Hyperstack doesn't sell "an H100." It sells three of them, SXM, NVLink, and PCIe, each at its own on-demand and reserved rate. Most H100-pricing roundups flatten that into a single number, which means they miss both the real Hyperstack H100 cost and the EU and North America-only footprint that comes with NexGen Cloud's infrastructure. Here's the full rate card, variant by variant, next to what the same GPUs cost on Spheron's marketplace.
TL;DR: Hyperstack H100 Cost vs Spheron (Nov 2026)
Hyperstack's cheapest H100 is the PCIe variant at $2.50/hr on-demand ($1.75/hr reserved), while its SXM variant runs $3.20/hr on-demand ($2.72/hr reserved); Spheron prices H100 SXM5 at $2.64/hr on-demand and $2.79/hr spot.
- Hyperstack H100 SXM: $3.20/hr on-demand, from $2.72/hr reserved.
- Hyperstack H100 NVLink: $2.60/hr on-demand, from $1.82/hr reserved.
- Hyperstack H100 PCIe: $2.50/hr on-demand, from $1.75/hr reserved, $2.00/hr spot.
- Spheron: H100 SXM5 runs $2.64/hr on-demand and $2.79/hr spot as of 03 Nov 2026, with no reservation required.
- Footprint: Hyperstack (NexGen Cloud) operates EU and North America data centers only; Spheron aggregates 5+ providers across a broader geographic spread. See on-demand H100 pricing on Spheron for the live rate.
Hyperstack H100 and H200 Rates: On-Demand, Reserved, and Spot
Hyperstack's rate card is the most granular H100 pricing on the market because it treats form factor as a pricing variable, not a footnote. The table below is pulled straight from Hyperstack's own pricing page.
| Variant | On-Demand | Reserved (from) | Spot |
|---|---|---|---|
| H100 SXM | $3.20/hr | $2.72/hr | Not published |
| H100 NVLink | $2.60/hr | $1.82/hr | Not published |
| H100 PCIe | $2.50/hr | $1.75/hr | $2.00/hr |
| H200 SXM | $3.99/hr | $2.79/hr | Not published |
A few things jump out once you see all four rows side by side. First, the on-demand-to-reserved delta is close to the same percentage across variants, roughly 15-30% off, which means the commitment discount isn't concentrated on any one form factor. Second, spot pricing exists on exactly one row: H100 PCIe at $2.00/hr, a 20% discount off that variant's own on-demand rate. If your workload needs SXM's NVSwitch fabric or NVLink's paired bandwidth and can tolerate preemption, Hyperstack simply doesn't have a spot number for you yet. Third, H200 SXM on-demand ($3.99/hr) costs roughly 25% more than H100 SXM on-demand ($3.20/hr), the premium for the larger HBM3e footprint H200 carries over H100.
Hyperstack also prices newer Blackwell-generation hardware on the same page: B200 at $6.00/hr on-demand (reserved from $5.10/hr) and B300 at $7.40/hr on-demand. Those sit outside this post's H100/H200 focus, but they confirm the pattern: NexGen Cloud publishes a real number for every tier it sells, on-demand included, rather than gating newer silicon behind a sales call the way some neoclouds do. Hyperstack bills on-demand instances per minute with no long-term commitment required, while reservation contracts are invoiced monthly.
If you're trying to decide which of the three H100 variants actually fits your job, before you even get to price, the H100 NVL vs SXM5 vs PCIe form factor guide breaks down the memory, NVLink bandwidth, and workload fit behind why Hyperstack prices these three differently in the first place. Multi-GPU training wants SXM's full NVSwitch mesh; large-context single-node inference favors NVLink's extra HBM3 per GPU; single-GPU fine-tuning is where PCIe's lower price actually pencils out.
What NexGen Cloud's EU and North America Infrastructure Costs You
Hyperstack is owned and operated by NexGen Cloud, and the pricing above comes with a geographic condition most roundups skip: Hyperstack's own documentation lists three live regions, NORWAY-1 in Vestland, Norway, CANADA-1 in Quebec, Canada, and US-1 in Texas, United States, not a global footprint. That's a deliberate choice, not a gap. NexGen Cloud markets Hyperstack toward compliance-focused buyers, GDPR-bound EU teams and enterprises that need a contractually narrower data residency story than a global marketplace can offer. If your legal team needs to point at a specific jurisdiction and confirm training data never left it, that's worth paying for.
It's also worth being precise about what it isn't: Hyperstack doesn't only operate in the EU. North America is in the footprint too, so a US-based team isn't locked out, but a team that needs, say, APAC capacity or wants its workload spread across the widest possible pool of regions is looking at a smaller map than a multi-provider marketplace covers.
That tradeoff shows up on both ends of the rate card. The reserved discounts above (H100 PCIe from $1.75/hr, H100 SXM from $2.72/hr) assume a monthly-invoiced contract, which is the kind of commitment that makes sense when you've already decided EU or North America residency is non-negotiable and you're optimizing cost within that constraint. If residency isn't the deciding factor, that same discount curve is available more broadly across the market, and it's worth comparing against what a marketplace model does to on-demand pricing instead.
For a side-by-side on the rest of the comparison, including VM hibernation and networking, which this post doesn't re-litigate, see the full Spheron vs Hyperstack comparison.
Spheron Price-for-Price on the Same GPUs
Spheron's pricing page lists H100, H200, A100, B200, and B300 on both on-demand and spot tiers, with GB200, GB300, and Vera Rubin available as reservation-only capacity for teams locking in rack-scale Blackwell or Rubin ahead of deployment. H100 SXM5 runs $2.64/hr on-demand and $2.79/hr spot as of 03 Nov 2026; H200 runs $5.52/hr on-demand and $3.42/hr spot over the same window.
The mechanism behind that number is different from Hyperstack's rate card. Spheron aggregates 5+ providers into one marketplace, so the on-demand and spot rates you see are what results when multiple providers compete for the same workload on the same GPU, rather than one company's own fleet pricing. That's the lever behind the spot discount: more supply bidding for your job, not a single operator's reservation tier. It also means you can deploy H100 on Spheron in minutes with no approval process and no minimum term, which is the direct answer to Hyperstack's monthly-invoiced reservation contracts: Spheron's discount path is spot and Stop/Park, not a signed commitment.
Where Spheron doesn't match Hyperstack is the parts of NexGen Cloud's pitch that have nothing to do with price. Spheron has no dedicated EU-only data residency guarantee the way Hyperstack's EU and North America infrastructure offers for GDPR-bound buyers, and there's no VM hibernation feature; cost control on Spheron runs through spot instances and Stop/Park instead of pausing a session mid-state. Spheron is also a newer platform than NexGen Cloud, so enterprise support tiers and formal compliance attestations are less built out today. What Spheron adds on the catalog side is breadth: RTX 5090, RTX 4090, and RTX 6000 Ada self-serve access that doesn't appear on Hyperstack's H100-and-up lineup at all, plus self-serve on-demand B200 and B300 rates where some competing neoclouds still gate newer Blackwell tiers behind a sales call. For a closer look at whether that H100-to-A100 cost delta matters for your training job specifically, the H100 vs A100 cost guide for LLM training walks through the throughput math.
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 03 Nov 2026; Hyperstack's rates reflect its published pricing page and may have changed since. Check current GPU pricing → for live rates.
Which to Pick: Hyperstack's H100 Cost vs EU Data Residency
Run the numbers against your own constraint, not just the lowest line on either card. If your legal or procurement team has already decided that EU or North America-only infrastructure is a requirement, not a preference, Hyperstack's variant-level pricing is the one that matters: pick PCIe for single-GPU fine-tuning at $2.50/hr on-demand ($1.75/hr reserved), NVLink for large-context inference at $2.60/hr, or SXM for multi-GPU training at $3.20/hr, and decide separately whether a monthly reserved contract is worth the discount for your contract length.
If data residency isn't the deciding factor, the comparison is really about commitment versus flexibility. Spheron's H100 SXM5 at $2.64/hr on-demand ($2.79/hr spot) comes with no reservation and deploys in minutes, which matters most for teams whose GPU needs change week to week, or who want to test a workload before signing anything. Hyperstack's reserved rates, especially PCIe from $1.75/hr, are hard to beat for a workload you already know will run for months at steady utilization, provided EU or North America capacity covers your region.
Neither rate card is "the" cheap option across every scenario, which is exactly why the variant breakdown matters more than a single average. If you want a wider shortlist before committing to either, 10 best Hyperstack alternatives lines up nine more providers against Hyperstack's own published rates, and a Crusoe pricing comparison is a useful second data point if your team is still pricing out reserved-leaning neoclouds generally.
If EU or North America residency isn't a hard requirement, Spheron's on-demand and spot H100 pricing skips the reservation contract entirely.
Frequently Asked Questions
Hyperstack publishes three separate on-demand H100 rates depending on the GPU's form factor: $3.20/hr for H100 SXM, $2.60/hr for H100 NVLink, and $2.50/hr for H100 PCIe, all billed per minute with no long-term commitment. Reserved contracts bring those down to $2.72/hr, $1.82/hr, and $1.75/hr respectively, invoiced monthly.
It depends on which Hyperstack variant you're comparing. Hyperstack's cheapest on-demand H100 (PCIe, at $2.50/hr) sits close to Spheron's marketplace rate, and Spheron's H100 SXM5 runs $2.64/hr on-demand and $2.79/hr spot as of 03 Nov 2026. Hyperstack's reserved PCIe rate (from $1.75/hr) is usually the lowest number on either card, but it requires a monthly-invoiced commitment that Spheron's on-demand and spot tiers don't.
Only on the PCIe variant, listed at $2.00/hr. Hyperstack doesn't publish a spot rate for H100 SXM or H100 NVLink. If you need spot pricing on those form factors, Spheron's spot tier ($2.79/hr as of 03 Nov 2026) is the one with a published number.
The price gap tracks the hardware. SXM runs in an 8-GPU NVSwitch server with the highest inter-GPU bandwidth, so Hyperstack prices it highest at $3.20/hr on-demand. NVLink pairs two GPUs over a bridge chip with extra HBM3 per card, priced at $2.60/hr. PCIe is a single-GPU card with no NVSwitch fabric, the cheapest of the three at $2.50/hr.
No. NexGen Cloud, which operates Hyperstack, runs data centers across the EU and North America, not the EU alone. That footprint is narrower than a marketplace that aggregates many more providers, which is the tradeoff to weigh against Hyperstack's compliance-first positioning.
