The NVL72 is 72 chips is 120 kW total for the rack. If you throw in ~25 kW for cooling its pretty much exactly 2 kW each.
(Quick, inaccurate googling) says there will be "well over 1 million GPUs" by end of the year. With ~800 million users, that's 1 NVIDIA GPU per 800 people. If you estimate people are actively using ChatGPT 5% of the day (1.2 hours a day), you could say there's 1 GPU per 40 people in active use. Assuming consistent and even usage patterns.
That back of the envelope math isn't accurate, but interesting in the context of understanding just how much compute ChatGPT requires to operate.
Edit: I asked ChatGPT how many GPUs per user, and it spit out a bunch of calculations that estimates 1 GPU per ~3 concurrent users. Would love to see a more thorough/accurate break down.
With that kind of singularity the man-month will no longer be mythical ;)
With varying consumption/TDP, could be significantly more, could be significantly less, but at least it gives a starting figure. This doesn't account for overhead like energy losses, burst/nominal/sustained, system overhead, and heat removal.
So at that point a DC replaces them all with ASICs instead?
Or if they just feel like doing that any time.
To be clear, I am comparing power consumption only. In terms of mining power, all these GPUs could only mine a negligible fraction of what all specialized Bitcoin ASIC mine.
Edit: some math I did out of sheer curiosity: a modern top-of-the-line GPU would mine BTC at about 10 Ghash/s (I don't think anyone tried but I wrote GPU mining software back in the day, and that is my estimate). Nvidia is on track to sell 50 million GPUs in 2025. If they were all mining, their combined compute power would be 500 Phash/s, which is 0.05% of Bitcoin's global mining capacity.
All-in, you’re looking at a higher footprint maybe 4-5kw per GPU blended.
So about 2 million GPUs.
Google is pretty useful.
It uses >15 TWh per year.
Theoretically, AI could be more useful than that.
Theoretically, in the future, it could be the same amount of useful (or much more) with substantially less power usage.
It could be a short-term crunch to pull-forward (slightly) AI advancements.
Additionally, I'm extremely skeptical they'll actually turn on this many chips using that much energy globally in a reasonable time-frame.
Saying that you're going to make that kind of investment is one thing. Actually getting the power for it is easier said than done.
VC "valuations" are already a joke. They're more like minimum valuations. If OpenAI is worth anywhere near it's current "valuations", Nvidia would be criminally negligent NOT to invest at a 90% discount (the marginal profit on their chips).
30 TWh per year is equivalent to an average power consumption of 3.4 GW for everything Google does. This partnership is 3x more energy intensive.
Ultimately the difference in `real value/MWh` between these two must be many orders of magnitude.
[1] https://sustainability.google/reports/google-2025-environmen...
You over-provision so that you (almost) always have enough compute to meet your customers needs (even at planet scale, your demand is bursty), you're always doing maintenance on some section, spinning up new hardware and turning down old hardware.
So, apples to apples, this would likely not even be 2x at 30TWh for Google.
More than a "Google" of new compute is of course still a lot, but it's not many Googles' worth.
AI that could find a cure for cancer isn't the driving economic factor in LLM expansion, I don't think. I doubt cancer researchers are holding their breath on this.