Low scenario / modelled
36%60% eligible × (1 - 1 / 2.5) = 36%For every $100 of long-context GPU-compute spend, $60 is eligible. Running that eligible share at 2.5x reduces the total modelled spend by $36.
Economic consequence
Modelled savings on the work WarpDrive can safely accelerate.
The model separates eligible long-context GPU-compute spend from everything that stays at baseline. It uses a conservative acceleration range inside the measured known-format result and shows the arithmetic in the same place as the claim.
01 / Reproducible model
No sliders, no hidden multiplier
36-56% modelled savings on eligible long-context GPU-compute spend.
The 2.5-5.0x scenario sits inside the measured 2.55-5.24x known-format range. Ineligible work remains at 1x and does not disappear from the model.
High scenario / modelled
56%70% eligible × (1 - 1 / 5.0) = 56%For every $100 of long-context GPU-compute spend, $70 is eligible. Running that eligible share at 5.0x reduces the total modelled spend by $56.
The simplest reading
Under these two scenarios, the same eligible work costs a modelled $64 or $44 per original $100 of long-context GPU-compute spend. This is not a forecast of total data-centre cost, revenue, customer demand or universal performance.
Public-pricing illustration
One million eight-H200 node-hours at public list pricing.
The example converts the percentage model into a large infrastructure number without treating CoreWeave as an ARES customer or partner.
$50.44M1,000,000 node-hours at $50.44 per eight-H200 node-hour
$18.16M36% modelled saving on the eligible GPU-compute portion
$28.25M56% modelled saving on the eligible GPU-compute portion
CoreWeave is a public-pricing example, not a customer or partner. Source: CoreWeave North America on-demand NVIDIA HGX H200, eight GPUs, retrieved 14 August 2026. Storage, networking and fixed charges are excluded. Pricing and availability can change. View CoreWeave pricing.
02 / Model discipline
What the number does and does not say
A workload model, not a data-centre forecast.
Eligible long-context GPU-compute
The portion of workload spend qualified for the declared acceleration scenario.
Ineligible work
Every workload outside the qualified boundary remains at 1x in the blended result.
Storage, networking and fixed charges
The model does not claim these costs fall when model work is reduced.
03 / Customer qualification
From illustration to contract
Real economics begin with a representative workload.
A commercial evaluation must measure the buyer's eligibility share, context distribution, receiver model, quality boundary, execution timing and actual infrastructure price before savings become a deployment claim.
Measure eligibility
Determine which workloads can enter the qualified acceleration boundary.
Measure the receiver
Keep context tier, passthrough treatment, accuracy and end-to-end timing attached.
Audit realised value
Convert a qualified technical result into buyer-specific cost or capacity economics.
Build the buyer-specific model.
Qualified investors and infrastructure operators can request the public Economics Brief or discuss a governed workload evaluation.