Really?!
Maybe I'm just a bit more cynical, but I've been around these corporate heads in the finance industry long enough to know that if they could create artificial scarcity, enough to keep cards mostly out of stock and prices high, for the new gen and the last gen, and make more money doing it, year over year, they would do so in a heartbeat! If supply of the next gen cards wasn't an issue, the price of the top last gen cards would have to drop at least another 25-50% below where they're at now.
As you say, they've already got the data.
I don't buy into this mentality. Marketing might, but I think it's BS. Having something continually sold out is bad for the business, because it means lots of people aren't able to buy your product. Not having sufficient supply at launch is different, because
it's almost impossible to have sufficient supply on any new high-end GPU. You get 80-150 chips per wafer, max. So even 10,000 wafers goes pretty fast if you're Nvidia. That's basically a whole month's supply. Having ~5% less than demand is one thing; having 5% of demand is a completely different story.
Please see the October 2nd 2022 Q&A for investors at GTC. Jenson and the CFO, her name escapes me, answered in response to a question on supply of GPUs that they would be undershipping GPUs in Q2 and Q3 and would not begin ramping until late Q4 and/or into Q1 of 2023.
This is being done in collaboration with their partners to clear inventory of 3 series cards in a, as they describe it, 'soft market'. Supply likely to ramp into Q1 of next year with the likely 'sold out' status drives holiday shoppers towards taking what they can get their hands on into Christmas.
https://investor.nvidia.com/events-and-presentations/presentations/default.aspx
Its also amusing that we seem to flip flop on card pricing. Analysts were quoting horrible price cuts to 3 series GPU' last year but the footnotes showed the pricing was taken from EBAY sales numbers, not MSRP. The average sales price on ebay last gen was probably close to what MSRP is now from NVDA but now its too expensive? Why wouldn't they charge as much as they are knowing their consumer is willing to pay. (Discounting the fact that their N4 monolithic die is likely a ton more expensive then Samsung's 8NM used in 3 series.)
Nvidia cut production requests
for the 30-series GPUs in Q2/Q3 — this already happened. It was done to help reduce 30-series inventory. Nvidia did not cut Ada orders as far as we know; it has never suggested such. But Ampere orders were to Samsung and Ada was for TSMC, so it couldn't use Samsung to create more Ada GPUs.
As for the pricing last year, which I tracked monthly, I used eBay prices precisely because everything was constantly sold out and there was no retail availability. The only way to reliably find stock on an RTX 30-series GPU last year was eBay, same for RX 6000-series. As of maybe ~May/June this year, we finally got to the point where there was inventory on shelves and online at retail. Then I could stop using eBay. There was never a suggestion from anyone responsible that the eBay prices were
good — quite the contrary. So just because people (miners often) were buying stuff on eBay for stupid prices doesn't mean those prices were acceptable. They were "too expensive" then and are "too expensive" now.
GA102 manufactured at Samsung Foundry, Nvidia gets up to ~86 usable chips per wafer. Let's assume the wafer cost is $4,000. That's $46.50 per chip. Wire-bonding, packaging, etc. then cost extra, but those are mostly going to be the same regardless of process node. AD102 manufactured at TSMC gets up to ~89 chips per wafer. Let's say triple the wafer cost, so $12,000. That's $135 per chip, plus the wire-bonding, packaging, etc. The real cost isn't the chips, it's all the R&D that goes into creating the chips in the first place. Nvidia likely spent billions creating Ada Lovelace, and probably 50% more than it spent on Ampere just because things were more complex going to N4 process and packing in lots more transistors.
Facebook
Twitter
Instagram