How does Palantir actually make their money? I know probably a lot though government contracts, but I've never heard of them out side of the comp sci field until now.
Palantir was one of the very early Silicon Valley firms that were going around selling businesses the idea that "machine learning" and "artificial intelligence" would transform the way they work. A lot of them did buy it, and Palantir ended up with lots of funding (over $2 billion) and a ridiculous valuation ($41 billion at its peak, somewhere around $11 billion now - depending on who you ask).
They tried to build several iterations of one-size-fits-all software to ingest and process data and come up with visualizations and other analysis that they could sell to businesses, but they all failed.
Now, like 15 years later, they are just one of the hundreds of consulting companies that build custom solutions for businesses and other organizations to crunch their data. Their only differentiator is the fat government contracts they get because of Peter Thiel's connections.
- but we aren't selling you a finished product (didn't we mention that?), just a base framework that needs customization. Cue fly-in-fly-out consultants for exorbitant daily rates.
- to lower costs, why not send your own staff to our "university" for thousands of dollars for a 5 day class?
- hire said staff to become consultants after customer pays for their training. They don't appear out of thin air!
- when customer finally ditches product, blame them for lack of investment and use contacts to complain at highest levels of bureaucracy.
- rinse and repeat.
A tried and tested formula.
Government is far less concerned about money-efficiency, and nobody working in government has a direct interest in profitability (like being a shareholder would). Hence those tactics work all-the-better.
Simply say "this is your budget for the year, if you spend more, we'll fire every one of you and sell the office".
And then hire an entirely new team of non-overlapping staff next year.
But not when you are not actually buying a product.
I don't remember exactly, but one example improvement was that they detected temperature fluctuations inside the reaction tank (e.g. upper side hotter than bottom) during bad production batches.
This was detected by visualizing the temperature across the reaction time on good vs bad batches.
It's an interesting find, but personally I think the work they have to do to get there is super boring.