GRGPURackAI infrastructureGet a box
Infrastructure

A utility cut is a transfer, not an outage.

Most hosting pages describe someone else's building. This one describes hardware we own, in Florida, USA, and you're welcome to ask us anything about it.

Power

Three sources. No transfer gap.

The usual arrangement is utility power with a diesel generator behind it — which means every outage is a race between your UPS runtime and how fast the generator spins up. We inverted it. The batteries are already carrying the load.

Status unavailable
Florida, USA
Solar
Battery
Grid
Rack
Fiber A — Carrier 1
Fiber B — Carrier 2

The monitoring agent hasn't reported recently. We'd rather show nothing than a stale reading.

Solar array

Generates into the facility and keeps the bank charged through daylight.

Battery bank

Sits under the load continuously. There is no gap to bridge because nothing has to start.

Utility feed

Present, but it is the backup — not the thing everything depends on.

The rack

Sees a steady supply regardless of which of the three is doing the work.

Network

Two carriers. Genuinely two.

"Redundant network" usually means two cables from one provider into one handoff. That protects you from a backhoe and nothing else — when the provider has a routing incident, both of your redundant links go down together at exactly the same second.

We buy from two separate fiber carriers. Different networks, different upstreams, different failure modes. One can go completely dark and the rack keeps serving.

Carrier A
Carrier B
Independent networks, not two drops from one provider. 1 Gbps unmetered as standard, 10 Gbps available.
Hardware

What's actually in the box.

One tenant per card

No virtualisation layer, no MIG slicing, no time-sharing. The card in your box answers to you and nothing else, so throughput doesn't wander based on who else is logged in.

NVMe and ECC throughout

Model weights load off NVMe, not spinning disk or network storage. System memory is ECC — a bit flip halfway through a week-long fine-tune is not a fun way to learn this lesson.

Your data stays on your disk

We don't image your volumes, we don't inspect your traffic, and we don't train anything on what you run. If you want the disk wiped on cancellation, say so and we'll do it and confirm.

Straight answers

What we don't claim.

You can check all of this, so there's no point pretending otherwise.

We're not a Tier III certified datacenter.

We're an owner-operated facility. What we have instead is power redundancy we control directly and can explain to you in detail, rather than a certificate about a building someone else runs.

We don't have thousands of GPUs.

We have the ones listed on the pricing page and we'll tell you exactly what's free. If we don't have what you need we'll say so instead of taking a deposit and sourcing it later.

We're not the cheapest per hour.

If your job runs for six hours a month, a spot instance will beat us and we'll point you at one. We're built for the workload that runs continuously.

We don't publish an uptime percentage yet.

We haven't been running long enough for that number to mean anything. When we have real measured history we'll publish it, including the bad months.

Tell us what you're running.

Model, context length, requests per second — that's enough for us to tell you which box you need, and whether the cheaper one would do.

No setup fee · Month to month · sales@gpurack.net