About two years ago, I was running an M&A advisory firm out of Nashville with a small team, a bootstrapped budget, and enough active deals to keep me up most nights running through checklists in my head. The kind of situation where you’re thinking about four clients, two capital raises, and whether your newest engagement letter has the right fee structure simultaneously while your six-year-old is asking you to watch him do a cannonball.

I didn’t have the capacity to hire the team I needed. I didn’t have the capital to outsource the work. What I had was a stubborn belief that if I could systematize the judgment layer of my business, not just the administrative tasks, but the actual thinking, I could operate at a scale that made no sense for a firm my size.

So I built it.

A mentor once asked me a question I have never been able to shake: “Daniel, are you doing it the best you can do it, or the best it can be done?” That question is the whole reason Icon’s infrastructure exists. I did not want to bring clients good advice. I wanted to bring them the best thinking available, at every position, on every engagement.

What I built is a digital staff modeled after the best thinkers in the world at each function. Imagine having a former Goldman Sachs CFO, a dealmaker who has closed 200 transactions, a strategic navigator who spent a decade as a short-seller finding every flaw before it became a crisis, and a world-class closer who understands exactly how trust-based sales work at the highest levels, all available for every client engagement. Not occasionally. Every time. That is what the infrastructure delivers. Each position is filled by an intelligence layer built around the mental models, frameworks, and decision-making approaches of the people who are genuinely the best in the world at what they do. My clients get that bench on every deal.

I’m not sharing this to impress you. I’m sharing it because of what it taught me about where every business on earth is headed, and how fast it’s coming.

The infrastructure operates across three layers. One layer handles the always-on operational work, processing signals and routing intelligence around the clock without requiring my involvement in every interaction. A second layer manages the judgment-intensive work, the kind that used to require flying in a specialist or pulling in an outside firm. A third layer runs on a schedule, handling systematic tasks and keeping everything current while I am focused on clients, on deals, and on my family.

The whole thing runs on servers we control completely, not managed cloud services someone else can reprice or shut down overnight. It routes between multiple AI systems depending on what each task requires, so we are never dependent on any single provider. And it has a self-improvement loop coded into its core, which means the system gets better at serving Icon clients the longer it runs. Not AI in general. Our AI. Trained on our philosophy, our standards, and our commitment to bringing world-class intelligence to every engagement.

Here is the lesson I want to spend the rest of this paper on: what I built for one Nashville firm is a preview of what is about to happen to your entire industry, your supply chain, your cost structure, your business valuation, and eventually, if you are willing to let your thinking go this far, the whole concept of work itself.

Buckle up. But not because it’s scary. Because it’s the most extraordinary moment in the history of commerce, and most people are going to miss it entirely.

Here’s a thought experiment that doesn’t require any imagination at all, because it’s already happening: a robot designs a component. A different robot manufactures that component. A third robot inspects, packages, and ships it. A fourth robot installs it in a product that a fifth robot assembled from materials that a sixth robot extracted from the earth.

No lunch breaks. No workers’ comp claims. No payroll taxes. No absenteeism. No union negotiations. No health insurance. No 401(k) matching. No HR department to manage all of the above.

When you remove human labor from a supply chain, and I mean the entire supply chain, from the mine to the manufacturing floor to the distribution center to the last-mile delivery, what happens to the cost of the goods being produced? It approaches the cost of the raw materials and the energy required to move them. And as renewable energy scales, even that cost trends toward zero.

This is not speculation. This is physics applied to economics. The deflationary pressure of full supply chain automation is as inevitable as the deflationary pressure of electricity replacing candlemakers. The only real question is the timeline.

Let’s talk about that timeline, specifically, what it means for the materials that build your business, your home, and your life. Construction materials get manufactured without human labor. Agricultural equipment operates autonomously from planting to harvest. Food processing runs lights-out. Freight moves without drivers. The price of a new home, a new car, a new piece of manufacturing equipment, all of it follows the same curve as computing power: steadily, inevitably, dramatically down.

Now here’s where it gets personal. If the cost of building materials approaches near-zero because robots extract, refine, and deliver them without human involvement, what happens to the cost of housing? What happens to the cost of food when agricultural automation scales from specialty crops to commodity grains? What happens to the cost of every physical good that currently requires a human to touch it at some point in its production?

The answer is they get dramatically cheaper. Not cheaper like a sale, cheaper like a structural, permanent, one-directional ratchet that doesn’t reverse. And here’s the thing most economic commentary misses: that’s a feature, not a bug. When the cost of necessities approaches zero, what you’re actually watching is the largest involuntary wealth transfer in human history, from the suppliers of labor to the consumers of goods.

Your monthly expenses go down. Your fixed costs go down. Your stress goes down. And if you own the right assets while this is happening, we’ll come back to which ones, your wealth goes up while everyone else is still arguing about whether it’s real.

I want to be precise here, because vague futures don’t build survival plans. Every industry follows a curve from human-dependent to AI-augmented to fully autonomous. Where your business sits on that curve today determines how much time you have to adapt, and what adaptation actually looks like.

The curve doesn’t move at the same speed in every sector. Some industries are being disrupted right now. Some have a five-year runway. A few have a decade or more, largely because regulation will slow the technology regardless of what it can technically do. Knowing which category you’re in is the most strategically important piece of information a business owner can have in 2026.

I’ll give you my honest read, sector by sector. This is a scenario matrix, not a prophecy. I’m making specific predictions because specific predictions are what business owners can actually use. I’m going to be wrong about some of the timing. I’m confident about the direction.

This is the chapter nobody else in the M&A advisory world is writing, because most advisors are too focused on today’s multiples to think about 2030’s. I spend every working day thinking about what businesses are worth and why, so let me tell you what keeps me up at night.

Business valuations in the lower middle market are built primarily on EBITDA multiples, how much is your earnings before interest, taxes, depreciation, and amortization worth to a buyer. Right now, a healthy services business with $2M in EBITDA might sell for 4-6x. A strong manufacturing business might command 5-7x. A tech-enabled services firm might see 8-10x.

Here’s what changes those multiples: the sustainability of the earnings. A buyer paying 6x is betting that the EBITDA will still be there in years four, five, and six of their ownership. If the earnings are structurally at risk, from a single customer, from a key employee, from a dependency that can be disrupted, the multiple compresses.

Now ask yourself: what happens to the multiple on a business where 60% of the cost structure is human labor in a role that will be automated in the next three to five years? The answer is it compresses, dramatically, and it starts happening before the automation actually arrives. Sophisticated buyers don’t wait for the disruption. They price it in advance. By the time the robots show up at your loading dock, your valuation has already taken the hit.

The businesses that will command the highest multiples in 2030 and beyond are not necessarily the ones with the highest revenue. They’re the ones with the most defensible earnings, earnings that don’t depend on labor that can be replaced, customers that can be replicated by an algorithm, or relationships that live only in one person’s head.

One more thing worth naming, and this is uncomfortable but necessary: if your business sells a commodity service that AI can replicate at a fraction of the cost, the question isn’t “how do I sell it for a good multiple before the disruption?” The answer to that question used to be good advice. It isn’t anymore. Sophisticated buyers already know what’s coming. You’re not going to sell your way out of a structurally obsolete business model to a PE firm that employs former McKinsey partners who track this stuff for a living.

The better question is: what do I build inside this business in the next three years that a buyer in 2029 will actually pay a premium for? That question has a very good answer. It’s just not the same answer it was in 2019.

I promised I wouldn’t make one big bold claim that I might be wrong about. So instead, here are three scenarios, each plausible, each internally consistent, and what the world looks like for business owners in each one. Your job is not to pick which one is right. Your job is to build a business that survives all three.

If you’ve been tracking the logic of this paper, you’ve arrived at a strange and genuinely exciting question: if manufactured goods trend toward zero cost, and AI-generated services trend toward zero cost, where does wealth go? What do you hold? What do you build? What’s worth owning in a world where the things that used to cost money don’t anymore?

The answer has four parts, and none of them should surprise you if you’ve spent any time thinking about what humans actually value when their survival needs are handled.

The first category is land. Specifically, productive land, farmland, timberland, land in geographies where people want to live. When manufactured goods are cheap and services are cheap, the fixed supply of desirable land becomes one of the most reliable stores of value on earth. This isn’t new, it’s the oldest wealth strategy in human history, and it comes back into favor every time manufactured goods deflate.

The second category is cash-flowing businesses in the window before their sector is automated. This is where Icon’s buy-side advisory work lives, and it’s one of the most interesting opportunities I see right now. A manufacturing business with $3M in EBITDA that has a 10-year runway before meaningful automation exposure is an extraordinary acquisition target at today’s multiples. You’re buying real cash flow at a price that doesn’t reflect the full value of the runway. The market hasn’t fully priced this in yet. That window closes.

The third category is experience, status, and genuine human connection. When every product is cheap and every service is automated, what commands a premium? The things robots cannot replicate: a handcrafted object with a human story behind it. A dinner with real conversation. An advisor who has built and sold companies and actually knows what they’re talking about. An experience that creates a memory. Hospitality, art, craftsmanship, and trusted human relationships appreciate in a deflationary goods economy.

The fourth category is AI infrastructure, the systems, data, and institutional knowledge layers that compound over time. I built mine for my firm. The same principle applies to any business. The company with two years of proprietary customer interaction data trained into its AI system has an asset that cannot be purchased, only built, and the building takes time. Every day you delay is a day your competitor with the built system gets further ahead.

I want to say something that most people in the AI space will never say, because it costs them recurring revenue: don’t build your business on a single AI provider.

I say this as someone who uses multiple AI models and genuinely finds them remarkable. I am not cynical about the technology. I think the major AI labs are building some of the most important tools in the history of human civilization, and I sincerely hope they all succeed. The best of these tools has changed the way I think, work, and build. There are systems I use every day that I would describe as legitimately life-changing.

And yet: the race between providers is ferocious, the lead changes every 90 days, and pricing models are evolving as these companies figure out how to monetize what they’ve built. The model that’s best for your use case today may not be best in six months. The pricing that works for your budget today may not work tomorrow. Any business that builds its entire operation on a single provider is making the same bet that companies made when they became completely dependent on a single cloud provider before the market got competitive, and some of those bets turned out fine, and some of them turned out to be very expensive.

The architecture principle I built my system around is portability. My institutional knowledge, the prompts, the workflows, the context layers, the decision frameworks, lives in structures I control. The AI model is the processor. My system is the program. When a better processor comes along, I swap it in. The program doesn’t change.

This is the model-agnostic approach: build your intelligence layer so that it can run on whatever model is best, cheapest, and most appropriate for a given task at any given time. Some tasks require the most capable reasoning model available. Others run fine on a faster, cheaper model. The smart architecture routes tasks to the right engine based on what the task actually requires, not loyalty to a brand.

Think of it this way: the world’s best race teams don’t build their competitive advantage around which engine manufacturer they’re loyal to. They build it around the engineering that wraps the engine, the aerodynamics, the pit strategy, the driver development, the data systems. The engine matters. The system around the engine is what wins championships.

Your AI infrastructure should work the same way. Be grateful for the engines. Don’t become dependent on any one of them.

Andrej Karpathy is a former director of AI at one of the world’s most consequential technology companies, a co-founder of one of the leading AI labs, and one of the most respected technical voices in the field. In 2024, he shared something online that had nothing to do with technical AI research, it was a description of his personal knowledge management system, and it went extraordinarily viral. Not because it was technically sophisticated. Because it was a permission structure.

He described how he uses a combination of personal knowledge tools, software that creates a linked, searchable, ever-growing external repository of his thinking, his notes, his research, and his experiences, to essentially build a second brain that augments his own memory and reasoning. The concept isn’t new. But coming from someone with his credibility, it gave thousands of people permission to take it seriously.

Here’s what the internet mostly missed about why this matters for AI: a second brain isn’t just a note-taking system. It’s the foundation of your AI’s intelligence.

Think about what makes one person’s AI assistant dramatically more useful than another person’s, even if they’re running the same underlying model. The difference is context. The AI that knows your business, your clients, your deal history, your voice, your preferences, your institutional knowledge, is incomparably more useful than the same AI starting from zero every time. And the way you give your AI that context is by building the knowledge layer that feeds it.

The tools for this range from sophisticated personal knowledge management software to a well-organized set of documents in a cloud storage system you already use. The specific tool matters less than the habit: consistently capturing your thinking, your decisions, your lessons learned, and your institutional knowledge in a structured form that your AI can access and learn from.

What I built for my firm is a version of this at the organizational level. Over 150 documents, frameworks, deal analyses, client profiles, voice samples, decision templates, and workflow specs, all of it feeding the AI layer that runs my firm. The AI that advises me on deals today is dramatically more valuable than it was a year ago, not because the underlying model improved (though it did), but because I spent a year feeding it the context that makes it specifically mine.

Here’s the practical version for a business owner who isn’t a tech person: start with what you know. Document your best client relationships, not just the contact info, but what they care about, what they’ve said, what they responded to, what they need. Document your processes, not the ones you want to have, the ones you actually use. Document your lessons, the deals that went sideways, the hires that didn’t work, the strategy pivots that did. Document your voice, the way you talk to clients, the phrases you use, the things you’d never say. Give your AI that material to work with, and it stops being a generic tool and starts being your specific operating advantage.

The building of your second brain is not a technology project. It’s a discipline. And the earlier you start, the deeper the moat.

This is the chapter that goes further than most people are ready to go, but you asked me to tell you where the puck is going, so here it is.

The Karpathy method is about humans building context for AI. What I coded into my own system is the next layer: the AI building context for itself. There is a meaningful difference, and it’s worth understanding.

Most AI implementations are static. You interact with them, they respond, the conversation ends, nothing is retained. The system is exactly as calibrated to your needs on day 365 as it was on day one. That’s using AI as a very expensive calculator, powerful in the moment, but not compounding.

The next tier is an AI system with a perpetual self-improvement loop. Every interaction, every output, every time you correct it, every time you rate a result, all of it feeds back into a continuously updated context layer. The system learns which approaches work for your specific use cases. It learns which clients respond to which communication styles. It learns the deal structures you favor, the risks you avoid, the language you use. It doesn’t just answer questions from your data. It gets progressively better at being your AI specifically, not smarter in general, but smarter at being yours.

This is not science fiction. The architecture to build this exists today. I know because I built it. And here’s what it taught me: the gap between a self-learning AI system and a static one compounds aggressively over time. After six months, the difference is noticeable. After a year, it’s significant. After three years, it is a structural competitive advantage that cannot be replicated by a competitor who starts fresh, regardless of which model they use.

The self-learning loop is what turns an AI tool into an AI asset. Tools depreciate. Assets appreciate. The question every business owner should be asking isn’t “which AI tool should I use?” It’s “how do I build an AI system that gets more valuable every day I run it?”

Ten questions. Be honest with yourself. Each “no” answer is a risk factor. More than three “no” answers means your business has meaningful structural exposure to the trends in this paper. That’s not a death sentence, it’s a starting point for a conversation.

I want to close where I started: with a field report, not a theory.

I didn’t build my AI ecosystem because I had a grand strategic vision. I built it because I had real problems, too many deals, too much routine work, too little time for the things that actually required me. I started with one automation that handled one task. Then another. Then I connected them. Then I added a layer of judgment. Then I built a feedback loop. Then I realized, somewhere around month twelve, that the thing I had built was genuinely changing what my firm was capable of.

That’s how this starts. Not with a complete architecture. Not with a CTO. Not with a large budget. With one problem and the willingness to use a new kind of tool to solve it.

Here’s a practical starting framework for business owners who want to begin:

One final thought on the model agnosticism principle: start with whatever tool you’re most comfortable with. Learn how to use it deeply. Build your context layer for it. And as you do, build that context in a format that doesn’t depend on staying with that provider forever. Your prompts in a document you own. Your knowledge in a system you control. Your workflows documented so they can be rebuilt on a new engine if the engine changes.

The preparation for the future is not complicated. It is not expensive. It is not reserved for technical people or large companies or people with Stanford degrees. It is available to any operator who is willing to pay attention, start building, and trust that the compounding works, because it does.

My AI is better than yours. Right now. But that’s not because I’m smarter. It’s because I started earlier, I built the context layer, I built the feedback loop, and I treated it as core infrastructure rather than an interesting experiment.

You can start today. And in eighteen months, you’ll be saying the same thing to someone who hasn’t started yet.

The puck is moving. I’d rather you be skating toward it.