Developers expect AI coding tools to speed up work. They don’t expect those same tools to quietly destroy their hardware. Yet that’s exactly what happened with OpenAI’s Codex.
One developer watched his SSD rack up 37 terabytes of writes in just 21 days. The culprit? A SQLite database stuffed with TRACE-level logs. Extrapolated, that pace hits 640 TB a year. For a typical 1 TB consumer drive rated at 600 TBW, it means burning through the entire warranted endurance in under 12 months.
The Scale of the Damage
Rui Fan, a project management committee member for Apache Flink, opened the GitHub issue that brought this to light. “On my machine, after about 21 days of uptime, the main SSD has written about 37 TB,” he wrote. “Process/file-level checks show Codex SQLite logs are the main continuous writer.” (The Register)
Another user ran Codex’s own analysis on the damage. It pegged the value lost on a Samsung 990 2 TB NVMe at $38.64. The same analysis suggested the bug had “plausibly burned low-single-digit millions of dollars of SSD endurance across users during the March-June window.” They based the math on $0.13 per TB written. Simple formula. SSD price divided by its TBW rating gives the cost per terabyte. Multiply by excess writes. The bill adds up fast.
But. The real story runs deeper than one number. SSDs don’t just wear from the bytes you see. They suffer from write amplification. Here, the SQLite database operated in WAL mode. Tens of thousands of insert and delete operations every minute. Each small logical write triggered far more physical activity on the flash. The log file itself didn’t grow huge. The churn inside it did the damage. 70 percent of the retained data was TRACE noise. Inotify events. WebSocket payloads. Routine file accesses. Data nobody would ever read.
This didn’t start last week. Related complaints appeared in the Codex repository as far back as April. Issue after issue flagged excessive SQLite WAL writes during streaming. The default logging target ignored the standard RUST_LOG environment variable. Users couldn’t easily quiet it. Telemetry came later, turned on by default except where prohibited by law. The local diagnostic logs, however, shipped with the first versions and stayed on.
OpenAI engineers trace the spike to a February change. They shifted app-server SQLite logs to TRACE level. More verbose than ERROR or WARN. The system then stored high-volume data in a format that generated far more disk activity than anyone anticipated. Ironically, Codex itself reviewed that pull request. The AI that writes code signed off on code that hammered storage.
Engineers at OpenAI knew. A spokesperson told The Register the company is aware and working on fixes. Several pull requests in recent days target the problem. One reduces noise in the persisted SQLite sink. Others adjust defaults while keeping diagnostic capability. Progress shows in the thread. Yet users still report issues. The ticket remains open. Workarounds circulate. On Linux or macOS, symlink the logs_2.sqlite file to /tmp. Writes go to RAM. The file contains no conversation data, so rebooting erases nothing critical.
The episode reveals something larger about agentic AI tools. They ship fast. They gather data aggressively. They sometimes overlook the physical costs on the machines running them. SSD endurance has improved. Newer models from Samsung and others push TBW ratings higher. Enterprise drives tolerate thousands of full writes. Consumer laptops? Not so much. A developer running Codex 24/7 on a MacBook or Windows notebook could face performance drops or outright failure within a year.
Write amplification isn’t new. Storage experts have warned about it for years. Databases with frequent small updates amplify writes dramatically compared to simple append-only logs. Codex combined that pattern with the noisiest logging level and a persistent store. The result was predictable once measured. One analysis found over 5 billion rows inserted over time while the retained count stayed around 500,000. The database constantly cleaned house. Each cleanup cost more writes.
Recent coverage echoes the original findings. Notebookcheck detailed the same 37 TB in 21 days and the write amplification from constant insert-delete cycles. (Notebookcheck) TechTimes reported the bug remains unpatched in some versions as of June 22. Korben noted that updates released June 22 wiped out around 85 percent of the excess logging. Developers should update immediately.
Discussions on X and Hacker News exploded over the past 48 hours. Engineers traded horror stories of unnoticed disk wear. Some ran checks on their own systems and found Codex among the top writers. Others pointed out the irony. An AI coding assistant that accelerates software development ends up accelerating hardware replacement cycles. The low-single-digit millions in lost endurance might seem small against OpenAI’s valuation. For individual developers buying SSDs out of pocket, $30 or $40 per machine adds up. Across thousands of users, it becomes real money.
OpenAI built Codex to handle complex coding tasks with GPT-5.3 under the hood. It analyzes code, suggests changes, and streams responses. All that activity generated events worth logging. The logging system simply captured too much, too often, in a format too expensive for constant use. Better defaults would have caught this. Raising the baseline level above TRACE. Filtering dependency noise. Avoiding raw OpenTelemetry payloads in the local SQLite file. These changes appear in the recent fixes.
So the bug gets patched. Users update. The immediate crisis passes. What lingers is the reminder. Every AI tool carries hidden infrastructure costs. Not just in the cloud. Not just in electricity. Sometimes right on the developer’s desk, wearing down the storage that holds their projects. Companies racing to ship agents and copilots must now factor in these physical limits. SSDs last longer than they used to. They don’t last forever. And when an application writes hundreds of terabytes a year for diagnostics alone, something needs to change.
Check your own drive stats. Tools like CrystalDiskInfo on Windows or smartctl on Linux will show the total bytes written. If you’ve run Codex for weeks, the numbers may surprise you. Update the tool. Apply the symlink workaround if you can’t wait. Then watch what OpenAI ships next. The company that wants to automate software engineering just learned a costly lesson about respecting the hardware that makes it possible.
WebProNews is an iEntry Publication