When Thomas Edison opened his commercial power station on Pearl Street in Lower Manhattan in 1882, he was selling electricity. Business owners bought Edison's electricity. His competitors were the gas-lamp companies, and Edison pitched his brighter, safer current as a product. Fifty years later, electricity was no longer a product — it was infrastructure. Nobody agonized over which company's electricity to subscribe to.
An essay by tech critic John Gruber on Daring Fireball revives that question. Is AI a product you sell, or a technology layer that dissolves into everything? Depending on the answer, your current AI subscription strategy points in entirely different directions.
The Signals the Market Is Already Sending
Look at the contours of the AI tools market in 2026 and you can see signs of convergence. OpenAI sells ChatGPT for $20 a month. Anthropic's Claude operates at a similar price point. Google Gemini and Microsoft Copilot compete in the same range. Compared to 2023, the feature gap between models has narrowed. When one vendor ships a new capability, competitors match it within weeks.
The direction of integration is just as telling. Apple Intelligence moved inside iOS and macOS — no separate purchase, built into devices people already own. Microsoft folded Copilot into the Office 365 subscription, and Google integrated Gemini into Workspace. If these companies believed AI was a standalone product, they would have shipped standalone apps. Instead, AI has been absorbed as a feature of products that already existed.
This is where Gruber's argument lands. In the late 1990s, AOL and CompuServe sold paid internet access in the United States. Both once commanded enormous subscriber bases. Almost nobody uses AOL today. As the internet converged into infrastructure, the differentiating value of access itself evaporated. If AI follows the same path, today's AI subscription services could go through a similar reckoning.
But the Analogy Has Holes
It's hard to accept this view at face value. There are several counterarguments, and some of them carry real weight.
The strongest is the model gap that exists right now. Electricity is physically identical no matter which power plant it comes from. AI models are not. They differ across platforms in reasoning ability, long-context handling, code-generation accuracy, and domain expertise. Anyone who has put each vendor's latest model to work on real tasks can feel the difference. As long as which model you use affects the quality of what you produce, AI remains, in large part, a product.
The timeline is another objection. Electric infrastructure took decades to settle. After Edison's power station came the standards war between direct current and alternating current, followed in turn by regulatory frameworks and massive cost declines. If AI needs a comparable stretch of time to converge into infrastructure, then buying it the way you buy a product stays valid for quite a while.
The revenue model differs too. Electricity is metered by the kilowatt-hour, and a kilowatt-hour is the same regardless of supplier. AI services are not like that. A given model's reasoning power, the creativity of its responses, the depth of its contextual understanding — each varies, and those variations are the basis of competition. A service that can't be uniformly metered doesn't easily become pure infrastructure.
What You Can Take From This Debate Right Now
Whichever way the argument ultimately tilts, there are decision criteria that solo founders and small teams can apply today.
Constantly swapping subscription tools does not raise your productivity. The flood of AI tool reviews has bred a kind of fatigue — the endless hunt for something better. But the more durable question is how to integrate the tool you already have into your workflow. If AI is converging into a technology layer, the asset worth building now isn't mastery of any particular tool; it's your own working methods and judgment for putting tools to use.
Platform dependence is worth auditing as well. If you keep your prompt structures, workflows, and output formats in shapes that aren't locked to a single service, they stay portable when the market shifts. The AI service you use today could change its pricing next year or get acquired by a competitor. The point is to decide in advance what you cannot afford to lose along with it.
In Korean working life, the more immediate problem is the subscribed-but-unused state. Buying an AI tool feels like an accomplishment in itself. But adopting a technology layer demands integration design to follow: deciding which stage of which task gets AI, and how its output will be reviewed. Without that design, the more subscriptions pile up, the wider the gap between your actual work and your tools grows.
And the difference between people who do this design well and those who don't has nothing to do with which AI tool they subscribe to. If anything outlasts technical skill in the AI era, it is the ability to judge how a technology should be woven into your own work. The path to building that judgment opens not through hunting for new tools but through using your current ones deeply. Your own approach and working methods for putting a tool to use transfer with you, even when the tool changes.
The companies that survived electricity's transition to infrastructure had one thing in common, and it wasn't owning electricity or buying it first. Ford integrated electric motors into its factory lines and drove down production costs. Countless factory owners of that era likewise figured out how to fit electricity to their own manufacturing processes. Whether AI takes ten years or twenty to settle in as a technology layer, what accumulates in the meantime is not a history of subscriptions to particular tools.
