Pika had better figure out how to monetize their product, because their compute costs are extremely high for an early stage startup. The Forbes article mentions "several hundred GPU's." So let's use AWS as a proxy. 30 x 8 GPU P4 clusters (giving them 240 GPUs total), using 1 year reserved instance pricing (it would be insane for a startup to lock in for more than 1 year, considering the P4 is the previous generation NVIDIA GPU) is over $6MM per year. How much is that for a startup? Well here's a reference point. In 2015, when I was at AWS, we did a $30MM+ multi-year deal (I think it was at least 4 years) with Airbnb, which was a much larger startup at the time with a "unicorn" valuation. If AI startups are shifting the compute costs so far up on the time scale that they've got costs larger than companies with 20X their valuation, how are they ever going to be profitable? Another way to look at it is that product-led-growth AI startups have exorbitant compute costs compared to non-AI PLG startups, and many of the latter cohort are still wildly unprofitable. What does that say about the former?
Pika had better figure out how to monetize their product, because their compute costs are extremely high for an early stage startup. The Forbes article mentions "several hundred GPU's." So let's use AWS as a proxy. 30 x 8 GPU P4 clusters (giving them 240 GPUs total), using 1 year reserved instance pricing (it would be insane for a startup to lock in for more than 1 year, considering the P4 is the previous generation NVIDIA GPU) is over $6MM per year. How much is that for a startup? Well here's a reference point. In 2015, when I was at AWS, we did a $30MM+ multi-year deal (I think it was at least 4 years) with Airbnb, which was a much larger startup at the time with a "unicorn" valuation. If AI startups are shifting the compute costs so far up on the time scale that they've got costs larger than companies with 20X their valuation, how are they ever going to be profitable? Another way to look at it is that product-led-growth AI startups have exorbitant compute costs compared to non-AI PLG startups, and many of the latter cohort are still wildly unprofitable. What does that say about the former?
The writeup about the GPT data extract was fascinating. Thanks for sharing that