Why Blackforge

One forge, no fine print.

How sixteen B200s in Iceland compare with renting from a hyperscaler or a GPU marketplace, line by line, including where we are not the right choice.

Topic	Blackforge	Typical hyperscaler	Typical GPU marketplace
GPU	NVIDIA B200, 180 GB, current generation	Newest GPUs often need a reservation or a quota increase first	Whatever each host has, often a mix of generations
Smallest rental	1 GPU, up to the whole forge	Top GPUs are often sold only as a full 8 GPU instance	Usually 1 GPU
Billing	Per second, 60 second minimum	Per second or per hour, depending on the instance	Per second to per hour, set by each host
The price you see	One published price per plan, changed at most once a month	List, reserved, savings plan and spot prices all differ	Set by each host, can change hourly
Egress	First 5 TB a month included, then 0.015 USD per GB	Charged per GB after a small free allowance	Varies by host
Where your data sits	Reykjanes, Iceland, in the EEA, under GDPR	The region you choose	Often a host you cannot audit
Sharing	A GPU belongs to one job at a time, never sliced	Dedicated per instance	Depends on the host
Interruptions	Only Ember work, with 30 seconds of notice	Spot capacity can be reclaimed at short notice	A host can go offline
What is on the record	A public monthly ledger: hours used and idle, every minute down, queue waits	Not published	Not published
Who answers	A person reads every work order; Anvil teams get a hardware line at any hour	Support tiers, sold separately	Community forum or tickets

The two right hand columns describe common public practice as of October 2026, not any one provider. Terms differ by provider and change; check each one's current pages.

What we do differently

Three things, on purpose.

A price that holds still

Our power comes from geothermal and hydro plants on a fixed price contract to 2031, and the outside air does most of the cooling. That is why one number per plan can stay put, and why a change is announced a month before it applies.

Every hour on the record

On the second business day of each month we publish how full the forge was, how long jobs waited and every minute it was down. If things get worse you see it there before you feel it on an invoice.

Whole machines

Up to 8 GPUs always land on one node and one NVLink 5 fabric. Sixteen take both nodes over InfiniBand. Nothing is sliced, shared or oversubscribed.

A worked example

A 72 hour fine tune on 8 GPUs.

Strike8 × 72 h × 3.792,183.04 USD

Hour bank8 × 72 h × 3.091,779.84 USD

Ember8 × 72 h × 1.39800.64 USD

Ember is cheapest when the run checkpoints and can be paused. Hour bank needs prepaid blocks of 1,000 hours; this run uses 576 of them. CPU, memory, scratch and traffic inside the forge are included in every line.

Honestly

When we are not the right choice.

You need hundreds of GPUs this week. F1 has sixteen. A second forge comes when it passes burn in, not on a date.
You need a full cloud around the GPUs. We have no managed databases, object storage product or serverless runtime. Bring those from where they already live.
Your users need single digit millisecond latency from far away. From London a round trip is about 38 ms; from San Francisco about 141 ms. Training does not care. Some inference does.
You need a compliance certificate today. We do not hold SOC 2 or ISO 27001 yet. See Trust and security for what we do have.

Still a fit? Tell us the job.

A person replies within one business day with start dates and a quote.

Write a work order