LinkedIn distilled OpenAI models into smaller in-house ones and kept AI serving costs flat as paying customers passed 25,000
What's happening? LinkedIn has kept the cost of running its AI Hiring Assistant roughly flat even as paying customers grew to more than 25,000 businesses, up from 20,000 in September, executives told The Information. Customers pay from several hundred dollars to over $1,000 per user per month.
How? VP of engineering Prashanthi Padmanabhan said LinkedIn distilled OpenAI models into smaller in-house models built on open-source bases (which ones were not disclosed) for candidate search, ranking and outreach, saving up to 67% versus what it paid to run OpenAI models a year ago. CTO Erran Berger said LinkedIn does not plan to raise GPU spending in the fiscal year ending June 2027.
The caveats: New features still launch on OpenAI models first, so total cost is flat rather than falling. OpenAI has cut prices since, and LinkedIn would not compare its savings against current pricing. LinkedIn also has an unusual advantage: through parent Microsoft's deal it can use OpenAI models without licence fees, paying only for compute, and is allowed to distil them, something OpenAI bars most other companies from doing.
Why it matters: It is a rare set of hard customer and unit-cost numbers for an AI agent product at scale, and a concrete example of the shift from frontier APIs to distilled, self-run models once a use case matures. That pattern protects app margins but chips away at model providers' inference revenue.
Read more: The Information
