digitalocean h100 pricing: per-second billing from $3.26/hr, 80GB H100 GPU 

A few weeks ago I was helping a friend debug a fine-tuning script that kept OOM-ing on a consumer-grade card. After the third crash he sighed, looked at me, and said the magic words: "I just want to rent one H100 for an afternoon, not mortgage my apartment." That sentence is basically the entire reason GPU Droplets exist. If you've been Googling "digitalocean h100 pricing" lately, you're probably in the same boat — you want the Hopper silicon, you want it now, and you want the bill to make sense at the end of the month. So let's actually walk through what DigitalOcean charges for an H100, what you get for that money, and where the cheaper entry points hide.

Why people keep landing on "digitalocean h100 pricing"

There's a reason this specific search string gets typed so often. The H100 is the awkward middle child of the current GPU lineup — powerful enough to actually train and serve real models, but priced like a serious investment. The big three hyperscalers quote H100 numbers that require a sales call, a commitment, and sometimes a fainting couch. Smaller providers quote prices that look great until you read the footnotes. So developers end up here, refreshing pricing pages, trying to find the version of the truth that doesn't involve talking to an account manager.

DigitalOcean's pitch is straightforward: published per-GPU-hour rates, no minimum spend, billed by the second. Whether that's actually a good deal depends on how you use it, so let's get into the numbers.

The H100 GPU Droplet, plan by plan

DigitalOcean offers the NVIDIA HGX H100 in two flavors — a single-GPU Droplet and an 8-GPU Droplet. Both run on the same Hopper silicon; the difference is scale and how much host hardware is bolted on around the chip.

NVIDIA HGX H100 (1× GPU) — the "let me just try this" option

This is the configuration you reach for when you want to test a script, run inference on a 7B model, or fine-tune something small enough to fit in 80 GB of HBM3. The on-demand price sits at $4.41 per GPU per hour, with no commitment. You get 80 GB of GPU memory, 240 GiB of host RAM, 20 vCPUs, a 720 GiB NVMe boot disk, and a 5 TiB NVMe scratch disk for staging training data. Out of the box it comes with Python, CUDA, and PyTorch pre-installed, so you're not spending your first billable hour fighting driver versions.

The trap here, and Reddit is full of people who learned this the hard way, is that a powered-off GPU Droplet still bills. Resources stay reserved, so the meter keeps running until you destroy the Droplet. If you SSH in, run a job, and close your laptop expecting the charge to stop — it won't. Five-minute minimum billing is great for experimentation; forgetting to clean up is not.

This single-GPU shape is aimed at individual developers, small teams doing R&D, and anyone whose workload genuinely fits on one card. If that's you, here's the door: 👉 Get started with an H100 GPU Droplet on DigitalOcean.

NVIDIA HGX H100×8 (8× GPU) — the "we're doing this properly" option

When one H100 isn't enough — say you're pre-training, doing distributed fine-tuning, or serving a model big enough that you need the NVLink fabric between cards — you step up to the 8-GPU node. Same on-demand rate of $4.41 per GPU per hour, which works out to roughly $35.28 per hour for the whole box. In return you get 640 GB of total GPU memory (8× 80 GB), 1,920 GiB of host RAM, 160 vCPUs, a 2,046 GiB boot disk, and a 40 TiB scratch disk. The eight GPUs are connected the way NVIDIA intends them to be — HGX topology, full bandwidth — so multi-GPU training doesn't crawl.

The honest math: at on-demand rates this node costs over $300 for a long day of tinkering, and people have posted bills in exactly that range. That's not a complaint about the price; it's just the reality of renting eight H100s by the hour. The 8-GPU Droplet is for teams with a real workload, a real budget, and ideally a reserved plan (more on that below).

If your project has graduated past a single card, here's where you provision it: 👉 Spin up an 8× H100 GPU Droplet on DigitalOcean.

The reserved price is where the actual savings live

This is the part most casual searches miss. The on-demand $4.41/GPU/hour is the headline number, but DigitalOcean also publishes a 12-month reserved rate of $3.26 per GPU per hour for the H100. That's roughly a 26% discount, and pricing is locked at creation. For a workload that runs continuously — a production inference endpoint, a long fine-tuning sweep, a research cluster that's basically always-on — the reserved rate is the one that actually matters. It turns the H100 from "expensive toy" into "defensible infrastructure line item."

Reserved plans require you to commit, so they're not the right call if you only need a GPU for an afternoon. But if you already know you'll be running H100 workloads for the next year, paying on-demand is just leaving money on the table. You can explore the reserved options through the same console: 👉 Check DigitalOcean H100 reserved pricing.

How the H100 stacks up against its siblings on the same platform

One of the nicer things about researching "digitalocean h100 pricing" on DigitalOcean specifically is that the H100 isn't the only card in the drawer, so you can compare apples to apples on the same bill.

The H200 is the natural upgrade. Same Hopper architecture, but 141 GB of HBM3e memory per GPU instead of 80 GB, which matters when you're trying to keep a larger model resident without splitting it across cards. On-demand it's priced higher than the H100, and the 12-month reserved rate lands at $3.40/GPU/hour. If your workload is bound by "the model doesn't fit," the H200 is worth the premium; if you're just doing throughput, the H100 still holds its own.

The newer Blackwell B300 sits at the top of the lineup, with 288 GB of GPU memory and a 12-month reserved price of $7.94/GPU/hour. It's the "if you have to ask, you probably don't need it" tier — big-memory, big-throughput, big-budget.

On the AMD side, the MI300X (192 GB) and MI325X (256 GB) are alternatives for teams that want more memory per dollar than H100 offers, with reserved pricing that can dip below the H100's. Whether that's a real win depends entirely on your stack — ROCm vs CUDA is not a small decision.

The short version: the H100 is the price-performance baseline. Everything above it trades money for memory or throughput; everything beside it trades ecosystem familiarity for capacity. Knowing where the H100 sits in that line is half the reason people search "digitalocean h100 pricing" in the first place.

What you're actually buying beyond the GPU

Pricing isn't just the per-hour number. A few things make the H100 Droplet easier to live with than the raw spec sheet suggests.

Billing is per-second with a five-minute minimum round-up, which is genuinely friendly for experimentation. You're not paying for a full hour because your job ran for seven minutes. The trade-off is the powered-off billing rule — a stopped Droplet still holds its resources, so the meter doesn't stop until you destroy it.

There's a 99% uptime SLA on GPU Droplets, which is the kind of number you can put in a procurement document without flinching. Availability is currently concentrated in North American data centers — New York, Atlanta, and Toronto — so latency is good across the continent, but if you need Europe or Asia residency you're out of luck for now.

Integration is the quiet win. GPU Droplets plug into the rest of the DigitalOcean stack — Kubernetes, the CLI, the API, Terraform — which means you can drop an H100 into an existing infrastructure-as-code setup without rewriting everything. For a team that already runs on DO, that's worth real money in saved yak-shaving.

A note on free credits and new accounts

A lot of older blog posts and coupon aggregators still claim DigitalOcean hands out $200 in credit for 60 days. That offer has changed. As of mid-2026, new accounts get a much smaller $5 credit valid for 90 days. It's enough to poke around the console, but it will not buy you a meaningful amount of H100 time — at on-demand rates, $5 is roughly an hour of single-GPU work. Treat any "free H100 for a month" claim you see online with suspicion; the current reality is a small trial credit, not a GPU free-for-all.

If you want to actually try the H100 with a referral credit applied, this is the signup path: 👉 Claim your DigitalOcean account credit and try GPU Droplets.

Who the H100 Droplet is really for

After walking through all of this, here's the honest read on who should care about "digitalocean h100 pricing" enough to actually pull the trigger:

• Individual developers and small ML teams who need real Hopper silicon for a script or a model that won't behave on consumer cards, and who want per-second billing instead of a sales call. The single-GPU H100 at $4.41/hr is the entry point.

• Teams with sustained workloads who can commit to a 12-month reserved plan and drop the rate to $3.26/GPU/hour. This is where the H100 starts to make economic sense for production inference or regular training runs.

• Existing DigitalOcean customers who already have Kubernetes, Spaces, managed databases, or other DO services and want to add GPU capacity without introducing a second cloud provider. The integration tax is zero.

• People who specifically need 80 GB of HBM3 on Hopper and don't need the extra memory of the H200 or the throughput of the B300. If 80 GB fits your model, the H100 is the price-performance sweet spot.

If you fall into any of those buckets, the pricing above is the real pricing — published, per-second, no procurement call required. Here's the link to provision your first H100 Droplet and see the numbers show up in your own billing dashboard: 👉 Deploy an H100 GPU Droplet on DigitalOcean.

The H100 isn't cheap, on DigitalOcean or anywhere else. But "cheap" was never the right yardstick for an 80 GB Hopper card. The right yardstick is whether the price is predictable, whether you can actually get one, and whether the bill at the end of the month matches what you thought you signed up for. On all three counts, that's the real answer behind the "digitalocean h100 pricing" search — and now you've got the numbers to do the math yourself.