Read the clock before you spend
Every programme we cover here expires unused credit. None of the official pages we checked will extend a date as a favour.
- AWS Activate. Credits have an expiration date, which you can see in the AWS Billing and Cost Management Console, along with the remaining balance and the services they apply to [source]. AWS's own explainer says credits usually expire within one to two years depending on the package, and recommends applying when you are ready to ramp up usage, so the value is not left on the table [source]. The Activate team cannot extend the date, even if credit is still unused [source].
- Microsoft for Startups. You must activate credits within 90 days of accepting the programme agreement, or the offer expires and you have to reapply [source]. How long you then have depends on the milestone: the initial $200 is available for up to 90 days from sign-up; after business verification, the $5,000 milestone is available for up to 180 days; at the $25,000 milestone you receive a one-time two-year period. Investor Network benefits, for an eligible startup approved that way, last two years from application approval. Unused credits do not carry forward, and extensions are not available [source].
- Google for Startups Cloud Program. The AI-startup structure Google publishes is up to $250,000 in year 1 at 100%, then 20% up to an additional $100,000 in year 2 (the standard non-AI cap it quotes is lower) [source]. Year-2 support is a partial contribution, not a second full pot at 100%, so work you can legitimately run in year 1 is usually the better use of the larger discount.
The published amounts, and who qualifies, are compared in Google, AWS and Azure credits for AI startups. AWS application steps are in AWS Activate credits: eligibility and how to apply. Live figures on the listings we track are Up to $350,000 in Google Cloud credits over 2 years (AI-first startups) for Google for Startups Cloud Program: AI track, Up to $5,000 (self-funded) or up to $200,000 (with an Activate Provider Org ID); invite-only $200K+ for AI startups at scale for AWS Activate Credits, and Up to $150,000 in Azure credits for Microsoft for Startups.
What the credits actually pay for
A useful rule, across all three, is that credits usually cover the provider's own consumption-billed products, and usually do not cover marketplaces, third-party models, or professional services. The official exclusions are stricter than that slogan.
AWS Activate credits offset eligible AWS services incurred during or after the billing cycle in which you receive them. They cannot be used for historical bills [source] [source]. You cannot choose which eligible service they attach to: they are applied automatically until they are consumed or they expire [source]. They are not applied to Amazon Mechanical Turk, AWS Managed Services, ineligible AWS Support, AWS Marketplace (with the exception of third-party foundation models used through Amazon Bedrock), AWS Professional Services, AWS Training, AWS Certification, Route 53 domain registration or transfer, cryptocurrency mining, or any upfront fee for services such as Savings Plans and Reserved Instances [source]. Credits are applied to Reserved Instances and Savings Plans in the same way as other eligible services; it is the upfront fee that is excluded [source].
Google for Startups credits apply to eligible first-party, pay-as-you-go Google Cloud products owned and built by Google. Subscription, pre-pay and annual-licence first-party products are excluded, as are all third-party products, including offerings in Google Cloud Marketplace or Model Garden [source]. Workspace, Maps and Ads, where they are offered, are issued as separate credits or discounts, not as a draw-down on the Cloud credit [source]. Credits may only be used by the original applicant and cannot be shared, split or resold [source].
Microsoft for Startups credits apply to eligible Azure services on a sponsorship subscription, billed at standard published pay-as-you-go rates [source] [source]. Microsoft Support plans, non-Azure products, third-party AI models (including Anthropic), partner services, and Microsoft Marketplace purchases are not covered and are billed separately [source]. GitHub usage is not covered by default; you have to connect an Azure subscription as a GitHub payment method before eligible metered GitHub usage can draw on the credits [source].
The usual ways credits get wasted
- Leaving them to expire. AWS will not extend a date [source]. Microsoft will not let unused credit carry forward [source]. If a large award is sitting unused in month 11 of a 12-month window, that is a planning failure, not a pleasant surplus.
- Paying cash for an excluded product. Third-party models, marketplace listings and support plans are the usual surprises. Microsoft is explicit that models sold by third-party providers, including Anthropic, incur separate charges even when you launch them from Microsoft Foundry [source]. Google is equally explicit about Model Garden and Marketplace [source].
- Putting the credit on the wrong account. Google deposits credit into the billing account you select, and says the Cloud account and billing ID must be administered by a user on the company domain [source]. Microsoft credits sit on a sponsorship subscription; usage on a different billing profile will be charged [source]. AWS credits cannot be transferred into another AWS account, and you may not sell, license, rent or transfer them; violating that policy leads to revocation [source].
- Applying too early, or too late. AWS recommends applying when you are ready to increase usage, because the clock starts once credit is issued [source]. Microsoft requires activation within 90 days of accepting the agreement [source]. Google's year-2 AI credit is 20%, not 100% [source].
- Assuming the meter cannot run past the credit. Microsoft says that when credits run out, the sponsorship subscription switches to pay-as-you-go so services keep running [source] [source]. You need a payment method on the account for that transition, and for any service the credit does not cover [source]. Set budgets and alerts before you start large training jobs.
A practical way to use a large award
- On day one, find the official balance, expiry date and eligible-services list in the provider's own billing console. For AWS, that is the Billing and Cost Management Console [source]. For Microsoft, Cost Management + Billing is the official place to confirm a sponsorship [source].
- Write down what is excluded. If your stack depends on a marketplace model, a support plan, or reserved-instance upfront fees, budget cash for those lines now.
- Spend first on the eligible work that is hardest to pause: training and inference on the provider's own GPUs and models, then the surrounding storage, logs and databases that have to live next to them.
- Turn on budgets and alerts. Microsoft's own guidance is to set monthly subscription budgets at 50%, 80% and 100% of expected credit burn, and to clean up idle disks, IPs and test clusters [source]. The same habit helps on AWS and Google.
- If the credit still cannot cover the GPUs you need, look at specialist GPU credits and rental rates rather than forcing the job onto an expensive excluded model. Runpod Startup Program and Vast.ai Startup Program (GPU credit matching) are listed on our compute-credits list, and we compare on-demand, interruptible and reserved rental terms in on-demand vs reserved GPUs.
Credits are one piece of a wider stack. They do not pay salaries, and they expire. How they sit next to grants, tax relief and equity is covered in the AI startup funding stack.
