- Authors
- Name
- 오늘의 바이브
Claude Stopped. So Did the Code.
In early March 2026, Anthropic's Claude went down for 48 hours. The service was completely unavailable. What happened next was revealing. Engineers at Meta, Netflix, and other big tech companies could not perform normal coding tasks.
An AI tool went offline for two days, and developers' hands stopped moving. This was not a minor inconvenience. It was structural evidence of technical dependency. In a survey of 1,000 developers, Claude Code ranked as the most-used AI coding tool. When the number-one tool disappeared, a number-one-sized gap opened up.
"Coding Without AI Is Just Slower"
The reaction from Gauresh Pandit, a senior software engineer at Meta, was telling. When Claude went down, he judged that "coding alone without AI would be significantly slower." A senior engineer at one of the world's largest tech companies perceives AI-free coding as the degraded state, not the baseline.
This is not about one person being lazy. The industry's structure has shifted. Claude Code, GitHub Copilot, and Codex are no longer extensions of the IDE. They are effectively the IDE itself. Autocomplete, debugging, refactoring, test generation -- AI handles it all. Remove the tool, and the entire workflow collapses.
Gergely Orosz, former engineering manager at Uber, flagged the same concern. As AI tool dependency grows, developers' foundational skills are atrophying. The 48-hour outage did not expose a weakness in the service. It exposed a weakness in the developers themselves.
Junior Developers Are Most at Risk
Skill atrophy does not hit everyone equally. The hardest-hit group is junior developers. Senior engineers spent years debugging, designing architectures, and reviewing code before AI tools existed. They have muscle memory. Juniors do not.
A significant number of junior developers entering the industry today started their careers alongside AI tools. They delegate debugging to AI before learning debugging principles. They paste error messages into AI before reading them. They get working code from AI before understanding why it works.
The problem is that this approach is efficient in the short term. AI-generated code usually runs. Tests pass. Code reviews often catch nothing wrong. But using AI without foundational skills is like doing math without understanding arithmetic. The calculator works fine until it goes offline for 48 hours. Then you have nothing.
Even Boris Cherny, Anthropic's head of Claude Code, acknowledged the problem. The person building the AI coding tool is warning about over-reliance on AI coding tools. That irony is hard to miss.
South Korea's Response: Doubling Down on Dependency
The situation in South Korea adds an interesting dimension. Rather than reducing dependency after the Claude outage, Korean companies moved to systematize it.
Startups began codifying AI tool usage in internal coding guidelines. "Actively use AI tools" became official policy. Some companies are now exploring whether to factor AI usage into performance reviews. The implication: if you are not using AI, you are considered less productive.
| Area | Response |
|---|---|
| Startups | AI tool usage codified in coding guidelines |
| Performance reviews | AI usage metrics under consideration |
| Semiconductor companies | Accelerating AI coding tool adoption |
| Samsung, SK hynix | Company-wide AI coding tool rollout underway |
Samsung Electronics and SK hynix are accelerating their adoption of AI coding tools. The direction is not wrong. AI tools do improve productivity. The problem is that there is no Plan B. What happens when the next 48-hour outage hits? Is there a multi-vendor strategy? Can teams maintain minimum operations without AI?
The Water AI Drinks, the Coffee Developers Drink
The Claude outage revealed another problem: the physical fragility of AI infrastructure. Data centers consume resources at a scale most people do not appreciate.
Every query to ChatGPT or Gemini consumes an average of 23.9 ml of water. A 500 ml bottle of water is enough for 21 queries. That water goes into cooling the GPUs in data centers. AI needs to think, thinking generates heat, and heat requires water to dissipate.
The scale is staggering. Microsoft's global water consumption surged 34% during the AI boom. Google set a target to return 120% of its water consumption to local communities by 2030 -- a goal that only makes sense if consumption is already alarming. In South Korea, large data centers in the metropolitan area are drawing on groundwater and municipal water supplies for cooling, prompting the Ministry of Environment to consider mandatory water usage reporting for data centers.
The exact cause of Claude's 48-hour outage was not disclosed. But the physical infrastructure that AI depends on -- power, cooling, networking -- is finite. The software lives in the cloud, but the cloud sits on the ground. Ground-level problems stop the cloud, and when the cloud stops, developers stop.
Dependency or Evolution?
There is a counterargument. When calculators appeared, people raised the same concern. "Mental arithmetic is dying" was a real warning decades ago. Humanity lost the ability to do mental math but gained the capacity for more complex mathematical reasoning. Maybe AI coding tools are the same kind of trade.
Fair point. But there is one critical difference. Calculators do not go down. They run on batteries. Their supply chain is distributed, with no single point of failure. AI coding tools, on the other hand, are tethered to a specific company's specific API. When Anthropic stops for 48 hours, every Claude Code user stops. When OpenAI goes down, every Copilot user goes down.
This is not a tool problem. It is an architecture problem. The entire development workflow depends on a single point of failure. In traditional software engineering, this design would never be accepted. Running critical infrastructure without high availability, redundancy, or fallback strategies violates basic engineering principles.
Yet developers are not applying these principles to their own workflows. They write fault-tolerance logic into their code while ignoring the fault tolerance of their own capability. The code has retry logic and graceful degradation. The developer does not.
What 48 Hours Left Behind
Claude's 48-hour outage started as a technical failure at Anthropic. What it revealed was a structural problem across the entire industry. When the most-used tool among 1,000 developers disappears for two days, how many of them can still work without it?
The answer is uncomfortable. Fewer every year. Juniors learned to code with AI from day one. Seniors do not want to go back to doing what AI handles for them. Korean companies are institutionalizing AI dependency. Data centers are pushing against the physical limits of their resources.
Forty-eight hours is not long. But what froze during those 48 hours was not a service. It was developers' self-awareness. When was the last time you checked how much you could accomplish without your AI tools? If you cannot answer that question, the next outage will feel much longer than 48 hours.
Sources