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 · Startups & Scale-ups
There are two expensive mistakes available to you. Buying hardware before you know your load, and renting a workload you have understood for two years because nobody did the arithmetic. Both cost roughly the same. We help you pick the cheaper one, with the model written down so the next round of the argument starts from numbers.
The constraint we design around first
Early on you pay a premium to stay flexible, and that premium is worth it because you do not yet know what you are building. The failure is not noticing when the uncertainty resolves and you are still paying for it.
Both are common enough to describe without naming anyone. They look different from the inside and identical on a cap table.
01
A seed-stage team buys a GPU server because the on-demand bill looked alarming. Six months later the model architecture has changed, the card is the wrong shape, and the capital is in a rack rather than in runway. The hardware was not the mistake. Committing capital against an unknown load was.
02
A platform assembled at speed is now carrying real revenue. Cost per customer is rising rather than falling, nobody can say which workload consumes what, and the architecture assumes a single managed service per function. The rebuild takes two quarters of engineering attention at exactly the point where that attention was promised to product.
03
Over-engineering for portability from day one. Two of every abstraction, a service mesh for four services, and a platform team of one who is now the only person who can deploy. The optionality is real and the cost is the product you did not ship.
The useful work is deciding which parts of your load are now predictable, and moving only those. Everything else stays rented, deliberately, with a review date.
Cloud versus owned
Comparisons that put an instance-hour rate beside a server quote are theatre. The honest model includes the people you would have to hire, the capital you would tie up, and the fact that half your load is genuinely unpredictable.
01
Owned hardware is priced per month regardless of use, so its economics are entirely a function of how busy it is. Below roughly half utilisation, owning is usually indefensible. Above sustained high utilisation on a workload whose shape is stable, renting is usually a premium you no longer need to pay. The interesting cases sit in between, which is why we insist on measurement instead of a spreadsheet estimate.
02
Egress charges are a small line until your product starts serving media, exports or model weights. More importantly, data accumulates somewhere and then pulls compute towards it. If your training corpus lives in one provider's object store and grows weekly, the cost of moving later rises every month you defer the decision. We quantify that as a number, not a warning.
03
Almost every startup has both. Inference serving a live product is steady and predictable within a band. Training runs, batch reprocessing and customer onboarding jobs are spiky by nature. The correct answer is nearly always a split: own or reserve the steady floor, rent the spikes. Sizing owned capacity for your peak is how facilities end up at 20% utilisation.
04
Venture equity is the most expensive money you will ever spend on a depreciating asset. Before recommending purchase we look at whether equipment finance, an operating lease or a vendor arrangement changes the answer, because converting capex to a monthly obligation often preserves the runway that matters more than the total.
05
Owning infrastructure means someone carries a pager, patches hypervisors, handles a failed drive at 2 a.m. and keeps the capacity model current. That is a real salary, plus a second one so the first can take leave. If you do not intend to hire it, the honest options are a managed arrangement where we carry it, or staying in cloud. Pretending it is free is the most common error in these models.
06
An instance is available in a minute. Hardware has a lead time, a rack, a power allocation and a commissioning window. If your growth curve means you need double the capacity inside a quarter, that lead time is a business risk and belongs in the comparison alongside the price.
Proof of concept to production
The point of staging is that each step is cheap to abandon. If a stage cannot be abandoned without losing the work, it was scoped wrong. Every transition has a written test, so nobody has to argue about whether you are ready.
On-demand everything, including the expensive accelerators. You are buying information about your own workload, not capacity. The exit test is narrow: can we state the resource profile of one unit of work, and does the approach do what the product needs. Do not optimise cost here. Optimise learning rate.
Production-shaped but small: a managed Kubernetes cluster or a handful of instances, one Postgres you did not build yourself, object storage, and OpenTelemetry instrumentation from the first commit. The exit test is a measured cost per customer or per thousand requests, with a stated growth assumption. If you cannot produce that number, you are not ready for the next stage regardless of how the product is performing.
Now the split becomes worth engineering. Reserved or committed capacity for the steady floor, autoscaling for the rest, and a clear separation between the workloads you understand and the ones you do not. This is also where availability targets stop being aspirational and get written down with the consequences of missing them.
Colocated or owned capacity for the workload whose shape has been stable for two or more quarters, sized to the floor rather than the peak, with burst retained in cloud. The test for entering this stage is that you can predict next quarter's base load within about 20% and you have the operational cover, ours or yours, to run it.
GPU access for AI startups
The interesting question is not which is cheapest. It is when each one stops being the right answer, because the signal that you have outgrown an arrangement usually arrives a quarter after the cost did.
| Mode | Suits | Relative cost per GPU hour | When it stops making sense |
|---|---|---|---|
| On-demand cloud | Experimentation, unknown architecture, short training runs, bursts | Highest | When the same job shape has run weekly for a quarter, or when you start queueing for availability in your region |
| Reserved or committed | Steady inference, recurring training cadence, predictable base load | Materially below on-demand for the committed portion | When the commitment term outlasts your confidence in the hardware generation, or when you are consistently exceeding the reservation |
| Colocated, hardware owned or financed | Sustained high utilisation, data residency requirements, latency to your own data | Lowest per hour once utilisation is high, plus power, space and cross-connect | When utilisation falls below the level that justified it, or when a new accelerator generation changes your cost per token faster than you can depreciate |
| Owned, on your premises | Development and lookdev-style interactive work, sensitive data, air-gapped requirements | Low per hour, high in facilities and attention | Almost immediately, at density. Office power and cooling become the binding constraint well before the compute does |
Accelerator generations arrive faster than a comfortable depreciation schedule. Buying a card outright is a bet that your cost per unit of work will still be competitive in three years against hardware that does not exist yet. Sometimes that bet is fine, particularly where data cannot leave the country or the building. It should be made explicitly rather than inherited from a spreadsheet that assumed a five-year life.
Before adding accelerators we look at what you are getting from the ones you have. Continuous batching and paged attention in a server such as vLLM, quantisation where evaluation shows quality holds, KV-cache reuse, and routing easy requests to a smaller model. On inference workloads these changes commonly move cost per million tokens further than a hardware upgrade would, and they cost engineering weeks rather than capital.
Portability and instrumentation
Full portability is an expensive insurance policy against an event that may never occur. The practical target is that leaving would be a project, not a rewrite, and that you know today what that project would cost.
Managed services are usually worth the lock-in when they replace work you would otherwise have to staff. The place to hold the line is anywhere that would force a rewrite of your own code rather than a change of configuration.
Every decision on this page depends on measurements that are cheap to collect early and awkward to reconstruct later. Six things, set up once.
Diligence and enterprise readiness
Investors ask about unit economics and concentration risk. Enterprise buyers ask about data handling, access control and what happens when you are acquired. Both conversations go faster when the answers already exist as documents.
The Essential Eight is a sensible frame for an Australian company and is cheap to adopt early: patching, multi-factor authentication, application control, admin privilege restriction and tested backups. Full ISO/IEC 27001 or SOC 2 certification is a deliberate commercial decision, usually driven by a specific deal. We will tell you when you do not need it yet.
Where customer data lives, which processors touch it, how long it is retained, how a deletion request is executed, and what crosses a border. Australian Privacy Principles apply to most of this. Government and health buyers will ask about residency specifically, and an answer that consists of a shrug ends the conversation.
Gross margin per customer with infrastructure attributed honestly, including the free tier, the support overhead and the workloads that run whether or not anyone is using the product. Investors discount numbers they cannot trace. A cost model with visible assumptions is worth more than a flattering one.
A diligence process will find the service only one engineer understands, the hand-configured host that is not in code, and the credential in someone's personal vault. These are findings that cost negotiating position. They are also straightforward to clear if you start before the term sheet.
When not to engage us
We would rather lose the engagement than take money for infrastructure a company does not need yet. It is also a better commercial position: the same founders come back when the load is real.
01
If the shape of the product is still moving, infrastructure spend is a distraction and hardware is a liability. Use on-demand, accept the premium, and put the capital into finding out what you are building. Come back when a workload has repeated often enough to have a shape.
02
Below a few thousand dollars a month, the achievable saving does not cover the cost of the analysis, let alone a migration. Right-size what you have, take the obvious commitment discount, and spend the attention on your product. We will say this on the first call rather than the third.
03
Owned infrastructure needs a person, or a managed arrangement where that person is ours. A team with no platform capacity and no budget for operational cover should not take on hardware. That is not a sales objection to overcome; it is a prediction about an outage in eight months.
04
Sometimes the bill is high because of an N+1 query, an unbounded retry, a nightly full table scan or a model that is twice the size the task requires. Buying capacity to cover that is the most expensive available fix. We will point at it, and you may not need us further.
Let's Talk
Three months of billing detail, a utilisation export and an honest forecast are enough to find the crossover point for each of your workloads. You keep the model. If it says change nothing, that is a legitimate result and we will tell you plainly.
Sydney · Melbourne · Canberra. Australian owned, Australian staffed.