Gruber argues AI will follow the same trajectory as the internet: nobody buys 'the internet' as a product, but every product worth using has it woven through it. He contends that companies leading with 'we use AI' as their entire pitch have an expiration date, and that integrated, invisible approaches like Apple's will prove more durable than 'AI-first' branding.
Gruber highlights the 'wrapper problem': if your product is a thin UI on top of an API call to Claude or GPT, you have a demo, not a product. He points to the 2023-2024 wave of GPT wrapper startups as evidence — survivors pivoted to owning a workflow, dataset, or distribution channel, while those that stayed as wrappers disappeared entirely.
The editorial reinforces that the moment an underlying model provider ships a better native interface or a new model makes existing prompt engineering obsolete, the moat of wrapper products evaporates. It frames this as already proven by market attrition among thin-wrapper startups.
Gruber uses the technology-vs-product framework to argue that Apple's characteristically slow, integrated, and invisible approach to AI adoption may prove more durable than the AI-first branding strategies of companies like OpenAI. His reasoning is that treating AI as a feature layer within existing products, rather than as the product itself, builds lasting competitive advantage.
John Gruber, writing at Daring Fireball, published a piece this week that landed with a 404-point score on Hacker News — fitting, perhaps, for an argument about things that are hard to find when you look for them directly. His thesis: AI is a technology, not a product. The distinction sounds semantic. It isn't.
Gruber's core claim is that the companies and developers who will win the AI era are not the ones building "AI products" — they're the ones building *products* that happen to use AI as a foundational technology layer. The parallel he draws is to the internet itself: nobody buys "the internet" as a product, but every product worth using has the internet woven through it. The same fate, he argues, awaits AI.
This lands at a moment when the market is saturated with startups whose entire pitch is "we use AI." Thousands of YC applications, hundreds of Product Hunt launches, and a non-trivial chunk of Series A decks all lead with the same premise: AI is the product. Gruber's argument is that this framing has an expiration date — and it's approaching faster than most founders think.
The technology-vs-product framing isn't just philosophical. It has direct consequences for how software gets built, funded, and adopted.
The wrapper problem is real. If your product is a thin UI on top of an API call to Claude or GPT, you don't have a product — you have a demo. The moment the underlying model provider ships a better native interface, or the next model makes your prompt engineering obsolete, your moat evaporates. This has already played out: remember the wave of "GPT wrapper" startups in 2023-2024? The survivors pivoted to owning a workflow, a dataset, or a distribution channel. The ones that stayed as wrappers are gone.
Gruber's framing explains why Apple's approach to AI — slow, integrated, invisible — may ultimately prove more durable than the "AI-first" branding that companies like OpenAI and Anthropic necessarily lead with as platform providers. When AI is a technology, it disappears into the product. You don't think about the AI in your camera's computational photography. You just think the photo looks good. The best AI features are the ones users never consciously interact with as "AI."
This also reframes the competitive landscape. Google, Microsoft, Apple, and Amazon aren't really competing to build the best AI product. They're competing to integrate AI most effectively into their existing product surfaces — search, productivity, devices, cloud. The standalone AI chatbot is a transitional form, not an end state. It's the equivalent of the early web portal: useful for demonstrating what the technology can do, but not where the long-term value accrues.
The counter-argument deserves airtime. Some would argue that foundational model companies *are* building products — that ChatGPT and Claude are genuine product experiences, not just technology demos. There's merit here. ChatGPT has over 100 million weekly active users. That's a product by any reasonable definition. But Gruber's point isn't that these don't exist as products today — it's that the category "AI product" will dissolve, the same way "internet product" dissolved. Netflix is a streaming product. Amazon is a retail product. They're both internet companies, but nobody calls them that anymore because the internet is assumed.
The internet analogy is useful but imperfect. Here's where it holds and where it breaks.
Where it holds: In the late 1990s, simply having a website was a competitive advantage. By 2005, not having one was a disqualification. The value shifted from "we're on the internet" to "we solve your problem, and obviously we use the internet to do it." AI is on a similar trajectory — the competitive advantage of "we use AI" is depreciating rapidly, and within two to three years, it will be table stakes rather than a differentiator.
Where it breaks: The internet was primarily a distribution and communication technology. AI is a capability technology — it can *do things* that weren't possible before, not just distribute things faster. This means AI-native products (products that literally cannot exist without AI) have a stronger claim to being "AI products" than any website had to being an "internet product." Code generation tools, real-time translation, and autonomous agents are qualitatively different from what came before. They aren't just faster horses.
But even granting this, Gruber's thesis holds at the macro level. The developers building code completion into VS Code extensions aren't shipping "AI products" — they're shipping developer tools. The teams building AI-powered fraud detection aren't in the AI business — they're in the fintech business. The technology fades into the background as the product matures.
If you're building software today, Gruber's framing suggests a few concrete moves.
First, decouple your product identity from your AI provider. If your pitch deck says "powered by GPT-4" or "built on Claude," you're advertising a dependency, not a moat. The defensible value is in the workflow you've built, the data you've collected, and the problem you've understood deeply enough to solve — not in which API you call. Model providers will leapfrog each other every six months. Your product needs to survive those transitions.
Second, invest in the boring parts. The AI call is the easy part. The hard parts are data pipelines, evaluation frameworks, error handling for non-deterministic outputs, and user experience that degrades gracefully when the model gets it wrong. These are engineering problems, not AI problems. Treat them accordingly.
Third, watch for the integration layer. The biggest opportunity in the next two years isn't building AI products — it's building the infrastructure that lets existing products absorb AI capabilities painlessly. Middleware, orchestration frameworks, evaluation tooling, and observability for AI-augmented systems. This is where the "AI as technology" thesis creates the most greenfield opportunity.
Gruber's argument is ultimately optimistic for practitioners, even if it's bearish for the "AI startup" category as currently defined. When AI becomes a technology rather than a product, the advantage shifts from those who understand AI best to those who understand their *domain* best and can apply AI to it effectively. That's good news for developers, domain experts, and product thinkers. It's less good news for the current cohort of AI-wrapper startups running on vibes and venture capital. The shakeout is coming. The builders who survive it will be the ones who always knew they were building products, not AI.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.