CoreWeave H100 and H200 Pricing 2026: Cost Per GPU-Hour vs Spheron

CoreWeave's current on-demand rate for an NVIDIA HGX H100 8-GPU node is $49.24/hr, which works out to $6.16 per GPU-hour (coreweave.com/pricing). H200 runs $50.44/hr per node, or $6.31/GPU-hour. Both are sold exclusively as 8-GPU bundles, there is no single-GPU option, and CoreWeave's own CFO told investors in Q1 2026 that the company is "largely sold out" of 2026 capacity while raising prices across every GPU generation it sells. This post breaks down what CoreWeave actually charges per GPU, why its pricing model forces an 8-GPU minimum, why its own site shows two different H100 numbers depending on which page you land on, and how the real per-GPU rate compares to Spheron's H100 and H200 pricing.

TL;DR: CoreWeave vs Spheron H100 and H200 (Jul 2026)

CoreWeave's on-demand H100 rate of $6.16/GPU-hour is more than 50% higher than Spheron's H100 SXM5 rate of $3.92/GPU-hour, and more than 3x Spheron's H100 PCIe rate of $2.01/GPU-hour. The gap holds on spot too: CoreWeave H100 spot at roughly $2.46/GPU-hour is close to Spheron's H100 SXM5 spot rate of $1.43/GPU-hour, but Spheron's rate applies to a single GPU while CoreWeave's minimum purchase is eight.

Pricing fluctuates based on GPU availability. The prices above are based on 5 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

CoreWeave's Published Pricing: HGX H100 and H200 Nodes, Per-GPU Rates

CoreWeave sells H100 and H200 exclusively as NVIDIA HGX 8-GPU nodes on its current pricing page. Each node bundles 128 vCPUs, 2,048GB of system RAM, and 61.44TB of local storage alongside the GPUs, and pricing is identical whether you provision in North America or Europe for on-demand.

On-Demand Per-GPU Rates: Why CoreWeave Prices by 8-GPU Node, Not Per-GPU

The math is simple division: a $49.24/hr H100 node divided across 8 GPUs is $6.16/GPU-hour, and a $50.44/hr H200 node divided across 8 GPUs is $6.31/GPU-hour. What matters more than the arithmetic is what it means for anyone who doesn't need all eight GPUs at once. A single-GPU inference workload, a small fine-tuning job, or a debugging session all get billed at the full node rate on CoreWeave's current tier because there's no smaller unit to buy.

That's a defensible design choice if your workload is genuinely a distributed training job that needs the full NVLink domain across all eight GPUs. It's a bad fit for the much larger set of teams running single-model inference or fine-tuning under 70B parameters, who end up paying for seven idle GPUs to get the one they need. Spheron rents H100 SXM5 by the single GPU starting at $3.92/hr on-demand, and H100 PCIe at $2.01/hr, so a one-GPU inference workload pays for one GPU. If you're still deciding between the two GPU generations for your workload, see NVIDIA H100 vs H200 for the hardware-level tradeoffs.

Spot Pricing on CoreWeave (H100 and H200)

CoreWeave's spot rate for the 8-GPU H100 node is $19.71/hr in North America and $19.51/hr in Europe, working out to $2.44-2.46/GPU-hour. That's roughly a 60% discount off the on-demand node rate. H200 spot runs $20.93/hr (North America, $2.62/GPU-hour) and $20.64/hr (Europe, $2.58/GPU-hour), a similar 58-59% discount off on-demand.

Spot capacity on CoreWeave, like spot anywhere, can be reclaimed without notice, and it's still sold only in 8-GPU units. Spheron's H100 SXM5 spot starts at $1.43/hr per GPU, available at single-GPU granularity, with H200 SXM5 spot at $1.76/hr per GPU.

Reserved and Committed-Use Discounts (Contact Sales Only)

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. CoreWeave doesn't publish the specific rate tiers or which commitment lengths unlock which discount; getting an actual number requires talking to an account manager. Applied to the $6.16/GPU-hour on-demand H100 rate, a 60% reserved discount would land around $2.46/GPU-hour, which is still above Spheron's $2.01/hr H100 PCIe on-demand rate with zero commitment.

What CoreWeave Doesn't Advertise

Two Pricing Pages, Two Different Numbers

CoreWeave still runs a legacy "Classic" pricing page at coreweave.com/pricing/classic that lists individual NVIDIA H100 PCIe at $4.25/hr and HGX H100 at $4.76/hr, both with 80GB VRAM. Neither number matches the $6.16/GPU-hour on the current pricing page for what is nominally the same GPU generation.

This isn't a typo. It's two live pricing surfaces on the same domain quoting materially different rates for H100 compute. If you're citing a CoreWeave price you saw in a comparison post, a forum thread, or an older review, check which page it came from. A lot of the "$4.76/hr CoreWeave H100" figures still circulating online trace back to the Classic page, not the rate CoreWeave actually quotes new customers today on its primary pricing page.

No Single-GPU Access: The 8-GPU Minimum

There's no way to provision a single H100 or H200 from CoreWeave's current pricing tier as a self-serve customer. The smallest unit is the 8-GPU HGX node. For teams whose actual workload is a single-GPU inference deployment, that means either overpaying for seven unused GPUs or working through an account manager to see if a smaller carve-out is available at all, with no published rate for it.

Average Contract Length and Prepayment

CoreWeave's business model runs on multi-year capacity commitments, not short-term rentals. On the company's Q1 2026 earnings call, CFO Nitin Agrawal reported a "weighted average contract length for new capacity remaining at approximately 5 years," and the company's contracted revenue backlog reached $99.4 billion, up nearly 50% sequentially and close to 4x year-over-year, with 36% of that backlog expected to be recognized within 24 months (Investing.com earnings call transcript). That scale of forward commitment is how CoreWeave finances its data center buildout, and it's also why the published on-demand and spot rates above exist mostly as a ceiling: the bulk of CoreWeave's actual GPU-hours are sold under negotiated, multi-year contracts, not at the rate a small team signing up today will get quoted.

"Largely Sold Out": What CoreWeave's Own Earnings Calls Say About 2026 Availability

"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 echoed the same point 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, not a market CoreWeave is actively trying to fill. Rising list prices alongside a "sold out" posture is a strong signal that CoreWeave's near-term capacity for new, non-contracted customers is tight, and that on-demand rates are more likely to climb than fall through the rest of 2026.

CoreWeave vs Spheron: Same-Tier GPU Cost Per Hour, On-Demand and Spot

Full Multi-Provider Table (CoreWeave, Spheron, Lambda, RunPod, Nebius)

Spheron figures below come from Spheron's live pricing API as of 5 Jul 2026.

CoreWeave is the most expensive on-demand H100 and H200 option in this table, both in absolute dollars and on a per-GPU basis, and the only one that refuses to sell fewer than 8 GPUs at a time. For the full 15+ provider table including AWS, GCP, Azure, and OCI, see the GPU cloud pricing comparison 2026. For how CoreWeave's H100 rate stacks up against AWS's P5 instances specifically, see AWS H100 pricing 2026, where AWS P5 on-demand runs $6.88/GPU-hour, higher than CoreWeave's $6.16 but in the same neighborhood: both hyperscaler-adjacent and priced well above the neocloud median.

Pricing fluctuates based on GPU availability. The prices above are based on 5 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Worked TCO Example (8-GPU Cluster, 720 Hours/Month)

H100, continuous on-demand, 30 days (720 hours):

• CoreWeave (8-GPU node, on-demand): $49.24 × 720 = $35,452.80/month
• CoreWeave (8-GPU node, spot): $19.71 × 720 = $14,191.20/month
• Spheron (8x H100 SXM5, on-demand): $3.92/GPU × 8 × 720 = $22,579.20/month
• Spheron (8x H100 SXM5, spot): $1.43/GPU × 8 × 720 = $8,236.80/month

H200, continuous on-demand, 30 days (720 hours):

• CoreWeave (8-GPU node, on-demand): $50.44 × 720 = $36,316.80/month
• CoreWeave (8-GPU node, spot): $20.93 × 720 = $15,069.60/month
• Spheron (8x H200 SXM5, on-demand): $3.70/GPU × 8 × 720 = $21,312.00/month
• Spheron (8x H200 SXM5, spot): $1.76/GPU × 8 × 720 = $10,137.60/month

Run an 8-GPU H100 cluster continuously for a month on CoreWeave on-demand and you're at $35,452.80. The same 8 GPUs on Spheron's on-demand tier run $22,579.20, a savings of $12,873.60 a month, before either side negotiates a discount. That gap alone covers a second small cluster.

The H200 spot comparison is even starker: CoreWeave's 8-GPU H200 spot node runs $15,069.60/month, while the same 8 GPUs on Spheron's H200 SXM5 spot tier run $10,137.60/month, a savings of nearly $5,000 a month on the cheapest tier either provider offers.

Hidden Costs and Billing Structure Differences

Per-hour billing. CoreWeave bills in hourly increments. A 40-minute job on an 8-GPU H100 node still costs a full hour at $49.24. Run ten such jobs in a day and you pay for ten hours while using roughly 6.7. Spheron bills per minute, so a 40-minute job costs 40 minutes.

No published storage or egress rates on the current pricing page. CoreWeave's node bundle includes 61.44TB of local storage, but persistent volume storage, object storage, and networking add-ons for production workloads are quoted separately through sales rather than listed on the public pricing page. Budget for a follow-up conversation before you commit to a number.

The 8-GPU floor is itself a hidden cost. If your actual utilization is 2 GPUs, CoreWeave's current tier still bills for 8. That's not a fee line item, it's a structural cost baked into the minimum unit of purchase, and it's the single biggest practical difference between CoreWeave and providers that rent by the individual GPU.

Reserved pricing requires a sales conversation with no published floor. "Up to 60% off" tells you the ceiling, not the price you'll actually get quoted, and CoreWeave's own earnings commentary about "largely sold out" 2026 capacity plus rising prices across GPU generations suggests that ceiling has been getting harder to reach. For more on when a reserved commitment actually pays off versus staying on-demand, see on-demand vs reserved GPU pricing, and for the financing structure behind CoreWeave's multi-year capacity contracts, see our piece on NVIDIA's neocloud backstop financing.

Which Provider Fits Your Workload

CoreWeave makes sense if: you're running a distributed training job that genuinely needs a full 8-GPU NVLink domain, you can commit to a multi-year contract (the weighted average new commitment runs around 5 years), and you have the procurement bandwidth to negotiate reserved pricing through an account manager rather than a self-serve checkout.

Spheron makes sense if: you need H100 or H200 access at single-GPU granularity, you want on-demand rates without a multi-year lock-in, or your workload is inference, fine-tuning, or anything under the scale where an 8-GPU minimum makes sense. Spheron's H100 SXM5 on-demand at $3.92/GPU-hour and H100 PCIe at $2.01/GPU-hour both undercut CoreWeave's $6.16/GPU-hour node rate, and H200 SXM5 on-demand at $3.70/GPU-hour beats CoreWeave's $6.31/GPU-hour by more than 40%.

If CoreWeave's pricing and contract requirements specifically are the issue rather than the hardware, our CoreWeave alternatives roundup and the direct Spheron vs CoreWeave comparison cover more of the tradeoffs beyond pricing. Deployment guides for connecting to a new instance are at docs.spheron.ai.

FAQ

How much does CoreWeave charge per H100 GPU per hour in 2026?

CoreWeave's current on-demand rate for an NVIDIA HGX H100 8-GPU node is $49.24/hr in both North America and Europe, or $6.16 per GPU-hour. There's no single-GPU option; you pay the full node rate regardless of how many GPUs you actually use. Spheron's H100 SXM5 starts at $3.92/hr on-demand for a single GPU, and H100 PCIe runs $2.01/hr.

What does CoreWeave charge for H200 GPUs?

CoreWeave prices its HGX H200 8-GPU node at $50.44/hr on-demand, or $6.31 per GPU-hour. Spot pricing on the same node is $20.93/hr in North America ($2.62/GPU-hour) and $20.64/hr in Europe ($2.58/GPU-hour). Spheron's H200 SXM5 runs $3.70/hr on-demand and $1.76/hr spot per GPU.

Why does CoreWeave only sell H100 and H200 in 8-GPU bundles?

CoreWeave's current pricing tier lists HGX H100 and H200 exclusively as 8-GPU nodes with 128 vCPUs, 2,048GB RAM, and 61.44TB local storage bundled in. There's no self-serve single-GPU option. It's a reasonable fit for distributed training that needs the full NVLink domain, but it means single-GPU inference or fine-tuning workloads pay for capacity they don't use. Spheron rents both GPUs by the single unit as well as in multi-GPU nodes.

Does CoreWeave offer spot pricing for H100 and H200?

Yes. CoreWeave's 8-GPU H100 node spot rate is $19.71/hr in North America and $19.51/hr in Europe, roughly $2.44-2.46/GPU-hour, about a 60% discount off on-demand. H200 spot runs $20.93/hr (North America) and $20.64/hr (Europe). Both remain 8-GPU-only and can be reclaimed without notice. Spheron's H100 SXM5 spot starts at $1.43/hr per GPU at single-GPU granularity.

Why do CoreWeave's two pricing pages show different H100 prices?

CoreWeave's legacy Classic pricing page lists individual H100 PCIe at $4.25/hr and HGX H100 at $4.76/hr, both well below the $6.16/GPU-hour quoted on the current pricing page for the same hardware generation. If a CoreWeave price you've seen elsewhere looks low, check whether it's sourced from the Classic page rather than what CoreWeave quotes new customers today.

CoreWeave's $6.16/GPU-hour H100 rate and $6.31/GPU-hour H200 rate are the highest on-demand numbers in this comparison, and they come with an 8-GPU minimum and no published single-GPU option. Add a company that's telling its own investors it's "largely sold out" of 2026 capacity while raising prices, and the case for shopping the rate elsewhere gets stronger, not weaker.

If an 8-GPU minimum and multi-year contract math don't fit your workload, single-GPU H100 and H200 access with per-minute billing is the more direct route.

Check H100 pricing on Spheron → | H200 SXM5 availability → | View all GPU pricing →