Ward Cunningham coined the phrase in 1992. He was at a software company called Wyatt Software, trying to explain to his colleagues why his team needed to refactor. The audience was finance people, so he picked a finance metaphor. It worked. Then it escaped the room. Then it spent thirty-four years getting misused so badly that Cunningham himself reportedly said the word he really wanted was *opportunity*.

That’s the air cover for what I’m about to say.

The thing you actually bought

When you buy a house, you take out a mortgage. That’s your debt. Principal, interest, term. You make payments. One day you pay it off, and the debt is retired. Done. Gone. The house is yours.

But the house doesn’t stop happening to you.

Year three, the front door starts squeaking. Year five, you’ve got a hairline crack running up the drywall because the foundation settled. Year seven, the dryer vent is so clogged your insurance company would like a word. The HVAC filter needs replacing every quarter or your energy bill turns into a second mortgage. The maple in the front yard sends a branch through your gutter and you spend a Saturday up a ladder with a chainsaw wondering where your weekend went.

That’s not debt. That’s maintenance.

And maintenance doesn’t retire. It cycles. Forever. Until you sell the house or the house falls down.

This is your codebase. The library that’s three major versions behind. The deploy script only Sarah understands, and Sarah’s on PTO. The flaky test that’s been “flaky” for nine months. The service running on an EC2 instance someone spun up in 2019 that nobody wants to touch because the IAM role attached to it has god-mode permissions and the original engineer left two reorgs ago.

You don’t retire that. You live with it. Or you maintain it. Pick one.

The “retire” problem

Listen to how people talk. “We’re going to retire our tech debt.” “We’ve got a debt repayment plan.” “We’re going to pay it off next quarter.”

Pay it off with what? Send a check?

Real debt has cash. You barter, you exchange, the balance goes down, the obligation ends. There is no cash for tech debt. There are only more engineers, more time, and another quarter of features that won’t ship because you’re up a ladder with a chainsaw.

A mortgage buys you a house that appreciates. Tech debt payment usually buys you back the velocity you had two years ago. You haven’t retired a liability. You’ve cleaned gutters. There’s no equity. There’s no closing ceremony. There’s only the next gutter.

Why this matters (in numbers)

Don’t get me wrong, the pain is real. The 2025 DORA report — the closest thing the industry has to a scoreboard — found that fewer than 1 in 10 teams achieve a rework rate below 2%.

Rework rate is tech debt by another name. It’s the unplanned deployment that fixes the bug introduced by the last unplanned deployment that fixed the bug. More than 9 in 10 teams are above the threshold. That’s not a debt problem. That’s a maintenance discipline problem.

The data is loud. We’re just labeling it wrong.

What Andy Grove would do

Andy Grove built one of the most operationally disciplined companies in history. He would have laughed at the financial metaphor.

There’s a story from Intel that’s been told a thousand times but never in this context. 1985. Intel is bleeding money on memory chips. Grove and Gordon Moore are sitting in his office. Grove turns and asks: *if we got kicked out and the board brought in a new CEO, what do you think he would do?* Moore says, without hesitation: *he would get us out of memories*. Grove replies: *why shouldn’t you and I walk out the door, come back and do it ourselves?*

Apply that to your codebase. If a new CTO walked in tomorrow and looked at what your team is doing with “tech debt” — the sprints, the offshore teams, the new-hire learn-by-suffering — would they say “yes, this is the right approach”? Or would they say “what is this maintenance backlog and why isn’t it run as a function”?

Another Grove gem that is relevant here. Intel’s facility maintenance teams were performing badly for years. Grove’s fix wasn’t to outsource them or schedule periodic deep-cleans. He instituted a competition — each building rated against the others, scores posted publicly. The cleanliness of every building improved dramatically. No bonuses. Just a scoreboard.

He didn’t ignore maintenance. He didn’t pretend it was beneath strategic attention. He instrumented it. Gave it a scoreboard. Made it visible. That’s the blueprint.

The three strategies that don’t work

When teams finally accept they have to “do something about the tech debt,” they reach for one of three strategies. All three are popular. All three measurably fail.

1. The Umbilical Outsource

Send it offshore. Cheap labor, big team, problem solved.

Except your FTEs are still tied to the work by an umbilical cord they can’t cut. Every PR needs domain review. Every broken pipeline routes back to the same three engineers who actually understand the system. Every “wait, why does this function exist?” turns into a thirty-minute call across timezones at hours nobody wanted.

You didn’t outsource the work. You outsourced the typing. The thinking stayed home. And now your senior engineers spend half their day reviewing code they didn’t write, in a context nobody on the other end can ever fully load.

The billing went down. Everything that matters got worse.

2. The Sprint Spike Theater

Rotate engineers across the roster. Pick a chunk of debt. Cherry-pick a fix. Ship it. Celebrate in the retro. Move on.

Looks beautiful in a dashboard. Means almost nothing on the ground. Because the engineer who fixed it learned something — and then the lesson died when they rotated back to features. Nobody else on the team knows what they learned. The pattern that caused the debt is still there. You fixed one squeaky door in a house with forty-seven of them, and you didn’t tell anyone where the WD-40 lives.

Pro tip: if your tech debt sprint doesn’t end with a writeup, a pattern, or a tool the rest of the org can use, you didn’t fix debt. You fixed a ticket.

3. The New Hire Trap

“Give it to the new hire. They’ll learn by going deep.”

Cute. Until you realize what they’re actually learning.

They’re learning this is what new hires get. They’re learning the senior engineers above them don’t want to touch this code, which is why it landed on their desk. And they’re learning the job market has companies that don’t punish curiosity with six months of janitorial work.

Then they leave. Six to nine months of salary in hiring and replacement costs, lit on fire. The debt is still there. The next new hire is going to hear about it from someone in the kitchen on day three.

What changes when you stop calling it debt

The word is doing more damage than you think.

When you call it debt, the mental model is: there’s a balance, we’ll pay it down, one day we’ll be free of it. Finance thinks in terms of payoff timelines. Product thinks it’s optional. Everyone keeps shipping on a foundation that’s quietly settling.

When you call it maintenance, the mental model changes. Maintenance is a line on the budget, not a negotiation. Maintenance has an owner. It doesn’t retire — it just runs. You don’t call a meeting to decide whether to change the HVAC filter. You have a schedule.

Start there. Put a maintenance function in your roadmap. Not a sprint. A function. Someone owns it. They have a scoreboard. And the scoreboard is output — lead time, rework rate, change failure rate — not tickets closed. Not hours billed. Not the size of the backlog you heroically defeated in Q3.

The conversation with product and finance changes too. When you stop saying “we need to retire tech debt” and start saying “our rework rate is 14% and our lead time doubled in two quarters,” something shifts. Product stops rolling their eyes. Finance actually listens. Because now you’re not asking for permission to do invisible work. You’re reporting on a system that’s visibly degrading, with numbers.

Stop handing it to people who can’t win. The new hire can’t win alone. The offshore team can’t win without context. Give someone a winnable game — real scope, real ownership, real visibility. Grove’s building maintenance teams had no bonuses and dramatically improved. All it took was a scoreboard and someone who could see the racetrack.

The codebase you have in two years isn’t a function of the sprint you scheduled today. It’s a function of whether you treated upkeep as the actual job or the thing you’d get to eventually.

You’re living in the house either way.

PS: I know it now needs to be update given the pace of change we’ve seen in this space, but I’ll save that for a future post.

Start a Substack

Welcome to the first edition of Cloudy with a Chance of Startups. I was going to publish a version of this back in January (2025). A colleague shared a vision for how they saw the year playing out for Startups and I (with enthusiasm but not as much conviction) quipped - “You’re not being bold enough, it’s gonna be the year we already reach rapid transformation!”. Now almost halfway through the year, my enthusiasm and conviction have both only grown.

If you’re reading my writing for the first time, this isn’t usually how I write. I’m usually very conversational but measured in my writing style. If you already know me, you know I write a lot, but had just taken a hiatus from public forums for a few years. You also probably know I love the movie (and book) Seabiscuit. So, given that this is such an exciting time, I’m borrowing the infectious energy of William H Macy’s Tick Tock McGloughlin and the complete opposite of his cynicism and hope you’ll follow along - or forgive this experiment and blame it on caffeine. Boy oh boy am I excited!

Remember Digital Transformation?

That thing every company spent the last decade scrambling to accomplish? Well, AI transformation is like that, but on steroids, Red Bull, and a shot of espresso. While digital transformation took years (and many companies are still working on it), AI transformation is happening at warp speed.

Let me take you back to the not-so-distant past. Digital transformation was this massive undertaking where companies had to:

• Move workflows to the digital worls

• Adopt cloud computing

• Rebuild their entire tech stack

• Retrain their workforce

• Completely rethink their business models

Now look at AI transformation. ChatGPT hit a million users in just FIVE DAYS after launch. By early 2025 (aka now), it had over 400 million weekly active users. That's adoption at a pace we've never seen before.

Why is AI transformation happening so much faster? A few reasons:

1. We're already digital. The foundation is built - companies have data, cloud infrastructure, and digital-savvy teams.

2. The barrier to entry is dramatically lower. You don't need to be a PhD to use AI tools. My entire family uses Claude almost daily! (Some of them used to still call me up to fix the zoom on their browser window or if the printer wouldn’t work till a few months ago.)

3. The benefits are immediately obvious. Unlike some digital initiatives where ROI took years to materialize, companies are seeing measurable results from AI implementation within months or even weeks. I know, there’s more to be unpacked here and this is not the most popular opinion. If you do it right though, and there is an ROI to be seen, you see it quick.

4. There's an arms race mentality. When your competitor is using AI to grow 2.5x faster, you can't afford to wait.

What AI Transformation Looks Like in the Wild

In an oversimplified and overly broad classification of how startups are using generative AI, it falls into three buckets.

Supercharging Products

Remember when adding a "mobile app" was the hot feature everyone needed? Now it's AI. Startups are embedding generative AI into their products in ways that genuinely delight users:

• E-commerce startups using AI for hyper-personalized recommendations

• Design startups letting users generate unique assets on demand

• SaaS companies using AI coding assistants to ship features faster

This category isn't just adding a chatbot - it's fundamentally rethinking what products can do.

Turbocharging Go-to-Market

This is where I'm seeing startups make ridiculous gains. When you're a small team trying to compete with big budgets, AI is the ultimate multiplier:

• Generating personalized marketing campaigns in minutes instead of weeks

• Creating entire content calendars with a few prompts

• Scaling sales outreach with AI-crafted messages that actually sound human

Operational Efficiency (on nitrous?)

This might sound boring, but it's where the money is:

• AI scribes saving thousands of hours and hundreds of thousands of dollars

• Reduce product documentation time and improve quality and coverage

• Finding insights from unstructured data and accelerate reporting

These aren't marginal improvements - they're transformative changes in how work gets done.

So where are we now?

We’ve gone from playing around to production

Here's what's fascinating about where we are in the AI adoption cycle. In 2023, most startups were just experimenting - building proofs of concept, running small pilots, using innovation budgets to play around.

Now? Companies are all-in. The number of startups with fully integrated, production-level AI systems nearly doubled between 2023 and 2024. Budgets are shifting from "innovation" to "core operations." This isn't experimentation anymore - it's existential.

Money talks (to AI even if it’s giving others the silent treatment)

They say (and I dunno who they are, but they say), if you want to know where things are headed, follow the money.

• Global VC funding in AI topped $100 billion in 2024 with 5 US-based startups accounting for over $30 billiong in Q4’24

• Almost a third of ALL venture capital went to AI companies

• Funding for generative AI startups increased 20x since 2020

The market for generative AI solutions is projected to go from $16 billion in 2023 to $143 billion by 2027. That's not growth - that's a rocket ship.

The Open Source vs. Proprietary Dilemma

One of the most interesting tensions I'm seeing is the choice between open weights and proprietary AI models. Startups are wrestling with the tradeoffs:

Open weights gives you:

• Chance to control infrastructure and costs

• More control and customization

• Transparency into how models work (to varying degrees)

But proprietary models offer:

• Simpler implementation

• Easier security and compliance

• More support when things break

There's no one-size-fits-all answer here. As we’ve seen waves of model releases unfold, startups have had to adapt from fine-tuning, continuous pre-training, and other techniques for better results only to find that the next major release of SOTA models give better performance out of the box. Most moats are centered around data, evals, and customer experience instead of the underlying generative AI model.

The Future is Agentic

I bet you’ve heard that a few times now. These aren't just chatbots - they're autonomous systems that can plan, execute, and achieve complex goals with minimal human supervision. They can use tools, access databases, and even direct other systems.

Imagine a sales agent that doesn't just respond to inquiries but proactively finds leads, conducts research, crafts personalized outreach, follows up, schedules meetings, and negotiates terms - or a software engineering agent that drafts user stories, builds new features, reviews code, deploys and maintains applications - all while you sleep.

This isn't science fiction - it's happening now, and startups that embrace agents early will have a massive advantage. We’re seeing more revenue per employee, well at least human employee. We’re seeing the rise of the Vibe workforce. We’re seeing an explosion of MCPs. We’re seeing agent to agent frameworks, agents orchestrating other agents, it’s really turning into Matrix Revolutions (did I already say science fiction?).

So what should you take away

from my new rant…

Digital transformation took a decade. AI transformation is happening in real-time, right now, whether you're ready or not.

The startups that thrive will be the ones that move beyond experiments to strategic implementation, that overcome the very real challenges of data readiness and integration, and that focus on augmenting their teams rather than replacing them.

I've never seen a technological shift with this much potential to completely reshuffle the competitive landscape. The next unicorns will be built on AI foundations, and they'll get there in half the time it took the last generation or less.

So, founders, the forecast for 2025 and beyond is cloudy with a chance of AI - and I suggest stepping out and enjoying a good soak.

What AI transformation are you seeing in your industry? Hit reply and let me know - I'm collecting stories for my next rant!

Subscribe now