Rihard Jarc's analysis frames the move from always using the priciest frontier model to routing most work through cheaper open-weight options as a volume booster that still funnels infrastructure spend through the big three clouds, leaving dedicated labs to absorb most of the margin squeeze while Amazon, Microsoft, and Google keep their sticky cloud economics intact.
I just published my article on how hyperscalers like $AMZN, $MSFT, and $GOOGL will benefit in the token optimization era and why I believe the companies will re-rate much higher.
https://www.uncoveralpha.com/p/why-token-optimization-is-a-gift
Even as per-token prices fall, total inference demand is expected to explode under Jevons-paradox dynamics, and every one of those tokens still crosses hyperscaler networks where operating margins have historically held in the mid-30s regardless of which model generated the output.
Frontier-model providers keep high gross margins only on the small slice of truly hard workloads that still require them, while the majority of cheaper requests bypass their APIs entirely and route through orchestration layers that capture little or no model margin.
No Digg Deeper questions have been answered for this story yet.
@RihardJarc “…might *not* be…, but it does…n’t mean…?”
Is that typo? Should it be “but it *does* mean”?
Great write up, by the way. Keep up the good work. 🤝
@BillAckman Bullish is the easy part, but it's the CapEx that's wild. We're seeing billions poured into H100s while the actual software revenue hasn't caught up to the hardware spend yet. It's a massive bet on future utility that hasn't hit the P&L.
@BillAckman Thanks for sharing Bill! 🙏
An important read on the hyperscalers and why we are bullish.
I just published my article on how hyperscalers like $AMZN, $MSFT, and $GOOGL will benefit in the token optimization era and why I believe the companies will re-rate much higher.
https://www.uncoveralpha.com/p/why-token-optimization-is-a-gift
@RihardJarc Current market weakness is temporary; after a necessary correction, the landscape will hit new highs.📈🔽 https://www.fenxizhiku.com/links/542.html
@RihardJarc Great post!
@BillAckman I think most are missing the point about hyperscalers like $MSFT $AMZN and others
They just broke their peg to $SMH semis and memory
Happens every cycle
Though this is why the CapEx factor may make 2026 a different type of break (or will it?)
https://offside-research.beehiiv.com/p/the-most-important-ai-divergence-of-2026
@RihardJarc As always, great stuff. Thank you Rihard!
@BillAckman I have lost faith $fnma $fmcc
@RihardJarc You aren't even factoring in the loss of hyperscaler revenues as a result of users moving to much lower margin open source models away from OpenAI and Anthropic. That's a BIG deal considering 50+% of AI revenues were supposed to come from frontier labs which now looks unlikely.
@Pregory1 Thx and thx for flagging the typo,correct is “does”.
@RihardJarc Great article. It's similar to a power grid, when you switch from a 500W halogen bulb (frontier model) to a 10W LED bulb (open-weights), it probably means you'll leave it on 24/7 and also install a bulb in every room. Hyperscalers will sell all that electricity
Great article @RihardJarc! So if frontier model labs(Anthropic and Open AI) suffer, it means models are becoming commodity. So training them to the moon might not be that good cost-benefit wise. Doesn't that make for a negative case on training GPUs for $NVDA? Also, if these frontier labs are suffering, won't that also reduce their appetite(due to cash issues) for making their custom silicon, which might make the case for $AVGO weaker?
@BillAckman The ackman cry is upon us
He has groveled
Its time to go full retard on hyperscalers
@RihardJarc Added to the readinglist
@RihardJarc People are totally missing the security risks of using open weight Chinese models. Open weight doesn't mean they won't insert subtle trapdoors into your code or steal your IP.
@RihardJarc Great post, given this do you have any opinion on Anthropic's current evaluation? Frothy or fair?
@BillAckman What about free local AI models?
@BillAckman Assumes that infrastructure layer is sticky, non-commoditizable part of the stack. Tollbooth analogy depends on the bridge being scarce. Serving open-weight tokens is the most commoditizable activity. If the model layer commoditizes, why doesn’t the toll?
@BillAckman But Microsoft is trading near its lows and feels like a dead stock.
