Compute Procurement Advisory

Compute procurement, structured to your requirement

We help inference platforms, AI labs, physical AI, and agentic platforms access the compute market efficiently — what capacity is sold in, the contract forms on offer, tenor, and cost.

We support your capacity planning, and structure commitments that map tightly to what the workload actually needs.

The four stages

STAGE 1

Sizing your requirement

The technical inputs are yours; we don't second-guess them. We structure them into the seven dimensions that decide what you can secure: SLOs, workload profile, traffic profile and mix, preemption tolerance, growth posture, placement and timing.

STAGE 2

Reading the market

What supply actually looks like right now: procurement increments, the qualified-host constraint, contract forms, tenor economics, generational cadence. Read monthly; priced for clients.

STAGE 3

Structuring the commitment

We structure the commitment against that market: which ladder, which denomination, what tenor, who carries the idle — and what's genuinely negotiable at your commitment size.

STAGE 4

Optimizing the cost structure

We turn the structure into numbers boards, lenders, and diligence will underwrite: margin drivers, utilization, commitment coverage, depreciation and residuals.

Start with your workload

Inference & agentic workloads

For teams running inference, RAG, chat, or agentic tool-calling systems.

See inference & serving procurement →

Training & post-training workloads

For teams fine-tuning, post-training, running RL, or training simulation and robotics policies.

See training & post-training procurement →

Market pulse

The current read on the supply side — refreshed monthly.

Market pulseAugust 2026
  • Vera Rubin NVL72 entered full production in June and ships this fall to eight named cloud partners; broad neocloud availability is tracking Q4 2026–Q1 2027, with custom rack configurations quoted at 40+ week lead times.
  • A scaled inference buyer just chose node-scale: the IBM–Together AI multi-year agreement (reported at ~$240M) deploys HGX B300 servers, not NVL72 racks — capacity landing Q1 2027.
  • Analysts expect H100 secondary values to pull back 10–20% as Rubin availability normalizes — residual-value assumptions across mixed fleets are live, not settled.
  • GPU-backed financing has arrived: the first senior secured debt facility collateralized by deployed GPUs and contracted customer cash flows closed this summer ($775M, SOFR+250).
  • PTU-style provisioned throughput has spread from hyperscaler platforms into the neocloud tier — token-denominated capacity is now a real rung, not a platform quirk.
  • Sub-six-month reserved terms on premium GPUs remain broadly unavailable; discounts concentrate at 1–2 year tenor, with prepayment norms beyond a year.
Directional market observations from public reporting, refreshed monthly. Client engagements work from the priced version.

Let's talk.

A 15-minute conversation about your company's situation and potential needs. No pitch deck, no sales process.

Book a call