Layer 2.5 · head to head
CoreWeave vs AWS (P5, P6, G6)
The 30-40% discount that is the neocloud pitch, against counterparty certainty.
| CoreWeave | AWS (P5, P6, G6) | |
|---|---|---|
| Can you leave | Portable. The workload is standard enough to move to another vendor without a rewrite. | Movable with effort. Expect to redo images, storage wiring and networking. |
| Counterparty | Publicly reporting, which is the point — you can read the filings. Carries substantial GPU-collateralised debt with maturities in the 2026-2028 window, and revenue is heavily concentrated in a few customers. Concentration cuts both ways: it funds the buildout and it is the risk. | Effectively none. This is the reason the premium exists. |
| What it is | Purpose-built GPU cloud. Owns or leases its own datacentre capacity and sells it directly. | A general cloud that also rents accelerators. Most expensive per hour, and the one your compliance team has already approved. |
| How you reach it | You get a cluster. Assumes you already run Kubernetes at scale. | You get virtual machines. The familiar cloud model. |
| Accelerators | H100, H200, GB200 NVL72, B200, A100, L40S | H100 (P5), B200 (P6), L4/L40S (G6), plus Trainium and Inferentia |
| Regions | US, EU, UK | Global |
| Pricing model | On-demand and reserved. Reserved is the real product; on-demand list price is high enough that it reads as a discouragement. | On-demand, spot, Savings Plans, and Capacity Blocks for reserved GPU windows. |
| Getting started | No formal minimum on-demand, but the commercial motion is contracts, not credit cards. | Existing AWS account. |
| Capacity | Large and contracted years ahead. Most capacity is spoken for by a small number of very large customers. | Constrained on the newest parts; Capacity Blocks exist precisely because on-demand cannot be relied on. |
| Ownership | Public (IPO completed). NVIDIA is an investor. | Amazon. |
CoreWeave
For: Teams that already run Kubernetes at scale and are buying committed capacity rather than experimenting.
The catch: List on-demand pricing is among the highest of the neoclouds — widely reported around $6/hr for an H100 where specialists sit near $2. You are not meant to pay list; if you are paying list, you are using it wrong.
Economics: Priced 30-40% under the hyperscalers on comparable committed terms, which is the entire pitch. Against the cheaper specialists it is not a price play at all — it is a scale, network and support play.
AWS (P5, P6, G6)
For: Anyone whose data, VPC, compliance boundary and team already live in AWS. The GPU price is rarely the deciding number.
The catch: Reported around $9.36/GPU-hour for a B200 Capacity Block against roughly $5.50 at Lambda. You are buying integration and counterparty certainty, and paying for both.
Economics: Egress and adjacency costs usually dominate the GPU line. Compare total workload cost, never the hourly rate alone.
Neither table row is a price
Deliberately. Published on-demand rates at this layer move weekly, and essentially nobody signing a real contract pays them — every serious buyer pays less than every list figure either of these companies publishes. Quoting one here would date this page within a month.
The GPU rental price index carries dated, sourced figures instead, and the durable finding there is the spread: the identical H100 rents from roughly $1.38 to $12.29 an hour depending only on who you rent it from.
The layers underneath both
Whichever you pick is renting you chips in a building that needs power. In 2026 that is the constraint that binds: Microsoft has disclosed an Azure backlog it cannot fill for want of megawatts rather than accelerators, and the US interconnection queue exceeds 2,600 GW with roughly 80% of projects withdrawing before they energise.
Layer 2 — Silicon · Layer 1 — Energy · The interconnection queue