Seven practices, one engineering team. We design, build, benchmark and operate the compute layer beneath your hardest problems — on-premise, hybrid or sovereign cloud.
A genomics pipeline and a trading engine are not the same machine. We start from your regulatory perimeter, your data gravity and your deadlines — then design backwards.
Insights & Case Studies
Two kinds of writing live here. Deployment stories describe systems we built, including what we got wrong on the first attempt. Technical notes and position papers argue a method, publish the measurement harness, and name the conditions where the argument fails. Nothing here is written by a marketing team.
Updated when something more useful is finished
A federal agency arrived with a model, a deadline and a shortlist of accelerators. The shortlist did not survive the first site survey. The hall could commit 780 kW, and the rear-door heat exchangers topped out at 32 kW per rack. The switchboard upgrade sat eleven months behind the project deadline.
The piece walks the decision in order. First the load study, and why we sized for sustained draw rather than nameplate. Then the two accelerators that fit the thermal budget, and the scheduling policy that stops a burn-in tripping a breaker. It closes on the interconnect compromise, and the throughput we accepted losing to hold the delivery date.
The library
Deployment stories are systems we ran. Analysis covers technical notes and position papers — arguments about method, with the harness attached. Topic filters follow the constraint, not the product name.
Two findings in the pre-assessment moved the jump hosts and split the audit sink away from the inference path. We publish both topologies and the reasoning that forced each change.
Signed manifests, a staging enclave and a rollback that works with no registry to reach. A year of operating the path produced three rules we now write into every isolated build.
A whole-genome workflow spent more wall-clock time waiting on file locks than on alignment. The teardown, the fix, and why we now profile the orchestrator before the compute.
Shares, decay half-life and QoS tiers decide who computes during a grant deadline. A worked configuration for a four-institution cluster, with the argument each parameter settles.
Grandmaster failure is an operational event, not a thought experiment, and timestamp obligations do not pause for weather. Measured drift across three oscillator classes, and the holdover budget we design to.
Queue analysis put the bottleneck in texture read amplification, not GPU count. Moving the working set onto NVMe-oF recovered most of the idle and cancelled a planned hardware purchase.
Batch shape, context length and time-of-day demand set the achievable ceiling long before the scheduler does. A utilisation model for vLLM estates that survives a real request mix.
A staged plan kept a Series A team on bursted capacity until the utilisation curve justified owning hardware. The crossover calculation, and the two signals that said stop renting.
Metadata operation mix, file size distribution and tolerance for a rebuild decide the filesystem. A decision procedure that reaches a defensible answer in an afternoon, including when both are wrong.
Fixity checking, format migration and a cost model a board could approve for twenty-five years. What we changed after the first annual audit found bit rot on one tier.
An honest reading of where the hardware sits, and why the migration register for classical cryptography is the only part of this with a real clock. Includes the Qiskit runs that lost to a good classical solver.
Cross-connect sequencing, a duplicate timing tree and a rollback that stayed live until Monday's open cleared. The runbook, the two rehearsals, and the step we had sequenced wrongly.
Burned-in pixel text, private tags and acquisition metadata all survive a standard scrub. What we test for, and the pipeline stage where the test belongs.
The OpenFOAM cases were small and the iteration loop was human-paced, so queue wait dominated runtime. Local hardware removed the wait for less than a year of cluster allocation.
Nothing published under that filter yet.
Editorial standard
Infrastructure writing is usually a sales asset with footnotes. These are the constraints we publish under, so you can tell the difference from the outside.
01
Every piece is drafted by someone who held the pager, ran the benchmark or attended the change window. A marketing editor may cut length and fix grammar. Nobody outside the practice adds a claim. The byline is a real person you can question in a tender interview, which is the point of putting it there.
02
If we print a latency, a throughput or a utilisation figure, the harness, configuration and raw run output exist and are available on request. Ask and you receive the actual files, not a summary slide. Where a figure comes from a client estate we cannot share, we say so beside the number rather than printing it anyway.
03
No manufacturer pays for placement here, reviews a draft, or sees a piece before publication. We hold partnerships with hardware and platform vendors and we name them on the relevant pages. None of those agreements includes editorial rights. A piece that reads like a datasheet has failed review, whoever wrote it.
04
A deployment story is published after the client signs off the wording, the figures and the level of attribution. Many run de-identified, because in a market this size a sector plus a scale plus a date identifies an organisation. Consent can be withdrawn, and a withdrawn story comes down without argument.
05
When a figure turns out to be wrong, the piece carries a dated correction note explaining what changed and why. Silent edits make an archive useless for anyone trying to check our reasoning. If a whole argument is superseded by better measurement, the old version stays up with a pointer to the new one.
Publication path
Most engagements never produce an article, and that is the expected outcome. The gate that stops the most drafts is the client's legal review, and it should.
If the only interesting thing about a project is that we won it, there is no article in it.
Before anything is drafted, there is a before-and-after measurement taken on the client's workload with a harness we can hand over. No measurement, no story.
Written by whoever did the work, including the wrong turns. A draft that hides the first failed approach is sent back, because the failed approach is usually the useful part.
A second engineer who was not on the project checks the reasoning and tries to reproduce the headline figure from the attached harness. Disagreements get noted in the piece.
The client marks up anything commercially or operationally sensitive. Topology detail, vendor names and absolute volumes are the usual casualties, and we rewrite around them.
Consent covers the final wording, every figure and the attribution level — named, sector-only, or fully anonymous. It is filed, and it is revocable at any time.
The piece goes up dated, bylined and versioned. Later corrections are appended in place so the original reasoning stays legible to anyone auditing it.
Reading the figures
Benchmark figures in vendor writing are usually true and useless, because the conditions are missing. This is the disclosure we hold ourselves to for each class of claim.
| Claim class | Always published with it | Usually withheld | How you verify it |
|---|---|---|---|
| Latency | Percentile, sample count, measurement point, clock source and whether the path was loaded | Client venue, cross-connect detail and strategy-identifying traffic shape | Request the capture method and the raw histogram. A median with no p99.9 beside it is not a claim we make |
| Throughput | Block size, access pattern, client count, filesystem and whether caches were dropped | Absolute dataset volumes where they identify the organisation | Ask for the fio or IOR job file. It runs on your hardware unchanged |
| Wall-clock improvement | The baseline configuration, the workload version and what else changed in the same window | Scientific content of the workload, and any unpublished research data | Ask which single change carried most of the gain. We will name it, or say we cannot separate them |
| Utilisation and efficiency | Measurement window, whether idle and drain time count, and the scheduler policy in force | Per-project allocation splits inside a shared facility | Ask for the accounting export definition. Utilisation is arithmetic on a definition, so the definition is the claim |
| Cost | What is inside the boundary — power, cooling, licences, staff time — and the amortisation period | Commercial rates and anything covered by a client's pricing confidentiality | Ask for the model, not the total. A cost figure with no boundary is a number without units |
| Availability | The measurement period, what counts as an outage and whether maintenance windows are excluded | Incident detail that would identify a client or a security posture | Ask for the SLA definition used. Excluded maintenance is where most published uptime figures live |
One request covers everything in a piece. Send the article title and the figure you care about, and the practice that wrote it answers directly.
Questions
If the answer you need is not here, ask. We would rather answer a hard question in writing than have a figure taken on trust.
Yes, for any figure we measured ourselves. You receive the harness, the configuration files and the raw output, so the run is reproducible on your hardware. Where the measurement was taken on a client estate, we can share the method and the shape of the result but not the client's data. We say which case applies beside the number rather than leaving you to guess.
No. No vendor pays for coverage, supplies copy, or reviews a draft before it is published. We do hold reseller and partner relationships with hardware and platform manufacturers, and those are disclosed on the relevant solution pages. None of them carries editorial rights. If a piece names a product favourably it is because the measurement said so, and the measurement is attached.
Three, in descending order of detail. A sector-only story naming no organisation. A technical note that carries the method with no engagement context at all. Or nothing published, with the write-up kept internal and shared only under NDA when a prospective client asks about comparable work. The third option is common and costs you nothing. We will never publish a de-identified story you have not read and approved, because de-identification frequently fails in a market this size.
Cite it freely with a link and the publication date. Ask first before reprinting a substantial extract. That is not about controlling the message. A figure quoted without its conditions becomes a claim we cannot stand behind. If you are citing a number in a tender response, request the method at the same time. It then sits in your evidence pack rather than in a browser tab.
Roughly two pieces a month, and the rate is set by how often a measurement worth arguing about turns up. The people who can write these are the people delivering the work, so a publishing schedule would either steal delivery time or produce filler. We would rather the archive stayed small enough that every piece in it is worth your time.
Both are welcome, and the second is more useful. If you think an argument here is wrong, send the measurement that shows it. A correction with your reasoning attached goes onto the piece, credited if you want the credit. Topic suggestions go into the queue and get written when a real engagement makes the answer testable rather than theoretical.
Three directions from here
Interrogate it
That request costs you nothing and commits you to nothing. It is also the fastest way to work out whether we measure the way your engineers would.
Every piece here was written in Australia by the team that built the system.