Thunder Compute Pricing 2026: H100 and A100 Cost vs Spheron

Thunder Compute pricing lands at $2.19/hr for an H100 PCIe instance and $1.09/hr for an A100 80GB, both billed by the minute with no data egress fees. Those are real, published rates, not a "starting at" teaser, and they undercut AWS and Lambda Labs by a wide margin on the same hardware. The Y Combinator-backed neocloud built its pitch around GPU virtualization tech that lets it run leaner than the hyperscalers, and the pricing page backs it up.

This post breaks down Thunder Compute's current rates, the fine print on storage and vCPUs, and exactly how much it saves against AWS P5, Lambda Labs, and Google Cloud A3 on identical chips. Then it puts Thunder Compute's single-provider pricing next to Spheron's aggregated marketplace, where the picture gets more interesting than a straight price comparison.

Thunder Compute Pricing 2026: Per-Minute Billing and Current Rates

Thunder Compute publishes flat, per-GPU hourly rates with no tiers, no reserved-instance paperwork, and per-minute billing on every instance. As of this post, the rates are:

Source: Thunder Compute pricing page, verified 16 Aug 2026.

GPU Lineup: RTX A6000, L40, A100 80GB, H100 PCIe

The lineup covers four tiers, and the spread between them is deliberate rather than arbitrary. RTX A6000 at $0.35/hr is the budget option for smaller fine-tuning jobs and inference that fits in 48GB of VRAM. L40 at $0.79/hr sits in the middle, useful for mid-size inference and rendering workloads. A100 80GB at $1.09/hr is the workhorse for training and larger-context inference. H100 PCIe at $2.19/hr tops the stack, but Thunder Compute doesn't list an SXM5 or NVLink variant. Every listing here is PCIe-based, which caps multi-GPU interconnect bandwidth relative to SXM configurations elsewhere in the market.

Thunder Compute also supports 1x to 8x GPU server configurations per instance, with vCPU and RAM scaling automatically as GPU count increases, so a single-GPU dev box and an 8x A100 training node both run through the same pricing model.

What Per-Minute Billing Actually Saves You

Per-minute billing matters most on short or bursty jobs, not on jobs that already run 24/7. Say you're running a 22-minute fine-tuning sweep on an A100 80GB. On a platform that bills hourly, you pay for the full 60 minutes: $1.09. On Thunder Compute's per-minute billing, you pay for the 22 minutes you actually used: roughly $0.40. That's a 63% difference on a single run, and it compounds fast if you're running dozens of short jobs a day, which is common for hyperparameter sweeps, notebook sessions, or CI-triggered eval runs. Thunder Compute frames this as saving "up to 40% on bursty workloads" compared to hourly billing, which lines up with how the math works out on jobs well under an hour.

The catch: per-minute billing only helps when your actual usage pattern is bursty. A model that trains continuously for 40 hours pays the same total either way, since the per-minute and per-hour totals converge once you're running full hours back to back.

Storage, vCPU, and Egress: Reading the Fine Print

The headline $/hr rate isn't the whole bill on any GPU cloud, so here's what Thunder Compute adds on top:

• Storage: 100GB persistent storage per GPU included free. Beyond that, $0.03 per additional 100GB per hour while the instance runs.
• Snapshots: $0.05/GB/month for long-term snapshot storage.
• vCPUs: 4 vCPUs included per GPU; additional cores cost $0.04/vCPU/hr.
• Egress: Not charged. Thunder Compute states plainly that it does not bill for data egress.

That last line is the one worth underlining. AWS charges $0.09/GB for data transfer out after the first 100GB free tier, on top of $0.08/GB/month for EBS storage, adding an estimated $45/month for a typical workflow moving datasets and checkpoints in and out. A team that trains a model on Thunder Compute and then repeatedly pulls checkpoints, exports datasets, or serves inference traffic back out to users never sees that line item.

How Thunder Compute Pricing Undercuts Lambda, AWS, and GCP

The single most useful thing about Thunder Compute's pricing is that it's a flat published rate you can put directly against a competitor's published rate on the same GPU. No sales call, no "contact us" gate. Here's how that plays out against the three providers most teams compare it to.

Thunder Compute vs Lambda Labs H100/A100

Lambda Labs is often the first stop for teams looking for cheaper H100 access than the hyperscalers, and it's genuinely competitive on its own terms. But Thunder Compute still undercuts it on both chips it offers as single-GPU instances:

Source: Thunder Compute vs Lambda Labs pricing comparison, verified 16 Aug 2026.

The A100 comparison has an important caveat: Lambda's A100 pricing is quoted for its SXM 8-GPU nodes, meaning you can't rent a single A100 from Lambda at that rate, you have to take the whole node. Thunder Compute's flexible 1x to 8x GPU configurations mean you pay for exactly the GPU count you need, which is where its real advantage over Lambda often shows up even before the per-hour rate difference. For a closer look at Lambda's own per-hour numbers across tiers, see Lambda Cloud H100 pricing.

Thunder Compute vs AWS P5 H100

AWS P5 instances are the default a lot of enterprise teams reach for simply because they're already on AWS, and that default costs a lot. A p5.4xlarge (a single H100) runs $6.88/hr, and the full 8-GPU p5.48xlarge runs $55.04 to $68.80/hr, working out to $6.88 to $8.60 per GPU-hour depending on region and reservation terms.

Source: Thunder Compute vs AWS P5 pricing comparison, verified 16 Aug 2026.

That's a $112.56 gap over a single 24-hour run on one GPU. Scale that to a week-long training job and the AWS bill is over $700 more expensive for identical hardware. For the fuller AWS on-demand and reserved pricing breakdown, including how it stacks up against a marketplace model, see AWS H100 pricing 2026.

Thunder Compute vs Google Cloud A3 H100

GCP's A3 instances bundle the H100 GPU with a required VM shell, and by the time you add that VM cost in, the per-GPU rate lands around $11.06/hr, with the full 8-GPU A3 High node running about $88.48/hr. Thunder Compute puts its savings against GCP at roughly 80% on that basis.

Source: Thunder Compute vs Google Cloud pricing comparison, verified 16 Aug 2026.

That 80% figure is Thunder Compute's own framing and it's steep because it includes GCP's bundled VM overhead rather than a bare GPU rate. If you want a neutral, provider-run breakdown of GCP A3 pricing across on-demand, spot, and committed-use tiers, Google Cloud A3 H100 pricing has the full math, including GCP's own spot rate, which comes in meaningfully lower than the on-demand figure cited here.

For the widest lens on where every major provider lands, from hyperscalers to neoclouds, the full GPU cloud pricing survey is the place to check before you commit to any one platform.

Pricing fluctuates based on GPU availability. The prices above are based on 16 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

Thunder Compute vs Spheron: Where the Trade-Offs Are

Thunder Compute is a legitimate budget option, not a bait-and-switch. Its rates are real, its billing is transparent, and it beats the hyperscalers by a wide margin on identical hardware. Where it gets more nuanced is against Spheron, which doesn't run its own data centers at all: Spheron aggregates GPU capacity from 5+ providers into one marketplace, and that structural difference changes what you're actually comparing.

Per-GPU Pricing Side-by-Side

Live Spheron rates, fetched from the Spheron GPU offers API on 16 Aug 2026:

On raw on-demand price, Thunder Compute wins clearly: its flat $2.19/hr for H100 PCIe beats Spheron's on-demand rate by about 17%, and its $1.09/hr A100 beats Spheron's PCIe on-demand rate by about 24%. But look at Spheron's spot column and the gap mostly closes. Spheron's H100 PCIe spot at $2.20/hr is a cent off Thunder Compute's flat rate, and its A100 SXM4 spot at $1.14/hr is within 5 cents. The trade-off for that spot pricing is the standard one: spot capacity can be reclaimed with little notice, so it fits checkpointed, interruption-tolerant jobs better than a production inference endpoint.

Pricing fluctuates based on GPU availability. The prices above are based on 16 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

Single-Provider Capacity vs Aggregated Marketplace Supply

Here's the part a per-GPU price table doesn't show. Thunder Compute runs on its own infrastructure and its own GPU virtualization stack, which is exactly what keeps its costs and its pricing simple, but it also means you're capped by one provider's inventory and hardware mix. If Thunder Compute's H100 capacity in a given region is tight, that's the ceiling.

Spheron's model is structurally different: it aggregates capacity from data center partners across multiple regions into a single marketplace, so you're not locked into any one operator's supply or GPU generation lineup. That's why Spheron can list H100 in both PCIe and SXM5/NVLink variants, something a single-provider platform's fixed catalog doesn't always support, and why its spot inventory tends to be deeper across regions. For distributed training that needs NVLink bandwidth across 4 or 8 GPUs, SXM5 access matters more than the per-hour rate difference on PCIe.

The practical version: Thunder Compute's simplicity is a real advantage when your workload fits its catalog and you want the lowest possible flat rate with zero setup friction. Spheron's aggregation is the advantage when you need capacity flexibility, NVLink-connected multi-GPU nodes, or provider diversity as a hedge against any single data center running short on inventory. Spheron vs Runpod walks through the same aggregation-vs-single-provider trade-off in more depth, and most of that reasoning carries over here.

Which Setup Fits Which Workload

Neither platform is the universal right answer. A solo developer running short fine-tuning jobs on a single A100 has little reason to look past Thunder Compute's $1.09/hr rate. A team scaling a training run across 8 NVLink-connected H100s, or one that wants spot pricing without betting the job on a single provider's uptime, is better served by H100 GPU rental on Spheron.

Thunder Compute's flat, per-minute rates make it a strong pick for short or budget-constrained single-GPU work. When the job needs NVLink bandwidth, deeper spot inventory, or capacity spread across more than one provider, Spheron's aggregated marketplace is built for that.

Check H100 pricing on Spheron →