0x00  Santa Clara, CA · Tier-III colocation · Owner-operated

Small, durable infrastructure, close to you.

DataROM is an edge hosting company. Private cabinets and managed hosting in Tier-III Santa Clara colocation, from a single rack to a row. You own the hardware. We procure it, build it and keep it running.

ADDR
DATAROM.ROM / MEMORY MAP
0x01
Colocation
0x02
GPU nodes
0x03
Managed infrastructure
0x04
Linux · SaltStack · Jenkins
POWER
A/B 208 V three-phase
COOLING
N+1
TRANSIT
10 G blended
XCONNECT
100 G

0x01  Services

Three services, one accountable party.

Most teams end up coordinating a reseller, an integrator, a colo and a network provider. We do all four, and the rack stays yours.

01

Colocation

Private cabinets in Tier-III Santa Clara colocation: A/B 208 V three-phase, N+1 cooling, generators, carrier-neutral fiber. We hold the colo contract and the remote-hands relationship. You get one invoice and one phone number.

02

GPU nodes

8-GPU NVIDIA HGX B200 and B300 nodes, specified to your workload, sourced through OEM and authorized channels, racked and burned in for 72 hours. The hardware is yours from the day it ships.

03

Managed infrastructure

Monitoring, firmware, incident response and first-line remote hands, with a monthly health report. Your engineers keep root. We keep it running.

0x02  Automation

Built and run from code.

Linux, SaltStack and Jenkins are services in their own right, on our hosting or on infrastructure you already run.

Linux

Linux administration

Image builds, PXE provisioning, kernel and driver baselines, patching and hardening for servers and GPU nodes.

SaltStack

Configuration management

Salt states and pillars for every host, so a rebuilt node comes back identical and drift gets reported instead of discovered.

Jenkins

Pipelines

Build, test and deploy pipelines, plus scheduled jobs for firmware checks and burn-in.

0x03  Case study

One HGX B300 node, fully racked.

What a single-node deployment contains. Every line is yours at cost plus our build fee. Nothing is bundled where you can't see it.

ComponentSpecEst. USD
HGX B300 serverGigabyte G894 or equivalent, 8× B300, 2× EPYC 9575F, 2.3 TB DDR5, 8× E1.S NVMe$460k–540k
Rack and PDUs42U 800 mm, 2× zero-U metered three-phase PDUs, C19$5k–10k
Network100 GbE top-of-rack, 1 GbE out-of-band, optics and DACs$5k–10k
Storage node2U, 50–100 TB NVMe, 100 GbE, for datasets and checkpoints$10k–20k
Firewall, KVM, sparesEdge firewall, IP KVM, spare PSU, fans, DIMMs, NVMe$7k–14k
Build and commissioningIntegration, firmware baseline, OS and CUDA stack, burn-in, documentationRequest price
Hosting, monthly15 kW private cabinet, A/B power, cooling, 10 G transit, 100 G cross-connect, remote hands, monitoringRequest price

Why colocation, not a closet

One B300 node draws 13 kW and rejects 45,000 BTU/hr. An office or warehouse needs a 5-ton cooling retrofit, three-phase service and a UPS before it can hold one. Colocation includes all of that, plus generators and 24/7 staff, for less than the electrical work alone.

Scales to a row

The rack is specified with 8U free and a 25 kW upgrade path, so a second node drops in without re-cabling.

0x04  Pricing

Hardware at cost. Fees you can read.

Three line items. No per-GPU-hour meter.

1 / Procurement

Request price

Flat percentage of hardware cost, capped. Covers spec, multi-vendor quotes, negotiation, PO management and receiving inspection. Hardware is invoiced to you directly or passed through at cost.

2 / Build and commission

Request price

Per node, one time. Racking, cabling, power and network design, firmware baseline, OS and driver stack, 72-hour burn-in, as-built documentation and handover.

3 / Managed hosting

Request price

Per 15 kW cabinet, 12- or 36-month term. Power, cooling, space, 10 G transit, cross-connect, monitoring, monthly health report and first-line remote hands.

Every quote is priced to what you want hosted. Tell us the hardware and the workload and we will send a bid. Hosting is the colocation cost plus a fixed operations margin, stated on the quote.

0x05  Process

Quote to first job in about six weeks.

Lead time is set by the server OEM. Everything else runs in parallel.

01 / Week 0

Workload call

Models, batch sizes, data gravity, utilization. We tell you if renting is still cheaper.

02 / Week 1

Spec and quotes

Side-by-side OEM quotes, colo quote, total cost of ownership.

03 / Weeks 2–5

Order and prep

You issue the PO. We reserve the cabinet and pre-build everything but the node.

04 / Week 6

Rack and burn-in

Node arrives, racked the same week, 72-hour burn-in, documented.

05 / Ongoing

Operate

Monitoring, firmware, incident response, monthly report.

0x06  FAQ

Straight answers.

Who owns the hardware?
You do, from the day it ships. It is on your balance sheet, under your warranty, and you can move it to another facility at the end of the hosting term.
Who has root?
Your team. We hold BMC and out-of-band access for hardware operations and are happy to be removed from the OS entirely.
What happens when a GPU fails?
We diagnose over BMC, open the OEM warranty case, meet the part at the facility and swap it. On-site spares for PSUs, fans and NVMe cover the common failures.
Can I visit?
Yes. Escorted access to your cabinet is part of hosting.

0x07  Request a quote

Tell us the workload. We'll send numbers.

A line-item quote for hardware, build and hosting, plus a rent-versus-own comparison at your utilization.

Santa Clara, California

We reply from a person, not a sequence.