Reserved capacity for training, fine-tuning and inference, sized to the workload, termed to your runway, and structured so it does not eat the capital you raised to build the model.
Your last cloud invoice made someone senior ask a question you did not have a good answer to.
You are queueing for on-demand capacity that is not there when the run is ready to go.
You have been offered a three-year commitment against eighteen months of runway.
Your training job and your inference fleet want different machines, and you are paying for one to do both badly.
Capex would wreck you, but the board does not want a depreciating asset on the balance sheet either.
Someone has quoted you the flagship generation for a workload that does not need it.
What we do
Sized to the workload. Termed to your runway.
Within one hourAcknowledged, every enquiry, every time
You contract directWith the supplier, never through us
No cost to buyersCompensation comes from the supply side, disclosed
Global coverageSourced wherever the position actually fits
Sizing
We size to the workload, not the invoice.
Nodes, not GPUs. Fabric matched to the job, since distributed training and production inference want different topologies. Prior-generation where it does the work.
Runway
We term it to your runway.
A commitment that outruns your funded horizon is a liability dressed as a saving. We recommend the term that survives your next raise, even when it is the smaller deal.
Capital
We keep the capital free.
Reserved and financed structures mean no hardware capex, no build-out, no depreciation risk. Time-to-compute in weeks rather than quarters.
Internal case
We give you the internal ammunition.
Reserved versus on-demand, buy versus build, in a form you can hand to a CFO without translating it first.
What we won’t do
The limits, up front.
We won’t put you in a term your runway can’t carry. We would rather do the smaller deal and still be your channel in two years.
We won’t oversell the silicon. Flagship parts for an inference fleet is a bad deal that photographs well.
We won’t promise a date we have not verified. Confirmed and indicative are different words and we use the right one.
We won’t quote you. No inventory, no position, nothing between you and the operator.
What else we handle
Somewhere to run itColocation, power and cooling if this needs a physical home.Space & power
Serving itOrchestration and multi-tenancy if you are offering capacity to your own customers.
Connecting itPrivate low-latency links when training data and compute live in different places.
Everything around the modelNetworking, storage, endpoints for the team.Hardware & capacity
What a deal looks like
From cloud bill to reserved cluster.
The conversation usually starts with a bill. A training team on on-demand cloud pricing crosses the line where reserved bare metal is cheaper somewhere around sustained utilisation of a few months, and by the time the bill makes the point, the team has already lost a quarter to queueing and preemption. The requirement is rarely exotic: a defined node count, InfiniBand or equivalent fabric, storage that can feed it, and a term that matches the funding runway rather than the vendor's preference.
We size from the workload backwards. Checkpoint cadence, dataset size, expected run lengths and the split between training and inference decide the node count and the fabric, not the biggest SKU on the price list. If your workload runs fine on H100 at a materially better rate than B300, we say so, because the exit matters as much as the entry: a right-sized cluster is one you can walk away from cleanly when the generation turns.
Financing routes exist for teams whose runway is venture-funded rather than revenue-funded. Structured correctly, a reserved cluster can sit off the burn line in ways an on-demand bill never can.
Straight answers
The questions this desk gets first.
When does reserved bare metal beat on-demand cloud?
As a rule of thumb: when utilisation is sustained rather than spiky, and the horizon is months rather than weeks. Spiky experimentation belongs on-demand. A training roadmap you can see two quarters of belongs on reserved capacity at a fraction of the hourly rate.
How fast can we be training?
Verified stock moves in days to weeks, not quarters. As a current reference point, current verified lines and their lead times sit on the live inventory; a racked cluster in a named facility typically lands in the eight-to-ten week range depending on the site.
What term should we sign?
The term the workload justifies and not a month more. Multi-year take-or-pay prices well but pushes generation risk onto you. We pressure-test the trade between rate and term against your actual roadmap before you sign, and we tell you where the real floor is.
Do you handle inference as well as training?
Yes, and they are sized differently. Inference wants latency, geography and right-sized cards; training wants fabric and density. Most teams need a mix, and buying the mix as one requirement is usually cheaper than buying the parts.
How we get paid
You pay the supplier direct.
Operators compensate us under referral agreements for qualified volume. You pay nothing incremental, the number you would get direct is the number you get, and that is exactly why we can tell you to buy less than you were about to.
Talk to the procurement desk
Here’s exactly what happens.
01Twenty minutes.
What you are training, how big, how long, and what your funded horizon looks like.
02We come back by a date.
With the two or three operators who fit, and the term structure we would actually sign.
