CoreWeave pricing in 2026: every GPU rate and what a node really costs

In April 2026, CoreWeave filed an 8-K with the SEC disclosing an expanded agreement with Meta: approximately $21 billion for AI cloud capacity through December 2032, deployed across multiple locations and including some of the first Nvidia Vera Rubin systems, per the filing.

Divide that evenly across the term and Meta committed to roughly $260 million of GPU capacity a month, from one vendor, before its own data centers spend a dollar. That’s arithmetic on the filed figures, not a disclosed rate.

Your commitment will start where Meta’s negotiators started, though: a public rate card that prices by the node and discounts up to 60 percent behind closed doors.

How much does CoreWeave cost in 2026?

On-demand GPU nodes run from $10.00 per hour for an 8x L40 to $68.80 for an 8x B200, with the workhorse 8x H100 node at $49.24. Every figure on the public card prices a full instance, not a chip, so the CoreWeave price you’ll actually compare against other clouds takes one extra division.

Here are CoreWeave’s North America rates as of 2026, normalized per GPU:

Note: GB200 and GB300 instances bundle two Grace Blackwell Superchips, four GPUs total. The inference column applies only to CoreWeave inference platform customers. European spot rates differ slightly: the H100 node runs $19.51 there versus $19.71 in North America.

What does CoreWeave do?

The short version of what CoreWeave does: it rents Nvidia GPU computing to companies that train or run AI models, the category the market calls GPU as a service. Founded in 2017 and listed on Nasdaq since March 2025, it builds data centers purpose-built for AI work and wraps them in Kubernetes-native orchestration.

Its customer roster explains the scale. Contracts with OpenAI reached approximately $22.4 billion cumulatively as of the September 2025 expansion, per CoreWeave’s announcement, and the Meta agreement added roughly $21 billion more this April.

CEO Michael Intrator said the OpenAI milestone “affirms the trust that world-leading innovators have in CoreWeave’s ability to power the most demanding inference and training workloads at an unmatched pace.”

The customer side of that deal agrees: OpenAI infrastructure VP Peter Hoeschele said CoreWeave is “helping us advance the frontier of intelligence.”

Analysts have graded the platform, not just the deal flow. CoreWeave reports a Visionary placement in Gartner’s 2026 Magic Quadrant for Cloud AI Infrastructure and a Platinum rating on SemiAnalysis’s ClusterMAX, which it says it has earned twice.

How is CoreWeave pricing structured?

You can buy CoreWeave compute three ways: on demand by the hour, spot at up to 60 percent off with interruption risk, and reserved capacity under negotiated multi-year contracts discounted up to 60 percent. The rate card shows the first two. The CoreWeave business model runs on the third.

No public term-rate table exists for reserved capacity. Where Google Cloud publishes its committed-use discounts in full, CoreWeave publishes a ceiling and routes you to sales. The Meta and OpenAI agreements are that tier at its extreme, and they tell you who the rate card is really for: everyone who hasn’t negotiated yet.

The second structural fact matters more to your budget. On Hopper and Blackwell HGX instances, you rent eight GPUs at a time whether your workload needs eight or three. NVL72 instances come in four GPU bundles. Only the GH200 rents as a single chip on demand, at $6.50 per hour, and inference platform customers get single GPU rates as a separate lane.

Teams get burned here in a predictable way. They budget from a per GPU number they saw quoted, then provision a node that bills at five times it. The unit you pay for is the instance, and the sooner your forecast speaks in nodes, the fewer surprises your first invoice holds.

What’s the real price of a GPU hour?

The list rate isn’t your price. Your price is the list rate divided by utilization, because a node bills at 100 percent whether your GPUs are working or waiting. CloudZero calls this the effective GPU-hour rate, and it’s the number that separates teams with defensible AI spend from teams with expensive surprises.

What utilization does to CoreWeave’s on-demand rates, in CloudZero’s arithmetic:

The comparison lands hard: a half-utilized H100 node on CoreWeave costs more per useful hour than a fully utilized B200. Most teams comparing providers argue about a dollar of list-rate difference while a multiple of that leaks through idle capacity, a pattern CloudZero sees constantly in GPU workload spend.

The fix isn’t a better sticker price, it’s knowing what each node actually did: per-workload attribution of the kind CloudZero documents for GPU spend in Kubernetes, where fractional GPU sharing makes the accounting hard. Rate, utilization, allocation. Get all three visible and the CoreWeave invoice stops being a fight.

What does a real deployment cost per month?

One 8x H100 node running around the clock costs $35,945 a month on demand: $49.24 times 730 hours. On spot, the same node runs $14,388, when capacity holds. At CoreWeave’s full published reserved ceiling, the arithmetic lands near $14,378, though your actual contract rate comes out of negotiation, not a table.

Scale up modestly and the numbers move fast

A four-node, 32-GPU training cluster runs about $143,781 a month on demand. A single 24-hour training run on one node costs roughly $1,182. The Blackwell flagship on the public card, the 8x B200 node, bills $50,224 a month on demand.

Storage and networking write their own lines

Object storage bills $0.06 per GB monthly on the hot tier, $0.03 warm, $0.015 cold, $0.0125 archive; distributed file storage runs $0.07. A 100 TB hot dataset adds about $6,144 a month under CoreWeave’s binary GB accounting. Public IPs cost $4.00 each monthly, and Direct Connect runs from $1,250 at 10G to $50,000 at 400G.

CPU support nodes fill out the rest of the bill

Training pipelines lean on them for data preparation, and CoreWeave’s on-demand CPU instances run $3.36 to $9.31 per hour. Modeling a full deployment before you commit is what CloudZero’s LLM cost calculator approach was built for.

What’s free on CoreWeave, and what isn’t?

Egress is free, and that one line separates CoreWeave from the AWS, Azure, and GCP bills most teams are used to. Data transfer out, transfer within CoreWeave, NAT gateways, VPCs, and storage IOPS all bill at zero. The managed Kubernetes control plane and the SUNK scheduler cost nothing either.

For AI work, that changes real decisions. Moving checkpoints out, serving outputs at volume, or migrating away entirely costs nothing on the network line, which deletes the classic lock-in tax. CoreWeave even runs a Zero Egress Migration program for inbound moves, a detail worth raising in any hyperscaler negotiation.

What still bills: GPUs, CPUs, storage by tier, public IPs, Direct Connect. The free items remove the gotcha lines. They don’t touch the big ones.

How does CoreWeave compare with other GPU clouds?

CoreWeave wins on published transparency, contracted scale, and free data movement rather than the lowest sticker. The CoreWeave competitors question splits three ways: hyperscalers, other AI-native clouds, and single-GPU rental shops, and each comparison turns on a different variable.

Against AWS, Azure, and GCP, the trade is committed-rate transparency versus network economics: hyperscalers publish term pricing CoreWeave does not, and they meter the egress CoreWeave gives away. CloudZero runs those numbers provider by provider in its cloud GPU pricing comparison.

Against Lambda, RunPod, and Nebius, the trade is granularity. Those clouds sell single GPUs on demand where CoreWeave sells nodes, which makes them the practical CoreWeave alternatives for bursty or sub-node workloads. And if what you need is tokens rather than infrastructure, managed AI pricing through model APIs or Bedrock skips the GPU question entirely. For the chip-level decision underneath all of it, CloudZero’s H100 GPU cost guide covers buying versus renting.

What changed in 2026

Blackwell landed on the public rate card. The 8x B200 node lists at $68.80 per hour with an $8.60 single GPU inference rate, the GB200 NVL72 instance at $42.00, and B300-class hardware shows spot pricing while on-demand buyers still go through sales, the usual tell for supply-constrained chips.

The contracts got longer and stranger. Meta’s expansion to approximately $21 billion runs through December 2032. Nvidia itself agreed to backstop CoreWeave in a deal reported at $6.3 billion, guaranteeing it will buy any unsold capacity, per CoreWeave’s own SEC 8-K filing on the order. That’s Nvidia underwriting demand for the same chips it sells CoreWeave, a hedge most buyers negotiating capacity elsewhere don’t get.

And the H100 price per hour conversation started aging. With H200 nodes at a 2.4 percent premium over H100 and Blackwell now orderable, 2026 budget cycles are the first where the chip generation decision comes before the provider decision.

How do you make CoreWeave spend defensible?

A rate card tells you what a GPU hour costs. It can’t tell you what a GPU hour earned, and that second number is the one your CFO will ask for. Node billing widens the gap: eight GPUs bill whether five idle or none do, so the invoice alone can’t say whether the money worked.

Getting to defensible means connecting the spend to what produced it.

CloudZero’s CoreWeave integration ingests that spend and allocates it to teams, products, and workloads alongside the rest of your cloud and AI bills, which turns the vague total into unit economics: cost per training run, cost per token, cost per customer served, the same lens CloudZero applies to inference economics generally.

That’s also the honest sequencing for savings. AI cost optimization levers like spot strategies, model right-sizing, and inference cost tuning move real money, but only after allocation tells you which workloads deserve the effort. The teams that treat this as an AI cost management practice, with budgets set before the invoice rather than after it, are the ones that keep scaling without a spending freeze.

Organizations like Toyota, Skyscanner, Grammarly, Duolingo, Upstart, and Coinbase use CloudZero to see where their cloud and AI spend goes, GPU hours included. Request a demo to see CoreWeave spend allocated by team and product, get a free cloud cost assessment, or take the self-guided product tour.

CoreWeave pricing FAQs

Author Spotlight

Lyne Carolyne

Lyne Carolyne has several years of experience in AI and cloud economics and brings that understanding into the content she creates. Outside work, she's an avid explorer.