Why are the Hyper-Scalers spending crazy amounts of money building out way beyond their current revenue?
1. Anthropic CEO stated they have on the books demand they cannot supply. With no further innovation, if they just doubled their capacity they will double their revenue. It's not a projection. It's already back-ordered demand. They just need more chips and the facilities to house them and they can print money.
2. There is a finite supply of GPU currently. In the early 90's MS used to hire up every big thinker and thought leader they could. If they didn't have a product for them, they would just stick them in research. They did this so they could have them when needed, and more importantly to deny that intellectual resource toa potential competitor. They'd rather have them off doing research for them rather than coming up with a competing product for an adversary. Even if the Hyper-Scalers lock the GPU they are buying in a vault and don't use them, they are denying a startup competitor from getting access to the core resource needed to compete in AI. Compute. So cornering the market of chip availability is smart anti competitive tactic and the big players with big money can do that.
3.
https://youtube.com/shorts/09TJOXmMBTM?si=L0S9PNQ2gAppR0MD All the AI chips come from Taiwan. When China goes for Taiwan in '27 it will take YEARS to replace that capacity in the West. Anyone already holding GPU will be like bars of gold.
4. They know the current AI design just gets smarter with more chips. Even without more advanced algorithms, they can just get smarter with brute force by applying more chips. There may be no limit to chip demand and production may always lag demand. Chips may be cheaper now than they ever will be as AI demand keeps exploding exponentially while supply stays fixed or craters. Buying chips now, might be getting a bargain even if you have to buy them on credit.
Pushing against this is the Chinese.
They can't make the chips themselves. They look at Taiwan and lick their chops. I don't think they will wait past 2027.
By their calculus, they either capture those facilities intact and enslave the engineers to work for them and become the Saudi Arabia of AI chips, or they destroy the facilities and prevent the West from reaching ASI before them, because whoever does first rules the world.
The Chinese have been the main leaders in LLM Compression.
The current Frontier model LLMs are MASSIVE and take insane amounts of compute to host and traverse. That's why they need so many chips and so much power and have a problem with excess heat.
However, a significant portion of that data is pure garbage. The estimates vary. The Chinese have been taking that first generation LLM and stealing from it (the 1st gen just stole content from the internet) and further compress it with fancy algo's and lots of cheap Chinese training labor. That would allow them to compete with much less compute that they could get away with buying on the black market. Hence DeepSeek.
Ironically they are doing the AI industry a favor. That is the approach small startups will need to take against giant Hyper-scalers who have limitless funding. Startups will never will a chip war against OpenAI and Google.
We just cant let the Chinese get to AGI or ASI first. Not unless you want to bend a knee and kneel to Xi.
