Compute mandate · CAP-02
GPU allocation and the contracts around it.
We hold no inventory and we are not the seller. What we provide is access — to allocation that never reaches an open market, and to terms a buyer negotiating one of these contracts every few years would not otherwise get.
End users only. This mandate is worked for the party whose workload runs on the hardware, or an advisor they name in writing — not brokers or resellers. Live platforms, scale bands and access policy sit on the GPU clusters & GPUaaS desk.
What a compute mandate covers
- —Configuration sanity check
Whether the cluster you have specified matches the workload you actually have. Over-specified fabric and under-specified storage are both expensive, and both are common.
- —Supply route selection
OEM direct, channel, systems integrator or secondary market — each has a different price, a different lead time and a different warranty position. The cheapest route is rarely the cheapest outcome.
- —Agreement negotiation
Allocation, delivery, remedies, spares, support terms and the software licensing attached to them. This is where the value sits.
- —Landing the hardware
The facility, power and cooling that has to receive it, sourced in parallel rather than after the fact. A cluster sitting in a warehouse is not an asset.
Where allocation actually comes from
Plans change constantly at this scale. When they do, hardware has to find a home quickly and quietly — and none of it appears on an open market.
- Cancelled and restructured orders. A committed order that no longer fits a changed roadmap, a financing round that moved, or a site that will not energise in time. The holder needs a buyer fast and does not want the market to know why.
- Incoming stock without a home. Inventory arriving against a commitment that has since shifted. It is landing regardless, and it needs a discreet buyer rather than a public listing.
- Partner positions across regions. Partners in every market we cover hold allocation that is never advertised. Reaching it is a matter of already being in those relationships, not of searching harder.
- Redeployment before end of life. Clusters coming off one workload with real life left in them, where the economics can be substantially better than new for the right requirement.
We hold none of this, and we are not the seller in any of it. We have no reason to steer you toward one route over another because we do not sell any of them. What we provide is the access and the negotiation — the introduction to the party who holds it, and the terms once you are talking.
The allocation that matters is rarely listed. It moves through people who already trust each other.
A snapshot of what is currently live is on the availability board — B300, HGX B300 and GB300 across the USA and Iceland, ready for service from Q4 2026.
Two honest caveats. Reallocated and cancelled-order hardware carries different warranty and support positions than a fresh OEM order, and that difference has to be priced rather than ignored — we work it into the terms. And where allocation genuinely is not available on your timeline, we will tell you that instead of keeping you waiting on it.
The terms that decide what you actually paid
Unit price is the number everyone negotiates and the smallest lever in the contract. In a market where lead times move and allocation is discretionary, the terms below decide your real cost.
- Allocation letter versus firm order. An indicative allocation is not a delivery commitment. If your business case depends on GPUs being live in a given quarter, the contract has to say so, with a consequence attached.
- Slip remedies. What happens when delivery moves. Price protection, a partial-delivery schedule, an escape clause, or nothing at all — the default is usually nothing at all.
- Price protection on a falling market. On a multi-quarter delivery schedule, a portion of your order lands into a different price environment. Whether you benefit from that is negotiable at signature and impossible to fix afterwards.
- Spares and RMA pool. At cluster scale, failures are a certainty rather than a risk. Who holds spares, how quickly they ship, and whether the pool is on your site or the vendor's determines your real availability.
- Fabric and networking scope. InfiniBand versus RoCE Ethernet is a significant cost and performance decision, and switches, cables and optics are frequently a larger line than buyers expect. What is included and what is quoted separately needs to be explicit.
- Software licensing. Enterprise AI software stacks, cluster management and support subscriptions are recurring and often bundled at list. They are negotiable, and they are rarely negotiated.
- Power and cooling assumptions. A rack spec that assumes direct-to-chip liquid cooling is not deployable in an air-cooled hall. If the facility cannot carry the density, the hardware discount is irrelevant.
- Residual and refresh. Trade-in credit, buy-back options and depreciation treatment change the total cost of ownership materially, particularly for buyers on a two-to-three-year refresh cycle.
- Buy versus lease versus reserved cloud. Sometimes owning is wrong. We will say so, because we have no position in the outcome either way.
Buy, lease, or rent — decided honestly
A vendor advising you on whether to buy has a preferred answer. We do not.
Most buyers at scale end up with a blend. Getting the split right is worth more than getting the unit price right on any one of them.
Asked on this mandate.
01Can you get GPU allocation we cannot get ourselves?
Often, yes. Plans change constantly at this scale, and when they do the hardware has to find a home quickly and quietly. We are pulled into cancelled and restructured orders, into incoming stock arriving against a commitment that has moved, and into partner positions across every region we cover. None of that reaches an open market, because the holder wants discretion rather than a public listing.
We hold none of it ourselves and we are not the seller. What we provide is the introduction and the negotiation. And where allocation genuinely is not available on your timeline, we will say so rather than keep you waiting on it.
02Which platforms do you work across?
H100 and H200 for current-generation training and inference. B200 and B300 where availability is live now. GB300 NVL72 at rack scale, including positions in the thousands of GPUs. And earlier generations where the reallocation economics are compelling. We are not tied to a vendor or a generation, and we have no reason to prefer one — we do not sell any of them.
03Are you the seller?
No. We hold no inventory and we do not sit in the transaction. You purchase directly from the OEM, channel or integrator, and we are paid by the ecosystem rather than out of your purchase order.
04We are not sure whether to buy or rent. Can you help before we decide?
That is the best moment to involve us. Before a route is chosen there is room to model all three honestly. After a purchase order is issued, the decision is made and we are only optimising the edges.
05Can you also source the facility for the cluster?
Yes, and we usually run them together. See the capacity mandate and power procurement — hardware without an energised home is a stranded asset, and the two timelines rarely align on their own.
GPUaaS & AI cloud negotiation
Reserved capacity and committed-spend agreements, read line by line before a multi-year number is signed.
Open the mandate →CapacityData center capacity sourcing
Retail and wholesale colocation up to 150 kW per rack, sourced globally and benchmarked against real comparables.
Open the mandate →EnergyPower capacity procurement
Interconnection position, PPAs and behind-the-meter options. Power is the binding constraint.
Open the mandate →Send us the cluster spec and the date you need it live.
We will tell you what the terms should look like, which route makes sense, and where the number in front of you sits against the market.