Blog
Notes from the machine room
Engineering write-ups, product notes and customer stories from the team that runs Ferrite. Real numbers, no launch-day gloss.
All posts
Engineering
21 Oct 2026
6 min read
Finding the one slow link in a 384-GPU job
A training run in ams-1 was 18% slower than it should have been. Here is how we traced it to a single optical cable, and what we changed after.
SO
Sam Okafor
Head of Networking
Product
7 Oct 2026
4 min read
Reserved blocks you can resize
Reserved capacity used to mean a fixed number of GPUs for a fixed term. Now you can grow or shrink a block each month and keep the committed price.
PN
Priya Natarajan
VP Infrastructure
Customers
23 Sep 2026
5 min read
How Halyard Speech trains without a platform team
A twelve-person lab building speech models for clinics moved from a shared university cluster to Ferrite and stopped queuing for compute.
LB
Lena Brandt
Head of Support
Product
9 Sep 2026
6 min read
Choosing between Forge 80 and Anvil 192
More memory is not always faster. A plain guide to which GPU suits fine-tuning, long-context training and inference, with the numbers we measured.
YP
Yuna Park
Head of Research
Engineering
26 Aug 2026
5 min read
Cooling hall B with lake water: one year on
Hall B in ams-1 is cooled by water from a nearby lake. A year in, here is what it did to power use, failure rates and our summer maintenance plans.
MH
Marcus Hale
Data Hall Lead, ams-1
Start in minutes
Your first 100 GPU-hours are on us
Sign up, add a card and launch a cluster before your coffee’s cold. No sales call, no minimum spend.