Blog

We sell accounts for every provider below, which means we have no reason to flatter any of them. Here is the comparison we would give a friend.

AWS — the broadest, rarely the cheapest

The widest service catalogue and the deepest community. If a tool exists, it integrates. If you break something at 2am, somebody has already written about it. That is worth real money.

What it is not is cheap or predictable. AWS bills are famously hard to forecast, and the platform will happily let you spend money you did not intend to. If your bill is the problem rather than your access, credit is the lever — not a different provider.

Azure — buy it for a reason, not by default

Two genuinely good reasons: you already live inside the Microsoft stack, or you need Azure OpenAI. Outside those, it is a competent and expensive way to do what cheaper platforms do more simply.

Google Cloud — the data one

BigQuery is excellent, Kubernetes is native territory, and Vertex AI is coherent rather than a bag of services. If your work is data or ML, GCP is a strong choice — and its credit works the same way AWS credit does.

Watch BigQuery costs. One badly-scoped query can scan a frightening volume of data. Set a budget alert before your first query, not after your first invoice.

Hetzner — the value outlier

Not really in the same conversation on price-to-performance. Hetzner is substantially cheaper per core, especially in Europe. If your workload is CPU-bound and region-flexible, it is very hard to argue against — and the savings are not marginal.

DigitalOcean — the pleasant one

The friendliest developer experience of the group and flat, predictable pricing. Nobody has ever been ambushed by a DigitalOcean invoice. Plenty of people have been ambushed by an AWS one.

Linode — the boring one, and that is a compliment

Linode is straightforward, globally distributed and reliable. It rarely tops a benchmark and rarely ruins a week. For a great many projects the second matters far more.

Oracle — the free tier nobody expects

Set aside the jokes: OCI has the most generous always-free ARM allocation in the industry, and Ampere instances are outstanding value per core. An excellent place to prototype before committing money anywhere.

The decision, compressed

• Widest ecosystem, hiring for it → AWS
• Microsoft shop, or Azure OpenAI → Azure
• Data/ML heavy, Kubernetes-first → Google Cloud
• Standard web app, cost matters → Hetzner, DigitalOcean or Linode
• Prototyping on nothing → Oracle
• Bill is the bottleneck, not access → stay where you are and buy credit

“It is what serious companies use” is not a reason. It is the single most expensive sentence in cloud procurement. Describe what you are building and we will name the platform we would choose — including when it is the cheap one.

There is a specific moment familiar to anyone who has opened a fresh AWS account with a deadline attached: everything works, the console loads, and the instance you actually need refuses to start.

A quota is a ceiling, not a budget

The single most common misunderstanding is treating a vCPU quota as a monthly allowance. It is not. It caps how much compute may run at the same instant.

Run one 32-vCPU instance and you have consumed a 32-vCPU quota entirely, for as long as it is up. Run four 8-vCPU instances and you have done exactly the same thing. Nothing is being “used up” over time — you are simply not permitted to exceed the ceiling at any given moment.

Why the ceiling is so low to begin with

New accounts get conservative quotas because AWS has no history with you, and abuse — crypto mining above all — arrives on new accounts at scale. It is a defensible policy that happens to make a brand-new account useless for real work on day one.

You can file an increase. It may take days. It may be refused without a reason you can act on. Neither outcome cares about your sprint.

Where it bites: the deploy

Here is the detail that catches people. A rolling deployment briefly runs the old and the new fleet simultaneously. If your steady state sits just under the ceiling, your deploy will not — and the failure arrives mid-release, in front of an audience.

How to size it in five minutes

1. List everything concurrent. Production, staging, CI runners, workers, cron, batch jobs.
2. Sum the peaks, not the averages. Averages conceal precisely the moment that breaks you.
3. Add about 30% headroom for spikes and deploy overlap.
4. Round up to a real tier. Under 32? Take the 32 vCPU account and stop optimising.

The regional trap

Quotas are granted per region. A 64 vCPU ceiling in Frankfurt gives you exactly nothing in Virginia. Discovering this mid-migration is a genuinely bad afternoon, and it is the most common expensive mistake in cloud procurement.

And the thing quota does not fix

If your instances launch fine and the bill is what hurts, quota is irrelevant to you. That is a cost problem, and the answer is credit, not concurrency. Here is how to tell which you have.

Buying credit is the easy part. Watching it evaporate in six weeks because nobody set a budget alarm is the part people do not talk about. These are the habits that actually decide how far a balance goes.

1. Tag everything, from the first launch

When half the balance is gone you will want to know which experiment consumed it. Retrofitting cost attribution after the fact ranges from miserable to impossible. Tag by project, by experiment, by person — whatever lets you answer “where did it go” in under a minute.

2. Use spot instances for anything interruptible

If your training job checkpoints properly — and it should — it does not need on-demand pricing. At credit-account scale the difference between spot and on-demand is not a rounding error; it is often the difference between finishing the project and not.

3. Profile one run before you scale out

The most expensive mistake we see is scaling a job that does not parallelise well. Twice the hardware, the same wall-clock time, twice the burn. Profile a single representative run first and confirm it actually scales before you fan it out.

4. Set alarms at 50% and 80% of balance

The complaint we hear is almost never “the credit ran out”. It is “the credit ran out unexpectedly“. Two alarms remove that entirely, and take four minutes to configure.

5. Shut things down

A forgotten GPU cluster over a long weekend is the single most expensive habit in cloud computing, and credit makes it painless — right up until the moment it does not. Ownership matters: somebody’s name should be against turning things off.

6. Know your expiry date

Credit carries a time horizon. A balance you cannot consume before it lapses is not a discount, it is a donation. Know the date on day one and plan your burn against it. If your realistic burn will not clear the balance in time, you bought the wrong tier — and the fix is to buy a smaller one next time, not to rush your spending.

The uncomfortable conclusion

If following this list means you would not consume a large tier before expiry, do not buy a large tier. Buy the one that matches your actual burn. We publish this guide knowing it steers people toward cheaper purchases, because a customer who wasted a balance once does not come back.

Tell us your monthly spend and we will size it with you. Or read whether you need credit at all.

Almost everyone who arrives at a cloud-account shop is trying to solve one of two problems, and a startling number of them cannot tell you which.

Problem one: it will not launch

You open the console, you try to start the instance your project was designed around, and it refuses. Not because the code is wrong — because the account is not permitted to run it. Your service quota is a ceiling on how much compute may run at the same time, and a fresh AWS account gets a deliberately small one.

This is a quota problem. Credit will not help you at all. What you need is a higher compute tier — a 32 vCPU account for most people, 64 when several environments compete, 128 for training and analytics at scale.

Problem two: it launches, and the bill is horrifying

Everything works. The instances start, the jobs run, the model trains. And then the invoice arrives and somebody senior wants a meeting.

This is a cost problem, and no amount of quota will touch it. What you need is credit — a usable balance applied to the account, which you draw down instead of paying Amazon from a card.

Why people get it wrong

Because sellers are not incentivised to distinguish them. “Premium account” sounds like it must be better at everything, so a customer with a quota problem buys credit, discovers their instances still will not launch, and concludes the market is a scam. They are not wrong to be angry; they were simply sold the wrong product.

The five-second test

• Does your workload run today? If no → quota problem. Buy compute.
• Does it run, but cost too much? If yes → cost problem. Buy credit.
• Both? Then buy both, and stop pretending one product will do. Heavy AI teams routinely need a high vCPU ceiling and a large credit balance.

The honest third answer

Sometimes neither. If your infrastructure will cost forty dollars a month, you do not have a cost problem worth solving with credit, and a $30 compute account is the whole answer. We say that to customers weekly and lose the larger sale each time — it is still the right call.

Send us your monthly burn and we will tell you which of the three you are, including the one where you buy nothing.