Sunny

← All weeks

What I’ve been reading (W33, 2026)

Where AI creates value and how early we still are

Distributed superintelligence, where AI value gets captured, inference as a fundraising story, Nvidia underwriting its customers’ bets, tech debt at negative interest, and a hedged fund that was really one bet

  1. The future is for everyone; a good way to think about how investing in personal intelligence will lead to a better future. meta.com/thefutureisforeveryone
  2. We’re still very early in understanding where the economic value of AI will ultimately sit. Token prices are high today partly because compute is scarce, but expanding infrastructure and rapid efficiency gains should push costs down over time. If multiple models become similarly capable and easy to switch between, the model layer could increasingly look like commodity infrastructure: essential, but with limited pricing power. More of the value may then be captured by the applications, workflows, proprietary data, distribution, and customer relationships built on top. The key question isn’t whether AI creates enormous value, but where that value ultimately gets captured. ben-evans.com/benedictevans/…/ways-to-think-about-token-pricing
  3. The point of inference (serving AI models) revenue is to convince investors to give you more money to buy more compute to train bigger models. If you’re spending most of your compute on inference, you’re kind of declaring that AI progress has stalled, because it’s not worth investing more in training, and your business is basically that of a cloud provider. The labs do not think this is true; they think they are serving models that will look extremely shitty within a year in order to build up the business case to continue training smarter models. dwarkesh.com/p/why-compute-might-get-10x-more-expensive
  4. Nvidia is taking on more financial risk because it is no longer just selling GPUs; it is now helping customers finance huge AI data centers and, in some cases, effectively promising that the hardware will still be worth something years later. That makes it easier for customers to buy more GPUs today, but if AI demand slows, data centers make less money, or old GPUs lose value faster than expected, Nvidia could end up taking losses just as its normal chip sales are also falling. In simple terms, Nvidia is starting to take on some of the risk of its customers’ bets, rather than leaving all of that risk with them. stratechery.com/2026/nvidias-risky-business
  5. Situational Awareness (situational-awareness.ai, an AI fund betting on physical and geopolitical rise of AI) lost roughly $35bn of value in July, a 67% drawdown, after building a highly leveraged portfolio around one thesis: that AI would advance very quickly. The interesting part is that the fund looked hedged, long the AI winners and short the companies AI would disrupt, but both sides were really the same bet. When the thesis moved against them, both sides lost at once. The lesson: you can diversify your positions without diversifying your assumptions. archive.is/JGuuQ