Google just put forty billion into Anthropic and the cloud wars look essentially over. Alex breaks down why this is really about maximizing economic value per token, why Google now controls roughly a quarter of all AI compute on the planet, and why TPUs are now designing TPUs in a recursive self-improvement loop that goes all the way down to the silicon. Then OpenAI counter-strikes with GPT 5.5: a massive jump on Terminal Bench 2.0 making Codex a real Claude Code competitor, plus 1% per month gains on Frontier Math Tier 4 meaning research-grade math is cooked in four years at current pace. Demis used to say five breakthroughs remained between us and AGI a decade ago. Alex's update from his Frontier Lab friends: it is half a breakthrough now, maybe zero. Plus: why Noam Brown thinks weights matter less than compute, why Anthropic is hunting compute everywhere, lunar AI data centers needing fewer regulatory approvals, fabs on the moon in maybe ten years not twenty, the OpenAI nonprofit-to-PBC trial getting jury-influence drama, Microsoft Recall as architectural atrocity, WorldCoin as Minority Report retina ID, professions cooked starting with law, mRNA vaccines as the actual Drexler nanorobots that finally arrived, and Henry Intelligent Machines turning every human into a tastemaker overseeing AI conglomerates. The singularity is not a point. It is an interval. We are right in the middle of it.

TIMESTAMPS

0:00 Cold open: Google commits $40 billion to Anthropic

0:09 GPT 5.5 drops, math is cooked

0:36 The three-way Western frontier race and the weights-versus-compute pivot

3:10 The average consumer is no longer in the equation

3:59 Where Kimi K2.6 and DeepSeek v4 actually fit

5:15 Sparsity as the path to a million-parameter diamond model

6:39 GPT 5.5 Terminal Bench 2.0 jump and Frontier Math Tier 4 cooked

9:20 Half a breakthrough left to AGI, per Demis

9:45 Google now controls ~25% of all AI compute, TPUs design TPUs

10:37 Powered land as the new strategic asset

11:36 The compute circular economy among the top ten labs

12:57 Lunar data centers and fewer regulatory approvals

13:26 Fabs on the moon in maybe ten years, not twenty

13:51 Anthropic's unifying principle: maximize economic value per token

15:31 OpenAI versus Anthropic as the next Sorkin Apple-vs-Microsoft

16:14 The OpenAI nonprofit-to-PBC trial drama and jury selection

18:40 Microsoft Recall and OpenAI Chronicle want to be in the OS

19:43 Privacy as a baked-in property of the silicon

20:30 WorldCoin as Minority Report retina ID

21:29 Hardware-level cryptography for camera chain-of-custody

23:08 The human-to-AI labor ratio trends to one-to-infinity

24:35 Professions cooked, GDPVal mapping all of knowledge work, law next

26:45 Of course this is about replacing doctors

27:32 Clinicians could not resist EMRs, they cannot resist superhuman AI

28:28 Clinicians hate EMRs but love AI, the surprising flip

29:02 Donor organs becoming unnecessary

30:09 mRNA vaccines as the actual Drexler nanorobots that finally arrived

31:57 CAR-T as the horse-and-buggy era of immunotherapy

32:36 Off-target effects as a secret weapon, drug repurposing on tap

33:29 Why AI table tennis took so long

34:13 Why did everything take so long

34:51 Henry Intelligent Machines and humans as tastemakers

36:50 Singularity as an interval, not a point

38:43 Revealed preference on superintelligence estimates

39:10 Where to follow Alex's daily takes: alexwg.org / Innermost Loop