- Service code
- BH-RES-01
- Family
- Research
- Cadence
- Quarterly market refresh; monthly for actively negotiated commitments
- Response
- Commissioned quote scoped within 2 business days; delivered to agreed schedule
Overview
Published GPU prices are the least reliable number in an AI budget. Sticker rates diverge sharply from effective cost once committed-use terms, egress, storage tiering, interrupt rates and utilisation are accounted for — and the gap varies by an order of magnitude between providers for nominally identical hardware. This is the research programme behind our public GPU pricing section: it produces the sample quotes published on this site and the client quotes delivered behind an access code.
What is covered
- Rate card collection and normalisation across hyperscaler, neocloud and regional provider tiers
- On-demand, 1-year and 3-year committed, and spot/preemptible price tracking per accelerator
- Effective cost modelling: utilisation, allocation versus reservation, interrupt and checkpoint overhead
- Non-compute pass-through: egress asymmetry, storage tiers, interconnect premium, idle-node charges
- Price-per-token and price-per-trained-token normalisation across model and batch configurations
- Supply-demand signals: lead times, allocation scarcity, regional arbitrage
- Contract term review: commit structure, ramp flexibility, true-up and early-termination exposure
Deliverables
- Normalised rate card comparison across providers for the accelerators in scope
- Effective cost model with your own utilisation and workload assumptions applied
- Commitment strategy recommendation with modelled savings and downside exposure
- Price trend and volatility index with forward view
- Negotiation brief: the specific terms worth contesting and the evidence to contest them with
Who this is for
Finance and platform teams committing significant accelerator spend who need effective cost rather than sticker price.
Specification
- Accelerator coverage
- Contemporary data centre GPUs across all major vendor families
- Pricing dimensions
- On-demand, reserved 1/3-year, spot, committed-use, private offer
- Cost basis
- Per GPU-hour, per trained token, per million inference tokens, per rack-month
- Delivery
- Behind an access code at /gpu-pricing/access.php
- Public samples
- See the sample quotes in the GPU pricing section
Related services
| Service | Scope |
|---|---|
| Inference Economics Benchmarking | Cost per served token, measured on your model and your traffic — not on a vendor benchmark. |
| Hyperscaler Capacity & Supply Forecasting | Where accelerator capacity will be available, when, and on what terms. |