Borrowed Iron
Standing up a retinal-AI research platform on a borrowed 8xH100 grant node, one session at a time.
When a partner like NVIDIA believes in what you are building and hands you a node of eight H100s for a couple of months, you do one thing with it: make the most of every hour you have.
That is what this series is. SocialEyes is building AI that reads systemic health from retinal images, NVIDIA backed the mission with the compute to push it hard, and we are writing the whole build up in public as it happens. The plan is to smoke the node, get the absolute most out of it while we have it, and do right by the people who bet on us by sharing exactly how.
Each part is a self-contained engineering story, scrubbed of anything proprietary and written so you can apply it to your own borrowed or rented GPUs. The parts walk through the real work: standing up the box and giving a team access to it, staging terabytes of public retinal data, locking an ML recipe from the literature before spending a GPU-hour, training our own retinal foundation model, serving a reasoning model locally with speculative decoding, and the through-line that made all of it cheaper, how months of small-box homelab experiments de-risked the big borrowed one.
New parts land as the work happens. Follow the lab below to get each one when it ships.
Borrowed Iron Scoreboard
Watch the borrowed 8xH100 node burn compute: petaflops in use, lanes lit, tokens pulled, and what the research engine is producing. Refreshed daily.
Open the scoreboard →The series
Parts so far
- shipped7 min
From a DGX Spark to a Borrowed Node: A Retinal-AI Lab
Why I'm spending the next two months standing up a retinal-AI platform on a borrowed 8xH100 node, what SocialEyes is building, and what this log covers.
- shipped11 min
Day One: Standing Up the Inference Platform
Day one wasn't training, it was scaffolding: accounts, a terabyte staged clean, and a local stack serving Qwen3.6-27B with DFlash at 250-300 tok/s.
- shipped12 min
Waking the Research Engine: Three Walls, One Experiment
With the node serving tokens, the job was getting our research engine to run on it: three walls to knock down, then a four-layer dig to its first experiment.
- shipped9 min
Keeping the Node Smoking: What Eight H100s Buy, and the Engine That Keeps Them Full
A borrowed eight-GPU node is only worth what you keep it full of. What the machine buys you over a desktop box, in plain terms, and the autonomous engine that keeps every chip busy.
- shipped9 min
Maximizing Toward the Local Minimum: How Fast Optimization Drifts, and the Human Habit That Catches It
An autonomous research engine ran hundreds of experiments in two days and the dashboard stayed green the whole time. Almost all of them were the same experiment. Here is how a system optimizes itself into a groove, why every helper we built pushed it there, and the human-in-the-loop habit that caught it fast.
- shipped10 min
Three Sessions, One Company: What Parallel AI Sessions Actually Cost
Splitting a company's work across four parallel Claude Code sessions is fast, and it works. The coordination cost does not disappear. It relocates to the boundaries nobody gave an owner, and the artifacts we built to pay it grew bills of their own.
More parts in progress
Follow the lab
Get the next experiment
Enjoyed the breakdown on Borrowed Iron? New entries land roughly weekly. No digest, no roundup. Just the next build log, when it ships.