Kruze serves over 175 AI startups, from massively funded players building their own LLMs to application layer startups to companies building AI for healthcare, drug discovery, vision, and more.
One thing we are noticing is that AI startups are expensive to build - much more so than traditional SaaS businesses. A huge component of these costs are compute costs. AI startups that we work with are spending, on average, 2x what a traditional SaaS business spends on hosting, compute, etc.
This is big - and results in major differences in the burn profile of an AI company vs. a traditional SaaS business.
AI vs Traditional SaaS
We analyzed over 125 AI and traditional SaaS startups (non-AI) that were earning over $1M a month in ARR and compared their growth, profitability and more (the charts in this article are based on an analysis of those companies). While AI startups in our sample were growing faster than the SaaS businesses, they also were burning more cash. Here is a look at the ARR and the operating income (or rather, losses) of the two groups over a year:
Comparing ARR and Operating Income
AI startups are growing revenue faster than SaaS companies, which is a big reason that VCs are excited to fund the category. This growth is driven by a number of factors …
However, on the flip side, AI startups are burning more - and the more that they grow, the bigger their losses. Whereas SaaS companies have done a better job keeping their costs and operating losses in line, AI startups are clearly growing their expenses even faster than they are boosting revenue.
So what is going on here? Why are AI companies costs increasing?
First of all, it’s not unusual for a growing startup to burn more as it gets bigger. Items like sales and marketing cost money to grow, and founders and VCs are often willing to trade a higher burn rate for faster growth.
Secondly, startups often need to invest in infrastructure like customer service, HR and other functions prior to hitting scale. This can depress operating income, and is often a good investment.
AI compute costs
There is another thing at play with AI companies - the cost of building and running models. AI takes a lot of compute and hosting. Some reporters are claiming that the next generation of models will take $1 billion to create. We see many startups doing it for less, but on the other hand, they are spending way more on compute and hosting than their traditional SaaS counterparts.
In fact, although AI companies represent just over 20% of Kruze’s 800+ startup clients, they account for more than half of all compute and hosting expenses!
Compute costs at AI startups
Over the past year, we saw our AI startup clients grow their compute/hosting costs by a CAGR of 300%, vs 53% for SaaS companies. More worrisome is the fact the compute costs for AI companies went from 24% of revenue to 50% of revenue (vs non-AI SaaS where it more or less stayed at about 18% of revenue over the course of the year).
This means that the more an AI company grew, the more it was spending on computer and hosting. If this trend continues, then AI companies will continue to see ballooning operating losses the more that they grow.
We are optimistic that they will eventually reach economies of scale, but it appears that we are still a ways away from the cost curve bending in the right direction. In our previous articles on how startups are using AI, we noted that well over half of all startups are paying for AI - it’s becoming a critical layer in many applications and workflows. Hopefully the costs of delivering artificial intelligence models will moderate so that companies can continue to afford to use them.
