Thanks for sharing your thoughts. I agree with many points you mentioned. Fiber/Roads/data centers etc. are good examples of infrastructure. The bold part above is the real question, that I asked earlier as well.You make some good points.If by your logic if we just replace GPU with the CPU then whatever you say also applies to CPU as per your logic. But clearly, the existing data center model is working well.
BTW, land doesn't depreciate. Also, whatever depreciates is a tax saving. NVIDIA H100 life-span range is between 5 to 10 years
Anyway, If we as BH believe that market knows the best then NVDA is top of the VTI.
The question we should ask is could this be highly profitable business? What do these people know that we don't know/understand otherwise so many companies and investor won't be investing in these.
This is infrastructure, not a good analogy but it is easier to understand, it is a highway. Future will be built on it.
The CPU vs GPU one is good. Kupperman's argument doesn't seem to rely on GPUs anywhere, and my rambling about Google's historical depreciated data center capex suggests there has been data center capex going on for decades that is generating very good returns at scale for some businesses. Google have their own flavour of proprietary GPU tech (they call them TPUs) used internally in their data centers, specialised for supporting AI / machine learning workloads.
I agree land doesn't depreciate, but Kupperman's estimate was the cost of the land and building only amounts to 25% of the overall capex, so the land value is maybe 10%-15%. The building will slowly depreciate, but the remaining 75% capex is in cooling, power systems, servers and GPUs/CPUs, all that stuff depreciates relatively quickly.
The life span of an individual H100 GPU may be 5-10 years, but the question is not how many years you can run it until that part fails, it is how long it makes sense to continue running a venerable H100 that's a few generations behind the state of the art, instead of the latest and greatest GPU which offers superior performance / watt.
We can regard data centers as infrastructure, but unlike earlier generations of infrastructure booms such as building railroads or laying fiber optic cable, the stuff in data centers becomes obsolete a lot faster. We might plan to get 20-40 years of life out of railway tracks. Similarly for fiber -- most of the cost of fiber is the labour of digging the trench, the cost of the fiber cable itself is insignificant.
Railway and fiber networks are both very valuable pieces of slowly depreciating infrastructure that have hugely benefited the economy and many businesses, but early investors in many railway and fiber projects lost everything as the early companies went insolvent after incurring all the capex, and their infrastructure was acquired cheaply by new investors with new companies who had a plan to turn a profit.
I have heard quite a few comparisons between AI data center capex and the telecommunications boom/bubble. Some of the comparisons are focusing on the circular financing deals. A different anecdote I have heard, in favour of AI, is that when fiber was being built out, nearly all of the installed fiber capacity was unused (dark), some of that capacity is still unused today. Whereas apparently when these new data centers are being brought online, their capacity is immediately being fully used. So arguably they're not being overbuilt.
There's still a question of if the current level of demand is sustainable, in the sense it corresponds to business models that generate enough actual value for customers with adequate profits for those running the datacenters. That's less clear. Easy to have a lot of demand if you're selling $5 bills for $1.
Thinking about it a little more, for fiber infrastructure, the bottleneck is the capex cost of digging the miles of trench or laying an undersea cable. The material cost of the fiber itself is insignificant, so you'd have been mad not to install 100x or 10,000x more fiber capacity into the hugely expensive trench or undersea cable than what you immediately needed. So that explains why so much of the fiber was "dark" when installed, the capital cost of the fiber itself was insignificant, it was not the bottleneck.
For these new data centres, I'm not sure what the bottleneck is. Power systems? Cooling systems? GPUs? Electricity generation capacity? All of the above? In any case, unlike fiber, all this stuff is not cheap. So that's a bit different.
As an example, I am pretty sure there would have been some great minds betting against the highway infrastructure when it was being built about 100 years ago and might have thought it is not going to be profitable.
It is perfectly fine and good to be skeptical. But, there are arguments on both the sides why these could be enormously successful.
There will be "commodity" effect in future as well. As of now, this infrastructure and tech is very high profit margin business because it is not commodity yet. e.g. 72% gross margin and 52% net profit margin for NVDA. There are very good reasons why it is $4.5T company at the top of the Total Stock Market Index Fund.
Which new companies will be born, which companies will thrive and which won't survive is hard to tell but I am absolutely convinced that it is going to have transformative effect on the society.
BTW, 2024 Chemistry Nobel Prize recipients include Sir Demis Hassabis and John Jumper, Google DeepMind.
In 2020, Demis Hassabis and John Jumper presented an AI model called AlphaFold2. With its help, they have been able to predict the structure of virtually all the 200 million proteins that researchers have identified. Since their breakthrough, AlphaFold2 has been used by more than two million people from 190 countries. Among a myriad of scientific applications, researchers can now better understand antibiotic resistance and create images of enzymes that can decompose plastic.
Statistics: Posted by babystep — Sat Oct 11, 2025 6:52 pm — Replies 323 — Views 30313