← index

the triple debt

19 March 2026 · Kris Kang

Most CXOs I speak with are focused on one type of risk: "What happens when AI token subsidies end?" It's the right question. But it's only a third of the problem. Companies adopting AI coding tools today are accumulating three debts simultaneously: financial, technical, and organizational. When the subsidy cliff hits, all three come due at once.

Debt 1: The Financial Debt

Distributors are starting to reset pricing: frontier models are being repriced closer to non-subsidized marginal cost, even as adoption and usage explodes. In the short-run, users may drift toward suppliers that offer more generous subsidies as a result. But build your house there, and you're in for a shock when the price resets to reality. Toy example: suppose your enterprise spends $1M on AI tokens today. The subsidy multiplier doesn’t need to be 10x for the problem to be real. Even a 2–3x increase can break your unit economics if usage has already scaled enterprise-wide.

But won't frontier model costs tend toward zero? Empirically, while the cost per token has dropped, the appetite for larger, more expensive context windows and frontier-class reasoning means the total bill is still going up.

My question back to CXOs: "have you priced this risk in?"

Debt 2: The Technical Debt

A recent Carnegie Mellon study [1] used a difference-in-differences methodology to measure the causal effect of Cursor adoption on open-source repositories. The findings are sobering: velocity spikes, then fades. Complexity grows persistently.

The accumulated complexity becomes a major driver of velocity slowdown. In other words: AI coding tools let you build faster, but what you're building is harder to maintain. The initial productivity gain is real. The long-term cost is also real. And most teams aren't measuring the second part. Note: this study covers repos that adopted Cursor between January 2024 and March 2025 that were still active in August 2025, before the latest generation of models. It's possible newer models (and Cursor's newest harness) narrow the gap.

My question back to CXOs: "Do you know the cost, and growth rate, of your tech debt?"

Debt 3: The Organizational Debt

This is the one almost nobody is talking about. 10x gains in individual productivity do not maketh 10x gains in team productivity. Teams that adopt agentic coding tools without rethinking their end-to-end software development life cycle (SDLC) are likely getting low ROI on spend, while incurring a debt of reliably inconsistent bad habits in AI use. For example, shallower PR reviews as throughput increases. The fix isn't just better tooling to magic the problem away. In fact, see Debt 1. If your SDLC hinges on subsidized models, then what happens when your cost base explodes?

My question back to CXOs: "Do your engineering managers know how to run an agent-forward team and SDLC with high ROI?"

The Compounding Effect

Here's what makes the triple debt dangerous: the three debts compound each other. Financial subsidies mask the true cost, so teams scale usage aggressively. Aggressive usage accelerates code complexity. Growing complexity overwhelms team capability and processes. And when the subsidy ends, you're left with an expensive, complex, under-reviewed codebase and a team that was optimized for individual throughput, not long-term enterprise value. Each debt is manageable in isolation. Together, they're a time bomb.

What to Do About It

If you're a CXO or engineering leader adopting AI coding tools here's what I'd suggest:

  1. Model the financial cliff.

    Take your current AI spend. Multiply it by 2-10x. If that number breaks your unit economics, you have a pricing risk that needs a mitigation plan today, not when the invoice arrives.

  2. Measure quality, not just velocity.

    Track static analysis warnings, code complexity, and review turnaround time alongside commits and merged PRs. Ask your Principal Engineers if there is a growing stench of bad smells. If velocity is up but quality metrics are flat or declining, you're borrowing from the future.

  3. Invest in upgrading your SDLC before you need it.

    People, processes, then tooling. The teams that will thrive with AI coding tools are the ones that take this seriously, not an afterthought.

  4. Don't optimize for throughput alone.

    The right metric isn't "how much code can Alice or Bob produce?" It's "how much maintainable code can we ship over time?" Those are very different questions, and AI tools that optimize for the first without guardrails for the second are creating a liability. The companies that get this right will have a genuine competitive advantage. The ones that don't will be the case studies in the next round of research papers.

[1] He, H., Miller, C., Agarwal, S., Kästner, C., & Vasilescu, B. (2025). "Speed at the Cost of Quality: How Cursor AI Increases Short-Term Velocity and Long-Term Complexity in Open-Source Projects."