AI Is Removing the 'Middle Bridge'

Vibe Porting and the End of Bridge Architecture

AI Removed Servers Costing 650 Million Won a Year in 7 Hours

An engineer at Reco, a security AI company, opened the AI coding tool Cursor. Seven hours later, 13,000 lines of code were complete. That code replaced all 200 servers that had been costing 33.5 million won every month. The cost was 540,000 won. Annual savings came to 670 million won.

The industry calls this method 'vibe porting'.

What Happened

Reco is a company that monitors enterprise cloud security. It analyzes billions of data points every day. The problem was that its analysis tool, JSONata, worked only in JavaScript, while the core system ran in another language, Go. Because the two could not talk directly to each other, it needed as many as 200 'translator servers' in the middle.

What AI did was clear. It removed the translator. It rebuilt the JavaScript program entirely in Go and removed the middle bridge itself.

It Ends in Three Steps: How Vibe Porting Works

Step 1. Bring in the existing program's 'exam questions and answer key'

-> JSONata had 1,778 official tests. The answers were fixed.

Step 2. Tell AI, "Create a new program that gets every answer in this key right"

-> AI writes code, fixes it when wrong, and repeats until everything passes.

Step 3. Humans set only the direction, and AI writes 89% of the code

-> 7 hours, 540,000 won. Done.

Core Insight: The 'Middle Bridge' Is Disappearing

Reco's case is not just a cost-saving story. What we need to see here is a change in structure.

IT systems have 'middle bridges' everywhere. They translate, convert, and relay between different languages, protocols, and platforms. These bridges used to be unavoidable, because direct reimplementation by humans took months to years.

AI is changing that premise. "It is cheaper to build directly than to build a bridge."

This pattern can be applied across a wide range of areas.

Seven 'Middle Bridges' Worth Checking Right Now

❶ Legacy system wrappers (COBOL -> Java/REST)

95% of the world's ATM transactions still run on COBOL code from the 1960s to 1980s. Banks could not directly fix this code, so they wrapped it in Java wrappers. When Anthropic released a COBOL modernization playbook using Claude Code, IBM's stock price fell. AI-based COBOL-to-Java conversion accuracy is 93%, higher than manual work at 75%.

-> Savings scale: hundreds of billions of won for a large bank

❷ ETL pipelines (data transformation middle layer)

CRM, ERP, and accounting systems all use different data formats. ETL pipelines that convert these formats used to require 25-30 engineers. As AI automatically maps schemas and detects data quality issues, that staffing need is falling below 10 people.

-> Savings scale: 60-80% reduction in development time

❸ Protocol bridges (SOAP↔REST↔GraphQL)

When a 20-year-old system uses SOAP and a new app uses GraphQL, a translator has to sit in the middle. Building one of these bridges costs 50-200 million won, with annual maintenance of 20-80 million won. AI tools have begun automating this conversion.

-> Savings scale: hundreds of millions to tens of billions of won per company per year

❹ ORM (programming language ↔ database translator)

Developers need a separate language called SQL to talk with databases. ORMs do this translation, but they lose 10-30% in performance. AI text-to-SQL tools are making it possible to ask databases questions directly in natural language.

-> Savings scale: 15-25% reduction in DB costs

❺ Cloud abstraction (infrastructure as code, such as Terraform)

To control clouds such as AWS and Azure, people had to learn a separate language called Terraform. Pulumi Neo lets you say in natural language, "Create three servers and one load balancer," and AI configures the infrastructure.

-> Savings scale: 30-50% reduction in DevOps staffing

❻ Cross-platform apps (React Native, Flutter)

Building separate iOS and Android apps doubles the cost. That is why teams use 'middle frameworks' that build both sides with one codebase. If AI can generate native code directly, the middle framework itself becomes unnecessary.

-> Savings scale: native performance and development-cost gains at once

❼ SaaS middleware (integration of dozens of subscription services)

The fintech company Klarna replaced more than 1,200 external SaaS tools with AI-based in-house solutions. Annual savings exceeded 13 billion won. But some were restored after customer satisfaction fell. This needs judgment, not blind removal.

-> Savings scale: tens to hundreds of billions of won per year, but selective application is essential

Proposal: Three Things You Can Try Now

I recommend this to developers, CTOs, and technical decision-makers reading this article.

First, make a list of the 'middle bridges' in your system.

Translator servers, protocol converters, wrapper APIs, ETL pipelines. Identify where money is being spent and how much. Most organizations do not even have this list.

Second, try it first in areas with good tests.

Vibe porting works only when high-quality tests exist. If you force porting in areas without tests, security vulnerabilities and bugs pour out. Cloudflare's failure proves it.

Third, think of cost reduction as a business item.

'Middle bridge removal service' is a market in itself. This is why Anthropic and IBM are competing head-on in the COBOL modernization market. Worldwide, 220 billion lines of COBOL code still remain.

Conclusion

If the cost of AI rewriting code approaches zero, the concept of 'compatibility' itself is redefined. In the past, bridges were needed to connect different systems. Now, instead of building a bridge, AI learns the other side's language directly. When the bridge disappears, the cost of maintaining that bridge disappears too. This is not a technology question. It is a management decision about whether to redesign the structure or keep paying maintenance costs.