People I talk to, active in the industry today, say: never buy used AI chips.
“the chips at the heart of the infrastructure buildout have a useful lifespan of one to three years due to rapid technological obsolescence and physical wear, but companies depreciate them over five to six years. In other words, they spread out the cost of their massive capital investments over a longer period than the facts warrant”
Perhaps a better way to say it is - a new generation of chips, more powerful, and with better power efficiency, will arrive before the old chips actually wear out. So the new chips replace the old ones before they are “dead” (depreciated). Sure some old chips wear out over time, but it’s a tiny percentage when compared to the larger fact of switching over as newer technology arrives.
or underwear!
Does it really matter if they get rode hard and put away wet, or simply become obsolete? Either way, longer depreciation schedules as compared to chip lifespans leads to risky business.
To organizations looking for capital efficiency, it does make a difference. “Obsolete” chips can still do the work, just slower and presumably less efficiently than newer ones. Not every task needs the latest and greatest model. In addition, using the models generally requires less compute power than training them does.
Because of that, obsolete but still functional chips can still have a useful economic life. Yet no-longer-functional chips generally become e-waste.
Regards,
-Chuck
That’s fair. I suppose it comes down to whether companies choose to continue using obsolete chips, or buy newer chips.
Kinda seems like we’re still in the run-up phase of AI, where US companies will continue spending massive amounts on the newest technologies.
I think you have it.
This was, at best, poorly worded:
Maybe. But possibly not. I’ve been reading that they are now measuring AI data centers, not in terms of GFLOPS or processing power, but rather in MW (Megawatts) or electric power consumed. It could be (I am not sure, but it sounds like it could be true) that it would be economical to discard old AI chips that consume 20% (or 50%, or whatever) more power and replace them with new, more efficient, AI chips … despite the cost of buying those new, more powerful, and more efficient, AI chips.
Or perhaps they might opt to sell those old AI chips to others who have different economics regarding capex (buying new chips) versus opex (paying the power bills). Maybe they would sell those old chips to someone operating in a place where excess power is available at very low prices (sometimes even zero during periods of excess power that has to be “burned” somehow).
This gives me a good chuckle…800 M active mostly not paying users???
Lot of good that does.
If this was your business, you’d be crying if you kept pouring money in for naught.
Another way the maths don’t work. First you charge nothing. Then you wearout the product extra early.
That’s the ticket!
NVDA has to be happy. Make a product that won’t last and charge again.
Excuse me please hold this bag!
Oh and to have this product at all you plan on building nuclear power plants. No cost there.
If someone calls your office and gets the AI operator, the caller immediately demands a human being. No savings there. 95% of corporations understand they are flushing money.
And yet another reason why I’m glad the solar panels start going up on my roof tomorrow. The project has an estimated 6% ROI over the warrantied life of the panels, assuming 3% annualized growth in electric rates. If AI and crypto demand squeeze generation and grid capacity, there’s a very real chance rate growth could be faster than that.
I am very much a fan of nuclear power. That said, it takes a lot of time and heavy up front financial investment to bring new nuclear capacity on line. Until that investment is completed, something has to give…
Regards,
-Chuck
I need to see the capex trajectory as well.
And also cumulative revenue divided by cumulative capex over time.
These are all wrong questions to ask.
Unless you’re, you know, going to run a business.
Since when is asking “what is my ROI?” a wrong question on an investment board?
When it’s said that is a wrong question, maybe that is a sign of a bubble.
Focusing on ROI and Capex when demand is growing exponentially is dumb.
You guys are running around “bubble talk” for last 20 years.
Is there evidence that demand is growing exponentially? Capacity probably is, but demand?
Most recent numbers I can find: AI Adoption Rate Trending Down for Large Companies - Apollo Academy
I am not sure how you define exponential growth, but I don’t think this is it.
Seems to me that the demand for AI is only projected to grow exponentially.
But sometimes projections don’t agree with the real world. When projections are far more optimistic than reality, that is when you get a bubble.
Perhaps summer 2025 is just an aberration but there does seem to be reason to be a wee bit skeptical that the revenues generated from the actual use of AI will come anywhere close to what appear to be some very optimistic projections that are driving the market.
There is no right question to ask.
This is not going to age well.
When you are down and out…again, might I say, you have a career in stand up.
Yes, adoption rate of firms is very different from demand. Review the last quarter earnings call, everyone from $GOOGL, $AMZN, $MSFT, … are saying we are not able to meet demand. Even better, they are reporting now, look for the demand commentary.
None of that would matter, if you are not going to keep an open mind. Someone looking at the chart you posted could drive a simple conclusion demand has been growing exponentially (3x) in the last 2 years, and still growing in most groups, and the consolidation you are seeing in one group is pretty normal, especially given the tariff related uncertainty. Someone looking at that graph and declare the demand is not growing is someone who already made up his mind.
One can present any amount of data, none of that is going to convince the folks who have made up their mind that we are in the middle of a bubble. I have seen that in erstwhile Berkshire board.
The stupidity of arguments they made against Amazon, is/ was a teaching moment. Even today, they are smarting how dumb Henry Blodget projections were. without even recognizing for a second, his projections were absolutely correct, in fact he was somewhat conservative on his long term projections. I remember, when chamath, for all his shenanigans subsequently, in 2016 predicted within a decade $AMZN will be a trillion $$$ company when it was $300 B Market cap. Everyone declared chamath should be sent to asylum. Yet here we are where $AMZN is a trillion $$ MCap. And still they celebrate the guy who declared $AMZN is worth $10 or $0.5 split adjusted. Everyone gets things wrong, but that single trick pony, boy it requires special talent to be wrong multi decade on most things and still be celebrated
So there are two set of investors, one set of optimists, who see blue sky and then there are others who are pessimists, too scared by past failures.
Neither group can convince others. In the long run optimists make money and generally happy, the pessimists, at least for those who see from outside, live a miserable life.
