Every time I open LinkedIn, Instagram, or X, I see the same headline: “AI Startup X raises $30 million Series A.” This isn’t just feed noise. AI startups captured more than half of all global venture capital dollars in 2025, the first time any single sector has done that. That made me wonder how much of that capital funds real technical differentiation versus a hot buzzword and a compelling story. Before writing a check, every investor should ask: What happens the day OpenAI, Anthropic, or Google ships your product for free?

Before asking this question, investors should understand the difference between a wrapper and a moat. Both words get thrown around in AI pitches and are rarely used correctly. Strictly speaking, “wrapper” describes a layer of a product, not the whole company. It refers specifically to the AI capability, which belongs entirely to the LLM provider, such as Anthropic. The startup formats its users’ requests and sends them to a rented model via an Application Programming Interface, or API, then formats the response it gets back. Think of an API as a waiter taking an order: the startup packages the customer’s request, hands it to the kitchen (the LLM vendor), and carries the dish back to the table. They don’t own the kitchen.

By this definition, almost every AI startup is a wrapper, since almost none of them train their own model. That’s normal, and not, on its own, disqualifying. The term isn’t useless just because it’s universal; it just means the interesting question is never whether a startup is a wrapper, but what else it is. A moat, on the other hand, is what makes a startup hard to replicate at some other layer of the product. You have a moat when a competitor cannot rebuild the entire product, even with access to that same rented AI model, without the data, integrations, infrastructure, and relationships the startup has built around it.

Signal 1: Is the Startup Building a Cumulative Moat?

Given that nearly every AI startup is a wrapper at the AI layer, the real work for an investor is figuring out whether there’s a moat somewhere else. First, look for a product that gets smarter with every use. If you call GPT-5 a million times, it doesn’t make your product smarter than a competitor’s, since OpenAI improves that model for every customer at once. For the product to actually get smarter, the improvements have to be stored somewhere the company owns. There are three ways a startup can do that: fine-tuning on proprietary data, retrieval-based knowledge accumulation, and feedback loops.

Fine-Tuning on Proprietary Data

Fine-tuning on proprietary data means periodically retraining a model on accumulated, company-owned interaction data so that it gets better at the specific task it’s meant to do. The process begins with an already-trained base model, like Gemini. The startup feeds a proprietary, expert-labeled dataset to the model, adjusting its parameters to master the specific task.

Say a startup builds an AI tool for reviewing insurance claims. Every claim it processes, along with the human adjuster’s final decision, becomes a new training example. Over thousands of claims, the model gets better at spotting the specific patterns that matter to that insurer, in a way a generic model never would. If the dataset is something only the startup has access to, such as years of real customer interactions, regulated-industry transaction data, or outcomes from real cases handled, then this is a moat. Even if a competitor copies the fine-tuning process, they can’t get the same data.

Crucially, fine-tuning doesn’t remove a startup’s vendor dependency. Startups usually still pay someone, either the vendor or a cloud provider, to actually run the fine-tuned model. A startup can be doing real, proprietary fine-tuning work while remaining dependent on the vendor’s underlying infrastructure. Neither of these facts is evidence that the startup lacks a moat.

Retrieval-Augmented Generation

A second way startups build this kind of moat is retrieval-augmented generation, or RAG. Unlike fine-tuning, retrieval leaves the model untouched and instead focuses on growing the database the model uses to answer questions. In practice, the system continuously logs user interactions and outcomes. Each time a new user has a question, the system searches the database for the most relevant matches and hands them to the AI as context before it responds.

Let’s take the example of a startup that sells AI customer support for e-commerce companies. On day one, the AI answers a ticket using only general knowledge, since it doesn’t yet know anything specific about that company. It then stores the entire resolution context, including what happened, whether a refund was issued, and the applicable policy. By ticket fifty thousand, a similar situation comes in, and now the AI retrieves actual past resolutions from that exact company, including edge cases nobody wrote into an official policy document. The answer gets better, not because the AI got smarter, but because the archive it’s pulling from got deeper. This is the moat, since the archive is built entirely from proprietary interactions competitors don’t have.

Proprietary Feedback Loops

Finally, proprietary feedback loops create defensibility through continuous quality improvement. Take the same e-commerce example: if the AI answers a support ticket, a human agent can review it and either approve, edit, or override it. That correction is itself a signal, and over thousands of tickets, patterns emerge showing where the AI tends to get things wrong. The startup can use these corrections in three ways:

• Fine-tuning data: Retraining the model on what went wrong versus the human correction.

• Retrieval context: Storing the corrected answer in the knowledge base so future queries pull the right version.

• Prompt calibration: Directly adjusting the system prompt instructions based on recurring patterns, without retraining.

This is a moat because this company’s users generate the correction data. A competitor can’t buy or scrape that.

Signal 2: How Hard Is It to Leave?

Start by checking integration depth: how difficult it would be for an existing customer to stop using the startup’s product and switch to another one. If a product offers no proprietary data and no deep integration, the customer can easily switch to another wrapper. However, if the product is deeply integrated into a customer’s existing systems, such as their CRM, single sign-on, or approval workflows, and has historical data or team habits built around it, switching becomes much harder. The AI itself might still be entirely rented. The switching cost comes from somewhere else: the customer’s data, workflows, and habits, not the intelligence layer at all.

Signal 3: Is It Built for One Industry, or All of Them?

A natural extension of workflow lock-in is vertical depth: whether the startup is built deeply for one specific industry rather than generically for everyone. A generalist AI tool, one that could technically serve any company in any industry, is much easier for a well-resourced competitor, or the model provider itself, to replicate than a product built around one industry’s specific workflows, data, and regulations. This is also one of the most common things investors are currently pushing founders on: a startup that says “we’re an AI tool for X” is a stronger claim than one that says “we’re an AI tool for everyone.”

Case Study: How Replit Proves the Framework

Now let’s test this framework against an extremely successful AI startup: Replit. For anyone unfamiliar, Replit is an AI-powered development platform where you describe an app in plain English and watch it get built, debugged, and deployed automatically, all inside the browser, with no local setup. It’s grown fast: in March 2026, Replit raised a $400 million round at a $9 billion valuation, tripling its value in just six months, and now serves over 40 million users. Replit doesn’t have its own frontier LLM, so what’s so special about it?

Replit runs a model-agnostic routing strategy across Claude, Gemini, and GPT, directing each task to the model based on cost and complexity. At this reasoning layer, Replit is renting intelligence, making it a wrapper by definition. Yet Replit still has a real moat, built on two things.

First, it owns small proprietary models for high-frequency micro-tasks. Inline completions run on distilled internal models, smaller models trained to mimic a larger one’s behavior at a fraction of the cost. This directly protects their margins: they’re not paying full API rates on every single keystroke, only on the reasoning calls that actually need it.

Second, Replit owns the non-AI infrastructure around the model: a browser-based cloud IDE, real-time collaboration, hosting, databases, and one-click deployment. If a competitor gained access to the same LLMs tomorrow, they would still have to rebuild an entire cloud development ecosystem from scratch.

Replit is a wrapper wrapped in a real product. The AI is rented, but the thing customers are actually locked into, their whole coding environment, is owned.

The Real Diligence Question

The sharper question isn’t “is this a wrapper.” Almost nothing passes that test. Nearly every AI startup rents its intelligence from OpenAI, Anthropic, or Google. The better question has two parts: which layers of this company does it actually own, and which does it rent? And is the owned layer something a customer would pay for on its own, even if the AI itself became a commodity tomorrow? Replit rents the reasoning and owns the environment. The next startup you look at might rent everything and own nothing, or it might own a dataset no competitor can touch. Either way, that’s the diligence question that actually tells you something. Not “do they have AI,” but “what happens the day the AI stops being the interesting part.”

If you're an investor who's had this conversation with a founder, I'd like to hear how it went.

Subscribe now

In this final post, I want to focus on a rarely discussed but fundamental aspect of AI: the physical infrastructure behind LLMs. It’s what allows us to create videos of dancing politicians or the thing your manager wants when he says “we need to go agentic” for the tenth time this week.

The $5.2 Trillion Shovel

By 2030, global investment in AI-ready data centers is projected to reach $5.2 trillion, according to McKinsey, as the world races to build the compute infrastructure required to power large language models. This is unprecedented. To put this in perspective, AI-related capital expenditures contributed roughly 1 percentage point to U.S. GDP growth in the first half of 2025.

To understand how massive that is, 1 percentage point of GDP is nearly double what the entire U.S. spends on coffee every year, or roughly half of the entire U.S. Department of Education’s budget. This has already surpassed the internet boom of the late 1990s, when IT investment peaked at about 0.5% of GDP growth. We are no longer just manipulating data. We have moved into building massive AI Factories to manufacture intelligence at an enormous scale.

The 1% Real Estate Crisis

I recently watched Eddington, the latest movie from Ari Aster starring Joaquin Phoenix. In the film, a giant, looming data center sits at the edge of a New Mexico town like a silent character. The film uses it as a metaphor for our brewing distrust. The data center is the ultimate winner while the townspeople descend into paranoia.

Sinister or not, what I realized watching that movie is that most people don’t realize that a data center is simply the physical home of the internet. For AI, these aren’t just warehouses for files. They are hyper-specialized power plants for math. However, we are currently facing a big problem. Despite a surge in construction, vacancy rates in primary markets have reached a record low of 1%. Over 60% of new builds are pre-leased years in advance. Access to space and power is now the #1 barrier to AI expansion.

Architects of Logic, Builders of Silicon

This lack of real estate is only one part of the problem. To understand the rest of the market, we must distinguish between the Architects and the Builders.

NVIDIA is a fabless company. This means they are the architects who design the complex blueprints for chips, but they don’t actually own the machines to build them. For that, the entire world relies on a foundry. A foundry is essentially an outsourced high-tech factory.

Google is also moving in this direction with its Custom Silicon strategy. By designing their own specialized chips, like the TPU, they avoid having to rely on NVIDIA and can improve efficiency. But even Google is fabless. No matter who wins the design war, everyone must wait in line for the same few foundries to physically manufacture the chips.

The Six Champions of the AI Stack

The AI stack is complex, so I selected one champion from each layer. These are the companies that own the essential parts of the industry.

The Manufacturing Edge

ASML: Lithography is the process of drawing with light. ASML uses massive lasers to etch microscopic circuit patterns onto silicon wafers. It is the only company on Earth that can do this at the scale required for AI.

The Master Builder

TSMC: It builds over 90% of the world’s high-end AI chips. It turns NVIDIA and Google’s blueprints into physical reality.

The Memory King

SK Hynix: AI chips need high-bandwidth memory, which is like a high-speed fuel line for data. SK Hynix is currently the lead partner providing this fuel for NVIDIA’s next-generation platforms.

The Cooling Leader

Vertiv: Air cooling is no longer enough for chips pulling this much power. Vertiv specializes in liquid cooling, essentially a radiator system that keeps the hardware from melting.

The Energy Frontier

Constellation Energy: It is the power plant of the internet. Constellation Energy recently signed a 20-year deal to restart the Three Mile Island reactor specifically to provide reliable and abundant nuclear power for Microsoft.

The Data Conductor

Broadcom: As clusters, which are networks of thousands of GPUs working together, grow larger, moving data between them becomes the bottleneck. GPUs are the specialized calculators that do the AI math, but they are useless if they can’t talk to each other. Broadcom builds the high-speed plumbing that keeps data moving between servers at lightning speed.

The Map is Ready: Moving to Execution

We analyzed the LLM’s logic, compared it to the incredible efficiency of the human brain, and mapped the massive physical infrastructure required to run it. We know the architects, the builders, and the power sources.

The blueprint is set. It’s time to move from theory to execution and start building.

The 20-Watt Miracle

The brain is composed of approximately 86 billion neurons. These are living cells that are powered by electricity and chemicals, such as dopamine and serotonin, to transmit information. In contrast, an LLM relies on parameters, which are numerical strengths of connections stored on silicon chips.

In terms of efficiency, the gap is massive. The brain needs approximately 20 watts to function, roughly the power used by a single dim light bulb, at all times. By contrast, just training a top-tier AI model can require 20–25 megawatts of power sustained over about three months. That is roughly equivalent to powering two to three Empire State Buildings non-stop for the entire duration.

Why Your Brain Never Stops Building Itself

The human brain is a master one-time learner. If we see a photo of a laptop once, our brains can rewire themselves through neuroplasticity, so we know what a laptop is forever. In contrast, an LLM must be shown millions of images and billions of sentences to learn patterns.

When an LLM is trained, its weights—the numerical strengths that dictate its logic—are frozen into a static snapshot. Humans, however, are adaptive even in complex settings. For instance, in the context of investing, we can listen to Trump introducing new tariffs and instantly adjust our strategy based on intuition. By contrast, an LLM’s core brain remains locked in the past. It cannot learn a new fundamental truth about the world once its training is complete.

You might ask: “But can’t I just feed it live tweets and speeches?”

Yes, but when the LLM processes live signals, we are handing it a temporary cheat sheet of current events. So, while it can follow the instructions on that sheet, it cannot update its underlying worldview based on the new information.

Probability is Not Intuition

In psychology, humans think in terms of two distinct systems: System 1 and System 2.

• System 1 is fast, instinctive, and emotional. This is when we tell ourselves, “Bitcoin is going up, I need to buy now!” when Bitcoin hits 100k.

• System 2 is slow, analytical, and logical. This is when we take a moment and calculate the actual risk.

As humans, we make decisions with System 1 and perhaps later justify them with System 2. For LLMs, there is no intention, only prediction. They operate by calculating the statistical likelihood of what comes next. When an AI picks Buy, it isn’t because it feels confident; it’s because Buy is mathematically the most likely word consistent with the data.

The Simulation of Experience

The brain knows the world through experience. It understands Risk because it has felt the tightness in the chest after losing money playing poker. It knows Success because it has felt the rush from its football team winning the Champions League. An LLM, by comparison, knows these things through association. To an LLM, Risk is just a token that frequently sits near the word Loss on its mathematical map. It is a simulation of meaning, not the experience of it.

Predicting the Future vs. Understanding the Present

Let’s consider an example: if an LLM says “Buy NVIDIA” and a human says “Buy NVIDIA,” and then the stock goes up, does the method matter?

As of today, the answer is yes, because of the gap between Correlation and Causality.

An LLM is a master of correlation: When Politician X says Y, Stock Z usually goes up. However, it lacks a world model to verify why one thing causes another. Humans, on the other hand, use context and acquired logic to pressure-test those links.

Having said that, we are already seeing the birth of Causal AI, which is designed to move beyond simple pattern matching and understand why things happen. While this work is advancing in research labs, we are not yet at the point where an LLM can evaluate qualitative factors like a CEO’s personality with the same judgment and nuance as an experienced investor.

The Infrastructure of the Trillion-Parameter Era

In my next post, we’ll look at the physical infrastructure and key companies behind these models to see exactly who is driving this revolution.

Our Words Are Just GPS Coordinates

LLMs do not read words the way humans do. Instead, they calculate statistical relationships within numerical patterns.

To do this, our language is converted into vectors. Vectors are arrays of numbers that show each word’s position in a huge, mathematical space. This lets the LLM compare many words and generate logical text.

There are two steps involved in this transformation process:

Step 1: Tokenization - Breaking the World into Bricks

First, the LLM splits our sentence into tokens. The computer cannot process the word NVIDIA as a whole, so it often breaks it into parts, like NVI and DIA. It then assigns each of these parts a specific ID number, which is the only language the math engine truly understands.

Step 2: Embedding Lookup - The GPS Map of Meaning

Next, the LLM finds the coordinates for each ID number in the Embedding Layer. The embedding layer acts like a map, where every word has a specific set of numerical coordinates based on meaning. Here, instead of referencing a dictionary definition, the model refers to a location that mathematically links the word to all other related concepts. For example, Profit is near Revenue, and Shortage is close to Price Increase.

The LLM then sees a long string of numbers (a vector) that mathematically represents Semiconductors, Jensen Huang, and Compute Power. This vector is not just a point in 3D space. The vector space often has over 4,000 dimensions, each capturing a different nuance of the word. Now, the LLM knows the contextual neighborhood of our question.

How Coordinates Become Context

The numerical coordinates move through multiple Transformer Blocks, like stops on a factory line. Instead of each layer doing a completely different job, every block performs a similar task: the LLM employs the Attention mechanism to re-scan our entire sentence and update the coordinates of our words.

With every layer the data passes through, the mathematical relationship between our words becomes more precise. By the time the data reaches the final layers, the model has polished the math enough to see the deep, logical thread connecting a mention of NVIDIA at the top of the page to a Strong Buy signal at the bottom.

A 50,000-Candidate Election

After the blocks finish their work, the LLM reaches the Output Layer to produce raw scores. Because the previous blocks have sharpened the coordinates of our sentence, the model can now hold a giant election across its entire dictionary of 50,000 to 100,000 words, assigning a raw score to every single one:

• Buy: 18.5 points

• Sell: 4.2 points

• Banana: -10.0 points

The model then converts these raw scores into percentages using the Softmax function. If Buy gets 89% of the vote, the LLM picks the winner.

However, we can adjust a setting called Temperature. A higher temperature allows the LLM to occasionally pick the second or third choice, which makes it feel more creative and less robotic.

The Predictive Feedback Loop

How does the LLM do this for paragraphs or pages?

It may seem like the LLM writes everything at once, but it actually predicts one word at a time. This is called Autoregression. After the LLM predicts Buy, it feeds that word back into its memory and repeats the process to find the next word.

• Loop 1: Buy

• Loop 2: Buy + NVIDIA

• Loop 3: Buy + NVIDIA + Because

The Stadium of Digital Students

You might wonder how the LLM produces results so quickly. Its speed actually comes from specialized hardware.

Think of the GPU as a stadium full of students all racing to solve math problems. To keep things fast, the model uses a digital scratchpad called KV Caching. Instead of re-reading our prompt for every word, it remembers its prior work and only recalculates the new information.

Ultimately, we should not view the LLM as a database of facts. Instead, we should recognize it as a mathematical simulation of human logic. By mapping the distance between ideas and weighing the importance of connections, it uses massive computing power to predict the most likely path of a conversation.

The Machine vs. Biology

Can AI truly replace a human mind?

Before we trust an AI agent with our money, we must compare it to the human brain. In my next post, we will look at the LLM and the brain side by side to see where the simulation ends and true intelligence begins.

Between my job and personal side projects, I realized two things:

• Most people are understandably concerned that AI will replace their jobs.

• The best way to plan for the future is to understand how these tools work so you can start building with them.

I’m starting this Substack because I want to bridge the gap between the black box of AI and the people who want to turn it into something useful.

This newsletter is a documentation of what I’m learning and building. In my posts, I’ll explain how AI actually works by diving into the technical details of LLMs and how they differ from the human brain. I’ll also introduce you to the dominant players—the companies building the models and the ones providing the massive infrastructure that powers them—so you can see the full picture of the industry.

As we progress through this series, I’ll share the projects I’ve built and ask you to try them out. The first of these is a financial advisor agent designed for beginners who want to invest but are overwhelmed by jargon. I constantly hear people say, “I want to buy stocks, but I don't know how to decide,” or “My friend told me to buy Nvidia before it goes up even more.” This tool is designed to cut through that noise. It explains how to select stocks and understand the metrics behind the process in plain English, using real-world examples and strategies.

Ultimately, I want to help you understand AI so you don’t fall behind. I believe that once you see how these systems are put together, you’ll stop seeing AI as a threat and start seeing it as a tool for your own ideas.

Everything here is free to read. I’m looking forward to sharing what I find and hearing your thoughts on what I should build next.

My first article is live. Click the button below to read it.

Stop Treating LLMs Like People

Subscribe now

Before we explore AI agents and their complex tasks, it’s important to understand what Large Language Models (LLMs) are and how they work. When we ask our LLM agent, “Should I buy NVIDIA stock today?” it may feel like we’re consulting the best financial analyst at J.P. Morgan. In reality, hitting Enter on our keyboard triggers a high-speed, multi-layered statistical prediction engine. The LLM runs our question through mathematical filters to determine which words are most likely to emerge.

The LLM does not think. It simulates logic through scale and statistics.

Why LLMs Treat Words Like GPS Coordinates

We often hear that an LLM is a neural network. What exactly is that?

A neural network is a massive mathematical system. While its name was inspired by how biological neurons fire, it is more helpful to see it as a vast, layered scoring mechanism that processes information in parallel.

Our brains use biological synapses to connect biological neurons, but an LLM uses artificial neurons. These are mathematical functions that weigh information before passing it along. Think of these neurons as tiny calculators on an assembly line. Each calculator examines incoming data, multiplies it by a weight, and decides whether to trigger the next calculator.

Here’s an example:

Gemini: Hi Zahed, where should we start?

Zahed: Help me become rich.

To process my request, the LLM first breaks my sentence into small elements called tokens. A token is a basic unit of text, such as a whole word or a part of a word. Think of tokens as the Lego bricks of language.

The tokens are then placed on a multidimensional map of meaning, known as an embedding. An embedding is a mathematical representation that helps the LLM understand word relationships. In this mathematical landscape, words don't have definitions; they have locations. Because of the model’s construction, the token for rich is placed in a high-dimensional neighborhood surrounded by wealth, assets, and investing.

This data then travels through dozens of Transformer blocks. You can imagine these as specialized refinement filters in an assembly line. Within each block, a Self-Attention mechanism acts like a high-speed spotlight, instantly scanning the entire sentence to see how words tether to one another. It’s the math that ensures the model knows rich refers to financial status, not the flavor of cheesecake, by observing the context of the words around it.

We should not think of knowledge in an LLM as a file on a hard drive. Instead, imagine it as a giant web of connections. Knowledge is encoded in the numerical values of billions of network parameters.

How Models Learn through Error Correction

An LLM starts its life as a blank slate of random numbers.

To become effective, the LLM must undergo training. This training is a massive, high-speed loop of trial and error. Here is how that process works:

• The Guess: The LLM receives a sequence of text and is tasked with predicting the very next token in that sequence. Using its current, initially random weights, it predicts the word.

• The Error: The LLM compares its guess to the actual word. The difference between the guess and the truth is called the Loss, a mathematical score that represents how incorrect the guess was.

• Backpropagation: The model works backward. It adjusts billions of internal parameters to ensure that if it saw that exact sentence again, it would make a smaller mistake.

We can see that when an LLM learns that NVIDIA is often followed by GPU, it does not understand a business model. It has simply altered its weights to maximize the probability of that sequence.

During training, the model uses gradient descent, a mathematical method that gradually changes its parameters to minimize prediction error. To understand this, imagine you are at the top of a mountain with your eyes covered, trying to find your way down. You might go up or down along the way. Eventually, you reach one of the valleys. This is what the LLM does during gradient descent. It follows the steepest path down the mountain, adjusting its parameters until it reaches a valley, where the errors are smallest.

Why AI Masters Patterns, Not Logic

Consider an LLM used for financial analysis. We feed it millions of reports and every relevant market document. While it may look like the model is studying the economy, it is actually identifying statistical, not economic, patterns.

• Pattern A: If (Revenue is Up) + (Data Center Growth is High) + (CEO is Promising) > Most likely next words are Strong Buy.

• Pattern B: If (Export Restrictions Tighten) + (Competition is Rising) > Most likely next word is Sell.

So, when we ask the LLM for a prediction today, it does not analyze the stock market. It performs a calculation. Our words pass through its web of tuned connections to see which word is the best mathematical fit. Out of thousands of possible words in its vocabulary, it picks the one with the highest probability of being the next piece of the puzzle.

The High-Dimensional Map

Now that we know how these models are built and trained, we need to see how they act under pressure. My next post will focus on the model’s behavior under stress: how the LLM transforms our words into math, navigates a map of meaning, and selects a token based on the final scores.