Here’s a good segment from the a16z podcast discussing how AI/ML companies are different than traditional software in terms of margins. Ultimately, what matters most is return on investment and not margins, but margins provide a basis for projecting the company’s future ROI.
Gross margins are calculated by total sales relative to the cost of servicing those sales, or Cost of Goods Sold (COGS). Unlike hardware, software often has low COGS, mostly cloud costs, but very high Sales, General & Administrative (SG&A) expenses such as the business IT infrastructure as well as high Research & Development (R&D) expenses. Once software is developed, it can be reproduced at very low marginal cost. Often, software companies lose money in the early years but when they scale up, the gross margins swamp the SG&A and R&D costs making a very profitable company.
Margins can be gamed. Sometimes, 20% of a software company’s SG&A expenses can be cloud costs that could increase with sales — perhaps better allocated to COGS. This means the company’s margins are perhaps higher than reality. Other companies may charge the cost of servicing “freemium” offerings to COGS when really that could be in SG&A, underreporting margins. Margins are misleading for companies like Booking.com, which have 90% margins but have to pay Google a huge share of that in terms of SG&A.
OK, so how is AI/ML different in all this?
It’s not clear to me that if you look at AI/ML companies whether they have the same margin structure as SaaS. There’s two reasons for this.
One is fundamentally a technical reason. The amount of computation and data handling costs actually increase over time. In traditional software developments, you build the software, optimize it, then the cost-per-compute over time reduces because the cost of chips gets cheaper, the cost of servers get cheaper, etc. There’s also a finite amount of work you can optimize.
In AI/ML, the accuracy of your solution depends on the amount of data, and the amount of data required increases so it’s not clear that you can get the same level of cost efficiencies. In order to improve your product you may need 10-times more data, and to improve it again 100-times more data. So the tail of complexity is quite a bit different.
If you have two companies trying to do vision applications — let’s say two companies building drones that can identify crops. One that’s been doing it a long time might have 96% accuracy and the other doing far less time might be 90% accuracy. Getting to those extra bits of accuracy takes much more investment which means you’ll always be fighting this data management-accuracy game. That will get harder as you go along, it doesn’t get better.
The second reason is, sometimes we see AI/ML companies taking operational costs from their customers and internalizing it. It’s almost the exact opposite. Let’s say I’m building an AI/ML company that automates filling in entries – -data entry. You have all these people doing data entry and I’m going to automate it. The customer says, that’s fantastic! I’ll take it off my book and you do it.
Then if you look at what the startup is doing, it is internalized this variable cost and it still takes some humans to do it. Even though they have good growth, revenue, and customers is because they taking someone else’s operational costs and putting it on their books. You can’t always make that go away. It could be an increasing cost even if you have software doing some of it. It will still impact your margins.
I’m personally not sure whether AI/ML companies have the same margins as software companies. I think they’re seeming to be lower. From the boards I’m on, we see 50% margins to be pretty standard.
By contrast, he stated that most SaaS companies see margins of 70-80 percent.
Leave a Reply