CoreWeave pricing teardown
https://www.coreweave.com/pricingCoreWeave publishes a rare, highly transparent price catalog for infrastructure: per-GPU-model and per-CPU-model hourly rates with on-demand and spot columns, per-region tables, free egress, and flat storage/network rates. It is a pricing reference sheet rather than a conversion-optimized page — no plan packaging, no recommended option, and the highest-end SKUs (GB300, B300 on-demand) default to 'Contact sales'.
Tier structure
NVIDIA GB200 NVL72
$42.00/hr on-demand
NVIDIA HGX B200
$68.80/hr on-demand ($34.11 spot)
NVIDIA HGX H200
$50.44/hr on-demand ($20.93 spot)
NVIDIA HGX H100
$49.24/hr on-demand ($19.71 spot)
NVIDIA A100
$21.60/hr on-demand ($9.65 spot)
NVIDIA L40S
$18.00/hr on-demand ($7.88 spot)
NVIDIA GH200
$6.50/hr on-demand
NVIDIA GB300 NVL72 / HGX B300
Contact sales
AMD Genoa (9454)
$7.78/hr on-demand ($3.32 spot)
AI Object Storage
$0.06 / $0.03 / $0.015 per GB/mo (Hot/Warm/Cold)
Value metric
usage-based: per-hour per-instance (GPU/CPU model), per-GB-month storage, flat monthly network fees
How limits scale
Escalation logic
There are no packaged plans; the 'tiers' are a hardware catalog where price scales with GPU generation and capability (GH200 up through HGX B200), with spot pricing as a discounted interruptible variant. Storage tiers (Hot/Warm/Cold/Archive) step down in price as access speed decreases, and committed-use contracts promise up to 60% off on-demand.
Psychological anchors
- Spot vs on-demand price contrast — Every GPU/CPU row shows an on-demand price next to a spot price roughly 40-60% lower (e.g. HGX H100 $49.24 vs $19.71), making the on-demand rate feel premium and nudging cost-sensitive users toward interruptible capacity.
- Committed-use discount callout — 'Reserved Compute Capacity' promises up to 60% off on-demand for committed usage, positioning the listed rates as a ceiling and funneling large buyers to a sales conversation.
- Zero-fee reframing (free egress, free networking) — Storage copy stresses 'no ingress, egress, or transfer fees', and NAT, VPC, internet transfer, and intra-platform transfer are all listed as Free, an anchor against hyperscaler egress charges.
- Frontier SKUs as prestige anchors — GB300 NVL72 and HGX B300 appear at the top of the table with 'Contact sales', signalling top-end capability without a public price.
- Entry-priced GPU for low-barrier trial — Cheap single-GPU options like GH200 ($6.50/hr) and inference single-GPU pricing give smaller teams a low-cost way in alongside the $40-70/hr HGX nodes.
What this page is optimizing for
Primarily enterprise/AI-team lead generation and capacity sales: public per-hour rates build trust and let buyers budget, but frontier SKUs say 'Contact sales', committed-use discounts are routed to sales, and the closing CTAs are 'Talk to an expert' / 'Contact us'. There is no instant checkout or plan selection on the page.
Red flags
- The page is a dense, scroll-heavy catalog of 12+ GPU rows duplicated across two regions and again for CPUs, with no filtering, sorting, or plan recommendation, making it hard to choose a configuration.
- Newest and most in-demand SKUs (GB300 NVL72, HGX B300, RTX PRO 6000 standard memory, some Turin/Emerald Rapids CPUs) show 'Contact sales' instead of a price, which undermines the otherwise transparent posture.
- The 'Inference Single GPU Price' column appears only in North America and footnote 2 says it applies only to inference-platform customers, which is confusing and easy to miss.
- No pricing calculator or example monthly cost, so buyers must do the math themselves across GPU, CPU, storage, and networking.
- The reserved/committed discount ('up to 60%') gives no tier thresholds or term lengths, pushing all high-value buyers into an opaque sales process.
- Little objection handling or trust content (no FAQ, SLAs, customer logos, or security badges) on the page itself, and CTAs are generic text links rather than per-row actions.
- Page text contains leftover template artifacts ('This is some text inside of a div block', 'SIMILAR TO TECH SPECS'), signalling polish issues.
Best-practices scorecard
What works · 2
- Visible prices — Most GPU, CPU, storage, and network prices are public, including spot rates, though a handful of top-end SKUs are contact-sales.
- Value metric matches usage — Hourly per-GPU/CPU and per-GB-month billing aligns directly with how AI workloads consume compute and storage, with free egress removing growth penalties.
Half measures · 3
- Annual discount offered — There is no monthly/annual toggle, but a committed-use discount of up to 60% is mentioned (without specifics).
- Tier differences are scannable — Spec columns (VRAM, vCPUs, RAM, storage) are consistent and comparable, but the tables are long, repeated by region, and have no filtering or comparison aids.
- Clear CTAs per tier — 'Contact sales' links appear on gated rows and 'Start building today' / 'Talk to an expert' at the bottom, but priced rows have no direct provision or sign-up action.
What's missing · 3
- Clear recommended tier — No plan or SKU is highlighted, badged, or recommended; the user must choose from a flat catalog.
- FAQ or objection handling — There is no FAQ; only footnotes on units and DX billing plus a short note on free egress.
- Trust signals present — The pricing section carries no logos, testimonials, SLAs, or security badges; only a Trust Center link in the nav.