75% AI Code and the Maintenance Tsunami Nobody's Budgeting For

The 75% Number Everyone's Celebrating Is Actually a Warning

Google confirms that 75% of its new code is now AI-generated. The headline reads like a productivity miracle. CEOs are forwarding it to their CTOs with a note that says "why aren't we doing this?" CTOs are forwarding it to engineering leads with a note that says "let's talk Monday." And I'm sitting in Jönköping, running a tech company that builds software for a living, thinking: who is budgeting for maintaining all of this?

Nobody. That's who.

This is not a productivity story. This is the beginning of a codebase obesity crisis. Organizations are generating technical debt faster than any human team can audit, review, or even comprehend. And the Swedish companies racing to adopt AI code generation without investing equally in AI code review and supply chain security are building on a foundation they will not be able to maintain in 18 months.

I'm going to explain why. And then I'm going to tell you what to actually do about it.

The Math That Should Keep You Up at Night

Let's do some simple arithmetic. If 75% of new code is AI-generated, and code generation speed has increased by roughly 3-5x (conservative, based on what we see in our own AI solutions work at HEIMLANDR), then the volume of code entering production is growing exponentially while the number of senior engineers who can review it remains flat. In many cases, that number is shrinking.

The New York Times is now reporting on what they call a "code overload" problem. That's a polite way to put it. What's actually happening is this: AI writes code that works. It passes tests. It ships. But it's often verbose, subtly redundant, and structured in ways that make future modification harder. Not broken. Just heavy. Layer that over 12 months and you've got a codebase that does what it's supposed to do today but fights you every time you try to change it tomorrow.

This isn't theoretical. We see it in client projects. Companies come to us at HEIMLANDR for MVP development and increasingly, they're bringing codebases that were rapidly generated by AI but are already tangled after a few months. The speed was real. The savings were real. But the bill is coming due and it's bigger than what they saved.

Technical Debt Has Always Existed. This Is Different.

Traditional technical debt accumulates because humans take shortcuts under pressure. We know this pattern. We've built entire methodologies around managing it. Refactoring sprints. Architecture reviews. Code quality gates.

AI-generated debt is different in three specific ways:

1. Volume. Humans taking shortcuts produce debt at human speed. AI produces it at machine speed. The ratio of debt generation to debt resolution has fundamentally shifted.

2. Legibility. When a human writes a shortcut, another human can usually read the intent. AI-generated code often lacks the contextual reasoning that makes maintenance possible. It works. But nobody on the team truly understands why it was structured that way, because it wasn't. It was statistically predicted.

3. Supply chain opacity. AI models have been trained on the entire internet. They reproduce patterns, dependencies, and sometimes outright code snippets from open-source projects with various licenses and various levels of maintenance. Manifest just launched a C/C++ SBOM generator specifically to address supply chain security blind spots. That's not a product launch. That's a distress signal from the infrastructure layer.

The Swedish Situation: Fast Adoption, Slow Governance

Sweden is a paradox right now. We punch above our weight in software development. Always have. Stockholm is Europe's second-largest tech hub per capita. We've got Spotify, Klarna, King. And Swedish companies are adopting AI code generation fast. Probably faster than most of Europe, because our engineering culture is pragmatic and our startups move quick.

But here's what I see from Jönköping that the Stockholm bubble sometimes misses: the governance and review infrastructure is not keeping pace. Not even close.

Swedish companies, especially in the Småland region and the broader mid-market, are deploying AI agent development tools and code generation at scale without corresponding investments in AI code review, automated security auditing, or SBOM compliance. The EU Cyber Resilience Act is coming. The AI Act is already here. And the average Swedish mid-sized tech company is generating code faster than their compliance team can even categorize it, let alone audit it.

Compare this to what's happening in the US. American big tech has scale. Google can generate 75% AI code because Google also has massive internal review pipelines, enormous testing infrastructure, and teams of hundreds dedicated to code quality. When a 50-person Swedish SaaS company tries to replicate that ratio without the review infrastructure, they're not copying Google's success. They're copying Google's input without Google's controls.

In Asia, particularly China and South Korea, the approach is different again. Heavy government investment in AI, but also heavy mandates around code auditing in critical infrastructure. Sweden's approach is more laissez-faire. That's usually our strength. Right now it might be our blind spot.

Breakit has been covering the Swedish AI boom with appropriate enthusiasm. I'd like to see them start covering the maintenance hangover. It's coming.

The Skills Gap Nobody Wants to Talk About

Anthropic just published research on how AI assistance impacts the formation of coding skills. This is the quiet bomb underneath all the productivity stats.

If 75% of code is AI-generated, and junior developers are spending most of their time prompting and reviewing AI output rather than writing code from scratch, then the pipeline of engineers who deeply understand codebases is narrowing. We are training a generation of developers who are excellent at directing AI but may struggle to maintain, debug, or fundamentally restructure the systems that AI builds.

This matters enormously for software development in Sweden. Our competitive advantage has always been deep engineering talent. Not the cheapest. Not the most numerous. The best, per person. If we lose that edge because our junior engineers never develop the muscle memory of building from the ground up, we're trading a long-term advantage for short-term velocity.

I'm not saying stop using AI. I use it every day. We build AI agent systems at HEIMLANDR for clients. AI code generation is real and it's powerful and it's not going away. But I am saying: if you're a CTO in Sweden right now and your junior devs are spending less than 30% of their time writing code without AI assistance, you are creating a future maintenance problem that no tool will solve. Because the tool needs someone who understands what it's doing.

Where This Goes: 2027-2030

Let me lay out the trajectory as I see it.

By late 2027: AI code review tools become as mandatory as CI/CD pipelines. Augment Code is already testing 10 open-source AI code review tools on a 450K-file monorepo. That's not a research project. That's the market scrambling to build the thing that should have existed before we turned on the firehose. Expect consolidation. Two or three tools will win, and they'll be required in any serious engineering org.

By 2028: The first major "AI code liability" lawsuits hit. A company ships AI-generated code that contains a vulnerability traceable to training data. Or an open-source license violation surfaces in production code that nobody on the team realized was there. The EU Cyber Resilience Act gives regulators the teeth to act on this. Swedish companies that haven't invested in SBOM tooling and provenance tracking will be exposed.

By 2029-2030: As we approach more capable AI systems, the dynamic shifts again. AI that can not only generate but genuinely audit, refactor, and maintain code changes the equation. But we're not there yet. We're in the gap. The dangerous period where generation outpaces comprehension. The decisions you make now about code quality infrastructure determine whether you survive the gap or drown in it.

The path toward AGI doesn't make this problem go away. It makes the stakes higher. Because an AGI-capable system maintaining code it generated is one thing. An AGI-capable system trying to maintain code that was generated by a less capable model, poorly reviewed by humans who didn't understand it, and patched repeatedly by other AI tools, is a nightmare scenario. The quality of what we build today is the foundation for what AI maintains tomorrow.

What to Look At

If you're a founder, CTO, or senior engineer reading this, here are specific things worth your attention right now:

OpenCode (149K+ stars on GitHub). An open-source coding agent that's gaining traction fast. Worth understanding not because you should necessarily use it, but because it represents the direction: agents that don't just generate snippets but operate across entire codebases. Understanding how these work helps you understand the review problem.

n8n (185K+ stars). Fair-code workflow automation with native AI capabilities. We use n8n in several client projects at HEIMLANDR. It's relevant here because it lets you build automated code quality and review workflows without being locked into a single vendor's pipeline. Self-hostable. Auditable. That matters when you're trying to keep control over what's happening in your codebase.

Manifest's SBOM tooling. Their new C/C++ SBOM generator addresses a real gap. If you have any C/C++ in your stack, or dependencies that do, you need to know what's in there. The supply chain attack surface is growing with every AI-generated dependency suggestion.

Everything Claude Code (166K+ stars). A performance optimization system for AI coding agents. This one is interesting because it focuses on skills, memory, and security for AI agents. If you're going to use AI agents to write code, at least optimize how they do it. This repo is a practical starting point.

What To Actually Do on Monday Morning

Vision is nice. Here's the practical list.

Audit your AI code ratio. Do you even know what percentage of your codebase was AI-generated in the last 6 months? If you don't know, that's your first problem. Measure it.

Invest in review before you invest in more generation. For every dollar you spend on AI code generation tools, spend at least 50 cents on AI code review, security scanning, and SBOM compliance. If that ratio seems high, you're underestimating the cost of the debt you're creating.

Protect your junior engineers' learning. Mandate time where they code without AI. Weekly. Not optional. This is not nostalgia. This is maintaining your organization's ability to understand its own systems in 2028.

Build your SaaS and product architectures for auditability. Modular. Well-documented. Clear boundaries between services. AI-generated code in a well-architected system is manageable. AI-generated code in a monolith with unclear boundaries is a time bomb.

Talk to your legal team about AI code provenance. If you're shipping products in the EU, the regulatory framework is tightening. Know where your code comes from. All of it.

Closing Thought

I build things. That's what we do at HEIMLANDR, from here in Jönköping. We build AI agents, we build SaaS products, we build for clients who want to move fast. I'm not anti-AI. I'm anti-stupid.

And right now, the industry is being stupid. Celebrating a 75% generation stat while ignoring the maintenance math is the kind of short-term thinking that creates long-term disasters. The companies that win the next five years won't be the ones that generated the most code. They'll be the ones that can still understand, modify, and trust their codebases when the pressure hits.

Build fast. But build like you're the one who has to maintain it. Because you are.

Fredrik Brunnberg is the CEO of HEIMLANDR.IO, building AI and software solutions from Jönköping, Sweden. This is the daily HEIMLANDR briefing. If you found this valuable, share it with someone who builds things.