CoreWeave vs Crusoe: A GPU Cloud Pricing Comparison 2026 for H100 and B200

Run a GPU cloud pricing comparison 2026 between CoreWeave and Crusoe and you'll find two opposite ideas about what a buyer should commit to before they even get a GPU. Crusoe prices H100 at $3.90/GPU-hr, checked out one GPU at a time, no minimum term required. Both companies are chasing the same large-cluster AI buyer. They've just built completely different pricing philosophies to get there.

The gap holds for H200 too, and then partially inverts at B200: Crusoe won't publish a number for it at all, while CoreWeave will, just at a price most single-GPU teams won't want to pay. This post lines up the full rate cards, the billing models behind them, and where each company's capacity actually sits going into the back half of 2026. For the deep-dive on either provider standing alone, see our CoreWeave H100 and H200 pricing breakdown and Crusoe Cloud pricing 2026 guide.

CoreWeave vs Crusoe at a Glance: A GPU Cloud Pricing Comparison 2026 Snapshot

CoreWeave publishes both an on-demand and a spot number for H100 and B200. Crusoe publishes an on-demand number for both, but its spot column reads "Contact sales" on every GPU it sells, H100 and B200 included. Neither structural gap is an accident, and both matter more than the headline sticker price once you know how much of the cluster you actually need.

Pricing fluctuates based on GPU availability. The prices above are based on 30 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

Rate Card Comparison: On-Demand, Reserved, and Cluster Minimums

Both providers publish self-serve pricing for most of their catalogs, which puts this comparison a step ahead of the usual "contact sales" fog that covers most of the GPU cloud market. Where they diverge is in what unit you're actually allowed to buy.

CoreWeave's 8-GPU-Only Node Pricing

CoreWeave's own pricing page lists HGX H100 and HGX H200, both metered as a single 8-GPU node with no smaller unit listed at any price.

There is no smaller on-demand compute unit to buy at any tier. A single-GPU inference deployment run as raw compute and an 8-GPU distributed training job pay the identical node rate, because nothing below an 8-GPU node appears on CoreWeave's on-demand or spot table. If your real utilization is 2 GPUs, you're still billed for 8.

Crusoe's Per-GPU Published Rate Card

Crusoe publishes on-demand rates per GPU: H100 (80GB HGX) at $3.90/GPU-hr, H200 (141GB HGX) at $4.29/GPU-hr, A100 SXM at $2.30/GPU-hr, A100 PCIe at $2.00/GPU-hr, L40S at $1.50/GPU-hr, and AMD MI300X at $3.45/GPU-hr. Every one of those is a normal self-serve checkout, not a sales exception, and every one is billed at single-GPU granularity.

Then the pattern breaks on the newest silicon. The provider whose whole pitch is a transparent per-GPU rate card goes quiet, and CoreWeave, the one gating everything behind an 8-GPU floor, is actually the one with a number you can read without picking up the phone.

Spot Pricing: Where the Two Providers Swap Places

Flip the transparency question around and CoreWeave and Crusoe trade positions.

Crusoe runs the opposite pattern. Its own pricing page carries a "Current spot" column next to every GPU model, and every row under that column reads "Contact sales," H100 and H200 included, even though both carry a published on-demand rate one column over.

So the discount lever that has a public number attached runs through different mechanics on each side. On CoreWeave, spot is already sitting on the rate card; a reserved capacity plan is the larger lever behind it. On Crusoe, spot access itself requires a sales conversation, same as a reservation.

Reserved and Committed Discounts

Crusoe's reservation model works differently. A reservation "represents an active reserved instance agreement, typically for a fixed quantity of a specific instance type (e.g. 128 A100-80GB GPUs) valid for a defined period of time," negotiated with sales. Neither company publishes a specific reserved rate card, so getting an actual number on either side still means a sales conversation. For a broader framework on when a commitment beats staying on-demand, see gpu cluster reservation and contract negotiation and serverless vs on-demand vs reserved GPU pricing.

Billing Model: Flat-Rate Commitments vs Flexible Single-GPU Access

The rate-card gap gets most of the attention, but the billing model underneath it is the part that actually decides what you pay for a workload that doesn't map cleanly onto a full node or a full month.

CoreWeave's Reserved Instance Model

CoreWeave's capacity-plan documentation is explicit about how its Reserved Instances work: "You pay a flat rate for every hour in the billing period, regardless of whether instances are running or idle." Sign a term and the meter runs whether you're training around the clock or the cluster is sitting empty overnight. CoreWeave also offers Flex Reservations, a two-part structure with a baseline holding rate that guarantees the capacity plus a separate usage rate that applies only while nodes are actively running, which is closer to a hedge against the flat-rate model than an escape from the 8-GPU floor underneath both options.

Crusoe's On-Demand Model

Crusoe's pricing page lists every GPU model against a single per-GPU dollar figure, with no unit-count column at all: pick H100, and the $3.90/GPU-hr rate is the same whether you're checking out one GPU or fifty. You check out a GPU, run it as long as the job needs, and stop. There's no node you have to fill and no term you have to justify before you start.

What the Idle-Time Gap Actually Costs

Take a 2-GPU embedding workload that needs to run continuously for a month, 720 hours. CoreWeave has no unit smaller than 8 GPUs, so you pay for the full H100 node regardless of how many of the 8 you actually touch, whether you're using 2 GPUs or all 8. Crusoe sells H100 one GPU at a time, so the same workload costs 2 x $3.90 x 720 = $5,616/month, a gap of roughly $29,837 a month that has nothing to do with the per-GPU rate and everything to do with CoreWeave's node floor.

Now flip to a committed CoreWeave Reserved Instance instead of on-demand. Reserve an 8-GPU H100 node and only actually occupy it 30 hours a week, nights, weekends, and idle time between training runs included, and CoreWeave's own documentation says you still pay for all 168 hours in the week, because the flat rate applies "regardless of whether instances are running or idle." Crusoe's on-demand model, and Spheron's per-minute billing, only charge for the hours you actually run the job.

Where Each Provider's Capacity Actually Sits Today

Pricing philosophy aside, the two companies are also at different points in their own build-out, and that shapes how much negotiating room you actually have on either side right now.

CoreWeave: Sold Out, Raising Prices, Platinum-Rated

CoreWeave's leadership has said its near-term inventory is tight for two straight quarters. On the Q1 2026 earnings call, CFO Nitin Agrawal told investors, "we pretty much remain sold out for our 2026 capacity. That is continuing to be true for us." By the Q2 2026 earnings call on August 11, CEO Michael Intrator said "pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs, while pricing for prior generation SKUs is at or above where it was years ago." Our CoreWeave price increase 2026 breakdown covers the SKU-by-SKU detail; notably, the public on-demand rate card in the table above hadn't moved as of this post, since the pricing pressure Intrator described is landing mainly in negotiated and reserved contracts.

SemiAnalysis's ClusterMAX 2.0 rating puts CoreWeave alone at the top: "CoreWeave retains top spot as the only member of the Platinum tier."

Crusoe: Nearly 5 Gigawatts Contracted, Gold-Rated

Crusoe announced on June 9, 2026 that its total development pipeline, "inclusive of its contracted projects, sites under active tenant negotiation, and sites in advanced development," exceeds 40 GW. SemiAnalysis's same ClusterMAX 2.0 rating places Crusoe one tier below CoreWeave: "Nebius, Oracle and Azure are the top providers within the Gold tier. Crusoe and new entrant Fluidstack also achieve Gold tier." Gold isn't a knock, it sits directly beneath the single Platinum slot CoreWeave occupies alone, and both providers are being taken seriously by the same enterprise buyers evaluating multi-year AI infrastructure commitments.

CoreWeave vs Crusoe: Which Fits Your Workload in This GPU Cloud Pricing Comparison 2026

CoreWeave fits if: you're running distributed training that genuinely needs the full 8-GPU NVLink domain, you're comfortable negotiating a multi-year contract to get anywhere near the advertised 60% discount ceiling, and Platinum ClusterMAX reliability matters more in procurement than the per-GPU sticker price.

Crusoe fits if: you need H100 or H200 access at single-GPU granularity without an enterprise sales cycle, a zero-minimum-commitment checkout genuinely matters for how your workload runs, and you can live without a published B200 number until you're ready to talk to sales anyway.

Neither fits if: you want single-GPU access to B200 today, per-minute billing with no minimum term, and no sales call required to see a live rate at any tier. That's the gap between them Spheron is built to sit in.

Where Spheron Fits Between the Two

Spheron's own pricing page advertises "Live rates across 5+ providers. No lock-in," billed with "per-minute billing granularity, so you only pay for the exact time you use," and states plainly there is "no minimum rental period." H100 is live at $2.65/hr on-demand and $2.04/hr spot; B200, the GPU Crusoe won't put a number on, is live at $7.20/hr on-demand and $3.90/hr spot, self-serve, no sales call. Where CoreWeave forces an 8-GPU floor even for a single-GPU inference job, Spheron sells the GPU count your workload actually needs, and Crusoe's own $3.90/GPU-hr H100 rate is still cheaper than Spheron's live on-demand rate today, a gap worth acknowledging plainly rather than hiding.

Where Spheron doesn't compete: CoreWeave is the better fit for distributed training that genuinely needs a full 8-GPU NVLink domain at Platinum-tier reliability and a multi-year contract. Crusoe's negotiated reserved contracts, backed by nearly 5 gigawatts of contracted capacity, may beat any on-demand number here for a team running real scale over a committed term. For the direct numbers against each provider individually, see Spheron vs CoreWeave and Spheron vs Crusoe. For the full 15+ provider price table including AWS, GCP, Azure, and every major neocloud, see our broader GPU cloud pricing rundown, and for a similar head-to-head against a third neocloud, CoreWeave vs Nebius breaks down that pairing the same way. Deployment guides for standing up an instance are at docs.spheron.ai.

Pricing fluctuates based on GPU availability. The prices above are based on 30 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

An 8-GPU floor on one side and a quote-only B200 on the other don't fit every workload. Single-GPU access at published, per-minute rates across the full catalog is the more direct route for most teams.

Check H100 availability → | B200 GPU pricing → | View all GPU pricing →