Over the years, I’ve lived through multiple waves of “revolutionary” technology: 4GL programming languages, “visual” programming languages, “agile” methodologies, SOA, various web frameworks, cloud computing, serverless, containerization, k8s, no code, and now “AI” coding. The Ruby ecosystem has had its share of “revolutionary” entrants, with Ruby on Rails and Chef being two that I’ve directly been involved with. Some of these, while they didn’t completely live up to their hype, are still deeply embedded in our day-to-day activities in software engineering. Others had impacts to our ways of thinking without being distinctly present in the tech we currently use.

Today, I want to talk about what might be the biggest subsidy in tech history – and why its end could be the best thing that ever happened to experienced software engineers.

Note: This all assumes that “AI coding” isn’t just a very good illusion that has duped a large portion of us. I’m not 100% convinced that it isn’t, but it’s a plausible enough illusion to study it in depth.

The $20 Illusion

Right now, you can get ChatGPT Plus or Claude Pro for $20 a month. Unlimited usage. State-of-the-art AI that can write code, debug problems, and handle complex reasoning tasks. It feels like we’re living in the future.

But here’s the thing: that $20 isn’t even close to covering the real costs.

I started researching this after noticing how my own usage patterns had evolved. Like many developers, I’ve built up instructions and context documents, listened to the “best practices”, feed entire codebases into context windows, and run complex multi-turn debugging sessions. The convenience is intoxicating. But when digging into the actual token costs, I realized perhaps all living in a fantasy.

The Mathematics of Delusion

Note: the estimates in the sections below are the output of Claude Sonnet 4.0. That content will be in quote sections.

What do the “best practices” AI usage actually cost at current API rates?

A typical power user session might include:

• System context and prompts: 50K tokens
• Code repository analysis: 200K tokens
• Multi-turn debugging conversation: 100K tokens
• Total: ~350K tokens per session

At current API pricing:

• GPT-4o: $0.88 per session
• o1: $5.25 per session
• Premium models: $26+ per session

If you’re doing 10 intensive sessions per day (which many serious developers do), you’re looking at:

• GPT-4o: $260/month in actual costs
• o1: $1,575/month
• Premium models: $7,875/month

So that $20 subscription is covering maybe 1-2% of the real infrastructure costs for power users.

The Subsidy Cliff

Like the early Internet, and then high flying Web 2.0 startups, but even more so… this relatively free lunch isn’t sustainable, and everyone knows it. Platform-as-a-service providers like Heroku and Fly.io used to give away “hobbyist” compute and disk space, but these days, you’re minimally going to be at parity with the lowest cost VPS, and possibly at least double for a barely functional dev environment. Gmail implied that your account’s quota would grow indefinitely, but now it’s capped at 15GB unless you’re willing to pay for more storage. Amazon, a site that sells merchandise to its customers, still has advertising for placement (which is really no different than how it plays out in physical stores).

The reality is, at some point, someone will expect their profit out of this. And the difference is scale. We’re not talking about a 2x or 3x price correction when the music stops. We’re looking at 10x to 50x increases to reach actual cost recovery, not to mention the profit squeeze that causes Google to not return “organic” search results on the first page for things like “solar battery bank“.

Conservative future pricing:

• Basic limited usage: $50-200/month
• Current “unlimited” usage levels: $500-2,000/month
• Enterprise-grade patterns: $5,000-20,000/month per user

This isn’t speculation – it’s basic math based on compute costs, infrastructure requirements, and the need for actual profitability.

The Great Stratification

Let’s look at a theoretical market when this pricing reality hits, splitting the market into three distinct tiers:

The AI Elite: Large enterprises and high-revenue consultants who can afford $500-2000/month for full AI capabilities. They’ll have access to AI that handles 70-90% of routine coding tasks.

The AI Middle Class: Most developers and mid-market companies working with strict token limits and pay-per-use models. Basic AI assistance, but nothing like today’s unlimited buffet.

The AI Have-Nots: Cost-conscious businesses and individual developers who return to pre-AI development methods.

This stratification is where things get really interesting for software engineering careers.

The Unexpected Beneficiaries

The counterintuitive part here is that experienced software engineers might be the biggest winners when the AI subsidy ends.

Think about it. Right now, AI democratizes many coding tasks. A junior developer with unlimited AI access can often match the productivity of a senior developer on routine problems. But when AI access becomes expensive and limited, that equation flips. More importantly, the more complex the system becomes, the more it benefits developers that are able to quickly understand and analyze the higher complexity.

The skills that will command premium pay:

• Complex system architecture that stumps AI
• Novel problem-solving and algorithm development
• Performance optimization at scale
• Integration of multiple complex systems
• Debugging the truly gnarly issues that AI can’t untangle

These are exactly the kinds of problems I’ve written about over the years – the obscure FIPS errors, complex PowerShell integration issues, and hardware compatibility problems that require deep understanding, not just pattern matching.

The New Math of Engineering Value

So perhaps there will be a premium for experienced engineers as follows:

Conservative estimate: 20-40% premium as companies compete for engineers who don’t need unlimited AI to be productive.

Moderate estimate: 50-100% premium as the gap widens between “AI-assisted” and “AI-independent” problem-solving capabilities.

Aggressive estimate: 200%+ premium for truly complex, novel problem-solving that remains solidly in human territory.

The irony is profound: AI was supposed to flatten the engineering job market by making everyone more productive. Instead, when the economics normalize, it might create the largest wage stratification we’ve ever seen.

Preparing for the Transition

If you’re a software engineer, this transition creates both opportunity and risk:

Double down on complexity: Focus on the problems that make AI struggle – cross-domain expertise, system integration, performance optimization, and novel algorithm development.

Invest in AI workflow mastery: Learn to be incredibly efficient with limited AI tokens. Prompt engineering and AI workflow optimization will be crucial skills.

Consider the subscription as career investment: If you can afford it, maintaining access to premium AI might be worth the cost for staying competitive.

The Bigger Picture

This shift represents more than just pricing changes – it’s a fundamental rebalancing of the technology market. The current AI boom has been artificially sustained by venture capital and corporate subsidies. When that ends, we’ll see who really benefits from AI and who was just along for the free ride.

For experienced engineers who can solve the hard problems that AI can’t touch, this might be the best career development in decades. The democratization of routine coding will end, but the premium for true expertise will skyrocket.

The gold rush is ending. But for those who know where to dig, the real treasure might just be getting started.

Another Bubble to Burst?

One last thing… if Santa Claus (AI coding is legit) and the Easter Bunny (costs stay flat) are both real: And this is actually a permanent democratization of access to software development skills, waiting in the wings is a sort of treadmill in which AI adoption advantages normalize back to zero. This could actually give smaller players a huge advantage, because they won’t have as much of a bottleneck in their organizations with regard to bureaucracy. And software engineers with product mindsets might be able to just build their own thing that competes with larger players.