I'm not sure that you and I actually disagree about all that much, and might be just arguing cases in different adjacent debates.
The realities of the modern hyper-sophisticated semiconductor fabrication business is that the chips aren't consistent with event a single logical "run" (or whatever the modern term of art is), and so you're binning and blowing fuses and all kinds of stuff to get even the same SKU in the ballpark as other cards with the same SKU (throw in MSI vs whoever, it just gets harder), and so clearly there is a mutually beneficial amount of "smoothing" that's more than zero (impossible) and less than anti-competitive (56% profit margins and a CEO saying "this is a gold mine" on camera).
And as someone who is clearly more current on the nuances of the gamer-oriented lineup, I'm inclined to take your word for it that gamers want a little more "smoothing" than ML people without a blank check (though I don't think I can endorse referring to the public as needing the same guidance as any other child, that's not me trying to chastise you, it's me noting that underestimating the public has burned a lot of smart people some rather, uh, recently).
It is complicated, and there are hard realities of the physics and tolerances of the manufacturing that complicate the situation, and there are hard questions in what constitutes a "major version" and what constitutes a "refresh" or a "rev". So to the extent you're saying that the vendors need some rope to give their customers something that isn't a casino by binning and labeling releases with an eye to some kind of non-ridiculous trendline in capability that isn't coming or whatever, yeah sure. I don't know enough about the pressures of a middle manager working on that stuff to offer hard policy prescriptions about it. I do know about the pressures of an upper-middle technology executive faces with this whole other bag of constraints over and above the already hard job about the "the Street is going to kill us if X isn't Y by Z for Q1": that's a problem I think you'd agree no one would actively solicit as part of their job description all else equal.
There's this different basket of stuff that's ethically and practically anti-competitive, anti-market, and at least used to be illegal most of the time (jurisprudence on how narrowly to interpret the anti-trust laws we do have has become an Olympic sport in narrowing their scope). That's the drivers I simply don't believe to be unfixable in a year or two, that's an unreasonably wide gap between a 2k card you can run PyTorch on and a 50-80k card (or whatever the real price is, I'm not sure anyone even knows in a public way) that's basically the next step up, that's the 4090 spending die area and TDP on vector units it doesn't have the GDDR6 to keep fed half the time. And finally, it's accepting margins that let you price cards at MSRPs you're willing to publish and meeting the demand for them at your stated price. This isn't an act of god chip shortage (in general, there are instances of that, particularly during COVID): I hate to pick on Meta here because the reason I'm using them as an example is that they're reasonably transparent about this (so keep it up guys, that's awesome), but IIRC their last publicly known block of H100 (Hopper v1 for lack of a better term) was 150k cards in one transaction, with a target of something like 700k "A100 equivalent compute units" by the end of 2024. I'm not surprised Lisa Su and Jenson Huang are inclined to do as much of that business as possible, but again, market failures left and right.
Maybe gaming and AI need to be different companies to get an efficient market, maybe export controls need to be fiddled with further, maybe a lot of things. Gamers should be able to get cards that play relevant titles at MSRP, startups and independent researchers should be able to get cards that work and don't require a blank check, and big AI labs should be able to get a reasonably level playing field in terms of access to the big iron. But a monopsony quasi-cartel of frenemies competing in some ways but with their fates intertwined in other, cornering the market on process nodes and still getting gouged by vendors having a "spirited debate" and not a UFC match. That's not getting us the benevolent global outcome of the locally selfish invisible hand. That's where sparing but well-refereed interventions in failed markets is what makes capitalism sustainable after the sugar rush wears off of having industrialized in the first place.
That's the stuff I think Intel is going to go on a rampage around if Gelsinger pulls off the turnaround.
The realities of the modern hyper-sophisticated semiconductor fabrication business is that the chips aren't consistent with event a single logical "run" (or whatever the modern term of art is), and so you're binning and blowing fuses and all kinds of stuff to get even the same SKU in the ballpark as other cards with the same SKU (throw in MSI vs whoever, it just gets harder), and so clearly there is a mutually beneficial amount of "smoothing" that's more than zero (impossible) and less than anti-competitive (56% profit margins and a CEO saying "this is a gold mine" on camera).
And as someone who is clearly more current on the nuances of the gamer-oriented lineup, I'm inclined to take your word for it that gamers want a little more "smoothing" than ML people without a blank check (though I don't think I can endorse referring to the public as needing the same guidance as any other child, that's not me trying to chastise you, it's me noting that underestimating the public has burned a lot of smart people some rather, uh, recently).
It is complicated, and there are hard realities of the physics and tolerances of the manufacturing that complicate the situation, and there are hard questions in what constitutes a "major version" and what constitutes a "refresh" or a "rev". So to the extent you're saying that the vendors need some rope to give their customers something that isn't a casino by binning and labeling releases with an eye to some kind of non-ridiculous trendline in capability that isn't coming or whatever, yeah sure. I don't know enough about the pressures of a middle manager working on that stuff to offer hard policy prescriptions about it. I do know about the pressures of an upper-middle technology executive faces with this whole other bag of constraints over and above the already hard job about the "the Street is going to kill us if X isn't Y by Z for Q1": that's a problem I think you'd agree no one would actively solicit as part of their job description all else equal.
There's this different basket of stuff that's ethically and practically anti-competitive, anti-market, and at least used to be illegal most of the time (jurisprudence on how narrowly to interpret the anti-trust laws we do have has become an Olympic sport in narrowing their scope). That's the drivers I simply don't believe to be unfixable in a year or two, that's an unreasonably wide gap between a 2k card you can run PyTorch on and a 50-80k card (or whatever the real price is, I'm not sure anyone even knows in a public way) that's basically the next step up, that's the 4090 spending die area and TDP on vector units it doesn't have the GDDR6 to keep fed half the time. And finally, it's accepting margins that let you price cards at MSRPs you're willing to publish and meeting the demand for them at your stated price. This isn't an act of god chip shortage (in general, there are instances of that, particularly during COVID): I hate to pick on Meta here because the reason I'm using them as an example is that they're reasonably transparent about this (so keep it up guys, that's awesome), but IIRC their last publicly known block of H100 (Hopper v1 for lack of a better term) was 150k cards in one transaction, with a target of something like 700k "A100 equivalent compute units" by the end of 2024. I'm not surprised Lisa Su and Jenson Huang are inclined to do as much of that business as possible, but again, market failures left and right.
Maybe gaming and AI need to be different companies to get an efficient market, maybe export controls need to be fiddled with further, maybe a lot of things. Gamers should be able to get cards that play relevant titles at MSRP, startups and independent researchers should be able to get cards that work and don't require a blank check, and big AI labs should be able to get a reasonably level playing field in terms of access to the big iron. But a monopsony quasi-cartel of frenemies competing in some ways but with their fates intertwined in other, cornering the market on process nodes and still getting gouged by vendors having a "spirited debate" and not a UFC match. That's not getting us the benevolent global outcome of the locally selfish invisible hand. That's where sparing but well-refereed interventions in failed markets is what makes capitalism sustainable after the sugar rush wears off of having industrialized in the first place.
That's the stuff I think Intel is going to go on a rampage around if Gelsinger pulls off the turnaround.