Neerav Mehta

Quant researcher / trader, software developer

[Neerav Mehta @neeravbm]

@a_living_fossil @orrdavid Based on the trajectory, I see that most current white collar jobs will disappear in the next 5-10 years. Maybe AI companies will earn 1/100th of what those employees earn today. That’s still a huge TAM.

[Neerav Mehta @neeravbm]

@a_living_fossil @orrdavid And I don’t even know how many developers are supporting MS Office vs building new products in Microsoft. If it’s a very small fraction, then that explains it. I don’t use Shopify so I can’t comment on it.

[Neerav Mehta @neeravbm]

@orrdavid Can’t same happen in AI inference?

[Neerav Mehta @neeravbm]

@orrdavid Need to simplify various factors. PCs / telecom have gone 1000x cheaper if you compare on performance basis (clock cycle, RAM size, bandwidth, etc.). What happened was that falling cost accelerated demand so much that they kept improving hardware at slightly falling prices.

[Neerav Mehta @neeravbm]

@orrdavid Does your argument hold for the previous capex cycles (PC, telecom/internet) where hardware became 1000x cheaper over time?

[Neerav Mehta @neeravbm]

@sjb987654321 @orrdavid This is true in enterprise software, where engineering accounts for only 20-25% of total cost.

[Neerav Mehta @neeravbm]

@orrdavid I agree with that. But this scenario is actually good for AI capex if there’s lot more competition, right?

[Neerav Mehta @neeravbm]

@orrdavid I see where you are coming from. You are saying that at 1/100th cost, there won’t be 100x software, which could be true.

[Neerav Mehta @neeravbm]

@orrdavid If I had to build all this without AI, I would just do basic algo on a single node with little analytics. AI gave me the ability to build 10x more in 1/10th the time and even lesser cost.

[Neerav Mehta @neeravbm]

@orrdavid As an example, I have been building infrastructure for my own algo trading. Currently my infra supports multiple nodes with failover, paced participation execution algo instead of just market orders, extensive analytics dashboard which I am constantly adding to.

[Neerav Mehta @neeravbm]

@orrdavid There’ll be way more software. You’ll be able to build features that you weren’t able to build earlier because of time and money.

[Neerav Mehta @neeravbm]

@a_living_fossil @orrdavid Agreed. I don’t think we have seen peak models yet. There is still a huge scope of improvement in just coding if the model builders can increase the context size. Reduction in inference cost will in fact increase demand.

[Neerav Mehta @neeravbm]

@SowingAlphaSeed Okay, maybe 0.8, which is still better than S&P. What you are saying is that no small company can survive because there is one dominant company in the same space, which I don't agree with. In fact, bigger hedge funds have a disadvantage due to size / slippage.

[Neerav Mehta @neeravbm]

@orrdavid @SowingAlphaSeed Agreed.

[Neerav Mehta @neeravbm]

@orrdavid @SowingAlphaSeed But even memory is a commodity so I don’t expect either memory or HDD companies to keep any unfair advantage over their competitors.

[Neerav Mehta @neeravbm]

@orrdavid @SowingAlphaSeed Yes, HDDs are not going extinct in near future.

As for the company valuations, right now all semiconductor companies are being priced high because of supply bottlenecks. You are right that memory companies will have more pricing power over time than HDD companies.

[Neerav Mehta @neeravbm]

@orrdavid At data center scale, even small reliability difference matters.

[Neerav Mehta @neeravbm]

@SowingAlphaSeed @orrdavid Agreed.