When Inference Gets Cheaper Than the VM It Runs On

The claim. A developer spent $20k of Claude Opus 5.5 tokens to port Microsoft's Go-based TypeScript compiler (TSGO) to Rust. The result, TSC RS, runs 12.5x faster than the JavaScript-based TS v6 and 2x faster than the official TS v7. Benchmark highlights include 4x faster than Bun's checker on a project with Effect patch.

The method. The port succeeded because the AI worked from Go, not JavaScript. The agent first faithfully ported Go's runtime—complete with goroutines and garbage collection—into Rust, then iterated for two weeks to achieve the speed target. The author never read a line of code. The approach sidesteps the difficulty of directly translating TypeScript's complex type system from JS to Rust.

The hidden agenda. Performance is not the ultimate goal. The author envisions a world where inference costs drop 10x per year while VM and compute costs rise. Running agents and tooling inside V8 isolates—like Cloudflare Workers—eliminates the overhead of OS-level virtualization. He demoed a full agentic dev loop inside an isolate using TS Rust and a fake filesystem, peaking at 260 MB RAM and negligible CPU.

The tradeoffs. The two-hop port (JS → Go → Rust) is not reproducible without Microsoft's Go version. The $20k cost, subsidized by subscriptions, is still prohibitive for most teams. The Rust code is not human-reviewed and faithfully reproduces existing TS bugs. Bun's checker, while faster on simpler projects, raises false positives on large codebases like tRPC and Sentry.

What leaders should watch. As inference commoditizes, infrastructure cost becomes the next engineering budget line item. Teams should evaluate whether their build and check tooling can run in lightweight sandboxes or browser environments. TS Rust is a technology preview; bet on Microsoft's TS v7 for production, but monitor the isolate-compute pattern for future agent economics.

The bottom line. The real signal is not a faster compiler but a shift in capital allocation: model intelligence is getting cheap, and the compute to run it is becoming the bottleneck. Engineering leaders who ignore this will face unexpected cost growth as agent adoption scales.