Who Really Wins When AI Writes Your Code
By LumaVista Team
Tuesday afternoon, around three. You’ve spent two hours pairing with a coding agent on a gnarly migration ticket — the kind that used to eat a full day. The agent drafted the schema changes, you caught two edge cases it missed, it wrote the tests, you fixed the one that asserted the wrong thing. Ticket closed. PR merged. Everyone’s happy.
Here’s what you probably didn’t think about: during those two hours, three different ledgers were being written. You only saw one of them.
Your ledger says: one migration shipped, three hours saved, flow state achieved. Your employer’s ledger says: velocity is up this quarter, the AI tooling budget is paying for itself. And the provider’s ledger — the company running the model — says: two hours of interaction data captured, several hundred accept/reject decisions logged, one more developer whose workflow now runs through us.
All three entries are accurate. That’s what makes this interesting. Nobody’s cheating anybody. But the three parties in every AI coding session are playing three different games, and it’s worth understanding all of them — especially the one being played with your keystrokes.
What the developer wants
The developer’s game is the simplest: finish the work, stay in flow, go home.
And the tools genuinely deliver on the feeling. In JetBrains’ 2025 survey of about 24,500 developers, 85% said they regularly use AI for development, and 90% reported saving at least an hour a week. Developers didn’t adopt these tools because a memo told them to. They adopted them because deleting boilerplate feels like winning.
The catch — and there’s a whole article’s worth of catch — is that the felt gain and the measured gain aren’t the same thing. When METR actually timed experienced developers on real tasks in 2025, the group using AI believed they’d been about 20% faster while measuring slower. That gap between perception and measurement has been the most stubborn finding in this field, and we’ll dig into it properly in part two.
For now, the point is narrower: the developer’s ledger is written in feelings and finished tickets. Both are real. Neither is the whole story.
What your employer wants
The company’s game is competitive: ship faster than the other firms shipping against you, at lower cost, without dropping quality — pick three, ideally.
By that logic, buying AI coding tools is close to mandatory. GitHub Copilot alone passed 20 million all-time users in mid-2025 and sits in roughly 90% of the Fortune 100. When adoption is that broad, declining isn’t a strategy — it’s a handicap. Your competitors’ developers have the tools, so yours need them too.
But look at what the company actually acquired. It didn’t buy a productivity gain; it bought a subscription to a productivity gain. The distinction matters. A compiler you bought in 2005 still compiles. A coding assistant is a metered service whose price, limits, and terms can change under you — and, as we’ll see in part three, they have, repeatedly, on a schedule that looks less like accident and more like strategy. GitHub said the quiet part out loud in April 2026 when it announced every Copilot plan would move to usage-based billing because the flat model was, in its own words, “no longer sustainable”.
The company’s ledger is written in velocity and spend. The dependency line item doesn’t show up on it — yet.
What the provider wants
The provider’s game is the biggest one, and it’s the one least talked about at standup.
The scale first. Claude Code went from launch to a $2.5 billion revenue run-rate in about nine months. Cursor’s maker Anysphere passed $500 million in annualized revenue in mid-2025 with revenue doubling roughly every two months, and was valued at $29.3 billion by November. AI coding is the fastest-growing software category anyone has ever measured, and every major lab is fighting for it.
Why so valuable? Because a coding session isn’t just revenue. It’s the richest training signal in the world. Every time you accept a suggestion, you’ve labeled it good. Every time you reject one, you’ve labeled it bad. Every time you take the AI’s code and fix it, you’ve handed over the correction — expert-annotated, in context, for free. Multiply that by millions of developers, all day, every day.
Your subscription fee is the smallest thing you pay.
The tell: look who’s exempt
If you doubt that the data is the point, don’t read the marketing. Read the pricing structure — specifically, who gets excluded from training.
Between late 2025 and mid 2026, the individual tiers of the major coding tools flipped, one by one, toward training on their users. Anthropic’s August 2025 consumer terms update asked Free, Pro, and Max users to choose whether Claude — including Claude Code — can train on their chats and coding sessions, with a five-year retention period if you say yes. GitHub announced that from April 2026, Copilot Free and Pro interaction data — prompts, code snippets, accept/reject signals — feeds model training by default unless you opt out. Cursor with Privacy Mode off may use your code “to train our models”. Windsurf’s current terms go furthest: the free tier trains on your data with no opt-out at all — you have to pay to be allowed to decline.
Now the other half of the pattern: business and enterprise tiers are exempt. All of them. Anthropic doesn’t train on commercial customers. GitHub explicitly excludes Business and Enterprise. Cursor lets team admins enforce Privacy Mode org-wide. Enterprise legal departments would never sign otherwise — so the carve-out exists, and it’s standard.
Sit with what that means. The same companies offer the same tool in two versions: one that learns from you, priced low, and one that doesn’t, priced high. The discount on the individual tier isn’t generosity. It’s payment — for your data, at a rate you never negotiated. Companies bought their way out of the bargain. Individual developers, mostly without noticing, are still in it.
The asymmetry nobody planned
Here’s the sentence this whole series hangs on: the provider is the only party in the triangle that gets stronger from every interaction.
The developer gets a closed ticket — valuable today, gone tomorrow. The company gets shipped features — valuable this quarter, table stakes the next, because its competitors are shipping AI-accelerated features too. Competition passes those gains through to customers surprisingly fast, which is good for the world and neutral for any single firm.
The provider keeps everything. The revenue recurs. The training data compounds into better models. The workflow integration deepens into switching costs — your muscle memory, your prompt patterns, your CI pipelines all shaped around one vendor’s behavior. And the better the model gets from everyone’s data, the more indispensable it becomes to each individual user. Every loop through the triangle transfers a little durable advantage to the same corner.
And no one is lying. This is the uncomfortable part — there’s no villain to expose. Developers rationally chase flow. Companies rationally chase velocity. Providers rationally monetize the most valuable interaction data ever collected while it’s legally and competitively available. Everyone is doing exactly what their incentives say. That’s precisely why the pattern won’t fix itself: durable arrangements aren’t built on deception, they’re built on aligned short-term interests with misaligned long-term ones. We’ve watched this movie before with databases, clouds, and CRMs — the difference this time is the speed, and the fact that the accumulating asset is your team’s daily judgment.
None of this means “stop using AI coding tools.” The productivity is real, the tools are getting better, and refusing them outright is career and business malpractice at this point. It means: know which game you’re in, because you’re in all three whether you know it or not. And the exit — genuinely capable open-weight models you can run on hardware nobody can reprice — is more real than it was a year ago. That’s part four.
What to do now
- Check your training toggle today. Five minutes: Claude settings if you’re on a consumer plan, GitHub Settings → Privacy for Copilot Free/Pro, Privacy Mode in Cursor. Know your current answer — most people are surprised.
- Find out which tier your work runs on. The enterprise carve-out means your employer’s data deal and your personal one are probably different. If you use a personal AI account for side projects with client code in them, that’s the gap.
- Read one provider’s data policy — the one you use most. Not the marketing page; the data-usage page. Twenty minutes that recalibrates how you think about the tool.
- Put a number on the dependency. One honest question: if this tool doubled in price or halved in limits next quarter, what would we do? If the answer is “pay it,” you’ve measured your switching cost.
- Follow the rest of this series. Part two is what the tool does to you (the psychology is stranger than you think). Part three is what happens to the bill. Part four is the exit.
The triangle isn’t going away. The subscriptions will keep flowing one direction and the training data the other. The only variable you control is whether you understand the game — and whose ledger your next two hours end up on.