GPU credits are not cloud credits
Several large cloud programs explicitly exclude GPUs and dedicated inference. If you train or serve models, the headline credit may not touch your real bill.
An AI-native team reads "up to $100,000 in cloud credits", plans a training run against it, and discovers at the worst possible moment that accelerated compute was never included.
This is not an edge case. It is written into the terms of several of the largest programs, and it is the single biggest gap between what founders think they have won and what they actually have.
The exclusion, stated plainly
DigitalOcean's startup program provides twelve months of credits with a $10,000 monthly spend ceiling, fifteen months of free Standard-tier support, and a bundle of partner perks. It also states that core credits exclude GPU Droplets, dedicated inference and third-party AI model hosting — those require a separate GPU credit application.
So the program is genuinely valuable, and it is valuable for the wrong half of an AI company's bill. Your API servers, databases, queues and object storage are covered. Your accelerators are not.
Anthropic has a different but equally consequential boundary: credits apply to the first-party Claude API through Claude Console only, and cannot be used on AWS Bedrock, Google Cloud Vertex AI or other third-party platforms. If your plan was to consume model credits through your existing cloud commitment so the two stack, that route is closed by design.
The general principle: credit programs are scoped to the provider's own first-party services, and accelerated compute is frequently carved out of the general pool even within that. Assume nothing is included until you have read the exclusion list.
Where GPU budget actually comes from
If sustained accelerator time is what you need, the relevant programs are different from the general cloud programs.
NVIDIA Inception is the obvious first stop and costs nothing: free to join, no fees, no equity. Requirements are incorporation, at least one developer, an official website, and being under ten years old. It gives preferred pricing on select NVIDIA hardware and software, free cloud credits from NVIDIA and partners, and free technical training. Consultancies, crypto companies, cloud providers, resellers and public companies are excluded. Membership also functions as a credential that improves your standing in other programs.
Modal grants one-time GPU credits with direct access to its engineering team, but the gate is investor-shaped. Seed to Series A needs either a VC from Modal's partner network or over $1M raised from any fund. Series B+ needs over $30M raised with a partner-network investor. Amounts are not published, and we do not guess at them.
Vultr offers up to $100,000 for migration plus long-term discounts of up to 35%, with executive sponsorship, architecture reviews and a dedicated account manager. Note the gate carefully: this program is aimed at companies with Series A or later financing and runs through the sales channel, not at first-time builders. Pre-seed teams should look at the free tier instead.
Together AI runs from around $1,000 to $50,000 depending on the referring partner, and is designed for teams serving open weights rather than training from scratch. For inference-heavy products this is often more useful than raw GPU time.
AWS Activate credits are redeemable on third-party models through Amazon Bedrock, and AWS lists additional credits for AI startups ready to scale — which makes Bedrock the one place where a general cloud credit does reach model inference, provided you are willing to consume models through AWS rather than first-party APIs.
The decision that follows
The carve-outs force a real architectural choice, and it is better made deliberately than discovered later.
Route one: consolidate on a hyperscaler. Consume models through Bedrock or Vertex so that one large cloud credit covers both infrastructure and inference. You give up first-party model credits — Anthropic's cannot be used this way — and you accept the platform's model catalog and pricing. In exchange, one credit covers your whole bill.
Route two: split the stack. First-party model APIs with their own credits, general infrastructure on whichever cloud program you won, and accelerated compute from a GPU-specific program. You capture more total credit, at the cost of running three billing relationships and three expiry dates.
For most seed-stage AI companies route two is worth more, because the first-party model credits are large relative to everything else and cannot be captured any other way. But it only works if you know from the start which credit pays for which workload — and if you have written down the three separate expiry dates. That is where most of this value gets lost.
Three questions before you plan against any credit
- Does this credit cover accelerated compute, or is it carved out? Check the exclusion list, not the headline.
- Can this model credit be consumed through my cloud, or only first-party? Anthropic is first-party only. AWS Bedrock accepts Activate credits.
- Is there a separate application for the GPU pool? For DigitalOcean there is, and it is a different process from the one you already completed.
Every entry in our database lists the exclusions next to the amount, because for an AI company the exclusions are frequently the more important number.