.@romeovdean and I wrote a blog post to teach ourselves about the AI buildout. We were surprised by some of the things we learned: 1. There’s a huge fab CapEx overhang - with a single year of earnings in 2025, Nvidia could cover the last 3 years of TSMC’s ENTIRE CapEx. In 2025,
@dwarkesh_sp @romeovdean there is less than $25B in total AI revenue across all companies for $500B of capex spend, that will likely depreciate even faster now 😬
@dwarkesh_sp @romeovdean Genuine question, the human brain seems to be orders of magnitude more energy efficient. Did you consider a scenario where the pressure leads to better algorithms decreasing the need for so much power?
@dwarkesh_sp @romeovdean the tech equivalent of flooring it in a Model T on a muddy road, convinced that more throttle will summon warp drive.
@dwarkesh_sp @romeovdean Can’t wait for my GPU allocation to show up on my utility bill
@dwarkesh_sp @romeovdean Great post. On your points about China: Compute power is software, firmware, and intellectual property. The West still dominates CUDA, EDA tooling, compiler stacks, and advanced packaging IP. Even if China builds the physical plants faster, the high-end stack remains
@dwarkesh_sp @romeovdean Would love to see an overlay of capital needs to stay solvent for OpenAI mapped on the %agi forecast chart
@dwarkesh_sp @romeovdean i've played civilization enough to know that production matters as least as much as scientific progress indeed the #1 thing usa has to fix rn
@dwarkesh_sp @romeovdean Amazing article. Congratulations!
@dwarkesh_sp @romeovdean I enjoy your podcast but you really should do some more research before writing things like this. Depreciating an asset is an accounting procedure. After equipment is fully depreciated, it does not have to be replaced. It continues to be used. $NVDA introduced the T4 chip in
@dwarkesh_sp @romeovdean The current lead time for gas turbines in the US is years long … to confirm your math
@dwarkesh_sp @romeovdean @grok how difficult is it to deploy Tesla megapack 3? How long will it take
@dwarkesh_sp @romeovdean What if hyperscalers don’t make money on their investments, e.g., due to too expensive power, chips/data centers becoming obsolete too fast, proprietary models losing against open source?
@dwarkesh_sp @romeovdean This is great, ty for putting together! Underscores that the stakes for the hyperscalers are now enormous, especially with all this debt being raised. They need, as highlighted, more revenue to support the spend - and it’s easy to be skeptical
@dwarkesh_sp @romeovdean Can't imagine what could happen if the bubble popped in the next few months.
@dwarkesh_sp @romeovdean I thought bro was just a talker. Bravo! Nice work.
@dwarkesh_sp @romeovdean When is your fab tour?
@dwarkesh_sp @romeovdean very nice work
@dwarkesh_sp @romeovdean This is just total especulation
@dwarkesh_sp @romeovdean The lead time for new transformer capacity, crgo steel production etc. is easily in 12-24 months and would need similar incentives