Amazon Web Services CEO Matt Garman thinks about startups more than you might expect. Garman, who stepped into the role in June, sees startups as customers, partners, and, in some sense, as Amazon's future.
"Startups are the enterprises of the future," he told Fortune. "Pinterest was a startup, Netflix was a startup, Rivian was a startup. X-energy is a startup...So, we want to be a part of helping them grow, helping them build on AWS, and helping them grow their company."
Every 10 minutes a startup around the world joins AWS Activate, a program that offers startups cloud credits and mentorship, among other things. The program was founded in 2013, and has since helped 280,000 startups, providing more than $6 billion in credits. Amazon says that, today, 80% of PitchBook-listed unicorns run on AWS.
Garman notes that AWS learns a lot from startups, who are often "at the bleeding edge of technology" and "looking around corners." This helps AWS improve its own services to better support startups and its broader customer base.
Garman has taken the reins at AWS at an interesting moment in the cloud business, as the AI boom drives up the need for computing power.
In December, AWS held its annual summit in Las Vegas, AWS re:Invent, which featured new partnerships with a wide range of startups. One of those startups, Redpoint-backed unicorn Poolside, announced plans to include its models on AWS Bedrock and leverage AWS Trainium chips. Jason Warner, cofounder and CEO of Poolside says AI training and inference have made securing reliable cloud computing horsepower more important than ever.
"These models are so large, and the resource requirements to train them and run them are so large," Warner told Fortune. "You can't just work with any old partner. This is actually one of those weird things where a startup really requires so much out of their other partner."
Luma AI's founder and CEO Amit Jain echoes the sentiment:
"Think about it this way: My business has only two inputs, talent and compute," Jain told Fortune. "Honestly, talent, I got it right. But the second thing is that you need to be able to find reliable compute, where it doesn't stand in your way... Video images are insanely compute-intensive, right? We think this next generation of AI is going to drive the majority of demand on compute."
Garman acknowledged the growing costs for startups, but predicted that costs would come down -- and in turn, create even more demand.
"Today, inference is just too expensive," said Garman. With smaller models, creative techniques like "distilling," and new inventions, Garman predicts that inference costs will drop by "an order of magnitude" -- and the amount of AI inference being done by companies will increase by three orders of magnitude.