GPU clusters and GPUaaS allocation. For the party running the workload.

No. This desk is open to the end user — the party whose workload runs on the hardware and whose balance sheet stands behind the contract — or an advisor that party names in writing.

We do not work with brokers assembling a position to resell, and we do not join chains where the same allocation has been shopped through four parties before it reaches us. The principled reason is that it is not what we do. The practical reason is that operators will not quote into a chain, so nobody in it ever sees a real number.

If you are an intermediary with a genuine named end user behind you, that is a different conversation and we will have it.

No. We are not a supplier — we do not buy, sell or hold anything, so there is no rate card of ours to send.

Any rate card you have been sent from elsewhere is close to worthless anyway. Price on a cluster moves with scale, term, prepayment, power cost at the site, cooling type, network fabric and how badly the operator needs an anchor that quarter. The only number that matters is the one a partner will sign against your requirement.

We could not send one if we wanted to. We hold no inventory and set no prices — every number comes from the operator or the channel partner, direct to you. What we do is get your requirement to the partners whose position fits it, tell you where their response sits against comparable deals we have seen, and negotiate the terms around it.

512 GPUs, and only with a credible path to production scale behind it. Below that the economics do not support advisory involvement on either side, and you are genuinely better off on public cloud on-demand until the workload is proven.

Above roughly 50 MW the binding constraint stops being GPUs and becomes energised power, delivery schedule, and cooling capable of rack densities over 120 kW.

We work that layer directly — powered shell and build-to-suit, interconnection position, PPAs and behind-the-meter options, liquid cooling retrofit against live halls — and assemble compute, capacity and power into one transaction instead of three separate searches on three timelines. That is the specific case for using us at this scale.

It turns on utilisation and refresh exposure.

Sustained utilisation above roughly seventy per cent on a workload you can forecast three years out favours owning — if you can secure the power and the hall, which is the part most buyers underestimate. Spiky or uncertain demand, or a workload that will migrate to the next platform inside eighteen months, favours a reserved GPUaaS commitment where refresh risk sits with the operator.

We model both against your actual load profile before recommending either, and we are paid the same way regardless of which you pick.

On GPUaaS and reserved clusters: first response within one business day, and 12 to 14 weeks from executed contract to service across B200, B300 and GB300 NVL72. Where the site and power position are already in hand that holds firm; where a hall still needs building, siting and energisation set the schedule rather than silicon. Hardware allocation moves on its own clock — it depends entirely on what has surfaced, so we tell you what is reachable rather than quoting a standard window.

Nothing directly on a buy-side mandate. Winning business carries a cost of sale that exists whether or not we are involved, and that is where we are paid from — so our involvement is funded by the transaction rather than added to it. We hold no inventory and take no position, so nothing we bring forward is something we need you to take.