$3 trillion passes through this code every day. 95% of US ATMs run on this code. 43% of the world's banking systems depend on it.
The code is called COBOL. And the average age of a developer who can work with it is 55 — with 10% retiring every single year.
220 Billion Lines, and Nobody's Learning It
COBOL was born in the 1960s. Over 60 years old — and somehow still alive. Here's why: the core logic of banking, insurance, airlines, and government systems was already written in COBOL, and migrating it to another language is typically a "50 engineers, 5 years" kind of project.
70% of Fortune 500 companies still run mainframes for core operations, and over 220 billion lines of COBOL are running in active production right now. Here's the problem: 85% of US universities dropped COBOL from their curricula after the 1990s. The talent pipeline dried up.
The consequences are showing up. 60% of organizations using COBOL say finding qualified developers is their single biggest challenge. 46% report the shortage is already affecting their operations. Meanwhile, skilled mainframe developers are commanding $45-$96 per hour.
GitHub Copilot Doesn't Work on Mainframes
There are plenty of AI coding tools out there. GitHub Copilot, Cursor, Claude Code — why can't you just use those?
Here's the thing: every one of those tools was built for modern development environments. Git, shells, REST APIs. But the mainframe world is different. Not files — Datasets. Not directories — PDS (Partitioned Data Sets). Not a terminal — a TN3270 emulator. You don't type commands; you press PF keys in panels.
Ask ChatGPT to write you some JCL and it'll produce something that looks right — but will almost certainly fail when you run it on a real z/OS system. The AI has never actually seen a mainframe environment. The HN community response summed it up: "US banks desperately need this yesterday."
| Standard AI Coding Tools | Hopper | |
|---|---|---|
| Terminal access | Not supported | Native TN3270 support |
| JCL authoring | Generates text only | Submits jobs, parses results |
| Error debugging | Guesswork | Reads and interprets JES spool directly |
| Dataset exploration | Not supported | Queries VSAM like SQL |
| CICS integration | Not supported | NEWCOPY execution + approval checkpoint |
What Hopper Does Differently
Hypercubic is a YC F25 startup founded by two former Apple ML engineers. Their product, Hopper, is a mainframe-native AI development environment.
The architecture is three things: a real TN3270 terminal emulator, mainframe-aware panels for datasets, jobs, members, and spool output, and an AI agent that can actually operate across all of it. The agent drives ISPF by panel ID, writes column-strict JCL syntax, submits jobs, parses JES spool output, and converts failures into structured diagnostics.
One HN tester put it well: "JES output parsing alone solves something that used to take a junior dev half a day — in seconds." COBOL editing, compiling, testing, and shipping to CICS — all within a single conversation.
Core Design Principle
The AI acts autonomously, but risky operations (CICS NEWCOPY, job submission) require human approval before execution. "AI works, humans review."
How to Start Today
- Download (free, no credit card)
Go to hypercubic.ai/hopper and download for your OS (macOS, Windows, Linux). Sign up and you're in. Current version: 1.3.1. - Connect your mainframe
Enter your mainframe credentials (hostname, port, user ID) to establish the TN3270 connection. No mainframe? Request a test z/OS account from their site. - Try your first agent task
Start with natural language: "Show me the PDS members inside the PAYROLL dataset." The AI opens ISPF and navigates directly. - Test automatic JCL authoring
"Write a COBOL compile job based on COMPILE.JCL." The agent produces syntactically correct JCL and waits for your approval before submitting. - Debug a failing job
Give it a job name that failed. The agent reads the JES spool, identifies the abend code, failing step, and source line, then explains the root cause.
Before deploying to mission-critical systems
Some in the HN community raised valid concerns about LLMs operating in financial core systems. Hopper's enterprise plan offers on-premises/VPC deployment and SOC 2 certification. Even so — validate thoroughly in dev/test environments before any production rollout.
Dive Deeper
Hopper Official Page Demo video + free download. Watch the AI agent operate a TN3270 terminal live. hypercubic.ai
Launch HN: Hypercubic (YC F25) Sharp questions from mainframe engineers, answered directly by the founding team. news.ycombinator.com
The $3 Trillion Code Nobody Knows How to Fix Deep-dive statistics on the COBOL developer shortage and market dynamics. metaintro.com
Hypercubic Insights Why legacy modernization projects fail — and how the AI approach differs. hypercubic.ai
Show HN: Agentic Interface for Mainframes Direct reactions from working mainframe engineers. news.ycombinator.com
