GuideCloud · 6 min
How do we get GPU capacity, and who actually has it?
Write down the chips, the count, the region and the date with Clectiv AI. Then GPU clouds, public clouds and data centers for your own hardware bid at once.
The Clectiv team ·
"We have capacity" is a promise about a date, not a fact about a chip. Nobody can tell you from a web page who has GPUs for your project. Write the ask down once, make everyone answer it, and you'll find out.
Who sells GPU capacity: GPU clouds, the big public clouds, or racks for our own?#
There are three routes to GPUs, and each one is sold differently. A GPU is the chip AI work runs on. You can rent them from a GPU cloud, rent them inside a public cloud, or buy your own and house them in a data center.
| Route | What you get | What you take on |
|---|---|---|
| GPU cloud (a neocloud: a cloud company built around GPUs) | GPUs rented by the hour or reserved for a term. Some providers rent a whole physical server, called bare metal. | The term, and the setup |
| Public cloud (AWS, Azure, Google Cloud) | GPU servers inside a cloud you may already run in | The same account and bill, and whatever that cloud will give you |
| Your own hardware | You buy the GPUs. A data center company supplies space, power and cooling. | The hardware, and a room built for it |
Here's who plays in each. Megaport rents bare-metal GPU servers, and Akamai sells cloud GPUs. AWS sells GPU servers inside a public cloud you may already use.
For your own hardware, Switch, CoreSite, Equinix and Digital Realty sell high-density space with liquid cooling. High-density means racks, the cabinets your servers sit in, that draw far more power than usual. Those are names of who plays, not who has room today.
Those are some of our cloud providers for this, and Clectiv AI runs the process with them for you: start a project.
Before you call anyone#
Chips, count, region and date go into Clectiv AI before you ask a single provider for GPUs. Clectiv AI runs the process from there: the providers that fit get it as a bid, they compete for it, and you choose. Call one first and you get the runaround: a rep, a qualifying call, then whatever that one provider has on hand.
What does "we have capacity" really mean?#
It means four different things, and a provider can say yes to the wrong one. Pin each down before you take the yes at face value.
- Reserved or on demand? Reserved GPUs are held for you for a term. On demand means whatever's free when you ask, and it can be gone tomorrow.
- What date do they show up? A date for GPUs that exist today isn't the same as a date for hardware still being built.
- Is it built or promised? A live cluster you can test beats a slide about one.
- Where, and how many? Region matters for your data and your users. Count matters because a big ask and a small ask get different answers.
Get the date in writing, and what happens if it slips. That's the line a "yes, we have capacity" call skips.
Nobody can tell you who has capacity from a page like this one. It changes with every ask. Write yours down with Clectiv AI, and the plan finds out who can deliver it.
Here's the thing: this has been done before. Clectiv AI runs the process we've run for buyers like you, so you start where they ended up. From there the plan goes straight to these providers, for the best pricing, terms and contract we can get.
How do we get in line without overcommitting?#
The same ask goes to several providers at once, through the plan. One provider gives you one answer and nothing to hold it against. Several answering the same ask give you a line you can leave.
Overcommitting means signing for more GPUs, or a longer term, than you'll use. Three questions guard against it:
- How many, and can it grow? Start with what you'll run, and ask for a path to more.
- How long is the term? A longer term locks you in for longer, so ask what you get for it.
- What if capacity slips? Do you pay while you wait, and can you walk?
One hint: ask to pay from the day the GPUs are delivered, not the day you sign. The plan asks the rest, and the team pushes on it before you decide.
Common questions#
Who has GPU capacity right now?#
Should we rent GPUs or buy our own?#
What's the difference between reserved and on-demand GPUs?#
How far ahead should we ask for GPUs?#
Do we have to sign for a year or more?#
How does Clectiv AI get this done?#
It starts in Clectiv AI. Tell it what you're training or serving, and by when. It scopes your GPU project with you and writes your plan: the chips, the count, the region, the date and the terms.
Then the Clectiv team takes your plan to the GPU, cloud and data center providers that fit, all at once, through our partner network of 620+ providers. They put in bids and compete for the work, the team pushes on terms with each, and you choose the winner.
Plenty of companies have stood where you are. Clectiv AI runs the process we've run for companies like yours, so you start where they finished. The plan goes straight to the providers that fit for the best pricing, terms and contract we can get on each.
It stays free to you, and it takes days, not months. The cloud page has everything we cover in cloud.
Once the capacity is lined up, read how to compare GPU cloud providers.
Sources and method. No prices appear in this guide, and nothing in it says who has GPUs, how many, or when. Provider names are examples of who sells each route, from Clectiv's provider directory, not a ranking and not a statement of availability.