So it seems like everyone is on the "bubble wagon", including most people on HN.
Personally for me, I'm on the "it's only 1995 wagon". There are so many ideas I want to do based on LLMs but they require very cheap tokens (better hardware, more efficient models, smarter smaller models, etc.) So to me, we're not even close to peak AI hardware demand.
In my humble opinion, peak AI demand isn't anywhere close. This is in a world where China is so starved for sanctioned Blackwell chips that they're still stuck on GPT4-level training hardware.
Not to mention that I believe every phone and computer will need local hardware powerful enough to inference capable models. That's a lot of chips.
Demand forces is measured in $$$'s all the bubble talk is there aren't enough demand in $$$'s to justify projected supply, but more significantly, the time dimension of building supply and deprecation of supply is not favourable. We can simultaenously go through a crippling bubble burst that gives all the big players huge haircuts and still live in a post bubble world where demand/supply continues to increase because hardware improves, i.e. chips X gen from now can do 1T worth of data center compute with 100m, but right now that 1T worth of data center might only return 200b - the economics of the current iteration is bubbly. Almost obviously bubbly because we're dealing sychronizing (really mismatch) between very expensive, capex heavy, fast to build but faster to depreciating compute, and also very expensive, capex heavy, slow to build infra. It's a deadly combo. It's an EXTRA deadly combo because the fast deprecating compute component (i.e. 80% of investment) makes are fast spoiling stranded assets. These are not 50/100 year hard assets like rail or fiber rollout where most of the work is simply civil engineering - moving a fuckload of earth that can be reaped for generations. Most of chips and bespoke datacenter infra are >5-10 year assets (it's not tulips but it's closer to tulips than rail), there's no reason to use them if gen +X chips makes their capex + opex completely obsolete / economically unviable.
That's a strange take. To attack the feasibility of building the datacenters.
I think it can be done, and it's not that impressive if someone does it.
Here, I'll explain: datacenter building seems to be a grind. Careful planning, bureaucratic deals with suppliers, waiting for specific timings. You know beforehand it needs to be overscaled, because you can't adjust for it after the fact. That's the hard part. Not physical limits.
What if you could turn those expectations upside down? I'll present just one of the ways of doing it: repurposing components.
If you can convince people that you can repurpose the unused datacenter components into something that can be sold directly to customers, then you can be much less careful in regards to buying shit.
I think (but I can't prove) that the Mac Mini M4 is exactly that. Made from components of brand new servers that never went into production and were soon-to-be obsolete for that purpose. That also explains why it has a carbon neutral sticker (some other part of Apple paid the carbon offset for it, but it's there somewhere). The Xeon PC-Gamer fever that happened some years ago follow a similar pattern, and it's probably the first instance of this practice in recent history. Snapdragon has hints of being from the same general practice as well. So, all the cool kids are doing it.
So, if you can consider that you can overscale, then there's no mystery here. All that cautious datacenter building bureaucracy suddenly becomes much simpler. It totally can be done. You have an out if the servers don't get used, and you get another shot at filling the datacenter capacity next cycle.
My guess will have its time to be proven true or false. If we see a lot of weird consumer hardware (such as the M4 mini), that looks a lot like server components packed into a consumer package, then it's very likely that this is happening.
Therefore, I see no reason to attack the feasibility of these datacenters. The true question here is how much of their capacity will be used. That determines how much of the unused parts will become repurposed products.
However, this whole thing depends on people needing to buy more hardware. We've already seen the cycles getting shorter. In this datacenter brilliance, I do not believe these geniuses are accounting for the consumer saturation ("my M1 is still amazing and I still don't need another computer for the next 5 years"). If that can all fall apart (economy-wise), it's likely to be related to this consumer saturation.
Of course, there are other ways of convincing people that they need to be less dilligent about building datacenters (to skip the careful planning). One of such ways is to promote cyberwarfare, which is something I already addressed in other posts. War can cut bureaucracies, really fast. Perhaps it's a combination of these multiple ideas that's in play here.
Anyway, I think there's a bubble but the portrayal of such bubble by specialized media (both tech and economoy) has been too naive and simplistic. It points to a scenario of a controlled burst of some kind. People are probably relying on the assumption that the ordinary population will buy the bubble argument _exactly as it is portrayed_.
Personally for me, I'm on the "it's only 1995 wagon". There are so many ideas I want to do based on LLMs but they require very cheap tokens (better hardware, more efficient models, smarter smaller models, etc.) So to me, we're not even close to peak AI hardware demand.
In my humble opinion, peak AI demand isn't anywhere close. This is in a world where China is so starved for sanctioned Blackwell chips that they're still stuck on GPT4-level training hardware.
Not to mention that I believe every phone and computer will need local hardware powerful enough to inference capable models. That's a lot of chips.
I think it can be done, and it's not that impressive if someone does it.
Here, I'll explain: datacenter building seems to be a grind. Careful planning, bureaucratic deals with suppliers, waiting for specific timings. You know beforehand it needs to be overscaled, because you can't adjust for it after the fact. That's the hard part. Not physical limits.
What if you could turn those expectations upside down? I'll present just one of the ways of doing it: repurposing components.
If you can convince people that you can repurpose the unused datacenter components into something that can be sold directly to customers, then you can be much less careful in regards to buying shit.
I think (but I can't prove) that the Mac Mini M4 is exactly that. Made from components of brand new servers that never went into production and were soon-to-be obsolete for that purpose. That also explains why it has a carbon neutral sticker (some other part of Apple paid the carbon offset for it, but it's there somewhere). The Xeon PC-Gamer fever that happened some years ago follow a similar pattern, and it's probably the first instance of this practice in recent history. Snapdragon has hints of being from the same general practice as well. So, all the cool kids are doing it.
So, if you can consider that you can overscale, then there's no mystery here. All that cautious datacenter building bureaucracy suddenly becomes much simpler. It totally can be done. You have an out if the servers don't get used, and you get another shot at filling the datacenter capacity next cycle.
My guess will have its time to be proven true or false. If we see a lot of weird consumer hardware (such as the M4 mini), that looks a lot like server components packed into a consumer package, then it's very likely that this is happening.
Therefore, I see no reason to attack the feasibility of these datacenters. The true question here is how much of their capacity will be used. That determines how much of the unused parts will become repurposed products.
However, this whole thing depends on people needing to buy more hardware. We've already seen the cycles getting shorter. In this datacenter brilliance, I do not believe these geniuses are accounting for the consumer saturation ("my M1 is still amazing and I still don't need another computer for the next 5 years"). If that can all fall apart (economy-wise), it's likely to be related to this consumer saturation.
Of course, there are other ways of convincing people that they need to be less dilligent about building datacenters (to skip the careful planning). One of such ways is to promote cyberwarfare, which is something I already addressed in other posts. War can cut bureaucracies, really fast. Perhaps it's a combination of these multiple ideas that's in play here.
Anyway, I think there's a bubble but the portrayal of such bubble by specialized media (both tech and economoy) has been too naive and simplistic. It points to a scenario of a controlled burst of some kind. People are probably relying on the assumption that the ordinary population will buy the bubble argument _exactly as it is portrayed_.