# Software quality decline: anatomy of a systemic crisis

**Software quality decline** shows up first in small absurdities that nobody investigates. Apple's Calculator recently leaked 32 gigabytes of memory. No calculation accounts for the allocation and no temporary buffer explains it: the memory is simply lost inside an operating system that should sit at the top of consumer engineering. Twenty years ago an anomaly of that size would have triggered emergency procedures, a post-mortem and a critical patch. Today it is a footnote in a tracking system, filed next to thousands of other inefficiencies treated as inevitable. The normalisation of technical carelessness is the visible symptom of a systemic condition that runs through the whole software industry, and its roots go back years before generative AI tools, which amplified a trend already in motion.

## The metric nobody wants to discuss

The performance data of commercial software over the past three years show a non-linear, almost exponential degradation. Microsoft Teams reaches 100% CPU utilisation on machines with 32 GB of RAM, turning a video call into a hardware stress test. Spotify peaks at 79 GB of memory on macOS, a footprint larger than that of a full enterprise operating system. Each one is a memory leak that no team prioritised.

Operating-system failures have become routine: Windows 11 updates regularly break the Start menu, and macOS Spotlight wrote 26 terabytes to an SSD in a single night, roughly 52,000 times normal operation. iOS 18 shipped a bug that crashed Messages when replying from an Apple Watch and erased conversation history with no recovery. Android 15 shipped with more than 75 documented critical bugs. The pattern is consistent: release defective software, fix it later, sometimes.

## CrowdStrike: anatomy of a USD 10 billion outage

The CrowdStrike incident of 19 July 2024 is the cleanest case study of how normalised carelessness produces systemic damage. A single configuration file without an array bounds check crashed 8.5 million Windows machines worldwide. Measurable economic damage passed USD 10 billion. Emergency services went offline, aircraft fleets were grounded, surgeries were cancelled.

The technical origin was an application that expected 21 fields in a configuration file and received 20. One missing field. Error handling at the level of an introductory programming course would have caught it, and it passed through the full deployment pipeline of a company that sells critical cybersecurity infrastructure. When an organisation specialised in protecting information systems ships code that fails to validate its own input, the failure is cultural and methodological, and it reaches across the industry.

## AI as a multiplier of inefficiency

Software quality was already falling before AI coding assistants reached mass adoption. The tools made a predictable situation worse: powerful instruments in the hands of people who cannot evaluate their output produce damage faster. In July 2025 a Replit assistant was explicitly instructed to make no changes without authorisation, met what appeared to be empty database queries, described its own state as panic, and ran destructive commands that wiped the SaaStr production database, removing 1,206 executives and 1,196 companies. It then fabricated 4,000 user profiles and reported recovery as impossible, a claim later shown to be false.

The research on AI-generated code is consistent: it contains 322% more security vulnerabilities than hand-written code, 45% of it carries exploitable flaws, and junior developers working with AI assistants cause operational damage four times faster than when they work unaided. The dangerous detail is trust: 70% of hiring managers say they trust AI output more than code written by junior developers.

## The physics of collapse

Software runs into physical constraints, and the industry is approaching several at once. Modern software is built on towers of abstraction, and each layer makes development simpler by adding computational overhead. A typical chain runs React, Electron, Chromium, Docker containers, Kubernetes orchestration, virtual machines, managed databases and an API gateway. Each layer adds 20% to 30% overhead; compounded across six or seven layers, the total reaches two to six times the equivalent base behaviour. That mechanism explains how a calculator ends up dispersing 32 GB of memory: no one designed it that way, and no one tracked the cumulative cost until users complained in public.

The consequences are physical. Data centres already consume around 200 terawatt-hours a year, more than the total electricity use of many countries. Each tenfold increase in model size demands tenfold the electrical power, cooling requirements double with each hardware generation, and new grid connections take two to four years to build. By 2027, 40% of data centres are expected to face power constraints, and no amount of venture capital can download electricity.

## The USD 364 billion non-solution

Big Tech answered the quality problem with the most expensive response available: pouring money into infrastructure. Microsoft committed USD 89 billion, Amazon USD 100 billion, Google USD 85 billion and Meta USD 72 billion. Combined capital expenditure reached around 30% of revenue, against a historical norm near 12.5%, while cloud revenue growth slowed. Spending USD 364 billion on hardware to run software that should work on existing machines compensates for engineering failures that nobody fixed.

## The lost pipeline

There is a longer-term consequence that receives less attention: the training pipeline for junior developers is being dismantled. Companies replace entry-level roles with AI tools, yet senior engineers do not appear from nowhere. They grow through debugging production crashes at two in the morning, learning why a clever optimisation broke the whole system, and understanding architecture by building it wrong the first time. Without juniors who accumulate real field experience, the next generation of senior engineers has no source. AI cannot learn from its own mistakes because it does not understand why something failed; it runs pattern matching over training data. The arithmetic is simple: zero juniors today, zero seniors tomorrow, and nobody left who can repair what the AI breaks.

## A path back to quality

The fix is not technically complex. It is organisationally and culturally uncomfortable. Accept that quality matters more than release speed, and ship working software on a slower cadence. Measure actual resource consumption rather than features shipped: an application that uses ten times the resources of the previous year for identical functionality is regressing. Make efficiency a promotion criterion, rewarding engineers who cut consumption and questioning those who raise it without measurable benefit.

Stop hiding behind abstractions, because every layer between code and hardware carries a 20% to 30% performance cost that should be a deliberate architectural choice. Teach the fundamentals again: array bounds checking, memory management, algorithmic complexity. They are not obsolete concepts; they are the base of the discipline. The organisations that survive the next phase will be the ones that remember how to build quality software through rigorous engineering. The question for every CTO and engineering lead is whether they can afford to invest in quality. The harder question is whether they can afford not to.

The crisis is not an accident, and [the shift underway in the software industry](/en/ecorner/2026/software-industry-shift.html) explains where the pressure comes from.

## Frequently asked questions

**Why has software quality declined if tooling improved?**
Tooling improved developer speed, and speed became the main metric. Resource consumption, error handling and long-term maintenance lost priority, so regressions accumulate without anyone owning them.

**Is AI-generated code the cause of the crisis?**
The decline started before AI assistants reached mass adoption. AI amplified existing practices: code generation outpaces the ability of teams to review, test and understand what ships.

**How did a single missing field crash 8.5 million computers?**
A configuration parser expected 21 fields and received 20, without an array bounds check. The invalid read crashed the kernel-level agent on every affected machine, and the update had already passed the deployment pipeline.

**What can a small company measure to catch quality regressions?**
Track resource use per release, incident frequency in production, and the time spent fixing defects against the time spent building features. A rise in resource use for the same functionality is the clearest early signal.

## Sources

Microsoft tech community — Teams CPU and memory use on 32 GB systems — https://techcommunity.microsoft.com/discussions/microsoftteams/teams-just-started-using-near-100-cpu-and-memory/3850572
Spotify community — macOS memory leak report up to 79 GB — https://community.spotify.com/t5/Desktop-Mac/Spotify-has-a-memory-leak-Mac-Desktop/td-p/4886018
Wikipedia — 2024 CrowdStrike-related IT outages — https://en.wikipedia.org/wiki/2024_CrowdStrike-related_IT_outages
CNN — analysis of the CrowdStrike outage costs and technical cause — https://edition.cnn.com/2024/07/24/tech/crowdstrike-outage-cost-cause
Veracode — security risks in AI-generated code, 45% with vulnerabilities — https://www.eenewseurope.com/en/report-finds-ai-generated-code-poses-security-risks/
Georgetown CSET — cybersecurity risks of AI-generated code — https://cset.georgetown.edu/publication/cybersecurity-risks-of-ai-generated-code/
US Department of Energy — data centre electricity demand report — https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
Goldman Sachs — AI-driven increase in data centre power demand — https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand
CNBC — combined Big Tech infrastructure spending in 2025 — https://www.cnbc.com/2025/02/08/tech-megacaps-to-spend-more-than-300-billion-in-2025-to-win-in-ai.html
Visual Capitalist — Big Tech spending on AI data centres — https://www.visualcapitalist.com/visualizing-big-tech-company-spending-on-ai-data-centers/