Something has changed. And it happened faster than almost anyone predicted.
For the past two years, the corporate world has lived comfortably between two camps in the AI debate. On one side: the optimists, promising exponential improvement, predicting that Large Language Models would get two, three, ten times better every year. On the other: the skeptics, pointing to hallucinations, limited context windows, weak reasoning, messy enterprise data, undocumented processes, and the looming plateau where scaling laws would hit a wall.
I’ll be honest: I was in the skeptic camp. When someone told me software developers would be 30% more productive by 2026, I said: “Sure, maybe by 2030. And only if we see a fundamental breakthrough in how these models are trained and deployed.”
Then something happened between Christmas and late January that I didn’t expect.
The Shift I Didn’t See Coming
It started quietly. In November and December, more and more people in my network began talking about Claude Code, not as a toy, but as the way they work now. These weren’t AI evangelists giving conference talks. They were working software developers telling me, matter-of-factly, that they had stopped writing code. They described setups where AI agents spawned sub-agents, executed multi-step tasks, and delivered working software. The developer’s role had shifted from builder to director.
Over Christmas, my brother showed me his setup. I decided to try it myself.
Then came the new models. Anthropic’s Opus 4.6. OpenAI’s GPT-5.3 Codex. And the development environments evolved in lockstep, more AI-native, more agent-friendly.
I migrated my blog to Hugo, a static site generator that runs on Markdown files. It requires terminal knowledge, specific document structures, custom metadata in frontmatter. In the old world, this is a weekend project for someone who knows what they’re doing. Instead, I downloaded an AI coding agent, pointed it at my blog directory, told it what I wanted. Within a session, I had the blog set up, customized, and all legacy articles migrated. Not a rough draft. The finished thing.
But the blog migration isn’t what prompted this article. What prompted it was the sentiment shift.
When the Forum Mood Flips, Pay Attention
If you’ve spent any time in developer communities (Reddit, Hacker News, specialized Discord servers) you know the rhythm. For the past year, opinions on AI coding were split roughly down the middle. Enthusiasts on one side, skeptics on the other. A balanced, sometimes heated, but fundamentally stable debate.
Sometime in January 2026, the balance broke.
The posts changed. The tone changed. Developers who had been cautious were suddenly writing things like: “I was using Claude as a glorified autocomplete for boring tasks. Now I trust it to write reliable code I don’t need to read line by line.” Or: “I gave Claude a file with ten complex tasks, a week’s worth of sales engineering work. It completed all of them flawlessly in twenty minutes.” Or, perhaps most tellingly, from a developer who pivoted careers entirely: “I finally gave up on my career aspirations in software development and landed an apprenticeship to become an electrician.”
One Reddit commenter captured the mood precisely:
“I would say a few weeks but yeah it’s changing so quickly. It feels like just as recently as 2025, most people were feeling apprehensive about coding agents. Then we went on vacation, came back to work in the new year, and the mainstream opinion suddenly shifted towards acceptance.”
Another observed:
“When Opus 4.5 was released to pro users I was first to jump on and didn’t understand why people were not freaking out in November. You couldn’t talk about it anywhere without getting downvoted to oblivion. I agree on the three weeks, that’s when it became a noticeable shift for me.”
This isn’t marketing. This isn’t a vendor pitch. This is the practitioner community, the people who actually write and ship software for a living, collectively arriving at a new conclusion: the thing the hype promised is actually happening.
The Evidence Is Stacking Up
It’s not just forum sentiment. Consider what Spotify shared on its Q4 2026 earnings call: the company’s best developers have not written a single line of code since December. They use an internal system called “Honk” built on Claude Code, where an engineer can fix a bug or add a feature to the iOS app from Slack on their phone during their morning commute, and have a deployable version pushed back to them before they arrive at the office.
Spotify shipped over 50 new features in 2025. They’re accelerating, not with more developers, but with AI-augmented developers who have fundamentally changed what “development” means.
Or consider Matt Shumer’s widely shared essay from this month, where the AI startup founder and investor laid it bare:
“I am no longer needed for the actual technical work of my job. I describe what I want built, in plain English, and it just… appears. Not a rough draft I need to fix. The finished thing.”
OpenAI’s own documentation for GPT-5.3 Codex states that the model was “instrumental in creating itself,” that it debugged its own training, managed its own deployment, diagnosed its own test results. Anthropic’s CEO Dario Amodei says AI is now writing “much of the code” at his company and predicts AI will eliminate 50% of entry-level white-collar jobs within one to five years.
The people building these systems aren’t making predictions about some distant future. They’re describing what happened last Tuesday.
What This Means for Your 2025 Business Case
Now, here’s where it gets uncomfortable for corporate planning.
Many organizations spent 2025 building careful, measured business cases around AI in software development. The thesis was sensible: a 20 to 30% productivity improvement (spread over a couple of years). The playbook followed logically:
- Revenue lever: Generate more output with existing development capacity. Ship more features, serve more customers.
- Redeployment lever: Shift some developer capacity to adjacent roles, more QA, more architecture, more product work.
- Cost lever: Gradually reduce headcount through attrition, maybe some targeted restructuring (though in Germany and similar markets, this is legally and culturally difficult).
This was a reasonable plan for a world where AI was a better autocomplete. A copilot that saved developers 20 minutes per hour.
That is no longer the world we live in.
The marginal cost of producing software is approaching zero. Not trending downward. Not declining gradually. Approaching zero. When a single developer can direct AI agents to build, test, and iterate on a working application in a day, work that previously required a team of five over several weeks, the entire economic equation changes.
Your 2025 business case isn’t wrong. It’s obsolete.
Two Strategic Responses Worth Considering
If marginal software development costs truly approach zero, the question isn’t “how do we save 25% on our development budget?” The question is: “What do we do with essentially unlimited development capacity?”
Here are two strategic directions that deserve serious consideration.
1. The Forward-Deployed Innovation Squad
Imagine you have 100 software developers. In the old model, a rapid prototype required a team of five working for several weeks. In the new model, each of those 100 developers can independently deliver a working prototype in a week.
That’s 100 customer-facing innovations per week.
The model works like this: Deploy developers directly into customer relationships. Not to build large contracted systems, but to solve specific pain points with rapid, custom-built solutions. Some will be rough. Some will have quality issues. That’s fine. The point isn’t production-grade software on day one. The point is validation. Does this solve the customer’s actual problem?
Where it does, you productize. You build a real offering around it, not because the coding is hard (it isn’t anymore), but because the management, operation, and accountability around the solution has value. The customer isn’t paying for code. They’re paying for someone to own the outcome.
Where you find recurring patterns across customers, you develop intellectual property. The competitive advantage isn’t that you can build it, anyone can build it now. The advantage is that you’ve already learned what to build because you were embedded in the customer’s problem.
This is the model of the software developer as innovation scout, not as ticket-processing factory.
2. The SaaS Reckoning
Here’s an uncomfortable truth that I notice in my own behavior: I’m tired of paying monthly subscription fees. And I suspect every CIO reading this feels the same way.
When building custom software was expensive and slow, paying $50 per user per month for a SaaS tool was an obvious trade-off. The vendor handled development, hosting, maintenance, updates, security, compliance. You couldn’t build it yourself, not at that quality, not at that speed, not with those economics.
But what happens when a developer can replicate 80% of a SaaS tool’s functionality in an afternoon? When companies start asking: “Why are we paying $200,000 a year for this project management tool when we could build exactly what we need in a week?”
We’re already seeing this in the forums. Developers canceling $10,000/year PagerDuty subscriptions. Rebuilding $1,000/month management software in a week. One commenter put it bluntly: “I am replacing every SaaS we use right now. There is simply no need for a Trello subscription if Claude can do it in an afternoon.”
Now, the counterarguments are real: security, compliance, scalability, multi-tenancy, ongoing maintenance. These are genuine costs that don’t disappear just because the initial build is cheap. As one skeptic noted: “When you buy SaaS, you’re not buying the code. You’re paying a very small fee for someone else to guarantee it’s always up, secure, and that bugs are resolved quickly.”
Fair point. But the calculus is shifting. And the stock market is noticing. Software-as-a-service valuations are under pressure, and the “AI-eating-software” narrative is gaining traction with investors.
The likely outcome isn’t that SaaS disappears. It’s that the market fragments: more players, less differentiation, lower margins, lower prices. The era of charging premium prices for commodity software functionality is ending.
The Big Question: Does the Market Expand or Contract?
This brings us to the critical strategic question that every corporate leader needs to wrestle with.
If your 100 developers now have the productive capacity of 100 teams, do you:
A) Need fewer developers, because there simply isn’t enough demand to absorb all that output? The work gets done with 20 people instead of 100, and the other 80 need new careers.
Or:
B) See demand explode, because software that was previously too expensive to build for niche use cases suddenly becomes viable? Every small business, every internal process, every customer workflow that was “not worth the development cost” is now buildable. The market for custom software isn’t shrinking. It’s expanding by orders of magnitude.
Both scenarios are plausible. And which one you believe in should fundamentally shape your corporate strategy.
If you believe in contraction, your playbook is managed downsizing: reduce headcount, reskill where possible, consolidate.
If you believe in expansion, your playbook is radically different: retrain everyone as AI-directed developers, push them into the field, solve problems at scale, capture the long tail of software demand that was never addressable before.
The honest answer is that different sectors, different company types, and different geographies will experience different versions of this. But the one thing you cannot do is stick with last year’s 20 to 30% productivity thesis and pretend the world hasn’t moved.
The Cognitive Load Bottleneck
One nuance worth flagging, because it tempers the euphoria: several experienced developers in the community have identified what may be the new bottleneck, not code production, but human comprehension.
“Yes, the big issue will be comprehension debt and intention drift. We will soon realize that, at scale, our cognitive load is the new bottleneck.”
“I relate so much to this idea of cognitive load being the bottleneck. I feel like I’m constantly asking my AI assistants, ‘Why did you recommend this vs. that?’ But I’ve seen how bad projects can turn out if you let your assistants outpace your understanding.”
“As long as we humans are accountable for the outcome and the output of software creation, we will have to throttle it to our bottleneck.”
This is important. The marginal cost of producing code may approach zero, but the marginal cost of understanding, validating, (selling) and being accountable for that code does not. AI can build fast. Humans still need to comprehend what was built, decide if it’s right, and take responsibility when it isn’t.
This is good news for senior developers and architects, the people who can evaluate, direct, and take ownership. It’s challenging news for organizations that thought they could simply replace human judgment with AI output.
A Call to Action
If you’re responsible for workforce planning, technology strategy, or corporate development, here’s what I’d ask you to do this month:
- Revisit your AI productivity assumptions. If your business case was built on 20 to 30% improvement in software development, test it against the new reality. Talk to your developers. Ask them what’s changed in the last 90 days. You may be surprised.
- Model both scenarios. What does your organization look like if developer demand contracts by 50%? What does it look like if it expands by 500%? Neither scenario may be exactly right, but thinking through both will reveal strategic options you haven’t considered.
- Experiment with the forward-deployed model. Pick a handful of developers. Give them access to the best AI coding tools available today. Point them at real customer problems. See what they can build in a week. The results will tell you more than any analyst report.
- Audit your SaaS portfolio. Not to cancel everything tomorrow, but to understand where you’re paying premium prices for functionality that’s becoming commoditized. This is a cost optimization opportunity that most organizations haven’t even begun to explore.
- Invest in comprehension, not just production. The organizations that will thrive aren’t the ones that produce the most AI-generated code. They’re the ones that can understand, validate, and govern it. Architecture, quality assurance, security review, accountability frameworks: these capabilities become more valuable, not less.
The sentiment shift in the developer community isn’t just forum chatter. It’s the canary in the coal mine. When the people who build software for a living collectively conclude that their profession is being fundamentally reshaped, not in five years, but right now, the rest of the enterprise needs to listen.
I was wrong about the timeline. The future arrived ahead of schedule. The question now isn’t whether to adapt, but whether you’ll adapt fast enough to turn disruption into advantage.
The wave is forming. The people on the beach are still sipping their drinks. But the water is already pulling back from the shore.