The Facts
On Monday, Databricks announced via the WSJ that they had acquired MosaicML for $1.3B.
This marks one of the first large acquisitions of the Generative AI era, and it wasn’t cheap — as Matt Turck pointed out in the aftermath, the acquisition comes out to ~$21M per employee at Mosaic.
Why it matters
When teams estimate the cost of training language models, you typically see math that looks something like “number_of_gpus * cost_per_gpu_per_hour * time.”
This acquisition reminds us of the real cost of business — the talent required to train LLMs well is incredibly scarce and incredibly expensive. Last week, another article estimated the median comp package offered to a software engineer at OpenAI at upwards of $900k. Training models at the largest scale is an incredibly challenging engineering effort.
MosaicML assembled a great team of talented ML engineers (and put together what seems to be world-class tooling to train language models). Having previously worked at a similar company to MosaicML (Determined AI, acquired in 2021 by Hewlett Packard), I’d be surprised if the market for MosaicML’s tooling justified a “valuation” of $1.3B. There aren’t enough teams training that large of models at scale.
Their talent probably does justify a $1.3B price tag in 2023. It’s an arms race, and Databricks wants to play. They definitely don’t want to get passed.
My thoughts
It’s also work revisiting the Mistral AI fundraise — a ~$250M valuation viewed by this same lens is ~$60M per employee. It will be a while before we get a good measure of the “fair market value” of the best AI talent, but I honestly wouldn’t be surprised if the answer is in the ballpark of these numbers. The prize is too large to make small bets right now.