Bread Paradox: Why Vibe coding can’t kill the SaaS software industry?

AI writing code scares software stocks, but cheap code doesn’t mean cheap end-to-end services. Research shows the number of major defects in AI-generated code is about 1.7 times that of code written by humans. Gartner observed that enterprise SaaS renewal price increases still generally fall in the range of 10% to 20%.
（Background: The worst start of the year! U.S. software stocks crashed because Claude Code got too popular）
（Context: When AI fills the SaaS moat, software companies only have three ways to survive）

Table of Contents

Toggle

• For five thousand years, humans have always chosen to buy
• AI makes code cheaper, but not the supply chain
• Thin products will die; the supply chain won’t

A bread machine can make a loaf of bread within three hours, with material costs of less than three dollars, and the machine itself costs no more than a hundred dollars. In theory, everyone could bake at home. But in reality, Americans still buy roughly ten million pre-baked loaves every day. This contrast, which tech writer Joan Westenberg calls the “bread paradox,” is exactly what we can use to test the panic that hit software stocks in January—when they wiped out about $300 billion in a single day.

Back then, Claude Cowork and Claude Code were all the rage, and the market started calling it “SaaSpocalypse” (the SaaS apocalypse). But if cheap recipes and cheap machines haven’t shut down the baking industry for five thousand years, then cheap code likely won’t either.

For five thousand years, humans have always chosen to buy

As early as 3000 BC, the ancient Egyptians were already running commercial baking shops along the Nile. The Romans industrialized the business; when Pliny wrote Natural History, Rome already had professional bakers’ guilds, dough-mixing machines powered by animals, and a logistics network that delivered bread to hundreds of thousands of city residents—almost none of whom had ever baked bread themselves.

In medieval London, bakers’ guilds received royal charters as early as the 12th century. Even bakers who sold loaves by light weight could be tied to sleds and dragged through the streets for public display. The pastoral fantasy of “every household baking its own bread” only exists, in practice, when bread is otherwise unavailable.

In 1928, Otto Rohwedder invented the commercial slicing machine. In 1961, the Chorleywood process compressed baking time from hours down to minutes. Today, Americans consume about 21 million tons of baked bread products each year and buy about ten million pre-baked loaves every day. Even if flour is nearly free, the bread machine costs under $100, and recipes have been passed down for five thousand years.

In an essay, tech writer Joan Westenberg points out that the answer lies in the economics “make-or-buy” decision. In simple terms, a rational person does DIY only when the cost is truly lower. Most people underestimate the hidden costs of “making it yourself”—buying ingredients, turning knobs and switches, waiting for fermentation, and doing the cleanup afterwards. Each step seems small when viewed alone, but repeating it for a lifetime ends up costing a lot.

Orwell once complained that industrialized British bread was “pale, soft, and tasteless,” but people still buy it because the psychological cost of buying bread is lower than baking it oneself.

AI makes code cheaper, but not the supply chain cheaper

The argument that “SaaS is dead” sounds tempting: with a good AI model plus a decent prompt, you can generate a customized CRM or analytics dashboard in an afternoon. Code is almost free, servers are cheap—why pay a monthly fee?

But when a company pays to subscribe to Notion, Jira, or Basecamp, what it’s buying is never the code itself. Instead, it’s the institutional knowledge accumulated over years by thousands of engineers, compliance staff, and security auditors; plus the integrated ecosystem, regulatory certifications, and the support infrastructure.

When you use AI to “build” a company for yourself, what you’re really getting is a bread machine. The ingredients are cheap, and the machine handles much of the work—but you end up being the baker: you have to handle maintenance, edge cases, and the security gaps that AI-generated code can introduce. Research shows that the number of major defects in AI-generated code is about 1.7 times that of code written by humans. Six months later, the person who wrote the system switches departments—no one else understands how it works anymore. And when things break at 2 a.m., no one can pick up the phone.

Gartner observed that recent enterprise SaaS renewal price increases are often in the range of 10% to 20%, exceeding the budget growth rate of most CIOs. This sounds like vendors are raising prices amid the chaos, but buyers don’t seem to plan to run. Avenir’s January 2026 report shows that 63% of enterprise procurement teams expect that existing software vendors will “benefit” from generative AI, while only 8% think they will “be harmed.”

The direction the market is betting on is clear: customers want existing services to evolve with AI—not to be replaced and rebuilt. Even Klarna, commonly cited as an example of “building to beat SaaS,” didn’t use AI to create a system from scratch to replace Salesforce. Instead, it switched to another SaaS stack plus some self-built components; its teams still rely on Slack within Salesforce to this day.

Thin products will die; the supply chain won’t

What should genuinely worry people are products that sell subscription access to features that can be copied with a single AI prompt: converting PDFs into tables, automatically generating meeting minutes, sending follow-up emails. These single-feature tools were thin by nature.

But SaaS companies with deep integrations, proprietary data, regulatory certifications, and years of business logic and partner ecosystems map to the full industrial baking complex. Even individual hobbyists can bake a loaf, but they have never threatened commercial baking because what bakers sell was never flour and recipes—it was the guarantee of stability, consistency, and “someone will take responsibility when something goes wrong.”

What will change next is the pricing model. As AI agents become a new type of software user, per-seat pricing will gradually give way to usage-based and outcome-based models. Those thin, single-feature products will die—and they should. They were never a real business. They only existed in an era when software development was so expensive that even trivial features could be charged monthly, repackaged as “a company feature.”

The real core logic of SaaS is paying for a solution rather than taking on the entire problem yourself. That logic has held from the Roman Empire to today, and it never depended on technological barriers. It depends on human nature: as long as the price is reasonable and trust remains, people will always prefer to pay someone else to handle the headaches.