This post contains affiliate links. If you purchase through our links, we may earn a commission at no extra cost to you.
Cloud GPU pricing is a mess right now. New Blackwell hardware dropped, AMD is finally competitive, and rates that were gospel six months ago are already outdated. So where does Vultr Cloud GPU land in 2026 — overpriced legacy platform or a legitimate contender for serious AI work?
Benchmark data on Vultr’s GPU instances for inference workloads shows the answer is more nuanced than most reviews let on. Vultr isn’t the cheapest option on every GPU tier. But for teams that need global coverage, AMD ROCm support, and a managed services stack around their GPU compute, it punches well above its weight.
TL;DR: Vultr Cloud GPU at a Glance
Best for: Enterprise MLOps, multi-region AI inference, AMD ROCm workloads, teams already on Vultr infrastructure
Not great for: Budget ML experiments on a tight budget (RunPod wins there), spot-price hunting on NVIDIA-only workloads
| Spec | Details |
|---|---|
| GPU Models | 10+ (NVIDIA A16 → B300, AMD MI300X → MI355X) |
| Lowest GPU price | ~$0.32/hr (A16 vGPU) |
| H100 SXM price | ~$2.30/hr per GPU |
| B200 price | ~$2.89/hr per GPU |
| Regions | 32 global |
| AMD support | Yes (MI300X, MI325X, MI355X + ROCm) |
| Kubernetes | Yes (VKE integration) |
| NVIDIA certification | Exemplar Cloud (April 2026) |
| SLA on GPU instances | No |
The GPU Lineup: From A16 to B300
Vultr’s GPU catalog is one of the widest available on a single platform right now — 10+ models spanning NVIDIA and AMD.
NVIDIA stack:
- A16 — 64GB vGPU, ~$0.32/hr. Entry-level, good for inference on smaller models and graphics-heavy VDI workloads.
- A40 — 48GB VRAM, ~$0.76/hr. A solid mid-range option for fine-tuning jobs that don’t need 80GB.
- A100 80GB SXM — ~$1.49/hr. The workhorse for serious training. Still competitive pricing in 2026.
- L40S — 48GB, ~$1.67/hr. Ada Lovelace architecture optimized for inference. Vultr’s committed-use pricing drops this to ~$0.85/hr on a 3-year term.
- H100 SXM — 80GB, ~$2.30/hr. Vultr has reduced pricing from older $2.99/hr rates — always check live pricing since this changes.
- GH200 — 282GB HBM3e. Arm-based Grace CPU paired with H100, excellent for memory-bound inference workloads.
- HGX B200 — 192GB per GPU, ~$2.89/hr. Vultr’s flagship Blackwell offering, the hardware that earned them the NVIDIA Exemplar Cloud designation.
- HGX B300 / GB300 NVL72 — Available on the roadmap; contact Vultr sales for configuration details.
AMD stack (a real differentiator):
RunPod doesn’t offer AMD GPUs at all. Vultr does, with full ROCm support pre-configured:
- MI300X — 192GB VRAM, 128 vCPUs, 2TB RAM, ~$1.85/hr
- MI325X — 256GB VRAM, 128 vCPUs, 3TB RAM, ~$2.00/hr. Preemptible tier introduced December 2025 for batch jobs.
- MI355X — 288GB VRAM, 128 vCPUs, 3TB RAM, ~$2.29/hr. Preemptible tier also available; contact sales for preemptible pricing.
The AMD numbers are genuinely impressive on paper. Benchmarks with the MI325X running Kimi-K2.5 (a 1-trillion-parameter model) show peak throughput hitting 952 tokens per second at 389.8W average power draw per GPU. That’s real, measurable performance — not marketing spec-sheet numbers.
All instances come pre-configured with Ubuntu plus CUDA/cuDNN (NVIDIA) or ROCm (AMD). You can deploy VMs, bare metal, or multi-GPU clusters. Kubernetes integration via VKE is available, which matters a lot for MLOps pipelines — more on that below.
Vultr Cloud GPU Pricing Breakdown
Vultr bills hourly with a one-hour minimum. There’s no upfront commitment required for on-demand, though the Committed-Use Discounts (CUDs) are worth knowing about.
On-demand pricing (per GPU per hour):
| GPU | VRAM | Price/hr |
|---|---|---|
| A16 | 64GB vGPU | ~$0.32 |
| A40 | 48GB | ~$0.76 |
| A100 80GB SXM | 80GB | ~$1.49 |
| L40S | 48GB | ~$1.67 |
| MI300X | 192GB | ~$1.85 |
| MI325X | 256GB | ~$2.00 |
| MI355X | 288GB | ~$2.29 |
| H100 SXM | 80GB | ~$2.30 |
| HGX B200 | 192GB | ~$2.89 |
Prices last checked: 2026-07-05. GPU pricing changes frequently — always check Vultr’s live pricing page before committing.
Committed-Use Discounts (CUDs):
- A100, H100, L40S, A40, A16: 20–35% off for term commitments
- 3-year L40S: ~$0.85/hr (roughly 50% off on-demand rate)
- Preemptible MI325X/MI355X: contact sales for pricing (introduced December 2025, designed for fault-tolerant batch training)
Note: GPU instances are explicitly excluded from Vultr’s standard SLA. That’s an honest disclosure on their part — keep it in mind for production inference with strict uptime requirements.
New accounts get $300 in free credit (valid for 30 days) — that’s roughly 200 hours of A100 time or 130 hours of H100 at zero cost. Useful for real benchmarking before you commit.
For a full breakdown of Vultr’s pricing across all product lines, see our Vultr pricing breakdown.
Performance: The NVIDIA Exemplar Cloud Certification
In April 2026, NVIDIA named Vultr an Exemplar Cloud after Vultr passed NVIDIA’s own performance benchmarks on 512-node HGX B200 hardware. That’s not a marketing badge — it requires passing specific performance thresholds across real model inference benchmarks.
Here’s what the B200 cluster achieved across a range of frontier models:
| Model | Benchmark Improvement |
|---|---|
| Nemotron-4 340B (BF16 → FP8) | 34.8% faster |
| Llama 3.1 70B (FP8 → NVFP4) | 35% reduction in latency |
| Llama 3.1 405B (FP8 → NVFP4) | 28.4% faster |
| Grok-1 314B (BF16 → FP8) | 31% faster |
These numbers come from 512-node B200 clusters at scale — not single-node tests. The gains are largely from quantization (BF16 → FP8 → NVFP4), which the B200’s transformer engine handles efficiently. For most teams fine-tuning 7B–70B models, the A100 or H100 remains the practical choice. B200 economics kick in at serious scale or when you need the full 192GB VRAM for massive model inference.
AMD performance is equally compelling for certain workloads. The MI325X’s 256GB VRAM is one of the few configurations where you can run 1T+ parameter models without tensor parallelism across multiple GPUs.
Vultr Cloud GPU vs Competitors
Let’s get specific. Here’s how Vultr stacks up on the GPUs that matter most, as of mid-2026.
H100 SXM (per GPU, on-demand):
| Provider | Price/hr | Notes |
|---|---|---|
| Vultr | ~$2.30 | 32 regions, AMD also available |
| RunPod | ~$2.69 | NVIDIA only, 24 regions |
| Lambda Labs | $2.99–$3.99 | US-centric, stable availability |
| Vast.ai | ~$1.50–$2.40 | Variable uptime, spot-style |
| AWS p4d | $5.00+ | Enterprise SLA, managed services |
A100 80GB (per GPU, on-demand):
| Provider | Price/hr | Notes |
|---|---|---|
| RunPod | ~$1.00 | Cheapest for budget experiments |
| Lambda Labs | $1.29–$2.49 | Pricing varies by config |
| Vultr | ~$1.49 | SXM cluster configs |
L40S 48GB (per GPU, on-demand):
| Provider | Price/hr |
|---|---|
| RunPod | ~$0.79 |
| Vultr | ~$1.67 |
HGX B200 (per GPU, on-demand):
| Provider | Price/hr |
|---|---|
| Vultr | ~$2.89 |
| RunPod | ~$5.98 |
Prices last checked: 2026-07-05.
The pattern is clear: Vultr is competitive on H100 and B200, loses to RunPod on A100 and L40S, and leads on B200 by a wide margin. Vast.ai can undercut everyone on H100 but with variable uptime and no guaranteed SLA — appropriate for batch jobs, not production inference.
One thing that rarely shows up in pricing tables: Vultr’s 32 global regions versus RunPod’s 24. If you’re running multi-region inference for latency-sensitive applications, that gap matters. Lambda Labs is largely US-centric, which eliminates it for global deployments.
See how Vultr compares to DigitalOcean on general cloud compute
What Users Actually Say: The Review Score Problem
Vultr’s public review scores look contradictory until you understand who’s leaving them.
G2: 4.3/5 — Reviews here skew toward technical users and developers. They consistently praise deployment speed, transparent pricing, and overall performance. The main complaint: support response time lags behind what you’d expect for the price point.
Trustpilot: 1.8/5 (541 reviews, 68% one-star) — This is the number that looks alarming at first glance. The complaints are real — billing disputes, account suspensions during verification, slow support resolution. But the review base here is dominated by general VPS users, not GPU users. VPS billing disputes and account verification issues are an entirely different customer journey than spinning up an A100 cluster.
Gartner Peer Insights — Enterprise practitioners, primarily positive on compute reliability. This audience is closer to the GPU use case.
My read: the Trustpilot score reflects real friction in Vultr’s onboarding for small-account VPS users — billing holds and slow verification are a genuinely bad experience at that tier. GPU workloads, where you’re typically pre-funding at scale, are a different journey. But support responsiveness is a legitimate concern regardless. Vultr’s compute is solid; their customer support is not their strong suit.
our full Vultr platform review for a deeper look at the overall Vultr experience beyond GPU.
Who Should Use Vultr Cloud GPU (and Who Shouldn’t)
Vultr Cloud GPU is the right call if:
- You’re running enterprise MLOps and need GPU + Kubernetes + managed databases + load balancers on one platform. The VKE integration makes this genuinely unified, not just co-located services.
- You need AMD GPU access — MI300X, MI325X, or MI355X with ROCm. RunPod is NVIDIA-only. If your stack is JAX on ROCm or you’re testing AMD’s inference throughput competitively, Vultr is currently one of very few cloud providers offering this.
- You’re deploying multi-region inference and need more than 24 regions. Vultr’s 32-region footprint beats RunPod and far exceeds Lambda Labs’ US-centric coverage.
- You need large-scale Blackwell clusters — the B200 pricing ($2.89/hr) versus RunPod ($5.98/hr) makes Vultr the obvious choice at scale.
- You’re already on Vultr for other infrastructure. A unified stack with shared networking, billing, and Kubernetes beats managing a separate GPU provider.
Vultr Cloud GPU is not the right call if:
- You’re doing budget ML experiments on A100 or L40S. RunPod’s $1.00/hr A100 and $0.79/hr L40S are significantly cheaper for short bursts.
- You need spot pricing for fault-tolerant batch jobs. Vast.ai’s variable-uptime instances undercut Vultr on H100. Vultr does offer preemptible MI325X/MI355X, but pricing requires a sales call.
- You require a GPU SLA. Vultr explicitly excludes GPU instances from their SLA. AWS, Google Cloud, and Azure offer managed GPU with SLA coverage if that’s a hard requirement.
- Your team is NVIDIA-only and price-sensitive on mid-tier GPUs. Lambda Labs has more stable A100 and H100 availability, and RunPod is cheaper on both.
Who Vultr Cloud GPU Is Actually For
Vultr Cloud GPU is a strong platform that gets unfairly dismissed by people comparing its A100 price against RunPod and stopping there. That’s the wrong comparison for most serious AI workloads in 2026.
Here’s the honest take: if you need the cheapest possible A100 or L40S for solo ML experiments, RunPod wins. Full stop. Vultr is more expensive on those tiers and there’s no getting around it.
But if you’re building a production AI stack — inference serving at multiple latency zones, MLOps pipelines with Kubernetes, AMD ROCm workloads, or large Blackwell cluster training — Vultr’s combination of 32 regions, unified managed services, AMD GPU access, and B200 pricing ($2.89/hr vs RunPod’s $5.98/hr) is a different proposition entirely.
The NVIDIA Exemplar Cloud certification is third-party validation that Vultr’s B200 infrastructure performs to spec at 512-node scale — relevant when you’re making infrastructure decisions at enterprise scale.
Support responsiveness is Vultr’s weakest point. Pre-fund your account, have verification documents ready, and don’t expect instant resolution at 2 AM.
For teams already running Vultr infrastructure across other services, adding GPU compute to the same stack is a genuine operational win — unified billing, networking, and Kubernetes, no multi-vendor complexity.
Bottom line: Vultr Cloud GPU is the pick for enterprise MLOps, AMD GPU workloads, multi-region inference, and large-scale Blackwell training. It’s not the pick for budget experimentation or NVIDIA-only spot workloads.
New accounts get $300 in free credit — that’s 200 hours of A100 or 130 hours of H100 at no cost. Use it to run your own benchmarks before you decide.
