AI pricing moved into the spotlight last week. The WSJ (paywall) said GitHub’s Copilot (summary no paywall) was losing money on most customers. The company charges $10 a month for access, and the WSJ claimed to have data showing some customers cost GitHub as much as $80 per month. Inference costs for Generative AI products change the pricing strategy conversation, but there’s more to consider.
In traditional digital products, each incremental customer costs very little to serve. The subscription pricing model makes sense for this cost structure. Microsoft Office costs very little for every new user, so each annual subscription is guaranteed to be profitable. Cloud costs are also very low per use, so there’s little danger of costs biting into margins.
Facebook user costs are hard to pin down, but it’s somewhere around $1 per user. The higher any Facebook member’s usage, the more Facebook makes by serving them ads. In the ad-supported business model, incremental usage costs are more than offset by revenue. This model fits better with Generative AI’s inference cost structure, but ads don’t fit into most Generative AI products.
How will pricing adapt to make Generative AI products profitable? First, we must understand what pricing reveals about the business model behind it.
Inference Costs and Subscription Pricing Models
You can tell a lot about a company based on its pricing strategy. Netflix only offers monthly subscriptions. There’s no annual option that comes with a discount. That tells us Netflix has a very low churn rate.
Disney+ offers a monthly subscription for $13.99 and an annual subscription for $139.99. That implies that Disney+ starts to see higher churn rates after 10 months. Annual subscription costs divided by monthly subscription costs equals the point where products start to experience higher churn rates.
Microsoft offers 365 with Office for individuals with a similar 10-month churn split monthly $6.99 and annual $69.99 pricing. For businesses, there’s only an annual option. That tells us business users churn at much lower rates than individuals, and businesses prefer annual pricing.
Copilot for 365 is an upsell for an additional $30/month. With traditional 365 apps and subscriptions, usage didn’t impact Microsoft’s margins much. Copilot could be a different story. Let’s say each request to the LLM powering Copilot costs Microsoft an average of $0.005. A user making 50 requests per day gives Microsoft $22.50 in margins, but a user making 200 requests per day costs as much as they pay monthly.
Adobe addresses this subscription pricing model challenge by throttling requests after users reach a threshold. It’s meant to minimize additional requests and incentivize users to be efficient with their monthly request quota.
Both approaches have perils.
Customer Acquisition Costs and Churn
Generative AI products are new, and there will be an adoption cycle. Streaming entertainment went through a similar cycle. Disney+ launched at a low cost to encourage users to adopt, then began raising prices to improve margins. Microsoft appears to be taking a similar approach. Will it work?