# AI capex — X 热门讨论 (2026-10-08 08:05 UTC)
## @masayoshisson (Masa’s Son) · 10-08 05:02 · ♥45 ↻4 💬3 I used to respect coatue a lot
but man, either this a desperate attempt to save the AI bags through the lab IPOs, or hiring dan ives and cathy wood would actually improve their investment judgment
couple highlights below
1) AI capex in 2028 $3tn: listen, I understand that AI is all fun and games if there’s no marginal seller to take the other side of the trade and when enough retards agree to just yolo certain companies because the trade is so hot that there’s always another idiot that’s willing to pay more (ie VC) and marks just keep going up - but, once the labs go public, which they must bc private markets tapped out and VC LPs have not seen distributions since 2021, they have to produce cash flow or at least have a perspective on doing that. and right now there’s about a zero chance they’ll be able to do that or pay for their announced compute commitments. which means the markets will have to fund the shortfall/ act as retarded as yolo AI VCs (while q3 plateuing ARR numbers are not even official yet, the credit market is already smelling blood, sooo it ain’t looking pretty). I don’t know what the number will end up being, but it certainly won’t 4x in 2 years lmfao
2) $META ad revenue 2x over 5 yrs due to muse: so that assumes muse winner takes all (google, OpenAI, anthropic, spacex won’t compete at scale), agents won’t run local (apple, OSS), buyer behavior changes from dopamine hit to autistic productivitymaxxing and merchants will just accept yielding customer ownership? lol ok
3) labs should be valued on EBITDA pre training: for one, I was quite disappointed you didn’t call it what it is: EBTITDA, but here we are. more importantly, do you really think that public investors are retarded enough to believe that if training spend were to be turned off, you wouldn’t create a more severe token supply glut pressuring prices even more (or alternatively earlier compute cancellations) with the consequence of instant commoditization leading to almost certain blowup?
while the chances are small, I remain hopeful - maybe the IPO window closes as more investors use their brains, and the mean reversion of excesses of private markets/ attempted californication of public markets remains contained on VC and a few select lenders books
probably not, but will be fun either way > 引用 @coatuemgmt: Public markets update: https://t.co/tLENVFAb06 https://x.com/masayoshisson/status/2108060440861278462
## @mredmond88 (Michael Redmond) · 10-08 00:15 · ♥31 ↻8 💬2 Love this argument: if you trim the economic sectors the administration influences through tariffs, the Iran conflict, and AI Capex incentives, inflation looks pretty good 😆 > 引用 @DavidZervosDC: Some fun charts to support the statements by @SecScottBessent in the oval today on yields and inflation—>
https://t.co/09K4heL4Vp via @YouTube
Four key measures of core inflation have all been trending LOWER!!! And long-term inflation expectations have been remarkably STABLE!!! https://t.co/l2hjFyf5jn https://x.com/mredmond88/status/2107988192020582757
## @kokko_coco (kokko | 古賀興司) · 10-08 05:41 · ♥33 ↻4 💬1 GSがTSMC決算のプレビューをしている。要点は、AI需要がGPUだけでなくCPU・ネットワークまで広っており、TSMCの成長期間が長期化している。だからCapExを猛烈に増やしても、N2/N3やCoWoSはなお足りない、ということ。AgenticAIによるサーバーCPU需要の増加は「過去1年間の重要な変化」とし、さらにTSMC経営陣はTokenコストが下がればAI利用量そのものが増えると考えており、GSも顧客や各種関係者との会話から需要の持続性に自信を強めている、との由。結果的に、売上成長率も設備投資も上昇。そして、それだけ増産してもなお足りず、利益率も維持できる。AI半導体サイクルが、AIコンピュート・インフラ全体を大量に作るフェーズへ移行し始めている、ということだろう。 https://x.com/kokko_coco/status/2108070157331451955