Crusoe Cloud Pricing 2026: H100 and B200 Cost Per Hour vs Spheron

Crusoe is the rare neocloud that actually publishes most of its GPU rate card instead of hiding everything behind "contact sales." H100, H200, A100, and L40S all have listed on-demand prices right on crusoe.ai/cloud/pricing. B200, GB200, and MI355X don't. That split is the story of this post: what Crusoe tells you upfront, what it makes you ask for, and how the published numbers actually compare against Spheron at the same GPU tier.

We're not going to cherry-pick the comparison. Right now, Crusoe's H100 on-demand rate is genuinely cheaper than Spheron's single-GPU H100 SXM5 on-demand tier. Spheron's spot pricing and its 8-GPU H100 NVL bundle both undercut Crusoe substantially, but the single-GPU on-demand comparison favors Crusoe today. If you want the honest math instead of a rigged one, this is it. For the broader feature and sustainability comparison between the two platforms, see our Spheron vs Crusoe breakdown.

Crusoe Cloud Pricing 2026: The Published Rates

Crusoe bills every GPU and CPU instance by the minute, and the company states plainly that it does not charge for network ingress or egress, inside a VPC or to the public internet (Crusoe Cloud pricing). That's a real advantage over hyperscalers that meter every gigabyte leaving their network. There are also no upfront setup fees for on-demand instances.

On-Demand: H100, H200, A100 Per-Hour Cost

Here's the published on-demand rate card, per GPU:

Source: Crusoe Cloud pricing page. H200 costs about 10% more per hour than H100, which tracks with the extra HBM3e capacity and bandwidth. If your workload needs the memory headroom for longer context windows or larger batch sizes, that premium is worth paying. If it doesn't, H100 remains the cheaper published tier. See our H100 vs H200 comparison for the throughput math behind that decision.

Pricing fluctuates based on GPU availability. The prices above are based on 11 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

B200 and GB200: Why There's No Listed Price

This is the section of Crusoe's pricing page that tells you the most by saying the least. B200 (180GB HGX), GB200 (186GB NVL72), and AMD's MI355X (288GB) all sit under "Contact sales" with no hourly figure attached (Crusoe Cloud pricing).

That's not unusual. Supply-constrained Blackwell-generation silicon gets routed through sales processes across most of the neocloud market, because allocation is scarcer than demand and providers want to match capacity to committed customers rather than publish a rate that might not hold. Crusoe's own newsroom points at why: the company's contracted AI infrastructure capacity reached roughly 4.9 gigawatts as of June 2026, against a development pipeline exceeding 40 GW, chasing what McKinsey estimates could be a 156 GW AI data center market by 2030 (Crusoe capacity announcement). Flagship campuses like the 1.2 GW Abilene, Texas site built for Oracle and a second 900 MW Abilene campus for Microsoft are built for exactly this kind of committed, large-scale demand, not walk-up self-serve capacity.

"The demand from the world's leading technology companies for AI infrastructure, quickly and at scale, has never been greater, and Crusoe is uniquely positioned to meet it," said Chase Lochmiller, Crusoe's co-founder and CEO, in that announcement.

The practical upshot for a smaller team: if you want B200 today without a sales cycle, Crusoe isn't where you get it. Spheron lists a live, self-serve B200 rate right now (see below), which is the more direct path if you need Blackwell-generation hardware without a procurement conversation. For a wider view of who else has published B200 rates, see our NVIDIA B200 cloud pricing across providers.

What's Not on the Pricing Page

The published rate card covers hourly cost. It doesn't cover contract terms, region availability, or what happens when you need hardware Crusoe hasn't put a price tag on. Those three gaps matter more than the sticker price if you're actually planning a deployment.

Reserved Capacity Terms: 6 Months to 3 Years, No Rollover

Crusoe's reservations are negotiated directly with sales rather than self-serve. According to Crusoe's own documentation, a reservation is "a fixed quantity of a specific instance type... valid for a defined period of time" (Crusoe Cloud reservations overview). Two terms in that doc matter more than the headline discount:

• Unused capacity does not roll over. Reservations aren't averaged over the billing month. If you reserve 8 GPUs and only use 5 for a stretch, you don't get credited or carry the unused 3 forward.
• Overage bills at on-demand (or your negotiated) rate. Usage above your reserved quantity is billed separately, calculated in the moment rather than smoothed across the month, so a short usage spike above your reservation triggers on-demand charges for the excess right away.

Crusoe doesn't publish exact contract lengths on the reservations doc. Industry reporting on neocloud reserved contracts generally puts terms in the six-month-to-three-year range, and that's consistent with how most GPU clouds structure committed capacity: short enough to match a training run, long enough that the provider can plan around it. The discount you actually get depends entirely on what you negotiate.

If you're trying to figure out whether a reserved contract or on-demand billing fits your workload better, the break-even math is the same one we've run for other providers: our serverless vs on-demand vs reserved GPU guide walks through the utilization threshold where reserved capacity starts paying for itself.

Region Limits: Crusoe's Five Live Cloud Regions

Crusoe currently runs five public cloud regions, each tied to a specific energy story (Crusoe Cloud locations):

Five regions is a narrow footprint next to a hyperscaler's dozens, but geographic sprawl was never Crusoe's pitch. The company built its business on stranded and clean energy sourcing rather than global reach, and each region above reflects that. If your workload needs a specific compliance region a hyperscaler covers and Crusoe doesn't, that's worth checking before you commit. If clean-energy sourcing per region matters for your own reporting, Crusoe is one of the few providers that documents it region by region.

The Sales-Quote Gap for B200, GB200, and MI355X

Getting access to Crusoe's newer silicon means going through the "Contact sales" or "Get a Quote" flow on crusoe.ai/cloud rather than a self-serve checkout. That's a real gap for teams that want to spin up a B200 instance today and compare the exact hourly cost against other providers before committing. You can request a quote without an existing relationship, but there's no public rate to shop against; you find out the number after you've already started the sales conversation.

Spheron's model is the inverse: B200 pricing is live on the pricing page right now, so you can compare exact numbers before talking to anyone.

Crusoe vs Spheron: Same-Tier Cost Comparison

This is where we stop talking about what's published and start comparing actual numbers, tier for tier. Spheron's rates below are pulled live from the Spheron GPU pricing API as of 11 Jul 2026, split strictly between on-demand (DEDICATED) and spot offers so we're not quoting a spot rate as if it were on-demand.

H100 SXM, Same Tier: Live Rates Compared

At the single-GPU on-demand tier, Crusoe's $3.90/hr is cheaper than Spheron's $4.41/hr, about 12% lower. That's the honest read: Crusoe wins this specific matchup right now. Spheron's H100 SXM5 spot rate at $2.94/hr flips it, running 25% below Crusoe's on-demand price, and the 8-GPU H100 NVL bundle at $2.06/GPU-hr on-demand is 47% cheaper still, though that bundle only makes sense if your workload can actually use all 8 GPUs. H100 GPU rental on Spheron lists all three tiers with live availability.

Pricing fluctuates based on GPU availability. The prices above are based on 11 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

B200: Published Self-Serve Rate vs Contact-Sales

There's no rate to compare here because Crusoe doesn't publish one. What you can compare is process: Spheron's B200 GPU rental page shows a live on-demand and spot rate today, no sales call required. If Crusoe's B200 quote comes in lower for your volume, that's a legitimate reason to pick Crusoe, but you won't know until you ask. Spheron's number is right there.

Real-World Monthly Cost at 200 Hours/Month

200 hours a month is a reasonable stand-in for a team running GPUs on weekdays during business hours, a common pattern for iterative fine-tuning and inference load testing. Single H100, on-demand rates:

• Crusoe on-demand: $3.90 × 200 = $780/month
• Spheron on-demand (SXM5): $4.41 × 200 = $882/month ($102/month more than Crusoe)
• Spheron spot (SXM5): $2.94 × 200 = $588/month ($192/month less than Crusoe on-demand)
• Spheron H100 NVL bundle, per GPU: $2.06 × 200 = $412/month (only economical if you need the full 8-GPU bundle)

Same math on H200: Crusoe's $4.29/hr on-demand comes to $858/month at 200 hours, cheaper than Spheron's $5.92/hr on-demand ($1,184/month), but Spheron's H200 spot at $3.36/hr ($672/month) runs 22% below Crusoe's on-demand rate. And on A100, the gap reverses again: Spheron's A100 SXM4 on-demand at $1.80/hr ($360/month) beats Crusoe's A100 SXM at $2.30/hr ($460/month) outright, no spot discount needed.

None of these numbers stay fixed for long. Spot and on-demand rates on Spheron move with real-time availability across 5+ providers, and Crusoe's published rates can change too. Run the comparison yourself before you commit budget.

Pricing fluctuates based on GPU availability. The prices above are based on 11 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Which Pricing Model Fits Your Workload

The tier-by-tier math above doesn't produce one universal winner, and that's the honest answer. It depends on which GPU and which billing mode you actually need:

• You need a single H100 on-demand and want the lowest published rate today: Crusoe's $3.90/hr currently beats Spheron's single-GPU on-demand rate.
• Your workload can checkpoint and tolerate interruption: Spheron's spot pricing across H100, H200, and B200 undercuts Crusoe's on-demand rate by 20-25%, and Crusoe has no spot tier to compete with.
• You need 8 GPUs in one node anyway: Spheron's H100 NVL bundle at $2.06/GPU-hr is the cheapest published rate for that GPU in this whole comparison.
• You need B200 or GB200 without a sales cycle: Spheron publishes a live rate. Crusoe requires a quote.
• You're running a multi-month training commitment at real scale: Crusoe's reserved capacity and 4.9 GW of contracted infrastructure are built for exactly that, and a negotiated reserved rate could beat any on-demand number here. You'd need a sales conversation to find out.
• You want to avoid rounding, minimums, and reserved lock-in entirely: Both platforms bill per minute on-demand, which is the right billing model for bursty or short-duration work either way.

For a broader gut-check on how to evaluate any GPU cloud provider beyond the headline rate, our AI buyer's guide covers the hidden costs and contract terms worth checking before you sign anything. And if you're weighing Crusoe against the rest of the market rather than just Spheron, the GPU cloud pricing comparison across 15+ providers puts all the published rates side by side. Deployment docs and SSH setup guides for Spheron instances are at docs.spheron.ai.

Crusoe's published rate card is honest as far as it goes, on-demand H100 at $3.90/hr is a real, comparison-shoppable number. It just stops at H100, H200, A100, and L40S. For B200 today, or for spot pricing that beats Crusoe's on-demand rate outright, Spheron's rates are live right now.

Check current GPU pricing on Spheron →