the ability to actually build things will matter more.
The useful exercise, then, is to work backward from those assumptions. If you believe the end state, what has to be true along the way? Where do the constraints move? What looks like a narrow problem today but becomes critical infrastructure ten years from now?
A few themes keep coming up.
1. Intelligence gets cheap
AI is making intelligence cheaper. Work that once required expensive human time can increasingly be done by software. Coding, research, analysis, design, planning, support and a growing list of other cognitive tasks are all getting cheaper to produce.
The more interesting question is what happens when intelligence is no longer the scarce input. Constraints move elsewhere: proprietary data, distribution, regulatory permission, electricity, skilled labor, manufacturing capacity and physical infrastructure.
There is a business model shift here too. Traditional software mostly helped humans do the work. AI can increasingly do more of the work itself. If that continues, customers may care less about buying tools and more about buying finished outcomes.
The product is no longer necessarily software for processing a claim, sourcing a component or resolving a support ticket. It may simply be the claim processed, the component sourced or the ticket resolved. That should blur the line between software and services, and push more technology companies toward selling the finished job rather than the tool used to produce it.
AI is also moving from producing information to taking action. Over time, we will increasingly ask it to spend money, schedule production, operate equipment, negotiate with suppliers or direct machines. As intelligence gets cheaper, the more important question becomes what we are willing to let it do.
2. Intelligence moves into the physical world
Most AI today still lives behind a screen. I doubt that is where it stays.
Over time, intelligence will find its way into drones, robots, vehicles, factories, warehouses, power systems, medical devices and military equipment. In that world, software looks less like a standalone product and more like the intelligence layer inside a broader system.
For much of the last two decades, software itself was often the thing being sold. Increasingly, the customer may instead be buying an autonomous aircraft, a factory, a power system, a logistics network or an industrial service, with software buried inside the product and responsible for much of its value.
A world with vastly more intelligent physical systems will need better ways to design, simulate, test, certify, secure, monitor and maintain them. You do not necessarily need to pick the winning robot, drone or autonomous vehicle. There is plenty of room to build the infrastructure they all eventually need.
3. Energy becomes a technology problem again
For most software investors, electricity has been something you could safely ignore. Over the next decade, it will be much harder to do that.
AI is power hungry. So are semiconductor fabs, advanced manufacturing, electrification and much of the infrastructure we want to build. At the same time, getting new power online remains slow. Generation is only part of the issue. Transmission, transformers, switchgear, interconnection, permitting and local grid capacity can all become bottlenecks.
We may be entering a world where intelligence gets dramatically cheaper while the electricity needed to run it does not. Access to power therefore starts to matter much more as a competitive and strategic advantage.
A question like “where can I actually get 100 megawatts by 2029?” starts to matter a lot more. So do the businesses that can answer it, accelerate it or find another way to deliver the power.
I am interested in generation, but also in the less glamorous layers around it: grid equipment, storage, cooling, power electronics, load management and infrastructure for finding and securing available capacity. “Speed to power” feels like it could become an important category in its own right.
4. The physical economy gets rebuilt
The U.S. spent decades optimizing around a world where manufacturing could happen elsewhere and physical capacity was relatively easy to access. That model is under more pressure now, particularly in areas tied to defense, aerospace, semiconductors, energy and critical infrastructure.
At the same time, many of the people who know how to build and operate these systems are aging out of the workforce. We want to build more with a labor pool that is not obviously getting larger, which should make automation and productivity gains increasingly important.
Factories will become more software-defined. Robots will do more physical work. AI will capture more of the knowledge sitting in the heads of experienced operators. One technician will be able to oversee more equipment. Supply chains will become easier to map and reconfigure.
Some of the best opportunities may look surprisingly mundane at first: finding the right supplier for a critical component, modernizing a 25-year-old factory, training enough electricians or technicians, qualifying a manufacturer, or keeping a plant running when the person who understands one critical system retires.
These are not new problems. What is changing is how important they are becoming, and what technology can now do about them.
Working backward
The themes reinforce each other. Cheaper intelligence makes automation better. Better automation pushes AI into the physical world. More compute and industrial activity require more power. Scarce labor makes automation more valuable. More complex machines create new requirements around safety, security and trust.
The common thread I keep coming back to is that intelligence is becoming abundant while many of the physical inputs to progress remain scarce.
The last technology cycle created enormous value by turning software itself into the product. The next one may look different. Software will increasingly sit inside machines, infrastructure, factories and services. The winners may not always look like software companies, even when software is the reason they can exist.
Rather than trying to predict exactly what the world looks like in 2035, I find it more useful to start with the outcomes that seem increasingly hard to avoid: much more intelligence, many more intelligent machines, much higher electricity demand, and a greater need to build physical infrastructure.