DigitalOcean GPU Droplet Pricing 2026: H100, H200, and Gradient Inference Cost vs Spheron

DigitalOcean GPU Droplets put H100 hardware on a familiar cloud platform at approximately $3.39/hr per GPU, per-second billing with a 5-minute minimum. The billing caveat matters: GPU Droplets charge while powered off. If your team keeps a Droplet alive overnight but idle, you pay for the full 24 hours. That single detail can double or triple your effective GPU cost compared to platforms that bill only for active compute. This post covers the DO GPU Droplet lineup, Gradient serverless inference pricing, hidden costs, and a direct cost comparison against Spheron's on-demand and spot rates.

TL;DR: DigitalOcean GPU Droplets vs Spheron H100/H200 (Jul 2026)

At 100% utilization, DO H100 HGX at $3.39/hr is ~69% more expensive than Spheron H100 PCIe at $2.01/hr (note: different form factors, DO's H100 is HGX/SXM and Spheron's is PCIe). At 50% utilization (idle 12 hours/day on a kept-alive Droplet), the effective DO rate is $6.78/hr vs Spheron's $2.01/hr for actual active hours.

Pricing fluctuates based on GPU availability. The prices above are based on 03 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

DigitalOcean GPU Droplet Lineup

DigitalOcean launched GPU Droplets in 2024 under their "AI-Native Cloud" positioning. The lineup centers on NVIDIA H100 and H200 hardware, offered in single-GPU and 8-GPU node configurations.

Note: DigitalOcean has also announced bare metal GPU options and is investing in expanded AI infrastructure, but self-serve bare metal GPU access is limited. Verify current availability at cloud.digitalocean.com when provisioning. The Gradient product line (acquired from Paperspace) sits alongside GPU Droplets for managed notebooks and serverless inference workloads.

DigitalOcean H100 GPU Droplet Pricing Per Hour

The GBH100x1 instance (1x H100 80GB, HGX/SXM architecture) runs approximately $3.39/hr. The 8x GPU node (GBH100x8) runs approximately $27.12/hr, which is the same $3.39/hr per GPU on a per-GPU basis. The per-GPU rate does not increase for the 8x node; both configurations bill at $3.39/hr per GPU.

Billing is per-second, which sounds granular but is undermined by the 5-minute minimum and, more critically, by the powered-off policy: a Droplet that is powered off but not destroyed continues to bill at the full per-second rate. DigitalOcean's billing documentation explicitly states that GPU Droplets reserve capacity in the data center even when powered off, and they charge for that reservation.

For teams running GPUs for 8 hours per day and keeping the Droplet alive overnight:

• DO charges: 24 hours/day at $3.39/hr = $81.36/day per GPU
• Spheron charges: 8 hours/day at $2.01/hr = $16.08/day per GPU (on-demand PCIe)
• Difference: DO costs ~5x more per active GPU-hour at 33% utilization

To avoid idle charges on DO, you must destroy the Droplet after each session and re-provision before the next. That adds provisioning time, loses in-progress state, and adds friction for teams running intermittent workloads.

Pricing fluctuates based on GPU availability. The prices above are based on 03 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

DigitalOcean H200 GPU Droplet Pricing

DigitalOcean H200 GPU Droplets launched in 2026 and are available self-serve at $3.44/GPU/hr. An 8x H200 SXM5 node runs approximately $27.52/hr. The same powered-off billing policy applies: destroy the Droplet to stop charges, because powering off without destroying still bills at the full rate.

Comparing DO H200 against Spheron H200 pricing:

• DO H200: $3.44/hr on-demand, no spot option
• Spheron H200 SXM5: $3.70/hr on-demand, $1.76/hr spot

On raw on-demand rate, DO H200 at $3.44/hr is about 7% cheaper than Spheron's $3.70/hr. That gap reverses once you factor in DO's powered-off billing for teams that don't run GPUs continuously. Spheron H200 spot at $1.76/hr cuts cost by nearly 49% versus DO's on-demand rate for workloads that tolerate preemption.

For training runs with checkpoint/resume: Spheron H200 spot at $1.76/hr is the more cost-effective path. For fully managed 24/7 inference workloads where you're already in the DO ecosystem and powered-off billing doesn't apply, DO H200 at $3.44/hr is a viable alternative.

Verify current H200 Droplet availability and pricing at cloud.digitalocean.com.

DigitalOcean Gradient Serverless Inference Pricing

DigitalOcean Gradient Deployments offer serverless inference without managing GPU instances. You submit inference requests and pay per compute-second or per-token rather than per GPU-hour. Published rates for open-source models range from roughly $0.18 to $0.99 per million tokens depending on model size: smaller models like Ministral 3 14B run around $0.20/M tokens, mid-size models like Llama 3.3 70B around $0.65/M, and larger models like DeepSeek R1 Distill 70B around $0.99/M. Some larger models with separate input/output pricing reach $1.10/M input and $4.40/M output. Verify current rates at digitalocean.com/pricing before budgeting.

Gradient abstracts away GPU management: no SSH, no container orchestration, no instance lifecycle to manage. The trade-off is less control over the inference stack and per-token billing that can be expensive at scale.

Break-even math against a self-managed H100 Droplet at $3.39/hr:

• Self-managed cost: $3.39/hr divided by tokens/hr processed
• If vLLM on H100 processes 2M tokens/hr (realistic for a 7B model at reasonable batch size): cost = $3.39 / 2,000,000 = $0.0017 per 1,000 tokens = $1.70/million tokens
• At Gradient's cheapest model (~$0.20/M tokens), self-managed H100 at $1.70/M effective cost is more expensive per token for low-throughput workloads; Gradient wins until throughput is high enough to amortize the fixed $3.39/hr cost
• Break-even: self-managed H100 at $3.39/hr matches Gradient's $0.20/M rate at ~17M tokens/hr sustained, and matches the $0.65/M average rate at ~5M tokens/hr. Production batched inference workloads on 7B-70B models can realistically hit the 5M tokens/hr threshold; the 17M threshold is harder but achievable with continuous high-batch-size serving.

For exploratory use or low-traffic API endpoints, Gradient's serverless model avoids idle GPU cost. For sustained high-throughput production inference, self-managed H100 on a per-minute billing platform is the more cost-efficient path. For teams that need a middle ground, the vLLM production deployment guide covers setting up self-managed inference on bare-metal H100s with auto-scaling.

DigitalOcean GPU Hidden Costs

Beyond the listed per-hour rate, four cost categories can significantly inflate total DO GPU spend:

Powered-off billing. As covered above, GPU Droplets bill at full rate whether running or powered off. A team using GPUs 8 hours per day but keeping Droplets for scheduling convenience pays for 24 hours. At $3.39/hr per GPU, that inflates monthly cost from $814/month (8hr/day) to $2,441/month (24hr/day) per GPU. Spheron charges nothing for idle time since billing stops when you terminate the instance.

5-minute minimum billing. Per-second billing has a 5-minute floor. Short evaluation runs or single-batch inference jobs under 5 minutes bill as 5 minutes. For iterative development workflows with many short runs, this adds up.

Block storage volumes. Persistent data volumes are billed separately at DigitalOcean's standard volume rates (approximately $0.10/GB-month for standard SSD volumes). For large model weights or training datasets stored on persistent volumes, this can be substantial. A 1TB volume for checkpoint storage adds roughly $100/month on top of compute.

Egress. Each Droplet includes a free monthly outbound data transfer allowance of approximately 15TB per Droplet. Transfers above the included amount are billed at $0.01/GB. Downloading large checkpoint files to local storage or moving data between regions can exceed these limits for active training runs.

DigitalOcean vs Spheron: H100 and H200 Cost Per Hour

Spheron H100 PCIe at $2.01/hr is ~41% cheaper than DO's H100 HGX at ~$3.39/hr at 100% utilization (note: different form factors, DO's H100 is HGX/SXM and Spheron's at $2.01/hr is PCIe). For any utilization below 100% on a kept-alive Droplet, the gap widens. The powered-off billing difference is the defining cost factor for most workloads. For H200, Spheron spot at $1.76/hr beats DO's $3.44/hr on-demand by 49% for training workloads that can use checkpoint/resume.

When DO's managed simplicity justifies the premium:

Teams already deep in the DigitalOcean ecosystem (Managed Kubernetes via DOKS, Spaces object storage, Managed Databases, DO's App Platform) benefit from consolidated billing and the DO control plane. A single vendor for the entire stack can simplify ops significantly. Teams building on DOKS with GPU inference as one component of a larger DO-hosted application may find the per-GPU premium offset by workflow integration. DO's GPU Droplets also work with their Load Balancers and VPC networking, which matters for teams running inference behind a DO Load Balancer.

When Spheron wins on price:

No Oracle/DO ecosystem dependency, bursty training jobs, intermittent inference workloads, or teams that need H200 spot access. Spheron's per-minute billing and no powered-off charges make it the clear choice for workloads where GPU utilization is below 100%. Spheron aggregates capacity from 5+ providers, giving broader GPU availability than any single-provider cloud. No quota process and provisioning in under 2 minutes make it easier to respond to urgent compute needs.

Break-Even Math: When DigitalOcean GPU Droplets Make Sense

Consider a team running one H100 GPU for ML workloads, comparing monthly cost at different utilization levels:

DO can only win on total cost if you run GPUs 24/7 with no idle time, and even then Spheron's on-demand H100 PCIe at $2.01/hr undercuts DO's $3.39/hr by ~41%. The only scenario where DO's platform can be the lower-cost option is a negotiated enterprise contract with significant volume discounts.

The one exception: if your team is already on DigitalOcean and the workflow integration savings in engineering time offset $1.38/hr per GPU, DO becomes defensible. At 5 GPUs running 8hr/day, the per-GPU gap is about $1.38/hr, which is $1,656/month in compute difference. Whether consolidated billing and DO control plane integration saves more than $1,656/month in team time is a judgment call.

DigitalOcean vs Spheron: Deployment Comparison

DO's managed Kubernetes via DOKS is a genuine advantage for teams building multi-service applications. Spheron's raw per-minute bare-metal pricing and broader provider coverage are better fits for compute-heavy ML workloads.

Spheron Live H100 and H200 Rates (Jul 2026)

For H100 SXM5 workloads on Spheron, Spheron H100 instances offer bare-metal access from 5+ providers with per-minute billing. For H200, the on-demand rate is $3.70/hr and spot is $1.76/hr, available without a sales process or quota application.

Pricing fluctuates based on GPU availability. The prices above are based on 03 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

DigitalOcean vs Other GPU Clouds

For the full provider comparison including AWS, Azure, GCP, CoreWeave, and Lambda, see the full GPU cloud pricing comparison 2026.

DigitalOcean GPU Droplets sit in the mid-range on raw per-hour price but are effectively the most expensive option for any team that doesn't run GPUs at 100% utilization, because of powered-off billing. Vast.ai verified hosts are cheaper on paper for H100 PCIe but introduce host reliability variance and container-only access. Spheron H100 PCIe at $2.01/hr with per-minute billing and no powered-off charges is the most cost-effective managed option for intermittent or bursty GPU workloads.

DigitalOcean GPU Droplets cover managed simplicity at a premium, powered-off billing included. Spheron aggregates bare-metal H100 and H200 capacity from 5+ providers with true per-minute billing, no idle charges, and transparent on-demand and spot rates.

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