I've addressed this point elsewhere in this thread and in a thousand other comments but it bears repeating: there are either zero or one groups of people who understand economics mathematically depending on whether or not Medallion is a a Ponzi scheme that's lasted 35 years (I tend to think they invented attention models at the same time LeCun invented working convolutional networks). If someone makes more money than the indices consistently by "using math" on markets they are lucky as hell (in a manifold compactness no fucking way sense), have some grey edge in the mix, or work at Ren.
Price discrimination is generally prohibited in naked form: it often if not usually induces market failures. So companies spare no expense working around it (there's a reason why ITA software people made a mint approximating NP-hard problems acceptably well and rather cheaply to compute airplane ticket prices).
"Market segmentation generally benefits consumers" is exactly the kind of Wien's Law first-order approximation that sounds compelling but doesn't fit the data.
It does not benefit consumers that a 4090 has the same amount of GDDR6 in it as a 3090, which is incidentally 2.6Gb shy of `dolphin-8x7b-v1-q4_k_m` (or whatever the NVIDIA equivalent is, I use Apple gear now because inference is the game now), while the price of that RAM dropped. The fact that an H100 or whatever the next Hopper iteration is costs more than an M-series BMW on paper says nothing about what Meta paid for 150k of them or whatever it was, and it certainly doesn't help someone who'd like a decent chat bot without sending their data to eyeball-scanner database guy. When you see a leading edge x090 for MSRP or less at Best Buy, it's a good idea to stock up if you're in the market over the last 5 years (London has been running ETH2 for what, a year and change now?). That hardly seems a subsidy to the consumer paid for by those spendthrifts in FAANG.
NVIDIA pioneered serious "GPGPU" engineering. 5-10 years ago their microarch and fab-sourcing and software and the whole show were legitimately differentiated, no one had done it. And they made a pile, and that pile seems "conflict-free". But anyone can make a fused multiply-add unit these days, EE undergraduates do it in Verilog. Intel makes PCI express cards that cost $300 and have higher memory bandwidth in the decoder-only language attention model use case. And their drivers are documented and work on everything.
Now did Lisa Su decide to "concentrate on the supercomputing market with the MI300XYZ" and Jensen decided to "concentrate on AI with Hopper" independently to a degree where the market is perfectly partitioned? Who knows, I certainly don't have proof one way or the other. But if someone made a call being like "I'm thinking of focusing on X but don't really see our differentiation in Y. How's Cathy?", it wouldn't be the fucking first time.
But even that isn't critical to the point: this, the here and now, isn't good for consumers, or society, or our industry, and unlike e.g. Stallman's melodrama, it fucking matters this time.