In 2025, four companies — Microsoft, Alphabet, Amazon, Meta — spent over $300 billion on AI data centers. The combined 2026 number is forecast at $725 billion, a 77% jump in a single year. For comparison, the entire global SaaS market in 2025 is roughly $295–370 billion depending on whose definition you use. The capital being poured into the thing that replaces software is now equal to or larger than the software market it competes with.
If you run a software business, "how do we add AI features" is the wrong question. The right one is whether your product would exist at all if you started the company today.
The Receipts on the Wrong Side
Chegg is the canonical example. The homework-help business — students paying $14.95/month for textbook answers — was structurally fragile already, but ChatGPT made it terminal. Chegg's revenue fell 39% in 2025 ($618M → $377M), the homework subscription business dropped 43% in the same year, and the stock is down 99% from its 2021 peak. The CEO told investors in late 2025 that Google's AI Overviews launch was "as material" to the collapse as ChatGPT itself.
Chegg did not lack AI features. They launched CheggMate, a GPT-4 study tool, in April 2023 — six months after ChatGPT's debut. They built AI tutors, AI study guides, AI essay help. None of it stopped the decline. The features were not the problem. The product was the problem. They were selling paywalled answers to questions ChatGPT was giving away free.
Stack Overflow followed a similar arc. Question volume collapsed almost immediately after ChatGPT's November 2022 launch — developers stopped asking on Stack Overflow because the AI was faster and trained on Stack Overflow's data. The 2025 Stack Overflow Developer Survey confirmed it: 84% of developers now use AI tools daily, and 79% rely on ChatGPT. Stack Overflow eventually licensed its data to OpenAI in 2024, but the licensing revenue does not replace the community engagement that drove the original product.
The pattern in both cases is the same. A workflow-automation business — Chegg automated finding textbook answers, Stack Overflow automated finding code answers — gets eaten when the underlying knowledge becomes free to query directly. The interface that used to mediate access stops being valuable when the access is direct.
The Receipts on the Right Side
Duolingo did the opposite trade. In 2023 they introduced Duolingo Max, a higher-priced tier built around AI — conversational roleplay with characters, AI grammar explanations on every wrong answer, AI-generated personalized lessons. They didn't bolt AI onto Duolingo Plus. They built a new product tier where AI was the product, and priced it above Plus.
The 2025 results: revenue crossed $1 billion ($1.01–1.02B annual), up over 50% year-over-year, with AI features driving 51% user growth. The bet was that language learning at any price point gets better with AI, and the customers who valued speed of progress would pay for the better version.
Adobe took a different but related path. Adobe Firefly, launched as a generative imaging model in 2023, has been embedded into Photoshop, Illustrator, Premiere, and the standalone Firefly app. As of Q3 FY2025, Firefly recorded 29 billion total generations with 40% quarter-over-quarter growth in video. Adobe's FY2025 revenue hit $24.05B, up 11%. The pivot here was structural: Adobe stopped treating creative software as the product and started treating creative output as the product, with AI as the engine for generating it.
Klarna sits in the middle of the framework. They took the disruption seriously enough to own it — their AI assistant handles two-thirds of customer service chats, doing the equivalent work of 700 full-time agents — and even after partially reinvesting in human support, the AI still handles the volume work. The pivot wasn't "we sell AI now." It was "we automated our own cost center before someone else automated it for us." A different strategic posture from Chegg, who tried to retroactively add AI features to a product the AI was making redundant.
The Strategic Question
Three moats survive an AI capex shift of this magnitude.
The first is data. Bloomberg's terminals survive because the data feed is proprietary and the licensing structure is decades old. MLS data for real estate survives for the same reason. If your customers cannot get your data from a public AI model, you have time.
The second is workflow with capture. The product owns a system of record that AI tools cannot easily reach into, and the friction of integration is what holds the position. ServiceNow, Workday, and Salesforce all sit in this category, though they are each spending heavily on AI features because the moat is shrinking.
The third — and most interesting — is owning the AI consumption layer itself. This is where Duolingo and Adobe sit. AI capability becomes a commodity; packaging that commodity for a particular user job is the product. The capex flowing into hyperscaler data centers builds the substrate. The product is what sits on top of the substrate, charging users for the application of the capability.
The wrong moat is workflow automation as a pure interface. Chegg's product was, structurally, "we make it convenient to look up textbook answers." Stack Overflow's was "we make it convenient to find code answers." Both moats vanished when the AI made the underlying capability free and direct. Any business whose pitch is "we automate X" is at risk if X is a knowledge-work pattern the model can reproduce.
How to Tell If You're Chegg
The diagnostic question is uncomfortable. If you started your company today, with full knowledge of ChatGPT's capabilities and access to frontier model APIs, would you build this product?
If yes, you have a real moat. Build harder, faster, ship more.
If no — if the honest answer is "we'd build something else, but we have customers and revenue so we're going to keep adding features" — you are Chegg in 2023. The features will not save the product. The strategic move is the pivot itself, not the feature roadmap.
In yesterday's piece on cloud support roles, I argued that AWS L1/L2 customer support is the first agent target because pattern-match-on-logs-and-escalate is what an LLM with tool use eats for breakfast. The same logic applies to your product. If the customer outcome you sell is "find the answer to X" or "summarize Y" or "automate Z workflow," ask whether a $20/month ChatGPT subscription plus a willing customer can produce 80% of the outcome you charge for.
If yes, the capex has already moved. Build like it.
