When Software Becomes Free, What Happens to the Economy?

Last weekend, I built an autonomous engineering team in 48 hours. Not a metaphor. Not a "vision." I sat at my desk in Dubai, talked to a chatbot on Telegram, and watched two AI agents coordinate a full development loop. One playing PM, the other playing CTO. They wrote code, reviewed each other’s work, raised pull requests, merged them, and sent me a notification when they were done.

Total infrastructure cost: under 30 cents a day.

I’ve managed engineering teams of 50+ across Apple, Netflix, KPMG, Tata. I know what that kind of capability costs in salaries, benefits, Jira licenses, and passive-aggressive Slack threads about code review etiquette. So when I tell you something fundamental has shifted, I’m not selling a course. I’m reporting from the building site. And the building site is getting cheaper by the week.

The Numbers Nobody’s Processing (Or: Why Your SaaS Valuation Is a Fiction)

Right. Let’s get the data on the table before the thinkpiece crowd shows up.

GitHub Copilot now has 20 million users. It writes 46% of all code for active users, rising to 61% for Java developers. Developers using it complete tasks 55% faster. Pull request turnaround at companies like Duolingo dropped from 9.6 days to 2.4. Ninety percent of the Fortune 100 have deployed it. This isn’t experimental any more. It’s infrastructure.

Google told investors that more than 25% of all new code at the company is AI-generated, reviewed by engineers, and shipped to production. That number hit 30% by early 2025. And Google still calls this "boosting productivity." Bless them. That’s like calling the printing press a "nice addition to the monastery."

Meanwhile, the lean AI startup movement has gone from cute to properly alarming. Cursor went from zero to $100M ARR in 21 months with 20 people. Midjourney hit $200M with 10. Bolt did $20M in two months with 15. Lovable, a European startup, hit $17M ARR in three months and its founder reckons his software will build "80% of all SaaS" by year’s end.

Sam Altman has a betting pool with his mates for the first one-person billion-dollar company. Y Combinator is actively funding solo founders with $500K and telling them they can build multi-billion-dollar businesses.

The Lean AI Native Leaderboard now tracks dozens of sub-50-person companies generating revenue that would’ve required hundreds of engineers five years ago.

McKinsey, who you’d think would be nervous about this given they sell human consultants by the kilogram, still stand by their estimate that generative AI could add $2.6 to $4.4 trillion annually to the global economy. Mostly through automating the kind of knowledge work that currently pays for a lot of school fees in Palo Alto.

This is not hype volatility. It’s repricing.

Right. So What Actually Gets Valuable? (Not What LinkedIn Thinks)

Here’s where every AI influencer with a ring light and a Notion template will tell you to "upskill" and "stay curious." Brilliant. Thanks. Very helpful. Let me just go upskill at the upskill shop.

The actual answer is more specific and a lot less comfortable.

For 30 years, software companies charged you for execution. Engineering leverage. Feature access. Hosting. Product velocity. You paid because building the thing was hard and expensive and required a room full of people who argued about tabs versus spaces.

But when AI writes the features, infrastructure auto-scales, APIs generate themselves, and interfaces get built during a coffee break, what exactly are you paying $99 a month for?

Money doesn’t vanish. It migrates. And I can see five places it’s going.

1. Trust Infrastructure (Where the Real Money Lives)

This one’s personal because I’ve been building it.

When code is cheap, trust becomes expensive. AI systems are brilliant at generating functionality. They are not, and I cannot stress this enough, secure, auditable, compliant, or legally defensible by default. There is a massive gap between "it works" and "it survives an audit."

I spent the weekend wiring up HMAC-signed communication between my AI agents. Append-only audit logs. Cryptographic verification on every operation. Path traversal prevention. Shell injection protections. A tiered command whitelist. Symlink protection on writes.

You know what none of that came free with the AI? All of it.

Every single line of security infrastructure was designed by a human who understands what can go wrong when autonomous systems start touching production environments.

As AI agents start influencing payroll, healthcare, logistics, finance, and government, the question stops being "can it do the thing?" and becomes "can you prove what happened, who authorised it, and why?"

That requires immutable audit logs, cryptographic signing, secure agent-to-agent protocols, access control governance, escalation checkpoints, human-in-the-loop validation, version traceability, and decision provenance.

Most AI-generated systems implement precisely none of this by default.

If you can build, integrate, or govern these layers, you’re not competing with free software. You’re monetising risk. In regulated industries, that’s a multi-billion-pound opportunity.

2. Domain Expertise Becomes Programmable (And That’s the Bit Nobody’s Ready For)

AI doesn’t kill expertise. It gives it a distribution channel.

The coding barrier collapsed. The thinking barrier didn’t.

The people winning right now aren’t the fastest prompters. They’re the deepest thinkers in a vertical.

Radiologists using AI to pre-screen scans while retaining diagnostic authority. Legal specialists building AI-assisted contract analysis. Healthcare veterans launching triage systems. Industry experts who spent 20 years learning edge cases, regulatory nuance, and failure modes that never make it into documentation.

All of that knowledge used to be trapped inside organisations, slowly leaking out when someone retired or got poached.

Now it’s programmable.

And programmable knowledge scales in ways that human knowledge never could.

The guy who understands why a particular building regulation exists in a specific emirate, and knows the three exceptions that nobody documents? He can now encode his entire career into a system that runs 24/7.

That wasn’t possible two years ago.

If you understand your vertical deeply enough to know where the bodies are buried, you just got a lot more valuable.

Not because you learned a new skill. Because the barrier between your knowledge and a working product just disappeared.

3. Orchestration Is the Actual Job Now

Here’s what nobody talks about at the AI conferences, probably because it’s boring and doesn’t fit on a slide: free AI agents without orchestration create absolute chaos.

Multi-agent systems introduce duplication, drift, security gaps, conflicting outputs, lost context, and runaway automation.

I know this because I spent an afternoon debugging why my PM agent and CTO agent were writing conflicting code to the same file. The answer involved a unicode canonicalisation mismatch in the HMAC signing layer. Python’s json.dumps was escaping non-ASCII to \uXXXX while Node’s JSON.stringify was outputting literal characters. Different strings, different hashes, angry bridge server.

Nobody tells you about the unicode bugs in the keynotes.

The future belongs to orchestrators. People who design workflow architecture, agent responsibilities, guardrails, escalation protocols, feedback loops, human checkpoints, and security boundaries.

A single operator with proper orchestration can now supervise what previously required teams of 10 to 50. That’s deflationary pressure on labour. It’s also enormous leverage for those who understand systems.

The job isn’t "write code." It’s "design coordination."

If that sounds like what a good engineering manager does, congratulations. You’re paying attention.

4. Distribution Wins. Always Has, Always Will.

If everyone can build, whoever owns the customer wins.

This is where incumbents still hold power. Embedded contracts. Brand trust. Procurement relationships. Regulatory approvals. The boring stuff that nobody makes YouTube videos about.

But smaller players can weaponise speed and intimacy. AI-native founders are already targeting micro-verticals, delivering systems tailored to specific workflows, iterating in days instead of quarters, and owning the customer relationship end to end.

Features don’t differentiate any more. Trust does. Responsiveness does. Reliability does.

Distribution has always mattered. In a world where software is free, it dominates.

5. Proprietary Data Is the New Land Grab

Models are improving fast. But models trained on unique, structured, proprietary data are defensible.

Raw data isn’t enough, though. Everyone has data. Most of it is rubbish.

What matters is structured, permissioned, continuously learning systems. Clean operational logs. Historical case data. Domain knowledge that’s been organised by someone who actually understands what it means.

Organisations that own this kind of refined data will compound advantage.

Everyone else will be training on the same public datasets and wondering why their outputs sound exactly like everyone else’s.

This is the new oil. But only if refined. Crude just sits there making a mess.

The Bit Where I’m Supposed to Inspire You (But Won’t)

Software development is being commoditised. The economic consequences aren’t hypothetical. They’re showing up in hiring data, revenue-per-employee ratios, and venture capital term sheets right now.

Over the next two years, this accelerates.

If your competitive advantage depends purely on building software features, you’re exposed.

If your skillset depends on manual execution of tasks an AI agent can do while you sleep, you’re exposed.

This isn’t alarmism. It’s trajectory.

I can see the curve because I’m literally riding it. Running an autonomous engineering team from my phone, for the cost of a dodgy takeaway coffee.

You don’t need to master all five layers. But you do need to pick one.

Own trust. Own orchestration. Own distribution. Own proprietary data. Own domain intelligence.

Execution is becoming abundant. Judgement, coordination, trust, and relationships aren’t.

Your move.

If you want real insights on AI, orchestration, and the future of work, without the inspirational LinkedIn bollocks, subscribe to the newsletter. No 10-step frameworks. No manifestation. Just a bloke in Dubai building things and telling you what actually happens.