Special series · Borrowed IronLive artifact

Borrowed Iron Scoreboard

What the borrowed 8xH100 node is doing right now: the compute it can throw, the iron it is built from, the tokens it pulls through an open model, and what the autonomous research engine produces with it. The series argues for renting frontier iron and running it hot. This is the live proof, refreshed daily.

← Back to the series

The iron
8x H100
GPUs
80GB HBM3 each
300 / 640 GB
GPU memory
HBM3, 47% committed
208 vCPU
CPU
2x Xeon Platinum, 104 cores
1.8 TB
System RAM
DDR5
17.6 TB
Fast storage
local NVMe

To scale: roughly 233 maxed MacBook Pros of raw AI compute, 5x a MacBook’s whole memory in GPU VRAM alone, and about 50x the memory bandwidth. Back-of-envelope, by peak throughput.

Borrowed iron · in use
0.0PFLOP/s69% of capacity burning
15.8PFLOP/s available
0 / 8
GPUs lit

all lanes working

0
Tokens pulled

last 24h

287
tok/s
Warm decode

served, single stream

9 / 66
Grant day

of the borrowed window

GPU lanes8 of 8 lit
0Qwen warm brain
1Job filler
2Encoder (filler)
3Encoder training
4Encoder training
5Encoder training
6Encoder training
7Encoder training

Eight lanes of borrowed iron. The warm brain answers ARIA, a filler keeps idle cards busy, the rest grind encoder training. Exact per-lane load fills in on the daily live read.

CPU lanes208 vCPU, 92% busy

Each lane bundles 13 of the 208 logical cores. The host CPUs stage terabytes of retinal data and feed the GPUs, so they run warm even while the iron does the heavy lifting.

Tokens pulled
70.6M
tokens
Per day

last 24h

517.7M
tokens
Since day one

and counting

Almost all of it is prompt. ARIA is a reader: it pulls that much context through our own open-source Qwen3.6 27B on GPU 0, kept warm around the clock at 287+ tokens a second on one card. Fan the queries out and the node pushes past 8,000 tokens a second flat out. No API meter running.

588
novels of text, every day
90
times through War and Peace, daily
~11 m
stack of printed pages a day, past a 3-story building
~2.5 yr
of nonstop human typing, every day
About 4,314 novels since day one. Rough figures at about 0.75 words per token.
What it would cost elsewhere

Those tokens ran on our own iron for $0 in API fees. The same 517.7M tokens on a commercial API, at list input prices:

Gemini 3.5 Flash
$777
$1.50/M input
Claude Opus 4.8
$2,588
$5.00/M input
GPT-5.5 Pro
$15,530
$30.00/M input
We paid $0.It is our own open-source model on our own GPU.
List input prices, Artificial Analysis / OpenRouter, June 2026. Output billed on top there; ours is near zero.
ARIA · autonomous research
662
Papers logged
48
Findings exported
14
Experiments

last 2 days

8%45%

config-path run-success, trending

Engine healthy

Live as of Jul 1, 2026, 05:00 AM UTC, auto-refreshed daily

Petaflops are FP8 dense tensor throughput for 8 H100 SXM cards. Numbers come from the project build log plus a daily read of the node. No credentials or private data are exposed here.

Follow the lab

Get the next experiment

Enjoyed the breakdown on the Borrowed Iron scoreboard? New entries land roughly weekly. No digest, no roundup. Just the next build log, when it ships.