CoreWeave GPU Pricing 2026: Per-Hour Rates

Quick Answer

CoreWeave GPU pricing 2026 per-hour: HGX H100 8x $49.24 on-demand / $19.71 spot = $6.16/$2.46 per GPU; HGX B200 8x $68.80 on-demand / $34.11 spot = $8.60/$4.26 per GPU; GB200 NVL72 4-GPU $42.00 = $10.50 per GPU. Reserved contracts up to 60% off, sales-negotiated. 8-GPU minimum on most nodes.

Last verified: Sep 16, 2026.

At a glance

• HGX H100 8x: $49.24/hr on-demand / $19.71/hr spot = $6.16/$2.46 per GPU
• HGX H200 8x: $50.44/hr on-demand / $20.93/hr spot = $6.31/$2.62 per GPU
• HGX B200 8x: $68.80/hr on-demand / $34.11/hr spot = $8.60/$4.26 per GPU
• GB200 NVL72 4-GPU: $42.00/hr on-demand = $10.50 per GPU
• GH200 1-GPU: $6.50/hr on-demand (only single-chip option)
• A100 80GB 8x: $21.60/hr on-demand = $2.70 per GPU
• L40S 8x: $18.00/hr on-demand = $2.25 per GPU
• Reserved contracts: Up to 60% off on-demand, sales-negotiated
• Egress: Free

Why CoreWeave is the neocloud leader for 200+ GPU training jobs

CoreWeave charges the most per H100 hour among major neoclouds, and the money is not going into the GPU — it goes into InfiniBand fabric and large contiguous clusters.

CoreWeave is purpose-built for multi-node training at scale. The 8-GPU HGX nodes ship with NVLink 4 interconnect (900 GB/s per GPU) and the cluster fabric is InfiniBand — what stops multi-node training from choking on gradient synchronization. If your job fits on one node, you are paying for a network you are not using. If your job spans 40+ nodes, the per-GPU rate stopped being the number you optimize (cloudzero.com, September 2026).

CoreWeave's business model runs on the third tier — multi-year reserved contracts that the published rate card does not show. OpenAI has cumulative agreements of approximately $22.4 billion as of September 2025; Meta $21 billion through 2032. These deals sit at the negotiated reserved ceiling, not the published on-demand rate. For everyone else, the published rate card is the starting point of a sales conversation, not the final price (usagepricing.com, June 2026).

Hopper pricing (HGX H100 / HGX H200)

CoreWeave's HGX H100 is the most expensive published on-demand H100 rate among major providers — but the spot tier and reserved contracts close the gap.

Verified July 2026 (coreweave.com/pricing, NA region):

• HGX H100 8x on-demand: $49.24/hr node ($6.16/GPU-hr); spot $19.71/hr node ($2.46/GPU-hr); 1-yr reserved up to 60% off ($2.46/GPU-hr target)
• HGX H200 8x on-demand: $50.44/hr node ($6.31/GPU-hr); spot $20.93/hr node ($2.62/GPU-hr); 1-yr reserved up to 60% off
• GH200 1x on-demand: $6.50/hr node (only single-chip option)

Comparing to Lambda 1-Click Cluster reservation (HGX H100 16-GPU at $6.16/GPU-hr; 256-GPU at $5.54/GPU-hr), CoreWeave on-demand is at parity with Lambda reserved at 16-GPU scale. The gap closes once you negotiate CoreWeave reserved — at the 60% ceiling, CoreWeave reserved H100 is $2.46/GPU-hr, lower than Lambda's $5.54/GPU-hr 256-GPU reservation.

Spot rates on CoreWeave H100 are $2.46/GPU-hr (NA) and $2.44/GPU-hr (EU), which is comparable to AWS spot at $1.95-$2.50/GPU-hr and Vast.ai interruptible at $1.49/GPU-hr. Spot is interruptible — capacity holds when available, reclaim risk applies (cloudzero.com, September 2026).

Blackwell pricing (HGX B200 / HGX B300 / GB200 / GB300)

CoreWeave's Blackwell-generation rates are at the higher end of the published rate card — but reserved contracts and GB200 NVL72 access justify the premium for rack-scale Blackwell.

Verified July 2026 (usagepricing.com):

• HGX B200 8x on-demand: $68.80/hr node ($8.60/GPU-hr); spot $34.11/hr node ($4.26/GPU-hr)
• HGX B300 8x on-demand: Contact sales; spot $35.84/hr node ($4.48/GPU-hr)
• GB200 NVL72 4-GPU on-demand: $42.00/hr node ($10.50/GPU-hr); no spot
• GB300 NVL72 4-GPU on-demand: Contact sales; no spot published

GB200 and GB300 NVL72 instances bundle two Grace Blackwell Superchips (4 Blackwell GPUs plus 2 Grace CPUs each) — the rack-scale building block for Vera Rubin deployments. A full NVL72 rack is 72 GPUs (18 of those 4-GPU instances), which at $756/hr = $544,320/month at full rack utilization. For Vera Rubin memory and compute specs (13.5 TB HBM3e, 1.4 exaflops FP4 per rack), the price per useful FLOP is competitive.

HGX B300 and GB300 NVL72 are gated behind contact sales for on-demand access — a structural disadvantage for self-serve Blackwell capacity. Spot is available on HGX B300 but not on GB300 NVL72.

Older generation pricing (A100 / L40 / L40S / RTX PRO 6000)

CoreWeave's older-generation rates are competitive for budget-constrained workloads.

Verified July 2026 (cloudzero.com):

• A100 80GB 8x on-demand: $21.60/hr node ($2.70/GPU-hr); spot $9.65/hr ($1.21/GPU-hr)
• L40S 8x on-demand: $18.00/hr node ($2.25/GPU-hr); spot $7.88/hr ($0.99/GPU-hr)
• L40 8x on-demand: $10.00/hr node ($1.25/GPU-hr); spot $6.27/hr ($0.78/GPU-hr)
• RTX PRO 6000 Blackwell SE 8x: $20.00/hr node ($2.50/GPU-hr); spot $11.09/hr ($1.39/GPU-hr)

The L40 at $1.25/GPU-hr on-demand is CoreWeave's cheapest published rate. The A100 80GB at $2.70/GPU-hr is the cheapest production-training tier. The L40S at $2.25/GPU-hr is the right balance of cost and capability for inference and 3D rendering workloads (cloudzero.com, September 2026).

Reserved contracts and the 'largely sold out' 2026 posture

CoreWeave's reserved contracts deliver up to 60% off on-demand — but the company is 'largely sold out' of 2026 capacity.

CoreWeave's 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. The implication: on-demand rates are stable but reservations for the remainder of 2026 are constrained; reservations for 2027 are taking bookings now.

Applied to the published rate card: a 60% reserved discount on the $6.16/GPU-hr on-demand H100 rate would land near $2.46/GPU-hr — the same as CoreWeave's spot rate with zero commitment. The real tension in CoreWeave's pricing is that 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 (spheron.network, August 2026).

For buyers willing to negotiate: the published on-demand rate card is the ceiling, not the floor. Sales engagement delivers meaningfully lower rates for committed multi-year volume, but the discount tiers by term length and GPU generation are unpublished — every customer negotiates separately.

Side-by-side CoreWeave vs hyperscaler and neocloud pricing

How to choose CoreWeave vs alternatives in 2026

The decision hinges on cluster scale, training duration, and budget tier.

1. Single-GPU workloads under 1 week: Skip CoreWeave. The 8-GPU minimum makes CoreWeave the wrong fit. Use RunPod Secure Cloud, Lambda, or Modal.
2. 8-32 GPU fine-tuning runs over 1-2 weeks: CoreWeave spot H100 at $2.46/GPU-hr is competitive. Accept interruptibility; checkpoint every 30 minutes.
3. Multi-month 200+ GPU training you can't afford to lose: CoreWeave reserved. Negotiate a multi-year contract with committed-volume discount. Target $2.46-$3.50/GPU-hr on H100 with 30-50% off on-demand.
4. Vera Rubin NVL72 rack-scale: CoreWeave GB200 NVL72 at $42.00/hr per 4-GPU instance ($10.50/GPU-hr) is the published path. Full 72-GPU rack at $756/hr = $544,320/month.
5. Budget-constrained workloads under 8 GPUs: Skip CoreWeave entirely. Lambda, RunPod, and Vast.ai deliver single-GPU rates that CoreWeave cannot match.

What enterprise buyers should do next

Three actions for organizations evaluating CoreWeave in 2026.

1. Engage sales before committing to on-demand. CoreWeave's published rate card is the starting point of a sales conversation, not the final price. Reserved contracts for committed multi-year volume typically deliver 30-60% off on-demand. For 200+ GPU workloads running 6+ months, reserved pricing is mandatory to avoid paying the published on-demand premium.
2. Model the 8-GPU minimum into your utilization math. If your workload uses 3 of 8 GPUs, you still pay for 8. Use the published on-demand rate card as the cost ceiling and negotiate from there. For workloads that genuinely need 8+ GPUs full-time, CoreWeave is competitive; for workloads that don't, the 8-GPU minimum is a structural disadvantage.
3. Take advantage of free egress. For data-intensive workloads, CoreWeave's no-egress policy saves $870-$1,200/month versus AWS. CoreWeave + S3-compatible storage is a meaningful TCO advantage over hyperscaler deployments at scale.

What to watch next

Three near-term datapoints. First, CoreWeave's Q3 2026 earnings call — the company has signaled list price increases across every GPU generation as 2026 capacity sells out. Expect H100 on-demand to push from $6.16 to $6.50-$7.00/GPU-hr by Q1 2027. Second, CoreWeave's reserved capacity for 2027 — Q4 2026 is when most reserved contracts for 2027 are booked; if you need capacity, engage sales now. Third, Vera Rubin GPU production ramping at CoreWeave, Google, Azure, OCI, and Nebius will relieve Hopper demand pressure; H100 spot prices will fall faster than reserved rates as Vera Rubin capacity comes online (NVIDIA Q2 FY27 earnings, August 26, 2026).