CoreWeave GPU pricing starts with a published rate card: HGX H100 at $6.16/GPU-hr, HGX H200 at $6.31/GPU-hr, HGX B200 at $8.60/GPU-hr, and GB200 NVL72 at $10.50/GPU-hr (coreweave.com/pricing). Those numbers are real, but they're not what most CoreWeave customers pay. CoreWeave's business runs on multi-year committed contracts, not hourly checkouts, and its own CFO told investors in Q1 2026 that the company is "largely sold out" of 2026 capacity while raising list prices across every GPU generation it sells. This post covers what's actually on the rate card today, including the Blackwell and NVL72 tiers (GB200, GB300, B200, B300) that don't get much airtime elsewhere, how CoreWeave's reserved-contract model changes the real number you'd pay, and where its pricing lands against Spheron and the rest of the on-demand market. For the deep per-GPU H100 and H200 breakdown, see our CoreWeave H100 and H200 pricing guide; this post's job is the contract structure and the GPU tiers that guide doesn't cover.
CoreWeave GPU Pricing Model: How Reserved vs On-Demand Contracts Actually Work
CoreWeave publishes two things on its pricing page: an on-demand hourly rate and a spot rate, both listed per node rather than per GPU. What it doesn't publish is the third and, by revenue, most important tier: reserved capacity sold under multi-year contracts. That's the part of CoreWeave's business that actually moves the needle, and it's worth understanding before you look at a single number on the rate card.
On-Demand and Spot Rates (What's Actually Published)
Every GPU on CoreWeave's current pricing page is sold as a fixed-size instance, not a single accelerator. Here's the full published lineup:
| GPU | Instance Size | On-Demand $/hr | On-Demand $/GPU-hr | Spot $/hr | Spot $/GPU-hr |
|---|---|---|---|---|---|
| HGX H100 | 8 GPUs | $49.24 | $6.16 | $19.71 | $2.46 |
| HGX H200 | 8 GPUs | $50.44 | $6.31 | $20.93 | $2.62 |
| GB200 NVL72 | 4 GPUs* | $42.00 | $10.50 | N/A | N/A |
| GB300 NVL72 | 4 GPUs* | Contact sales | N/A | N/A | N/A |
| HGX B200 | 8 GPUs | $68.80 | $8.60 | $34.11 | $4.26 |
| HGX B300 | 8 GPUs | Contact sales | N/A | $35.84 | $4.48 |
| RTX PRO 6000 Blackwell | 8 GPUs | $20.00 | $2.50 | $11.09 | $1.39 |
| A100 | 8 GPUs | $21.60 | $2.70 | $9.65 | $1.21 |
| L40S | 8 GPUs | $18.00 | $2.25 | $7.88 | $0.99 |
| L40 | 8 GPUs | $10.00 | $1.25 | $6.27 | $0.78 |
| GH200 | 1 GPU | $6.50 | $6.50 | N/A | N/A |
*Each GB200/GB300 NVL72 instance bundles 2 Superchips, and each Superchip pairs 1 Grace CPU with 2 Blackwell GPUs, for 4 GPUs total per instance (source: coreweave.com/pricing).
A detail worth flagging on its own: HGX B300 has a published spot rate ($35.84/hr) but no published on-demand rate at all. You can apparently get reclaimable B300 capacity through self-serve checkout, but a guaranteed B300 node requires a sales call. That's an unusual gap, and it says something about how tight Blackwell supply still is even for a provider CoreWeave's size.
Reserved/Committed Discounts: The Part CoreWeave Doesn't Publish
CoreWeave offers "up to 60% discounts over on-demand prices for committed usage, making it more cost-effective for predictable workloads," according to a pricing review from Thunder Compute. What CoreWeave doesn't publish is which commitment length buys which discount tier, or how that varies by GPU generation, region, and quantity. There's no rate card for this the way there is for on-demand; you get a number after a sales conversation, not before one.
Applied to the numbers above, a 60% reserved discount on the $6.16/GPU-hr on-demand H100 rate would land around $2.46/GPU-hr, roughly what CoreWeave's own H100 spot rate already costs with zero commitment. That's the real tension in CoreWeave's pricing: the best reserved rate you can negotiate is in the same neighborhood as the on-demand spot price of a smaller neocloud, except you had to sign a multi-year contract to get there.
Pricing fluctuates based on GPU availability. The prices above are based on 12 Aug 2026 and may have changed. Check current GPU pricing → for live rates.
Real CoreWeave GPU Pricing by Model: H100, H200, GB200, and B200
The table above covers what's published today, but the two tiers worth separating out are Blackwell/NVL72 (the newest, least-covered hardware) and Hopper plus the older generations (which most workloads still run on).
Blackwell and NVL72 Pricing (GB200, GB300, B200, B300)
GB200 NVL72 is the cheapest way onto NVIDIA's rack-scale Blackwell platform through CoreWeave: $42.00/hr for a 4-GPU instance, or $10.50/GPU-hr. Scale that to a full NVL72 rack (72 GPUs, 18 of those 4-GPU instances) and you're looking at roughly $756/hr, or about $544,320 for a continuous 30-day month at full rack scale. That's the real number behind "renting an NVL72 rack," not a marketing abstraction. For the memory and compute specs behind that rack (13.5TB HBM3e, 1.4 exaflops FP4), see our Vera Rubin NVL72 cloud availability piece, which covers where rack-scale pricing is headed next.
HGX B200 is the standard 8-GPU Blackwell tier: $68.80/hr on-demand ($8.60/GPU-hr) and $34.11/hr spot ($4.26/GPU-hr). GB300 NVL72 and HGX B300 are both gated behind "contact sales" for on-demand access; B300 at least has a published spot rate ($35.84/hr, or $4.48/GPU-hr across the 8-GPU node), but GB300 has no self-serve price of any kind yet. If you're deciding between the two Blackwell rack generations rather than just CoreWeave's price for either, our GB300 vs GB200 NVL72 pricing and availability guide breaks down the real premium, which analysts peg at roughly 10-30% over GB200, not double. For broader B200 pricing context beyond CoreWeave, NVIDIA B200 cloud pricing across providers lines up the rest of the market.
Hopper Pricing (H100, H200) and Older Generations (A100, L40S, GH200)
HGX H100 runs $49.24/hr ($6.16/GPU-hr) and HGX H200 runs $50.44/hr ($6.31/GPU-hr), both 8-GPU-minimum, both with no self-serve single-GPU option. CoreWeave also still runs a legacy "Classic" pricing page at coreweave.com/pricing/classic that lists individual H100 PCIe at $4.25/hr and HGX H100 at $4.76/hr, materially below the current page's $6.16/GPU-hr for what's nominally the same hardware generation. If you're citing a CoreWeave H100 price you saw somewhere else, check which page it came from; we go deeper on that discrepancy, plus the full spot and per-GPU breakdown, in the CoreWeave H100 and H200 pricing post.
Older and adjacent generations round out the catalog: A100 at $21.60/hr ($2.70/GPU-hr), L40S at $18.00/hr ($2.25/GPU-hr), L40 at $10.00/hr ($1.25/GPU-hr), RTX PRO 6000 Blackwell at $20.00/hr ($2.50/GPU-hr), and GH200 as the one true single-GPU listing on CoreWeave's page at $6.50/hr flat. None of these carry a published reserved rate either; the 60% ceiling from Thunder Compute's review is the only public number for any of them.
Why CoreWeave Is Built for Enterprise Commitments, Not Hourly Renters
The Contract Math Behind CoreWeave's $104B Backlog
CoreWeave's revenue backlog hit $99.4 billion at the end of Q1 2026, up nearly 50% sequentially, with the company's CFO Nitin Agrawal reporting a weighted-average contract length for new capacity of "approximately five years" (Q1 2026 earnings call transcript). Roughly 36% of that backlog is expected to convert to revenue within two years, and about 75% within four, according to a Zacks analysis of the same call carried on TradingView. By the end of Q2 2026, the backlog had grown to $104.2 billion, up 246% year-over-year, on quarterly revenue of $2.6 billion that itself grew 112% year-over-year and 24% sequentially. CoreWeave raised its full-year 2026 revenue guidance to $12.4-13.2 billion on the back of it, and the stock jumped 13.16% in after-hours trading to $102.21 (Q2 2026 results coverage).
That backlog is built from named, multi-billion-dollar contracts. OpenAI's commitments now total up to roughly $22.4 billion across three expansions signed in 2025 (CoreWeave's own announcement). Meta's spend runs up to about $35.2 billion through December 2032, after an April 2026 expansion stacked on top of a September 2025 deal (24/7 Wall St. coverage of the Meta expansion). And NVIDIA itself is on the hook for up to $6.3 billion as a "backstop" buyer of unsold CoreWeave capacity through April 2032 (Motley Fool coverage of the NVIDIA backstop deal). That NVIDIA arrangement is also a vendor-financing structure, not just a customer relationship, and it's worth understanding if you're trying to figure out why a CoreWeave rate can move independently of demand for compute itself. Our piece on NVIDIA's neocloud backstop financing covers how those arrangements filter down into what you pay.
The practical read for a buyer: CoreWeave's actual GPU-hour sales are dominated by these negotiated, multi-year deals, not by whoever shows up and clicks "on-demand." The published rate card exists mostly as a ceiling and a reference point for negotiation, not the number most of CoreWeave's revenue is transacted at.
What "Largely Sold Out" Means for a New, Non-Contracted Customer
"We remain largely sold out of our 2026 capacity, with prices increasing across the board from Ampere to Hopper to Blackwell."
Nitin Agrawal, CFO of CoreWeave, Q1 2026 earnings call
CEO Mike Intrator said much the same on the same call: "We are virtually sold out. Likewise, we have placed our POs, we have secured the infrastructure." Read plainly, that means the on-demand and spot rates in the tables above describe pricing on whatever capacity is left over after multi-year contract customers are served, in a market where CoreWeave has been raising list prices, not cutting them. CoreWeave earned SemiAnalysis's Platinum ClusterMAX rating for a second consecutive ranking in November 2025, remaining the industry's sole Platinum-rated provider (CoreWeave's announcement of the rating), which explains part of the demand: it's a genuinely well-run cloud, not just a scarce one. But "well-run and largely sold out" is a hard combination for a new, non-contracted customer trying to book capacity on short notice.
Cost Comparison: CoreWeave vs Spheron and Other On-Demand Neoclouds
| GPU | CoreWeave On-Demand $/GPU-hr | CoreWeave Spot $/GPU-hr | Spheron On-Demand $/GPU-hr | Spheron Spot $/GPU-hr |
|---|---|---|---|---|
| H100 | $6.16 (8-GPU minimum) | $2.46 | $2.75 (1-GPU) | $2.20 |
| H200 | $6.31 (8-GPU minimum) | $2.62 | $4.79 (1-GPU) | $3.31 |
| B200 | $8.60 (8-GPU minimum) | $4.26 | $9.36 (1-GPU) | $5.34 |
| GB200 NVL72 | $10.50 (4-GPU minimum) | N/A | Reservation only, no live rate | N/A |
Spheron figures come from Spheron's live pricing as of 12 Aug 2026. On Hopper, the gap is straightforward: Spheron's H100 on-demand rate undercuts CoreWeave's by more than half, and it's available one GPU at a time instead of eight. B200 is the more honest, less flattering comparison. CoreWeave's per-GPU B200 rate ($8.60) is actually a bit cheaper than Spheron's single-GPU on-demand rate ($9.36), because CoreWeave amortizes across a full 8-GPU node. Run the math on a full 8-GPU B200 cluster for a month (720 hours) and CoreWeave comes out ahead: $68.80 × 720 = $49,536 versus Spheron's $9.36 × 8 × 720 = $53,913.60. The catch is you have to actually use all 8 GPUs to get there; use fewer, and CoreWeave still bills the full node while Spheron scales down with you.
For the full multi-provider view including AWS, GCP, Azure, and Nebius, see the GPU cloud pricing comparison 2026. Nebius specifically is the closer neocloud-to-neocloud comparison on Hopper pricing; our CoreWeave vs Nebius H100 and H200 breakdown covers that matchup in detail, including Nebius's shorter, sub-annual reservation terms against CoreWeave's five-year weighted average.
Pricing fluctuates based on GPU availability. The prices above are based on 12 Aug 2026 and may have changed. Check current GPU pricing → for live rates.
Who Should Actually Use CoreWeave vs a GPU Marketplace
This is a genuinely workload-dependent call, not a "marketplace always wins" pitch. Here's how we'd frame it:
CoreWeave makes sense if you:
- Can commit to a multi-year contract; the weighted-average new commitment runs around five years
- Need a full NVLink domain or NVL72 rack-scale deployment for distributed training, not single-GPU inference
- Have procurement and legal bandwidth to negotiate reserved pricing through an account manager rather than a self-serve checkout
- Value CoreWeave's Platinum ClusterMAX reliability rating enough to plan around its capacity constraints
A marketplace like Spheron makes sense if you:
- Need on-demand access at single-GPU granularity without an 8-GPU minimum spend
- Want per-minute billing and zero commitment instead of a sales cycle
- Are running inference, fine-tuning, or any workload under the scale where an 8-GPU node makes sense
- Need Blackwell capacity now and can't wait through CoreWeave's "largely sold out" 2026 posture; Spheron's GB200 NVL72 reservations and B200 on-demand instances are both live today
If CoreWeave's contract requirements specifically are the sticking point rather than the hardware, our CoreWeave alternatives roundup and the direct Spheron vs CoreWeave comparison go deeper on the non-pricing tradeoffs: Kubernetes requirements, egress, and lock-in. And if you're still weighing reserved against on-demand as a general question rather than a CoreWeave-specific one, on-demand vs reserved GPU pricing covers the utilization math that decides which one actually pays off. Deployment guides for getting a new instance running are at docs.spheron.ai.
If an 8-GPU minimum and a five-year contract don't fit your workload, on-demand access to H100, B200, and GB200 NVL72 at single-GPU granularity is the more direct route.
Rent GB200 NVL72 → | B200 on-demand pricing → | View all GPU pricing →
Frequently Asked Questions
On CoreWeave's current published rate card, an HGX H100 8-GPU node runs $49.24/hr ($6.16/GPU-hr), HGX H200 runs $50.44/hr ($6.31/GPU-hr), HGX B200 runs $68.80/hr ($8.60/GPU-hr), and GB200 NVL72 runs $42.00/hr per 4-GPU instance ($10.50/GPU-hr). But CoreWeave's CFO has said the company remains largely sold out of 2026 capacity, and the bulk of its revenue comes from multi-year committed contracts, so the rate card functions more as a ceiling than a price most customers actually pay.
CoreWeave prices GB200 NVL72 at $42.00/hr per instance, where each instance bundles 2 GB200 Superchips (4 Blackwell GPUs total, 186GB VRAM), working out to $10.50/GPU-hr. A full NVL72 rack (72 GPUs) would need 18 of those instances, or roughly $756/hr, about $544,320 for a continuous 30-day month. GB300 NVL72 isn't self-serve priced at all; CoreWeave routes it straight to a sales conversation.
CoreWeave advertises 'up to 60% discounts over on-demand prices for committed usage,' but doesn't publish which commitment lengths unlock which discount tier. Applied to the $6.16/GPU-hr on-demand H100 rate, a 60% reserved discount lands around $2.46/GPU-hr, which would actually undercut Spheron's $2.75/GPU-hr on-demand H100 rate, but only after a multi-year commitment. Spheron's $2.75 rate requires zero commitment and no negotiation. Getting an exact CoreWeave number requires a sales conversation, not a checkout page.
CoreWeave's HGX B200 8-GPU node runs $68.80/hr on-demand ($8.60/GPU-hr) and $34.11/hr on spot ($4.26/GPU-hr), the same 8-GPU-minimum structure as its H100 and H200 tiers. Spheron's B200 runs $9.36/hr on-demand and $5.34/hr spot at single-GPU granularity, so CoreWeave's per-GPU rate is actually a little cheaper if you genuinely need all 8 GPUs, while Spheron wins if you need fewer.
CoreWeave fits teams that can commit to a multi-year contract, need full NVLink or NVL72 rack-scale domains for distributed training, and have the procurement bandwidth to negotiate reserved pricing through an account manager. A marketplace like Spheron fits teams that need on-demand or spot access at single-GPU granularity, want to avoid a multi-year commitment, or are running inference and fine-tuning workloads that don't need an entire 8-GPU node.
