[mike_hearn]

Not the OP but Zitron makes clear errors:

• He seems to think that the moment Nvidia release new hardware, all existing hardware becomes worthless. It doesn't and there are plenty of tokens being served by old GPUs. This makes all his calculations about how quickly datacenters have to pay off useless.

• All his numbers about costs, revenues etc are guesses or attempts to work backwards from off the cuff and frequently inconsistent comments by tech executives. They could easily be very far off.

• He doesn't seem to understand that datacenters have never been full of hardware on their opening day. A lot of his attacks revolve around this confusion - he learns that an opened datacenter isn't yet at full load or fully equipped with GPUs and thinks that means it's been delayed. I remember when Google first opened their facility in the Dalles, it took years for it to completely fill with machines.

[cmiles74]

> All his numbers about costs, revenues etc are guesses or attempts to work backwards from off the cuff and frequently inconsistent comments by tech executives. They could easily be very far off.

Agreed, but I'd argue that Ed doesn't have much else to work with. I'd like to see journalists take this tack and start asking these executives to either back up their statements or back down from them. They should be held accountable for their statements.

Even if we dial down these numbers by a magnitude they are still insanely large and the AI companies do not seem to be making enough money to balance things out.

> He seems to think that the moment Nvidia release new hardware, all existing hardware becomes worthless. It doesn't and there are plenty of tokens being served by old GPUs. This makes all his calculations about how quickly datacenters have to pay off useless.

I agree that older hardware from Nvidia doesn't become worthless when Nvidia releases new, more powerful hardware. I have to point out that it certainly loses a great deal of value and that's not nothing.

> He doesn't seem to understand that datacenters have never been full of hardware on their opening day. A lot of his attacks revolve around this confusion - he learns that an opened datacenter isn't yet at full load or fully equipped with GPUs and thinks that means it's been delayed. I remember when Google first opened their facility in the Dalles, it took years for it to completely fill with machines.

Is that really the case? I mean, I read about the build out of these data centers being delayed all of the time. I read this last week and it seems roughly in line with Ed's ravings:

> A JPMorgan analysis last month found that more than 60% of data-center capacity planned for completion in 2027 isn’t yet under construction, and another 7% is delayed.[0]

[0]: https://www.msn.com/en-us/news/technology/america-s-data-cen...

[JacobAsmuth]

H100s installed 4 years ago are more expensive to rent now than they were on day 1. It is not at all clear that older hardware is losing its value in a world where the next gen model is smarter and faster due to improved training+inference algorithms (e.g. custom kernels) but runs on the same hardware.

[LogicFailsMe]

It's either new GPUs make the old ones worthless or old GPUs make the new ones too expensive because they're still useful, it depends which ranter you're reading at the time.

Just like Michael Burry kept comparing NVDA to CSCO and now he doesn't do so anymore now that NVDA's P/E is ~31 and CSCO's is ~41. Funny that.

[ai_critic]

It helps if you look at Zitron's work history and experience. He's a hype man and a games journalist. His opinions on this are whatever sells, not exactly whatever is correct.

This is alarmingly obvious whenever he talks out of his depth about things like how companies actually use AI and reason about business decisions.

[michael-ax]

accuracy and precision are not the same thing. he's delivering one, you're asking for the other. no?

[ai_critic]

To put it more bluntly: he provides neither in his pursuit of rage views.

[sumeno]

They don't immediately become worthless, but they don't last all that long either

https://www.tomshardware.com/pc-components/gpus/datacenter-g...

[CuriouslyC]

This doesn't match my experience, in academia I saw ~40-45% utilization NVIDIA GPU clusters that went 6 years with <20% failure rate. Might be a TPU thing?

[CraigRood]

I'm FAR form an expert on this, but I believe that the operating costs such as power + cooling form a big part of the lifecycle. I have no doubt that at some point within the 6 years that are being booked, that replacing entire working racks won't be more cost efficient.

[CuriouslyC]

That is current practice, yes. The economics of replacing racks then selling the old ones to people who will salvage and resell working components works out better than trying to repair/retrofit in place.

[dofm]

> He seems to think that the moment Nvidia release new hardware, all existing hardware becomes worthless.

I am the OP and I totally agree with you on this one point. In fact the progress being made by open weights models strongly suggests that some of this hardware has much more of a life.

The overarching point he makes about incomplete data centres is that the current offering is running successfully on that very incomplete capacity, right?

What he is saying is that he cannot believe the demand exists to fill any of the unbuilt stuff, but much of it is still commitments that are going to have to be paid for, unless they can be backed out. He points to Nadella essentially confirming there will be overcapacity.

He also makes an interesting point that people tend to think "I can't get a GPU right now" means "there is intense, live demand for GPUs in data centres" when in fact the reason you can't get one is buy-and-hold. Including much of that new replacement hardware: it is being bought even the old stuff would (let us stipulate will) do the job.

I think he (or someone who interviewed him) recently said it reminded them less of the dot com boom and more of the Chinese real estate bubble.

[mike_hearn]

Future demand is unknowable. He might be directionally correct but wrong in magnitude, or right about everything, or wrong about everything. Unfortunately people who call bubbles never make their claims falsifiable or do anything else to build confidence, like take short positions. Zitron attacks the very notion that he might put skin in the game like that as obviously crazy.

I don't know to what extent we can say the current offering is running successfully. Anthropic have had visible capacity constraints for 18 months now with lots of throttling and quota capping going on. Those are good signs that demand does exceed supply at the current price point.

Additionally, Mythos has not launched publicly and one reason seems to be that it's too slow/expensive to make widely available, i.e. is capacity constrained.

But supply/demand is always in equilibrium, in some sense. So you could argue that it's currently balanced, or would be if priced correctly. That tells you little about future demand though.

[dofm]

All of this is fair, but it's also important to weigh your criticisms of Zitron's claims against the absolutely unsupported claims being made by Altman on a regular basis. They never show their working in even the way he does.

FWIW on capacity constraints, my gut instinct is that like every other startup these AI companies are are really only now beginning to do the serious efficiency work, because they had money and resources to throw at scaling without it; never optimise too early is pretty much a startup mantra.

[mike_hearn]

Sure, I'm not claiming that anyone who isn't Zitron must therefore be completely reliable.

I think all the labs have done a lot of efficiency work for a long time, tbh. You can see the evidence in their papers, open source releases and product design choices like model routers. They know they need to reduce their cost base a lot to become profitable.