Why Michael Burry isn’t always right?

Scion disclosed a 1 million-share equivalent put on Nvidia valued at about $186.6 million, and a 5 million-share equivalent put on Palantir worth roughly $912 million. The filing shows trades from the last quarter, and Scion's current holdings may differ from the disclosure.

Before you get panicked like the rest of the market, do realize that these were trades from last quarter and we do not know when those put options expired.

In a cryptic post on X captioned "Move along," Burry shared a highlighted page from a book. The text described the 2000 telecom bubble, noting an "extreme imbalance" where "less than 5 per cent of US telecoms capacity was in use" and "Thousands of miles of expensive fibre optic networks remained 'unlit' beneath the ground."

Short note to Michael Burry’s comparison to fibre optic networks, and why it isn’t comparable:

- Fiber optic networks are not generational
Fiber optic network are not generational like how CPU and GPU will grow from 20 years ago to today’s generation. It is highly unlikely for it to be “unlit” because each generation of the chip enables more complex application and systems. This is historically true as not just AI alone, our cloud compute demand has not shrunk nor stay stagnant.

- Fiber optic network do not grow, and not easily replaccable
Secondly, you won’t be able to replace optic network the same as replacing servers. We don’t see new version of fiber optic being replaced every year. Telcom doesn’t crack open the network to “replace” or “upgrade” this system as regularly as servers.

Based on Moore’s law, compute power has been growing along side the applications. This is definitely not the case with fiber optic network.

- Lease-based system in-place:
   OpenAI and Oracle are well aware of the risks of building out data centers as this isn’t the first time Oracle is doing this. Which is why they have been looking into leasing chips instead of buying chips out right. This allows for Nvidia to keep developing newer chips and technologies without OpenAI and other companies incurring high up-front costs. This is also why Nvidia is investing in OpenAI.

- Future-proofing:
The contracts between OpenAI and Nvidia are not based on an agreement to buy a specific generation of chips. but it is based on Gigawatts. Both parties understand that the technology will keep evolving. The current compute demand is already at its limits so it is not possible to say that it will be “unlit”. Datacenter infrastructure is not even a new concept.

- Multi-functional unlike fiber optics:
Datacenters are not single-function single-purpose infrastructure. It is not entirely confined to the limits of AI alone. And even within the applications of AI, it is not solely confined to today’s AI models. Alot of bitcoin mining companies like Coreweave have since converted their machines for AI, and they similarly also faced criticism in the past for excess compute. This isn’t new.

The current Blackwell chips are still in high demand, and is the most advanced technology. Even 2-3 years from now, when the chips have generated enough revenue to recoup its costs, they are not going to become useless or “unlit” right away. These chips can still be used for less-demanding applications after 3-4 years (as applications will get more demanding), they are not going to be “unlit” or useless.

Even in the gaming industry, alot of the older Nvidia models still have high resale value especially for high-end chips. Like how GeForce RTX 3080, released 5 years ago, is still fairly desired in resale market.

Simply put, fiber optics cannot be repurposed for other types of applications.

In other words, comparing fiber optics network to Compute Demand is like comparing Apples and Oranges.

Note: Trolls/Haters/Speculators in comments will automatically be blocked. This is to ensure a healthy discussion for those who genuinely need sound advice not speculations. If you wish to troll/speculate, do it elsewhere