A coding agent for every engineer, without an exploding token bill.

GB300-class workstations serve local open-weight models for most of the work; frontier cloud models are called only for the tasks that need them. Engineers work in VS Code with a local harness and a specification assistant.

SectorMulti-billion-revenue companies across several industries · engineering teams

StatusIn production · Not yet published

EvidenceOperator report

Story
Stack
Load
Measurements

THE SETTING

Who uses it, and where.

Multi-billion-revenue companies across several industries · engineering teams. Environment: software engineering.

Users. Engineers per company and concurrent sessions per workstation: not yet published.

THE STACK

What runs on the machine.

Local open-weight models such as NVIDIA Nemotron Super and Moonshot Kimi, served on the GB300 workstations.
A VS Code coding agent with a local execution harness on the engineer’s machine.
A specification assistant that reads a complete specification set (the example shown holds 352 requirements) and acts on it, for instance rewriting requirements that point to tables, clauses or annexes so that each one is self-contained.
Runtime and exact model versions: not yet published.

Network posture. Hybrid: local inference on the workstations for most requests; frontier models reached only when needed. How the routing decision is made, and what may leave the company: not yet published.

THE WORKLOADS

What it does, day to day.

Everyday coding assistance.
Specification analysis and rewriting.
Specification-to-code and test generation.
Escalation of hard tasks to a frontier model.

OUTCOME

What changed.

Reported · not yet measured

Reported by Understand Tech: frontier-model spend limited to the tasks that need it; the rest served locally at a fixed cost, with source code and specifications kept inside the company. Share of requests served locally and cost per engineer: not yet published.

EVIDENCE & LIMITS

How to read this record.

The same pattern is deployed at several companies that are not named. Reported by Understand Tech; no measured figures published yet.

Measured figures (latency, throughput, concurrency, cost) are published only from a documented run on an identified machine. None is attached to this record yet.

LESSONS & FAILURES

What the team learned.

Not yet published. Three lessons from the project team, including what did not work, would complete this record.

RELATED RECORDS

Go deeper.

Application record: coding tools with a local model The machine: Dell Pro Max with GB300 Estimate the cost of local tokens

Author: Understand Tech · reviewed 2026-09-16. Customers are not named at their request.