CoreWeave pricing

About

CoreWeave is a publicly traded (Nasdaq: CRWV) cloud-computing company that rents NVIDIA GPU and CPU compute to AI labs, hyperscalers, and enterprises. It has one of the more unusual origin stories in AI infrastructure: founded in 2017 as Atlantic Crypto, an Ethereum-mining operation run by three former commodities traders (Michael Intrator, Brian Venturo, and Brannin McBee). When crypto prices collapsed in 2018, the company bought up GPUs from distressed miners, rebranded to CoreWeave in 2019, and pivoted to renting that compute as a service — the same NVIDIA cards that turned out to be exactly what AI training needed.

That bet made CoreWeave one of the defining “neoclouds.” Revenue reached roughly $1.9B in 2024 and $5.13B in 2025 (up about 170% year over year), and the company went public on March 28, 2025 at $40.00 per share for an initial valuation near 23B. The catch is concentration: Microsoft was around 67% of 2025 revenue, OpenAI has contracted up to about $22.4B in total commitments, and Nvidia — CoreWeave’s primary GPU supplier — disclosed a roughly 7% stake in May 2025. CoreWeave’s economics are driven far more by a handful of large reserved contracts than by self-serve GPU rentals.

For current pricing, see CoreWeave’s pricing page.

Pricing summary : How CoreWeave’s pricing model works

CoreWeave is pure usage-based with a heavy commitment layer: you pay by the instance-hour for compute, with three purchasing motions stacked on top of each other:

1. On-Demand instances — self-serve, no commitment, published per-instance list rates for GPU and CPU nodes. Best for burst capacity.
2. Spot instances — interruptible capacity at roughly 40-60% off on-demand (e.g. HGX H100 spot at $19.71/hr vs $49.24 on-demand). Published openly.
3. Reserved / committed capacity — “up to 60% discounts over our On-Demand prices for committed usage,” quoted by sales. This is where the large AI-lab and hyperscaler contracts live, and it is the bulk of revenue.

Storage and networking are metered separately. Storage is openly priced (object storage from $0.06/GB/mo Hot down to $0.0125/GB/mo Archive; distributed file storage $0.070/GB/mo), and there are no ingress, egress, or data-transfer fees — a meaningful differentiator versus hyperscalers.

Two further dimensions sit alongside the compute meter. Rate cards are published per region — North America and Europe both list the same on-demand instance prices, but spot rates differ slightly by region (HGX H100 spot is $19.71/hr in NA and $19.51/hr in Europe). And a legacy “CoreWeave classic” rate card still runs at coreweave.com/pricing/classic, where instances are priced à la carte: a per-hour GPU component (NVIDIA H100 PCIe $4.25, A100 80GB $2.21, RTX A6000 $1.28) plus $0.01/hr per vCPU and $0.005/hr per GB of RAM, with CPU-only instances billed per vCPU from $0.0125 to $0.035.

What makes this different: CoreWeave publishes a full portfolio-wide rate card — on-demand, spot, and inference single-GPU rates — for GPUs that most rivals gate behind sales calls, yet its actual revenue overwhelmingly comes from sales-quoted reserved contracts with a few mega-customers. The published card is a credibility and self-serve on-ramp; the business is large committed deals. It also sits at the pure usage pricing end of the spectrum, with GPU-hour pricing as the primary unit and committed-use pricing carrying the revenue.

Pricing by product

GPU compute (on-demand, spot, inference)

On-demand GPU list prices (North America region), per instance. Most nodes are 8 GPUs, so the per-GPU rate is roughly the node price divided by 8; GB200/GB300 NVL72 instances are built from 2 Superchips (4 Blackwell GPUs plus 2 Grace CPUs):

Inference single-GPU pricing carries an explicit caveat: “Pricing based on GPUs is exclusively applicable to the CoreWeave inference platform customers. Please contact your account executive for more information.” Reserved/committed usage is “up to 60% discounts over our On-Demand prices,” sales-quoted.

A separate Europe rate card publishes the same on-demand instance prices with slightly different spot rates — HGX H100 spot is $19.51/hr in Europe versus $19.71/hr in North America, HGX B300 spot $36.70/hr versus $35.84/hr — and the Europe card omits the inference single-GPU column entirely.

CPU compute (on-demand + spot)

The two Standard Memory shapes are spot-only on the public card — on-demand is “Contact Sales.” Europe CPU spot rates differ slightly (e.g. AMD Genoa 9274F spot $3.41/hr in Europe versus $3.50/hr in North America).

Storage, networking and Kubernetes

*SUNK’s free listing “Excludes CPU Resource Usage required to operate SUNK” — the scheduler is free, the CPU instances it runs on are not. Direct Connect prices exclude cross-connect costs (Dedicated) and Equinix Fabric / Megaport port and virtual-connection charges (Virtual); CoreWeave adds those to the quoted price if it orders and manages them for you. Storage is metered in binary gigabytes (GiB).

CoreWeave classic (legacy à-la-carte card)

A separate legacy rate card at coreweave.com/pricing/classic prices instances component-by-component instead of per node: total cost = GPU component + $0.01/hr per vCPU + $0.005/hr per GB of RAM. A valid configuration needs at least 1 GPU, 1 vCPU and 2 GB RAM (plus 40 GB of root-disk NVMe for Virtual Servers).

CPU-only classic instances are billed per vCPU with RAM included: Intel Xeon Ice Lake and AMD EPYC Milan $0.035, Intel Xeon Scalable and AMD EPYC Rome $0.03, Intel Xeon v4 $0.02, Intel Xeon v3 $0.0125. Classic storage is priced separately from the modern card — File and Block Storage at $0.04 per GB/month (HDD) and $0.07 per GB/month (NVMe), Object Storage at $0.03 per GB/month.

Sales motions across products: on-demand and spot GPU/CPU instances are self-serve at published rates; reserved/committed capacity, Direct Connect, GB300/B300 on-demand, the Standard Memory CPU shapes, Archive object storage, and inference-platform single-GPU pricing are all sales-led and quoted.

Hidden costs : What CoreWeave users actually pay

CoreWeave’s networking is unusually clean (no egress, no transfer fees), but a realistic bill stacks several metered items beyond the headline GPU rate:

The biggest real-world cost driver is the reserved commitment itself: the discount (up to 60%) comes with a contractual obligation to pay for capacity over a multi-month or multi-year term, so idle committed GPUs are a sunk cost. The second is always-on storage — object and file storage meter continuously, independent of whether any instance is running. On-demand buyers should also watch the per-instance framing: the headline number is for a whole 8-GPU node, not a single GPU.

Want to estimate your own CoreWeave bill? Use the CoreWeave pricing calculator to model your costs based on usage patterns.

Pricing evolution : CoreWeave pricing history and changes

Cadence

Tracked range: 2024–present. CoreWeave’s IPO S-1 and investor disclosures are the cleanest public record of its economics; the live rate card is the source of current list prices.

Notable changes

• 2024 — Pre-IPO Hopper-era cloud: on-demand and reserved NVIDIA H100/A100 capacity, pitched at 30-60% below hyperscalers. Reserved contracts (sales-quoted) were the bulk of revenue.
• March 2025 — IPO on Nasdaq at $40.00/share, ~23B-dollar valuation. Disclosures revealed Microsoft at roughly two-thirds of revenue and a multibillion-dollar OpenAI commitment.
• June 2026 — Portfolio-wide rate card publishes on-demand, spot, and inference single-GPU prices across Blackwell, Hopper, and Ada/Ampere cards: HGX H100 $49.24/$19.71 (on-demand/spot), HGX B200 $68.80/$34.11, GB200 NVL72 $42.00. Reserved usage up to 60% off, sales-quoted.
• 2026-07-17 — CoreWeave launched a $2.6B term loan to finance further GPU acquisition. It is capacity financing rather than a pricing event, but it is the mechanism that lets published rates stay flat while the fleet grows: the fixed cost of the fleet is carried on the balance sheet, not passed through the hourly card.
• 2026-07-21 — Packaging reshuffle with zero published GPU price movement. Every North America GPU on-demand and spot rate held (HGX H100 $49.24/$19.71, H200 $50.44/$20.93, B200 $68.80/$34.11, GB200 NVL72 $42.00, A100 $21.60/$9.65), as did all storage and networking rates. What changed: two spot-only General Purpose (Standard Memory) CPU shapes were added with on-demand marked “Contact Sales,” the $15,000/mo 100G Virtual Direct Connect tier went “Not Available,” and the reserved-storage capacity discount footnote was removed. Europe CPU spot drifted a few cents in both directions.

The trajectory is from a quiet sales-quoted neocloud to a public company that publishes a full, transparent rate card while still earning most of its money from a few enormous committed contracts. The July 2026 update is the first evidence that the published surface itself is an editable variable: CoreWeave now adjusts which lines a buyer can price without a call more readily than it adjusts the prices on those lines.

The July 2026 packaging reshuffle in detail

Three edits landed together on 2026-07-21, and each moves the boundary between the self-serve card and the sales conversation:

• Two new CPU shapes are spot-only. The AMD Turin (9655P) General Purpose (Standard Memory) shape at 192 vCPU / 768 GB lists spot at $3.94/hr in North America ($4.04 in Europe), and the Intel Emerald Rapids (8562Y+) shape at 64 vCPU / 256 GB lists spot at $2.86/hr ($2.98 in Europe). Both read “Contact Sales” for on-demand — the only CPU shapes on the card where the non-interruptible rate is gated. For a buyer, that means the cheapest published CPU capacity is now also the least reliable capacity, and pricing a steady-state CPU workload on these shapes requires a quote.
• 100G Virtual Direct Connect was withdrawn. The $15,000/month tier now reads “Not Available,” leaving 10G at $1,500/mo as the only self-serve Virtual DX speed. Anyone who needs more than 10G of dedicated connectivity is pushed to Dedicated DX at $12,500/mo (100G) or $50,000/mo (400G) — which is cheaper per gigabit but adds cross-connect costs CoreWeave quotes separately. The practical effect is a gap in the middle of the connectivity ladder, not a price rise.
• The reserved-storage discount is gone from the public card. The storage table’s asterisk (“Discounts are available for reserved storage capacity for AI Object Storage and Distributed File Storage”) no longer appears, so up-to-60% reserved compute is the only commitment discount CoreWeave advertises. Storage commitments may still be negotiable in a contract, but a buyer modelling a multi-petabyte footprint can no longer assume a published lever exists — the storage line in a five-year TCO model should now be built at list.

None of this changes what a running H100 node costs. It changes how much of a CoreWeave bill a buyer can construct before talking to an account executive — and on 2026-07-21 that share went down.

What’s unique : CoreWeave’s distinctive pricing mechanics

1. Three-motion stack on one rate card. CoreWeave publishes on-demand, spot, and inference single-GPU prices for the same hardware, then layers up-to-60% reserved discounts on top. Few neoclouds expose all three tiers openly.

2. Zero egress and transfer fees. No ingress, egress, internal-transfer, VPC, or NAT charges — removing the most unpredictable line items in a hyperscaler GPU bill and pricing connectivity almost entirely into the instance-hour.

3. Published list price, sales-quoted reality. The open rate card is a self-serve on-ramp and credibility signal, but the company’s economics are dominated by sales-quoted reserved contracts with a handful of mega-customers — a deliberate split between the shop window and the back room.

4. The published surface is curated, and it moves in both directions. On 2026-07-21 CoreWeave added two General Purpose (Standard Memory) CPU shapes that carry a spot price but a “Contact Sales” on-demand rate, while simultaneously withdrawing the $15,000/mo 100G Virtual Direct Connect tier and the reserved-storage capacity discount footnote. Not one published GPU, storage, or networking price changed. The lever CoreWeave pulled was coverage — how much of a bill a buyer can assemble unaided — which is a quieter form of pricing power than a rate change and much harder to spot in a diff.

Strengths & weaknesses

Billing UX : CoreWeave billing controls and transparency

• On-Demand vs Spot toggle on the public rate card — every GPU and CPU shape lists an “On-Demand Price (Per Hour)” and a “Spot Price (Per Hour)” side by side, so the interruptibility discount is visible before you deploy. Spot capacity is governed by a separate Spot Terms of Service linked from the pricing page.
• Reserved Compute Capacity — the only lever that is not self-serve: “CoreWeave offers up to 60% discounts over our On-Demand prices for committed usage,” with a contact-us path rather than a checkout. There is no published commit ladder, so the discount curve is quote-only.
• Region selector on the rate card — the pricing page publishes distinct “REGION: NORTH AMERICA” and “REGION: EUROPE” tables for both GPU and CPU instances, so buyers see region-specific spot rates rather than a single blended number.
• Zero-fee networking primitives — NAT Gateway, VPC, data transfer between CoreWeave and the internet, data transfer within CoreWeave, and object-storage egress/IOPS are all listed at Free, which removes the usual per-GB reconciliation line from the bill entirely. Public IPs are the exception at $4.00 per IP / month.
• Free control plane, metered substrate — the CoreWeave Kubernetes Service control plane is Free and SUNK is “Free*”, but the asterisk states it “Excludes CPU Resource Usage required to operate SUNK” — scheduler licensing is free, the CPU instances underneath still meter.
• Fleet LifeCycle Controller, Node LifeCycle Controller and Observability — named platform components that surface node health and utilization, which is where instance-hour spend is actually attributed.
• Two live rate cards — the modern per-instance card and the legacy “CoreWeave classic” à-la-carte card (GPU component + $0.01/hr per vCPU + $0.005/hr per GB RAM) both remain published, cross-linked from each other’s footer.

Strategic wins : Why CoreWeave’s pricing decisions worked

1. Undercutting hyperscalers with a transparent, egress-free card

By publishing per-instance rates 30-60% below AWS/Azure/GCP and eliminating egress fees, CoreWeave turned price and predictability into a wedge for AI workloads that move huge datasets. See how AI companies structure pricing.

2. Reserved commitments as the revenue engine

The up-to-60% committed discount converts spiky GPU demand into multi-year contracted revenue — the mechanism behind the Microsoft and OpenAI deals and the predictable backlog investors reward. Related: outcome-based pricing trends.

3. Spot + inference tiers to monetize spare capacity

Publishing spot prices (~40-60% off) and an inference single-GPU rate lets CoreWeave fill capacity between committed contracts, smoothing utilization. The 2026-07-21 card extends the tactic to CPU: the two new General Purpose (Standard Memory) shapes are published at a spot rate only, so the interruptible price does the demand-shaping while the firm-capacity rate stays negotiable. That is a clean way to sell surplus without setting a public floor on the reliable version of the same shape. See choosing the right usage metric.

Areas to improve : Gaps in CoreWeave’s pricing approach

1. Per-instance framing obscures the per-GPU cost

A $49.24/hr HGX H100 sounds expensive until you realize it is eight GPUs (about $6.16 each); the rate card would be easier to compare if it showed per-GPU rates alongside node prices. See bill shock and cost unpredictability.

2. The best prices are still behind a sales call

On-demand and spot are public, but the up-to-60% reserved discounts — where most spend lands — are quoted, so buyers cannot self-estimate a committed contract without engaging sales.

3. Concentration risk is a pricing risk

With one customer near 67% of revenue, CoreWeave’s leverage in self-serve pricing is limited; the published card matters less than the terms negotiated with a few anchor tenants, which makes the public rate card more marketing than margin.

4. Withdrawn SKUs and discounts disappear silently

The 2026-07-21 update removed the reserved-storage capacity discount footnote and flipped 100G Virtual Direct Connect from $15,000/mo to “Not Available” with no dated note, no successor SKU, and no grandfathering statement — a buyer who built a TCO model in June has no way to learn from the page that two of its assumptions expired. Two concrete fixes: stamp withdrawn SKUs with a withdrawal date and a pointer to the replacement path (here, Dedicated DX at $12,500/mo for 100G), and either publish a storage-commitment ladder or state plainly that storage commitments are quote-only. Silent removals are the cost-model equivalent of bill shock — the number does not move, but the plan built around it breaks.

Key takeaways

1. CoreWeave is pure usage-based with a dominant commitment layer — hourly on-demand and spot, plus up-to-60% reserved discounts that carry most of the revenue. For the underlying model, see the introduction to usage-based pricing.
2. It publishes a fuller rate card than most neoclouds — on-demand, spot, and inference single-GPU prices openly listed — but the real economics are sales-quoted reserved contracts, and the July 2026 update showed the published/quoted boundary is something the vendor edits at will rather than a fixed disclosure standard.
3. No egress or transfer fees is a genuine differentiator versus hyperscalers for data-heavy AI workloads.
4. Prices are per instance (mostly 8-GPU nodes), so headline numbers must be divided down to compare per-GPU against rivals like Lambda or RunPod.
5. Customer concentration is the real story — Microsoft ~67% of revenue and OpenAI’s ~$22.4B commitment mean a few mega-contracts, not the public card, define CoreWeave’s pricing power.

UBP implications

1. A public rate card and a sales-quoted core can coexist — and the line between them is itself a pricing lever. CoreWeave shows you can publish on-demand/spot prices for credibility while running the business on negotiated committed contracts, so the transparency is an on-ramp, not the monetization. Its 2026-07-21 update moved two CPU shapes’ on-demand rates behind “Contact Sales” and dropped an advertised storage discount without repricing a single line, which means change-detection that only watches numbers will miss the most consequential edits.
2. Commitment discounts convert volatility into backlog. Up-to-60% reserved pricing turns spiky GPU demand into multi-year contracted revenue, the metric public-market investors reward — a reusable pattern for any capacity-constrained usage business.
3. Removing punitive fees (egress) can be worth more than a lower unit rate. Zero data-transfer charges remove the least predictable line item in cloud bills, which buyers value even when the per-hour GPU rate is not the absolute cheapest.

Sources

• CoreWeave Cloud pricing (accessed 2026-07-21)
• CoreWeave classic pricing — legacy à-la-carte rate card (accessed 2026-07-21)
• CoreWeave Docs — Billing & Pricing (accessed 2026-07-21)
• CoreWeave Docs — Capacity plans (reserved, on-demand, spot) — defines reserved capacity as “a flat rate for every hour in the billing period, regardless of whether instances are running or idle” (accessed 2026-07-21)
• CoreWeave Docs — Inference billing — “You pay per GPU-hour based on the instance type,” node-based vs GPU-based billing (accessed 2026-07-21)
• CoreWeave Docs — H100 InfiniBand instance spec — confirms the HGX H100 node is 8 GPUs, so the $49.24/hr list rate is ~$6.16/GPU/hr (accessed 2026-07-21)
• CoreWeave Docs — GB200 NVL72 instance spec — confirms 4 Blackwell GPUs plus 2 Grace CPUs per instance (accessed 2026-07-21)
• CoreWeave Investor Relations (accessed 2026-07-21)
• CoreWeave announces pricing of IPO (accessed 2026-06-15)

Bottom line

CoreWeave is a pure usage-based GPU cloud with a commitment-heavy core: it publishes on-demand, spot, and inference single-GPU rates openly (HGX H100 nodes at $49.24/hr, about $6.16/GPU/hr; spot ~60% cheaper; storage from $0.0125–$0.07/GB/mo with free egress), yet earns most of its money from up-to-60%-off reserved contracts that are sales-quoted. Its July 2026 update proved the point in miniature: no published GPU price moved, but the newest CPU shapes arrived spot-only, 100G Virtual Direct Connect was withdrawn, and the reserved-storage discount stopped being advertised — the card got narrower, not dearer. The numbers that matter most aren’t on the rate card — they’re in the IPO disclosures: $5.13B of 2025 revenue, ~67% from Microsoft and a ~$22.4B OpenAI commitment. The published card is the on-ramp; the mega-contracts are the business. Browse the pricing blueprint for more fully-researched company profiles, or compare CoreWeave against other Infrastructure, Compute & MLOps companies.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

Packaging reshuffle: spot-only CPU shapes in, 100G Virtual DX and reserved-storage discounts out

Jul 2026

No published GPU price moved between the June and July 2026 cards (HGX H100 $49.24/$19.71, H200 $50.44/$20.93, B200 $68.80/$34.11, GB200 NVL72 $42.00, A100 $21.60/$9.65 all held), but the shape of the card changed. Two General Purpose (Standard Memory) CPU shapes were added spot-only — AMD Turin 9655P at $3.94/hr and Intel Emerald Rapids 8562Y+ at $2.86/hr in North America — with on-demand marked Contact Sales. The $15,000/mo 100G Virtual Direct Connect tier became Not Available, leaving 10G at $1,500/mo, and the reserved-storage capacity discount footnote was removed, making up-to-60% reserved compute the only publicly advertised commitment discount.

Published portfolio-wide rate card with spot + inference tiers

Jun 2026

June 2026 list pricing publishes on-demand, spot, and inference single-GPU rates per instance: HGX H100 $49.24/$19.71 on-demand/spot, HGX H200 $50.44/$20.93, HGX B200 $68.80/$34.11, GB200 NVL72 $42.00. Reserved usage up to 60% off (sales-quoted). Storage and networking priced openly with free egress.

IPO on Nasdaq (CRWV) at $40/share

Mar 2025

CoreWeave went public March 28, 2025 at $40.00/share, ~37.5M shares, a ~23B-dollar initial valuation. The S-1 disclosed heavy revenue concentration (Microsoft ~62-67%) and a multibillion-dollar OpenAI contract, reframing the company from private neocloud to public infrastructure provider.

Hopper-era GPU cloud pricing, pre-IPO

Jan 2024

Through 2024 CoreWeave sold on-demand and reserved NVIDIA H100/A100 capacity, pitching 30-60% savings vs hyperscalers. Reserved/committed contracts (the bulk of revenue) were sales-quoted; revenue was 1.915B for the year.