These photos were taken about 59 minutes into Jensens Keynote, check it out. He said as much and with the same number of exclamation points. It’s a little confusing when he doesn’t update the $1T number; but, I don’t think that should excuse the Analysts from doing the smallest bit of math🙄
Best
Jason
Thanks @WillO2028 for highlighting the two slides. The 2nd one of 100x compute required for Inference i would assume is partially CEO talk. But sure the R1 (and o1 before that) breakthrough certainly changed the paradigm of inference. But lets’ keep the 100x of what we thought of in 2024 aside.
So from 1st slide-that claim of 3.6M Blackwell GPUs sold just in 2 months vs 1.3M in Full year seems absurd. Some back of the napkin calculations if that is true:
2024:
Revenue
2025:
That is absurd. So i am assuming the 3.6M is the deals made with CSPs for the FY. That might increase a little bit but not a whole lot. As the planning and contracts will be done for the whole year. Re-doing the calcs based off of this:
Adding them all up, total Rev for 2025: 264B (+100% YoY). That is still crazy. So what am i missing. Any feedback on assumptions are welcome. Some potential wrong assumptions:
I’m thinking Jensen is counting the fact that Blackwell is sold out for the rest of the year, was a bout a month ago.
So there’s that. Even thought the revenues may not, are likely not, collected fully I’m thinking that none of this is CEO speak and instead a reasonable assessment of the current state of play in this space. Did you listen to Jensen around the 59 minute mark?
There is about ten minutes of explaining why he’s using the 100x more compute figure.
Best
Jason
There are two points that we are talking about:
Regarding Future, Jensen said we need 100x more compute than what we thought we needed back n 2024. As per his explanation, the reasoning paradigm, or more precisely the success of Reasoning paradigm as demonstrated by R1 has changed the ball game. I do not disagree with the assessment just that the 100x seems a little hard to believe. I might be wrong. Without revealing too much, but i work very closely with the Reasoning on/off paradigm and how it impacts models etc.., That is all what we have been doing since December. Either way i wanted to err on the conservative side when looking for future projections, in informing my Nvidia investment decision.
More importantly, in my reply above, I focused mainly on 1. Current i.e. revenue estimated for 2025 based on the number that Nvidia already sold 3.6M Blackwell GPUs. If this number is true, and totally ignoring the 100x increased predictions for future, the revenue estimates for 2025 seems on track to double which is really good at this scale. Do my calculations make any sense?
Your response, yourself closely working with AI, says volumes about why the share price is so low, IMO. I get that your optimizing spend by toggling between train of thought and not (edit: although now you’d likely be better off using Nvidia’s Dynamo to optimize this). I believe many reasonable people are thinking the same way about this.
I don’t believe anyone has the line of sight Jensen has, with 12 months at 1.5x volume (2 chips per Blackwell, he’s counting each one when he says 3X) completely sold out, 900 libraries locking in CUDA, and a global supply chain/partners, all unparalleled in history.
[Edit:I’ve gone back through the Keynote and listened to others take and 
… I added the following edit out respect to the point of this thread … scaling growth like we’re seeing with Nvidia is mind blowing which makes investing in it difficult.
The “last 12 months spent producing Hopper” refers to all of 2024, where 1.3 million GPUs (and thus 1.3 million chips, assuming 1 chip per GPU) were made and sold.
To compare production fairly, we’d need to know the production rate (chips or GPUs per month) for both periods, not just the sales totals. For example:
Hopper production rate: 1.3 million chips ÷ 12 months ≈ 108,000 chips/month.
Blackwell production rate: 7.2 million chips ÷ 2.5 months ≈ 2.88 million chips/month.
That’s about 26.6 times the monthly chip production rate (2.88 million ÷ 108,000), but this assumes all sales came directly from production in those exact periods, which isn’t necessarily true—some GPUs could’ve been stockpiled. Huang didn’t provide production rates, only sales figures, so we can’t definitively say production is “twenty-six times” or any specific multiple without more data. Anyone?.]
How could the average investor zoom out to a 1-3 year horizon with all this growth scaling exponentially😳. I spend more time watching Jensen Huang interviews than anything else for this investment.
Nvidia 31% of my holdings 
Just to clarify. Jenson is not saying $1 trillion in Nvidia revs by 2028, but $1 trillion in AI factory total spend by 2028. The portion of which Nvidia intends to grow that is theirs.
They could grab more of the spend and let’s say 50% (Microsoft only got a portion of PC spend but they kept trying to grow that portion). So that would be $500 billion of data center revs in 2028 + their other product categories. 50% is just a guess, but quite possible.
Nvidia, as an example, would not want Vertiv’s lower margin cooling infrastructure portion. They can take more and more of the networking budget say from Broadcom or ALAB.
Either way, enormous economics. But unless something changes, not $1 trillion to Nvidia (yet).
Jenson has been far more accurate in his forecasts and predictions than critics who are still announcing Nvidia killers but then admitting Nvidia still has no real competition.
