Prices set the scale of any credit. DigitalOcean lists on-demand GPU Droplets at $0.76 per GPU-hour for an RTX 4000 Ada, $2.59 for an AMD MI300X, $4.41 for an H100 and $4.47 for an H200. At the H100 rate, a $5,000 Google Cloud research award buys roughly 1,100 GPU-hours, enough for fine-tuning but nowhere near pretraining. AWS Activate Portfolio, at up to $200,000, is the only common program sized for a real training run.
Restrictions follow the credit source. AWS research credits limit EC2 to on-demand and spot instances, so reserved GPU capacity is out. Google research credits go only to faculty, PhD students and postdocs and expire 365 days after redemption. TPU Research Cloud participants must share results through publications, open source code or blog posts and give Google detailed feedback. AWS Activate requires a Paid plan; a Free plan account is ineligible. Read the terms before you size the job.
How to get the most from GPU and training credits
Upgrade before requesting GPU quota
Google won’t let a Free Trial billing account add GPUs or request a quota increase, so a team planning a training job should click Activate on the Welcome page first. Upgrading keeps the unused part of the $300 inside the original 90 days, and only then can you file the GPU quota request, which is a separate approval that can take days. Start it the same afternoon, because the clock on your credit doesn’t pause while you wait.
Price the run before applying
The Google research credit form requires a pricing calculator estimate with its URL, and AWS sizes faculty awards to the proposal and documented service usage. Work out GPU type, hours and storage, then multiply. A 48 hour fine-tune on 8 H100 class GPUs at DigitalOcean rates is about $1,700 of compute before storage and transfer. Reviewers reward numbers that add up, and a vague ask gets a small award.
Use spot where the terms allow
AWS Cloud Credit for Research explicitly allows spot instances, which cut the cost of interruptible training jobs. Google research credits cover compute without a stated instance restriction. Checkpoint every few hundred steps, run on spot or preemptible capacity, and keep on-demand for the final evaluation pass. The same credit then covers two or three times the experiments, which is the difference between one paper and three.
Destroy idle GPU Droplets
Powered-off Droplets still accrue charges on DigitalOcean because the hardware stays reserved, and at $4.41 per hour an H100 left off over a weekend costs more than $200. Snapshot the disk, destroy the Droplet, and rebuild from the snapshot on Monday. The same rule applies to any cloud GPU instance that’s stopped rather than deleted. Stopped is not free. Deleted is.
Publish or lose the TPUs
The TPU Research Cloud is free, but it comes with obligations: share the supported research publicly through peer reviewed papers, open source code or blog posts, give Google detailed feedback, and work within the Google AI principles. Teams that can’t publish because of commercial restrictions should apply for startup credits instead and treat TRC as a research-only resource. Don’t apply hoping the publishing rule won’t come up.
Plan for the AI startup step down
Google covers 100% of eligible costs up to $250,000 in year one for AI startups, then 20% up to another $100,000 in year two. A training pipeline built in year one on free compute becomes an 80% out-of-pocket cost in month 13. Schedule the expensive pretraining or data generation inside year one and keep year two for inference, which is cheaper and steadier. Front-load the heavy lifting.
can i use google cloud free trial credits on gpus
No. Adding GPUs to VM instances is one of the things a Free Trial billing account can’t do, alongside requesting quota increases and creating Windows Server VMs. The Always Free tier also excludes GPUs and TPUs. To use GPUs you upgrade to a paid billing account, which keeps any unused trial credit within the original 90 days, then request GPU quota for the region you need.
how much do gpu droplets cost on digitalocean
On demand, $0.76 per GPU-hour for RTX 4000 Ada, $1.57 for RTX 6000 Ada or L40S, $2.59 for AMD MI300X, $3.80 for MI325X, $4.41 for NVIDIA H100 and $4.47 for H200. H100 Droplets come with 1 or 8 GPUs at the same per-GPU rate. Reserved pricing with longer contracts can go as low as $1.91 to $7.94 per GPU-hour depending on hardware.
how do i get free tpus for research
Apply to the TPU Research Cloud through the express interest form on its about page. Accepted researchers get no-cost access to a cluster of more than 1,000 Cloud TPU devices. Google doesn’t publish a per-participant allocation or a fixed duration, and it expects you to publish results, release code or write publicly about the work, plus give detailed feedback. For smaller jobs, Google also offers free TPUs inside Colab without an application.
do aws activate credits cover gpu instances
AWS Activate credits are promotional credit for AWS services, and Amazon Bedrock and EC2 are named uses, so GPU instances on a Paid plan are covered. The promotional credit terms exclude Marketplace charges, upfront fees for Reserved Instances and Savings Plans, and enterprise support. Portfolio credits reach up to $200,000 with an accelerator or VC referral, and an invite-only AI tier goes above that for startups that are ready to scale.
does digitalocean hatch cover gpu droplets
Not with the core credits. Hatch core credits don’t cover GPU Droplets, Dedicated Inference, Serverless Inference, third-party hosted models or Paperspace. GPU credits are a separate pool, not guaranteed, offered to select customers after a GPU credit application review. Hatch applications are paused until an updated program launches in early 2027, so in the meantime GPU Droplets bill at the on-demand rates.
Google Cloud