I would argue that most ML processing by most "big data" companies also has marginal negative output for each additional GPU. Certainly anything to do with social media or adtech and a lot of analytics in general is a net negative for the world, while consuming electricity and processing hardware.
Even if the security budget is zero-sum, the amount of actual security added to the network with each additional GPU is constant.
Security is an easy argument to reach for when you want to justify something stupid.
Yes, the hash rate of the network increases, thus the difficulty for an attacker to perform a double spend. But, arguably, that's an already insane level of resources, many orders of magnitude higher than what a bitcoin-like payment system would reasonably require.
So even if you consider the network socially useful, pouring thousands of times more hashing power that the network actually requires to perform its task securely is still a waste of resources.