AI Breaks the Mold: Why High Costs Are Now a Feature, Not a Bug

Princeton computer science professor Arvind Narayanan just dropped an essay declaring that AI has escaped the commodity trap. This isn't just academic talk — it’s a direct challenge to the idea that these models will get cheap and interchangeable. Last week, in episode one-oh-six, we talked about the fierce debate over the cost of models like Fable-5; Narayanan’s argument today is that this high cost is not a bug, it’s the entire business strategy. The conversation on AI has fundamentally shifted. The question is no longer "is this a bubble?" but "who captures the value in the long run?" And the answer, according to this thread, is a new kind of enterprise lock-in that we are ALL underestimating. Here's what else is moving today, Friday, July tenth.

First, and this connects directly, a thread from Sonia Randhawa is picking up steam. The claim? Microsoft has started replacing OpenAI and Anthropic models with its OWN, in-house "MAI" models inside Excel and Outlook. If you're using Copilot in the productivity suite, you might not be using a GPT model for much longer. This is the enterprise lock-in theory playing out in real time. It’s not about offering the best model, it's about owning the entire stack, from the operating system right up to the AI assistant embedded in the spreadsheet you use every day. Microsoft isn't just a customer of OpenAI; it's building the pipes to eventually cut them out. This is a power play, plain and simple. Then you have the other side of the AI coin. Not the corporate strategy, but the pure, chaotic, financial degen side.

A post from Crypto Rover is making the rounds, highlighting an AI filtering bot built by a user named AIBullss. This bot scans the new Robinhood Chain, a permissionless layer-two, looking for the next massive crypto play. The headline example? An anonymous trader turned eighty-six dollars into one-point-six MILLION dollars on a token called CASHCAT. That's an eighteen-thousand-six-hundred-X return. The AI bot's purpose is to find the next one, sifting through thousands of new tokens to spot patterns that signal a potential explosion. The creator spilled the details for free. And the takeaway is brutal: in this new world, you're either early, or you're exit liquidity. This isn't about long-term enterprise value. This is about using AI as a weapon for high-frequency, high-risk alpha.

And finally, there's the pushback. A post from Zvi Mowshowitz confirms the AI Protest is happening tomorrow, July eleventh, in San Francisco. It starts at noon. This isn't just online chatter anymore; it's hitting the streets. The protest is a physical manifestation of the anxieties that have been building for months around the speed of AI development, job displacement, and the concentration of power in a few companies. You have people asking for ten-thousand-dollar microgrants to fund safety research, and you have activists organizing to make their voices heard outside the corporate campuses. It shows that while the technologists and financiers are carving up the future, a growing part of the public is demanding a say in what that future looks like.

So you have these three currents all swirling at once. The grand strategists figuring out how to lock you into their ecosystem for the next decade. The gunslingers using AI to make fortunes overnight in the crypto casinos. And the activists trying to pump the brakes and ask if anyone has thought this through. Now let's go deeper on that first story. Because Arvind Narayanan's essay, titled "Up the Stack: How AI’s Escape From the Commodity Trap Risks Enterprise Lock-in," is the framework that makes sense of EVERYTHING else happening today. For the past two years, the big debate has been a false binary. You had one camp screaming "AI is a bubble!" and another camp arguing "No, foundation models are just becoming a commodity." The commodity argument went like this: eventually, there will be dozens of good-enough models, open-source will catch up, and the price for intelligence will drop to near-zero.

It’ll be like cloud computing — you just pick the cheapest provider. Narayanan says that entire frame is now obsolete. He's making a different claim. Frontier models are NOT becoming a commodity. And the reason is that the very things that make them powerful also make them defensible moats. Here's the logic. First, there's scale. Building and training a state-of-the-art model like GPT-5 or Fable-5 costs billions of dollars in compute. That’s a barrier to entry that only a handful of companies on the planet can cross. This is not something a startup can just spin up in a garage. This is infrastructure on the level of a national power grid. Second, there's proprietary data and reinforcement learning. The best models get better because they are used by millions of people, generating a constant firehose of data on what works and what doesn't.

This is Reinforcement Learning from Human Feedback, or RLHF. That feedback loop is a proprietary asset. A model that isn't deployed at scale can't get the data it needs to improve, so the leaders just keep pulling further ahead. And third — and this is the absolute key to the lock-in argument — is the move "up the stack." The game is no longer about selling raw API access to a model. The game is about integrating that model so deeply into essential software that it becomes inseparable. And that brings us back to Microsoft. The report that they're swapping in their own "MAI" models for OpenAI's inside Office is the strategy made manifest. Think about it. Your company spends a year building custom workflows, training employees, and integrating your data with Microsoft Copilot.

You're using it in Teams, in Excel, in Outlook. The AI becomes part of your company's institutional memory. Then, one day, Microsoft decides to raise the price. Or change the terms. What are you going to do? Rip it all out and switch to Google's ecosystem? Re-train your entire workforce? It's a nightmare. The switching costs become astronomically high. You are locked in. This isn't like switching your cloud provider from AWS to Azure, which is already hard enough. This is like trying to switch your company's collective brain. The AI isn't just a tool the employees use; it becomes the connective tissue between them. So what does it all add up to? The war for AI dominance is not a feature-to-feature battle between models. It is a platform war. The goal is not to have the smartest model, but the STICKIEST ecosystem.

Apple did this with the iPhone and the App Store. Microsoft did this with Windows and Office. Now, the same playbook is being run for artificial intelligence. They are building walled gardens, and the walls are made of your own data and workflows. The debate we were having last week about the cost of Fable-5? It misses the point. The high price isn't just about covering compute costs. It's about establishing a premium, defensible position at the top of the market, forcing deep integration, and then, once the customer is dependent... squeezing. So, as we look at the week ahead, the conversation has permanently changed. The threads you need to watch are no longer just about model performance benchmarks or new capabilities. That's the old game. The new game is about strategy.

It's about ecosystems. It's about lock-in. When you see a new AI feature announced, the question to ask is not "what can this do?" The question is "what does this bind me to?" When a company offers a cheap or free AI tool, the question is not "what's the catch?" but "what are the switching costs in two years?" The crypto bots and the street protests are just side-shows to the main event. The real action is in the quiet, deliberate architectural decisions being made inside Microsoft, Google, and Amazon. They are not just building AI. They are building the infrastructure of dependency for the next generation of business. This week's chatter proves the focus has shifted from the magic of the technology to the cold, hard calculus of market control. The race is on, not to build a better brain, but to build a stickier cage.