What to do when the music stops (and the free credits run out)

It sometimes feels that everyone (and their mother) has a new AI startup that they’re working on - we think that’s great and are excited for all of the innovation to come. Major cloud platforms (e.g. AWS, Azure, and GCP) are more than happy to help AI startups along this journey, showering them with hundreds of thousands of dollars of free cloud credits to get started. Yet we know, from our mothers, that there’s no such thing as a free lunch. In this article, we want to unpack the motivations behind cloud platforms offering free credits and what startups can do once the music stops (and they’re forced to pay for their compute).

Free Cloud Credits By Provider

Each of the big 3 cloud providers offers hundreds of thousands of dollars worth of free credits to get started. Each provider has a different set of qualifications, but largely if you’re a venture backed startup, founded in the last 10 years, and have a website, you qualify.

AWS Activate offers up to $100k, Azure offers up to $150k, and Google provides up to $250k. For a new startup, hopping between cloud platforms is lucrative - with over $500,000 in credits at their disposal, keeping costs low as they build their product and find product-market-fit.

Just like UberEats & DoorDash (Dashpass) offer access for a few months to get started, cloud companies hope that the startups that use their credits to build everything on top of their platform so that in the future it’ll be difficult to leave. A few hundred thousand dollars is a high cost of customer acquisition (CAC), however, not for an AI company that will spend (tens/hundreds) of millions over the company’s lifetime value by staying on the platform (LTV).

We advise startups to take full advantage of the cloud’s generosity, but there are a few key things to think about ahead of time:

1. When to Start the Credits - Most of the cloud providers will give an expiration date to their free credits, typically within the first year. Make sure that you apply for credits when you feel confident you’ll be able to fully maximize them. For instance, a company applies and gets $100k worth of credits, but doesn’t start training its model until 6 months later. By that time, you only have 6 months to burn $100k; smarter to wait.
2. Building Your Infrastructure to Stay Agnostic - To be able to recognize the full value of credits, you want to build your workflows/systems to be agnostic across compute providers. This will also serve you well to ultimately leave the big 3 once all of your credits expire, providing you with more optionality.
3. Oliver Twist “Can I have some more?” - Even when your free credits run out, it never hurts to move onto another provider and then go back a few months later to ask for more credits…chances are they will be willing to grant your company another round of credits.‍
4. Don’t Waste Free Resources - One of my favorite quotes is “Don’t go broke saving money.” Even with $250k in free credits, it’s important to exercise restraint and build cost-effective processes. For instance, don’t burn all of your credits on on-demand H100s; you can make those credits last twice as long by training on A100s and leveraging spot-pricing.

What to do When the Credits Expire

After the free credit party music stops, before committing to a single cloud provider to work with, take a moment to look around at all of your other options. While it might be tempting to stick with one of the large providers who was most generous giving you free credits, you’ll quickly see the greed behind the generosity when you start receiving your first cloud bills in the mail. Sadly, this occurred recently with Stability AI - they had to close down earlier this year after racking up $100m+ in cloud bills.

At Build AI, we want to see as many AI companies as possible succeed and keep their costs low so they can continue to innovate more and more. We fundamentally believe that there are different kinds of optimal compute infrastructure for different kinds of workloads (e.g. hybrid approach). For example, it might make sense to use AWS Sagemaker for inference, have some GPUs on-prem for pre-training/smaller runs, and leverage Build AI for larger training runs. One size does not fit all - so why use one cloud for everything?

Build AI 

Build AI doesn’t offer hundreds of thousands of free cloud credits - but we also don’t charge an arm and a leg later on. We have consistent, lower prices, which is a reflection of the computing workloads we service (batchable/interruptible use cases) and our approach to take advantage of their unique characteristics.

We’re developing data centers in parts of the country with cheap/green power - places like West Texas. We also pause AI models being trained with us at their checkpoints - turning off our data centers during the day (~5-8pm) when the grid is most strained, when power prices are the highest / the power is predominantly coming from fossil fuels. We’re a team of specialists that wants to make AI training less expensive, more accessible, and better for the planet. 

Lower cost & lower environmental impact - start training your AI models with Build AI today. To get started, please complete our short questionnaire detailing your requirements.