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.
Industries · Defence & National Security
Classification and isolation come before performance. No package repository, no licence server, no vendor telemetry, no remote hands. Every convenience modern infrastructure assumes has to be replaced by a procedure that a person can carry out, log and prove. That is the engineering problem.
The constraint we design around first
Commercial infrastructure assumes continuous connectivity for patching, licensing, identity, time, telemetry and support. Remove all six and most reference architectures stop working. We design for their absence from the first diagram rather than compensating later.
The failures are rarely dramatic. A cluster runs for eleven months and then a certificate expires with no path to renewal. A hypervisor needs a licence check-in. A vendor agent buffers telemetry until the disk fills. These are the faults that actually take isolated systems down.
We test that by building the enclave, severing the link and leaving it. A design that cannot survive a year of deliberate disconnection is not finished.
Air-gapped and cross-domain operation
An air gap that nobody can cross is easy and useless. A gap that everybody crosses informally is worse than none. The engineering sits in making the authorised path convenient enough that nobody invents an unauthorised one.
01
Unidirectional gateways enforce direction in hardware rather than configuration, so a misconfiguration cannot create a return path. The design consequence is that nothing can acknowledge receipt, so protocols that expect a handshake do not work. Transfers need forward error correction, sequence numbering and out-of-band reconciliation. We size the link for the loss the receiver can tolerate, not the bandwidth of the wire.
02
Where a bidirectional path is genuinely required, the accepted pattern is a protocol break with inspection and release control between the two sides. Low to high import is a technical problem. High to low export is a governance problem: someone must hold release authority, the decision must be recorded, and the format must be constrained enough to inspect. We keep export formats deliberately narrow, because a format nobody can parse is a format nobody can clear.
03
Removable media remains the most common transfer mechanism and the most common source of incidents. The controls are procedural and unglamorous: registered and serialised media, a standalone inspection station, scanning with signatures that themselves arrived by a reviewed path, two-person integrity for high-consequence transfers, and a custody log that reconciles. Write-once media removes an entire class of accidental back-flow.
04
Isolated estates fall behind on patching unless the pipeline is built for it. We stand up internal mirrors of operating-system repositories, container registries and language package indexes, populated from a curated snapshot. Signatures are verified against upstream keys before the bundle is admitted, a staging enclave takes the update first, and the software bill of materials is ingested so the enclave can answer what it contains without asking the internet.
05
Certificates, logs, distributed consensus and forensic reconstruction all depend on time. Inside an isolated facility a satellite reference may be unavailable, and in some environments it cannot be trusted. The answer is a local grandmaster with a high-stability oscillator and specified holdover, distributing by PTP with PPS to the systems that need tighter bounds. Drift is monitored and the holdover figure is stated in the design, not assumed.
06
Centralised monitoring services are unavailable, so the enclave keeps its own collection, retention and alerting. That makes retention sizing a real capacity question rather than someone else's problem. Log export for cross-domain review goes through the same release process as any other data, which means the log schema has to be designed for clearability.
Ruggedised and deployable compute
A rack that performs well in a data hall may be unusable on a vehicle. Power is finite and dirty, ambient temperature is whatever the day provides, and the chassis is being shaken. Useful work per watt matters more than peak throughput.
Backwards from the platform it rides on. The available generator or vehicle supply sets the power budget. The power budget sets the thermal load. The thermal load and the ambient temperature set whether the enclosure can be sealed. Only then does accelerator choice become a sensible conversation.
MIL-STD-810 is a set of tailored environmental test methods, not a pass mark. A supplier saying "810 compliant" has told you nothing. Ask which method, which procedure, which severity profile, and whether the unit was tested in the configuration you intend to field. The same applies to electromagnetic requirements and to vehicle and aircraft power standards: the tailoring is the information.
Deployed nodes cannot assume a link home. Models, indexes and reference data are staged locally before deployment. Work is queued for store-and-forward, and the node produces useful output while disconnected. Where a model is involved, we pre-compute and pin the artefacts rather than depending on a download that will not happen.
| Environment | Effect on compute | Design response |
|---|---|---|
| High ambient temperature | Sustained thermal throttling, silent loss of accelerator clocks, shortened component life | Derated power envelope, conduction cooling, thermal headroom verified at the stated worst-case ambient |
| Altitude | Reduced convective cooling and lower dielectric strength | Sealed liquid or conduction paths, fan curves recalculated, altitude-specific thermal testing |
| Shock and vibration | Connector back-out, solder-joint fatigue, module unseating, storage failure | Solid-state storage, retained modules, staked and torque-marked fasteners, isolated mounts |
| Sand, dust and salt fog | Filter blockage, abrasive wear, corrosion of contacts and heat sinks | Sealed enclosure with heat exchanger, conformal coating, no external airflow path |
| Humidity and condensation | Tracking across boards during rapid temperature change, corrosion on power-on | Controlled warm-up sequence, desiccant and breather design, coated assemblies |
| Dirty or transient power | Brown-out resets mid-write, supply damage, filesystem corruption | Input conditioning to the applicable vehicle or aircraft power standard, ride-through energy, journalled and power-loss-protected storage |
| Electromagnetic environment | Interference with co-located systems, susceptibility to nearby emitters | Emissions and susceptibility testing to the tailored requirement, filtering, bonding and cable discipline |
Sovereign supply chain and personnel
Two questions run through every defence engagement. Who touched this hardware between the factory and the rack, and who is allowed to touch it now. Both are answered with records, not assurances.
Control is a question about the contracting entity, the ultimate owner, and where the people doing the work sit. Local incorporation with offshore technical administration is not sovereign delivery. Export control adds a second layer: Australian defence trade controls and foreign regimes can both restrict who may receive technology, and the answer differs by item and by nationality of the person.
Evidence — ownership chain, staff locations, export-control assessment per item
Security clearances in Australia are sponsored and issued through the government vetting authority at levels from baseline upward. A clearance is necessary but not sufficient: access still requires a need to know, briefing into the relevant compartment, and a record of both. Roster design matters because a fault at 3 a.m. cannot be handed to whoever is awake.
Evidence — clearance level and currency, compartment briefings, access register
We buy through authorised distribution only, never the grey market, and record serial numbers at receipt against the purchase order. Platform integrity is checked before a unit is trusted: firmware measured at boot, hardware root of trust enabled, signed firmware update paths only, and any pre-installed management agent removed or disabled deliberately rather than by default.
Evidence — serialised receipt records, firmware measurement baselines, build manifests
Australia's Defence Industry Security Program sets membership levels across governance, personnel, physical and information security, and defence contracts commonly require membership at a level matched to the work. Membership is held by the entity, applies to nominated facilities, and has to be maintained. It is a prerequisite to be checked at qualification, not a marketing line.
Evidence — membership level, covered facilities, currency of the annual obligations
Facilities, emanations and end of life
Physical security zoning, emanation control and disposal are treated here at a descriptive level only. Implementation detail in these areas is controlled information and is not published. We discuss specifics with cleared personnel under the relevant agreement.
Australian protective security policy describes graded physical security zones, with construction, access control, audit and equipment requirements rising by zone. Sensitive compartmented work carries additional facility requirements, and approved products and accredited installers are involved. For infrastructure design the practical consequences are ordinary engineering: where cabling may run, which equipment may share a rack, how maintenance access is supervised, and how spares are stored.
Electronic equipment radiates, and unintended emanations can carry information. Mitigation involves separation of processing at different classifications, treatment of cabling and power, filtering, and in some cases shielded enclosures. Specific measures, thresholds and assessment methods are controlled and appear nowhere on this site. We plan for the space, power and cable-route implications early, because they are difficult to retrofit and they change the rack layout.
Degaussing does nothing useful to flash memory. Cryptographic erase depends on a key actually being destroyed and on the drive's firmware behaving as documented, which is a vendor claim rather than an observation. For media that has held classified information, sanitisation guidance generally points to physical destruction by approved means. We plan destruction into the purchase, including failed drives still under warranty, because a warranty return is a data-egress path.
Every drive and every device carrying non-volatile memory is serialised into an asset register on arrival, before it is racked. Disposal reconciliation is only possible if the register was right at the start.
Media inherits the highest classification it has processed, including caches, journals, hypervisor swap and management-controller storage that inventories often miss.
A failed drive stays inside the accredited boundary. No warranty return, no vendor diagnosis off site. Spares strategy is priced with that in mind, which raises the spares holding and lowers the surprise.
Destruction follows the applicable guidance for the classification and media type, using approved equipment or an approved provider, with two-person witnessing and a certificate that names serial numbers.
Certificates are matched back to the register, unaccounted items are treated as an incident rather than a discrepancy, and the closed register becomes part of the decommissioning record.
Allied interoperability
Australia's trilateral and allied arrangements, including work under AUKUS, involve collaboration on advanced capabilities. Discussed here at a general level only. The infrastructure difficulties are unglamorous and consistent.
None of them are compute problems. All four are usually discovered after the architecture has been agreed, which is the expensive moment to find them.
01
Partner nations use different markings, caveats and release conditions. Labels must be translated explicitly, and the translation has to be agreed by the release authorities rather than inferred by a system.
02
Access decisions depend on nationality, clearance, compartment briefing and programme role. Those attributes originate in different national systems, and the authoritative source for each has to be named in the design.
03
Whether a person may see a technical artefact is a legal question, not only a security one. Changes to export-control arrangements between partners alter what is permissible, so the control point must be configurable rather than hard-wired.
04
Collaboration does not suspend national requirements about where data may rest. Replication topology, backup targets and disaster-recovery sites each need a jurisdictional answer before the first transfer.
In isolated environments the interesting failure is never the compute. It is an expired certificate, a licence check-in, or a transfer nobody had authority to approve.
Straight answers
Where the honest answer is "not without verification", that is the answer given. It is a shorter conversation than the alternative.
That depends on facts about our entity, our facilities and our personnel that this page deliberately does not assert. Ask us directly and you will get a written answer covering membership level, cleared personnel available for the work, and facility arrangements, with reference details you can verify with the issuing authorities. Treat any vendor website that answers this question with a badge as unverified.
Yes, but not on its own. Isolation removes remote access as an attack path and leaves transfer discipline, supply chain and insider risk as the dominant concerns. That is why most of the work on an isolated build is procedural: media handling, custody, two-person integrity, release authority and reconciliation. Isolation without those controls produces a false sense of safety and a slower patch cycle.
Inference, generally yes. Weights are staged as versioned artefacts, served locally, and pinned so behaviour does not change without an authorised update. Training and fine-tuning are harder, because data curation, evaluation and the tooling ecosystem all assume connectivity. The realistic pattern is training in a connected enclave at a suitable classification, then a reviewed transfer of the resulting artefact. We will say plainly when a workflow does not fit the isolation you require.
By assuming no part leaves. Failed media is destroyed on site rather than returned, which changes the commercial arrangement and needs to be negotiated with the supplier before purchase. Spares holdings are larger, and the cost shows up in the business case instead of in an incident. Remote diagnosis is not available, so documentation and local fault-finding capability have to be part of the delivery.
Power and heat, always. A field generator and a sealed enclosure put a hard ceiling on sustained accelerator load, and that ceiling is well below a data hall. The productive response is to reduce the work: quantised and distilled models, smaller context, batching, and moving anything that can tolerate delay to a rear enclave. Anyone promising rack-equivalent performance on a vehicle is describing a burst, not a duty cycle.
Not in public, and not in a tender response beyond what the relevant authority has released. In a cleared setting we discuss capability, method and the people who would do the work. If a competitor's public material describes their defence deployments in detail, consider what that predicts about their handling of yours.
Let's Talk
We can talk capability, method and constraints in an unclassified setting before anything else is arranged. You will get a clear statement of what we can evidence today and what would need to be established first.
Sydney · Melbourne · Canberra. Australian owned, Australian staffed.