This week, Google released EmbeddingGemma 2—a 308M-parameter open-source embedding model under the Apache 2.0 license, allowing free commercial use and self-hosting.
More worth discussing than the new model itself is a comment by well-known developer Simon Willison on Hacker News: he systematically lays out, for the first time, the hidden costs that closed-source embedding models impose on enterprises.
What This Is
An "embedding model" can be thought of as AI's coordinate converter: it turns a sentence into a string of numbers (a vector), letting machines find similar content by "meaning" rather than "keywords"—enterprise AI knowledge bases and the "search-before-answer" mechanism in intelligent customer service both run on this under the hood.
EmbeddingGemma 2 is an upgrade from Google's previous generation, focused on small size, multilingual support, and open source. In plain terms: free, commercially usable, and privately deployable.
Industry View
Simon's core argument is straightforward: embedding models have a unique "sunk-cost trap." Once an enterprise adopts a closed-source solution, it pre-computes and stores millions of company documents as vectors. The day the provider deprecates an old model and pushes a new one, the old vectors become incompatible with the new model—enterprises either pay to recompute all the data or migrate elsewhere. This hidden cost can run from hundreds of thousands to over a million dollars.
He notes that in April 2024, OpenAI voluntarily covered users' costs to migrate embedding data, but he believes enterprises cannot count on every provider doing the same.
The counterargument deserves equal attention: An open-source license does not equal "worry-free service." Who maintains it? Who patches security vulnerabilities? Who do you call when something breaks? For traditional enterprises without an AI engineering team, the real barrier to "running open-source models yourself" may actually be higher than "paying for SaaS." Apache 2.0 offers freedom, not after-sales service.
Impact on Regular People
For enterprise IT: If your company is evaluating or has already procured an "AI knowledge base" or "intelligent customer service" tool, you can now start asking vendors: which company's embedding model sits on the back end, and under what license? Closed-source isn't wrong, but you must write "deprecation and data-migration clauses" into the contract.
For your career: In the short term you won't feel any change, but whether your company's internal AI assistant searches accurately depends on these models running under the hood. Knowing a bit more gives you one more professional question to ask when communicating with IT down the road.
For the consumer market: The biggest beneficiaries of open-source embedding models are mid-sized SaaS vendors and independent developers—they can build AI search features at lower cost, and the various small tools you use will get smarter in the coming years.