Another possibility we haven't considered or discussed yet is what percentage of dies are being allocated to 6900XT cards as opposed to 6800/XT models. The yields on TSMC 7nm have been reported as being exceptionally good so there's a fair chance that most of the dies are coming back good. The 6800 having the full Infinity Cache and all of the memory controllers also lends some support to this.
Of course this doesn't mean that they don't get binned for a variety of other reasons such as hitting target clock speeds and economic factors, but in this case the only thing that matters right now is whether those full dies can hit the required clocks. Results with the 6800/XT shows that most chips can be pushed well beyond the reference limits, so it's not hard to imagine a majority of the full dies don't need to the weakest performing hardware disabled.
Normally you still might bin your chips simply because you don't expect you'll be able to sell that many of your flagship card at the price you want and lowering the price doesn't increase sales enough to make up for the loss of additional profit on each sale at the higher price. However, we're not operating under normal circumstances at the moment, so there's no reason to disable any shaders on any but the worst performing full dies.
There was the other rumor that AMD wasn't going to have AIB 6900XT cards at launch. I don't know if it's another case of those being delayed or by how much if they are, but there is an article from earlier this week on a few sites about an ASRock 6900XT
so there are some AIB cards in the works for sure. However, if you do believe that AMD is being greedy or playing dirty tricks, the best move (the ultimate bait and switch) is to devote as many dies as possible to the 6900XT and ideally keep as many of those for yourself.
The actual market price for the 6800XT (and everything else in its class) is so inflated that you'll be able to sell any amount of 6900XT cards at launch even if the $1,000 MSRP normally wouldn't be that good of a value. AIB 6800XT cards on eBay are getting bids up to $1,050 so anyone willing to spend $900 or more would gladly get a 6900XT at $1,000.
Even people who would not normally buy one probably might simply due to perceived lack of supply and uncertainty of when they can buy a 6800XT/3080 at MSRP. There are so many 3090 backorders that you can probably pick up some of those people as well even if you somehow manage to exceed your normal customer base's demand. Given the frenzy of the past month the scalpers will certainly buy up every last card that doesn't go to anyone in the former groups.
All of this still makes perfect business sense even if AMD isn't malicious, but if you want to be one of those people that believe they're up to no good then this is also exactly what would allow them to profit the most. I don't know if the material costs for the reference 6900XT are substantially higher than the other Navi 21 cards, but assuming they aren't, every reference 6900XT sold is $350 more in AMD's pocket.
There's a recent Tom's article
that puts AMDs wafers that aren't dedicated to Sony or Microsoft at about 10,000 per month. Most of this obviously goes to Zen 3 production since it's the most profitable, but they also need to dedicate some to supplying existing product lines that use TSCM 7nm assuming they don't have any inventory on hand that can be used to free up wafers. It's hard to say how much is being used for those other products, but it probably isn't zero.
So let's assume that AMD sets aside 500 wafers per month which is 5% of their supposed free wafers. With the size of Navi 21 you get right around 100 dies from each wafer. The good news is that with the suspected good yields, around 75% of those should be defect free. Assume that you only artificially bin about 10% of those to ensure performance. This leaves you with about 66 dies per wafer that could be sold as a 6900XT.
Normally you probably can't sell 66% of your dies as a flagship card, especially not at $350 over your value-oriented card. Under normal circumstances you might be happy to sell 25% of the die as the flagship model. I did some exceptionally casual analysis to arrive at this figure, but without a more thorough analysis I don't know how true it is today, but the products used to arrive at the result are similarly situated to the 6800XT and 6900XT in terms of price and performance difference. But remember, we're not in normal circumstances.
So if we assume that AMD is evil and is able to get 66 6900XT dies per wafer and that normally they couldn't expect to sell more than 25 of them at the $1,000 price they're asking, this means that 41 additional 6900XT per wafer will be sold due to some extremely diabolical planning and capitalizing on circumstances. Assuming 3 months of production at 500 WPM, that's a little over 60,000 extra cards and at an extra $350 each, another $21 million in the pockets of AMD.
Now there are a variety of ways in which this analysis could go wrong, but there are enough factors at play that it starts to turn into a Fermi problem. So if AMD were in fact utilizing this strategy (whether you think it's utterly dastardly or just good business sense) their extra profit is probably in that ballpark and not off by more than an order of magnitude in either direction and quite possibly constrained by a factor of two in either direction.