For about a decade, software was built on money that cost almost nothing, and a lot of our architecture quietly assumed it always would. A sharp essay on the new cost of capital and why financial friction is reshaping competitive advantage explains what happens when that assumption breaks: business models that survived on easy refinancing get exposed, while companies that turn effort into cash efficiently start pulling away. Swap the word "company" for "codebase" and it reads like a code review of the systems most of us maintain.
The Discount Rate Just Moved Against You
On September 16, the Federal Reserve raised its benchmark rate by a quarter point to a range of 3.75 to 4 percent, its first increase since July 2023. In October it held steady, with two voting members pushing for another hike. Long-term borrowing costs moved even more: the 10-year Treasury yield, which traded around 1.3 percent five years ago, climbed this fall to its highest level since 2007.
Why should someone who writes code care? Because every roadmap is an investment decision with a discount rate baked in, whether or not anyone writes it down. Picture a platform rewrite whose first $1 million in savings on cloud and on-call costs lands three years from now. When money is nearly free, that million is worth about a million today. At a 12 percent hurdle rate, which is modest for a venture-backed company, it is worth roughly $712,000, and every quarter of slippage shaves off more. Projects with long, fuzzy payoffs sailed through planning in 2019. Today the first question in the room is "when does this pay for itself?", and "it will scale better" no longer counts as an answer.
The price of an engineering hour changed too, and not only because of salaries. From 2022 through 2024, U.S. tax rules forced companies to spread software development costs over five years instead of deducting them right away, which left some barely profitable software companies owing tax on income that existed mostly on paper. The One Big Beautiful Bill Act, signed on July 4, 2025, restored immediate expensing for domestic R&D through a new Section 174A. Research done outside the U.S. still has to be amortized over 15 years, though, so an hour billed by a contractor abroad can cost an American company more after tax than the same hour billed at home.
Capital Didn't Disappear. It Moved.
All of this can sound strange next to headlines about record funding rounds, until you check where the money went. An OECD analysis found that AI firms captured 61 percent of global venture capital in 2025, about $259 billion out of $427 billion, and that mega deals made up roughly 73 percent of the AI total. Crunchbase counted OpenAI and Anthropic alone at 14 percent of all venture dollars that year. Some market commentary this fall even listed heavy bond issuance by AI companies among the forces pushing long-term yields up.
So cheap capital still exists, but only for a short list of frontier labs and the companies pouring concrete for their data centers. Everyone else, including plenty of AI startups, lives in a high-rate world where investors want proof that a business can carry itself. Paul Graham's old question about whether a startup is "default alive" or "default dead" is back on the agenda, and because engineering is often the largest line in the budget, the answer is frequently decided in architecture reviews.
Complexity Has a Carrying Cost
In March 2023, Amazon's Prime Video team published a post that made a lot of architects squirm. Its video quality monitoring system, assembled from serverless components orchestrated by AWS Step Functions and Lambda, hit a hard scaling limit at around 5 percent of its expected load. The team rebuilt it as a single-process service and cut infrastructure costs by roughly 90 percent.
37signals ran a version of the same experiment across an entire company. After pulling its apps off AWS servers, it reported shrinking its yearly cloud bill from $3.2 million to $1.3 million in 2024, and said the roughly $700,000 it spent on Dell hardware had paid for itself during 2023. In 2025 it bought 18 petabytes of its own storage to leave S3 as well. David Heinemeier Hansson's caveat deserves equal billing, though: companies with wild swings in traffic, or ones still figuring out what they are, still have good reasons to rent.
Neither story proves that monoliths beat microservices or that the cloud is a rip-off. What they show is that architecture is a financing decision. Elasticity is an option you pay for whether or not you exercise it, and a distributed design is a bet that future scale will justify today's overhead. When money was free, those premiums vanished inside growth curves. The essay has a name for the business version of this, "expansion theater": growth that looks impressive while it piles on complexity and thins out margins. The engineering version is a Kubernetes cluster for a contact form, or an event bus for an app with 300 users.
Notice also how slowly the 37signals bill came down. The savings arrived only as reserved instances and multiyear commitments expired, much like a company with old fixed-rate debt feels protected right up until the day it has to refinance. Your cloud contracts, vendor licenses and "temporary" infrastructure run on the same schedule. They reprice when they renew, which is usually the first time anyone reads the invoice closely.
Tech Debt Finally Has a Real Interest Rate
The debt metaphor was born in finance software. Ward Cunningham introduced it in a 1992 experience report about WyCash, a portfolio management system for managers of cash portfolios, partly to explain to his boss why the team kept refactoring. His line that "shipping first-time code is like going into debt" has survived three decades of overuse.
When McKinsey surveyed CIOs about the real size of their tech debt, the metaphor got an actual balance sheet. Respondents estimated that tech debt equals 20 to 40 percent of the value of their entire technology estate before depreciation, and that 10 to 20 percent of the budget meant for new products gets diverted to dealing with it. One former CIO at a major cloud provider described cutting the share of engineering time spent paying that "tax" from 75 percent to 25 percent.
In a zero-rate world, rolling debt over forever looked rational, on corporate balance sheets and in codebases alike. What changed is the currency of the interest. Tech debt is paid in engineering hours, and for a team under a hiring freeze those hours are the scarcest resource it has. Martin Fowler points out where the metaphor breaks: you only pay interest on messy code when you actually have to work in it. That flaw turns out to be useful, because it tells you where to spend. Debt in a module nobody has opened since 2021 is a cheap loan you can leave alone. Debt in the checkout flow you touch every sprint is a payday loan.
Run a Zombie Test on Your Stack
Economists have a blunt definition for zombie companies. In OECD and Bank for International Settlements research, a zombie is a firm at least ten years old whose operating earnings have failed to cover its interest payments for three years in a row. The BIS has linked their rise since the late 1980s to reduced financial pressure, partly a side effect of falling interest rates. Cheap money keeps zombies alive, in the economy and in your service catalog.
The same test works on code. Count the hours spent keeping a service alive (upgrades, incidents, flaky tests, dependency bumps) as its interest, and the hours spent shipping new value through it as its earnings. A service whose interest has outrun its earnings for three straight quarters is a zombie, however elegant its design doc was. An afternoon and a spreadsheet are enough to build a rough balance sheet for your systems:
- Price each service monthly. Add cloud spend, on-call hours and incident time, then divide by a unit users care about, such as requests, active accounts or orders. Unit cost tells you more than total cost because it survives growth.
- Calculate interest coverage. Pull a quarter's worth of tickets and commits, split them into "keeps it running" and "makes it better," and compute the ratio. Below 1 for three quarters in a row means you have found a zombie.
- Count your inventory. Long-lived branches, half-rolled-out feature flags and finished work waiting for a release train are capital tied up at zero return. Your lead time is your cash conversion cycle.
- Attach a payback period to every big proposal. "This migration pays for itself in 14 months" survives a budget meeting. A slide that says "future-proof" usually doesn't.
- Treat rewrites as refinancing. A rewrite makes sense when the new rate, meaning the ongoing maintenance load, is clearly lower than the old one and you can afford the closing costs of the migration.
- Retire one zombie per quarter. Deleting a service, an endpoint or a stale flag is one of the few kinds of engineering work with a near-guaranteed return.
Boring Is a Moat You Can't Buy
None of this is an argument for starving engineering. Underinvesting creates its own debt, and the teams that pull ahead when money is expensive usually keep spending, just with sharper questions attached. The essay's closing point, that durable value comes from a business that works under pressure rather than one that depends on favorable conditions, holds just as well for software. A system that is cheap to run, quick to change and easy to delete is an advantage a competitor can't simply buy with its next funding round.
Rates will move again, in one direction or the other. Teams that treat the cost of capital as a design constraint, right next to latency and uptime, won't have to rewrite their strategy every time the Fed meets.
So here's a question for the comments: which service in your stack would fail the zombie test today, and what's keeping it alive?