sequenced.ai
Articles/Models & infrastructure/Blueprint//7 min read

Firmus turns a GPU reservation into a training-cluster decision

Firmus combines GPU compute, storage and scheduling. For a training team, the telling question is how quickly a job can restart with a usable checkpoint.

By Sequenced deskAI-assisted, source-led · how we work
Visit Firmus website ↗
GPU cloudCompute choiceInstances and larger reserved configurations.
Bare metalDedicated capacitySingle-tenant clusters are available by reservation.
SlurmJob schedulingManaged clusters use a reservation route.
WEKA NVMeTraining storageParallel file storage for data and checkpoints.
Firmus mark
Firmusfirmus.co · independent research

Represent this company? Verify your work email to access its workspace, or send the desk a factual correction.

A research team can reserve eight GPUs and still lose much of its training window waiting for data or reconstructing a failed job. Firmus sells GPU cloud and dedicated clusters alongside the file storage, networking and scheduling those jobs need. The most revealing buying exercise is to trace one training run from its first data read to the moment it can restart. That exposes which pieces of the infrastructure the reservation actually secures.

In brief
  1. 01What Firmus operates GPU infrastructure, from individual compute configurations to dedicated multi-node clusters with high-speed networking.
  2. 02Who benefits Research and infrastructure teams with a known training job, a reproducible environment and constraints on time or location.
  3. 03What to verify Which hardware and storage the team can book together, where they will run and how much useful work the allocation delivers.

01 / ProductHow an AI factory becomes something a researcher can book

Firmus’s compute catalog describes both virtual machines and bare metal. Its displayed H200 configuration illustrates the difference between a GPU name and a usable system. The example contains eight accelerators, each with its own memory, along with a host, system RAM and network connections. A model that fits only when its parameters are spread across devices still depends on software that can divide the work. The total memory shown for a node cannot be treated as one ordinary allocation.

On the bare-metal route, the buyer reserves dedicated infrastructure. The public offer describes single-tenant clusters and InfiniBand links for multi-node work. Cloud services add managed Slurm scheduling by reservation, preconfigured CUDA libraries and observability. These are meaningful distinctions for someone bringing a distributed training script. That person needs to know how jobs get allocated, what software the environment supplies and who can investigate a run that stalls after leaving a single node.

Firmus also develops physical infrastructure, including the Australian Project Southgate program. The program explains why power, cooling and site design feature so prominently in the company’s description. It is a different source of evidence from the GPU availability table. A campus development or planned electrical capacity gives context about the operator’s direction. To schedule work, a team needs a confirmed location, machine configuration, start date and reservation.

02 / AudienceThe strongest use case has a deadline and a repeatable workload

Imagine a team adapting a document model to recognize layouts particular to its industry. It trains periodically when a labeled data release is ready. The engineers know their container image, their dataset size and the validation metric they want to preserve. Their real constraint is finishing a qualified run within a four-day window. A dedicated period with the required storage can be more valuable to them than a lower hourly figure attached to capacity whose start date is uncertain.

A small team still choosing between public model APIs is in a different position. It may need to learn whether its documents produce usable predictions before deciding how to operate a cluster. A single-GPU experiment can answer some compatibility questions. It will provide very little evidence about a future multi-node network or a shared file system under load. Write down which uncertainties the small test can resolve and leave the rest in the specification for the larger evaluation.

03 / WorkflowA proposed trial that includes an interrupted job

Begin with one representative data shard and the production training image. Measure the time to copy or mount that shard, load the first batch and produce the first valid checkpoint. Preserve the data version, image digest and job configuration with the output. Then repeat the run using the intended number of workers for a short interval. Compare validation output before focusing on speed so that a changed numerical setting cannot quietly masquerade as a faster training path.

Now consider the difference between a healthy GPU and a useful cluster. One worker reading files slowly can keep others waiting at a synchronization point. A storage write that is harmless on a single node can take a significant share of the interval when every worker saves at once. Include a step in the trial that writes a normal checkpoint while other work is active. Record training progress during that period, the size of the saved artifact and the time until it can be read back.

The storage catalog distinguishes an on-demand WEKA NVMe parallel file system from distributed RDMA storage available by reservation for cluster-scale work. Specify the one used in each observation. A result achieved on a particular reserved storage configuration is poor evidence for what happens if another team books only the compute. Ask how data reaches the region before the allocation begins, how it leaves at the end and how long the system can retain the checkpoint between those dates.

Finally stop the proposed test at a planned boundary and restore from the saved state. Verify progress against the uninterrupted reference run and time the path back to useful training. This is a recovery exercise for a consenting team’s own job. It gives a purchasing lead a number that an isolated throughput chart cannot supply. If the recovery consumes most of the remaining window, change the checkpoint policy, the reservation length or the delivery promise before committing to a full run.

04 / PricingOrder capacity, storage and operating support on their actual terms

ComponentPublic availability descriptionInput to the quote
GPU instancesOn-demand or reserved configurationsGPU count, location and elapsed allocation time
Dedicated clusterBare metal by reservationNode topology, term and support coverage
Storage and schedulingWEKA on demand, distributed RDMA storage and managed Slurm by reservationVolume, retention and reservation overlap

Access and commercial routes consulted 19 November 2026. Firmus compute catalog, storage options and cloud services.

The pages above describe ways to obtain infrastructure and direct customers to a commercial inquiry. They do not supply an all-in dollar tariff for the proposed cluster. Request a quote for the exact test window and the full run. Price data staging, checkpoint retention, interconnect and operating support explicitly. Ask whether each charge follows time allocated, storage provisioned, traffic moved or another unit. Without these definitions a single blended GPU-hour figure can disguise a considerable storage commitment.

For an internal comparison, divide the quoted cost of the entire window by the number of accepted training runs it can deliver. Count time spent restoring a checkpoint or waiting for the data to arrive. Then change one assumption at a time, such as retaining the dataset for a second training cycle. This arithmetic can show why paying for persistent storage might reduce the cost of a series of jobs even when it makes the first job’s invoice larger.

05 / DistinctionsIts commercial scale and its cloud catalog answer different questions

In October 2026, Firmus announced an agreement with Thinking Machines Lab for infrastructure in Asia using GB300 systems. The announcement identifies the partner and the intended deployment. The general compute page still describes GB300 and Vera Rubin NVL72 capacity as coming soon. Both are relevant to a buyer. A negotiated anchor-customer agreement is evidence of a serious infrastructure relationship. A smaller buyer should obtain independent confirmation of what it can reserve, at which site and on which date.

WEKA’s storage Blueprint helps distinguish the underlying file-system technology from the particular service Firmus offers around it. CoreWeave’s Blueprint gives a comparison with another operated GPU cloud. Send competing operators the same description of the training job, especially its largest simultaneous read or write. Ask for the same restore exercise. Two offers with similar compute specifications may behave very differently at the point where training has to resume.

06 / QuestionsWho owns a failure that crosses the storage and compute boundary?

Agree in advance which evidence both sides can inspect. A training log might show a worker waiting without revealing whether another worker has failed, an input file is unavailable or a network event has interrupted a transfer. Name the support owner who can coordinate the investigation. Ask which telemetry remains available after the reservation expires, since a team analyzing an interrupted run may need the evidence after its compute allocation has ended.

Also confirm what happens if a workload finishes early. The economic value of a reservation changes if unused time cannot be released, transferred to another approved team or used for a separate experiment. In the other direction, clarify what a one-day overrun means if the next customer needs the cluster. These are scheduling and commercial questions that follow directly from buying a bounded training window.

07 / DecisionMatch the reservation to the bottleneck

01

A first compatibility test

Run one representative shard and verify model output, software versions and data loading before increasing the GPU count.

Use the smallest configuration that answers the compatibility question.
02

A repeatable multi-node training run

Bring the measured job and recovery profile. Reserve the required network, scheduler and storage for the same dates as compute.

Judge the quote by completed, recoverable work.
03

A future requirement for new hardware

State the model, location, availability window and acceptable substitutes in writing. Distinguish general catalog access from capacity allocated under a named partnership.

Commit against an agreed delivery plan.
What should we explore next?

A business worth understanding.

Suggest your business or one you find interesting. Tell us what you want to understand about its product, positioning, design or workflows.

Suggestions are free. Selection and publication stay with the desk.

Sources

Continue reading

All in this category