I don't disagree. But nVidia's market cap is based on their share price, not their actual revenue or profitability - which will remain unaffected by any drop in share price provided they're not over-leveraged with debt.
With a substantial drop , the profitability will likely sag some.
They are not entirely decoupled. Three things. A substantial portion of the demand is driven by "FOMO" (fear of missing out). The stock price increases is substantive contributing factor to stoking that fear. Tech companies throwing billions at Nvidia like drunken sailors on shore leave after 4 month long cruise. Some of what is going on here is buying up data center chips so that almost nobody else can get them. That is going to lead to higher profits than a properly functional (i.e., rational) market. Once the hype ( "company X is loosing simply because they are not buying NVidia training chip services" ) dies down , the price power also fades a bit.
Second, Nvidia paying customers large sums of money to buy their gear that suggests that customer really , really, really, don't want to pay that much for this stuff. That is a bit of 3 card monty where Nvidia books it as 'profits' in year n and then shovels the money back to the companies to buy more stuff in year n+1. Even if they get to keep that cycle going for 2-3 years, it is also a 'leaky' cycle. (gets lots more leaky if there is a mania collapse).
Third, All the large hyperscalar services vendors have their own inference SoC projects. AWS Inferentia (and Trainium ) , AZURE Miaia. Google Tensor somewhat predates Nvidia's data center efforts . China somewhat banned them (the USA is kneecaping them so two way effort there).
The more ML models get standardized , the easier they will be to port. CUDA has pragmatically a moat for Nvidia . Moats for fixed fortifications. "Fixed fortifications are monuments to the stupidity of man" -- George Patton
That moat isn't all that uniformly deep. The specific 'syntactic sugar' elements of CUDA can be wiped away with a decently trained ML tool to HIP or SYCL (or some other standardized portable format). The underlying algorithm will just come over.
The people that will lose their money when the LLM bubble bursts are the idiot speculators that don't actually understand the technologies or companies they're investing in. They're not really investing, they're gambling so they probably deserve it.
Nvidia doesn't has as much traction in inference as they do in training. ( auto ML market ... Nvidia isn't untouchable dominant. Smartphone market ... nope. PCs ... most deployments not there. etc. ) The notion of creating models so big that they 'have to' go to mega datacenters is also flawed. Making larger and larger models at some point hits diminishing returns.
That really isn't true at this point. The market indexes are really even weighted indexes. Many indexes are weighted to capture higher market cap stocks that are moving up. So when irrational speculative moves push too hard on a stock for long enough , it then tends to capture the index market funds also. The corporate board could be complete horse poo ( e.g, Tesla) and indexes will just keep blindly buying it up at distortional numbers because it hasn't tanked yet.
en.wikipedia.org
