Back to deep dives

SemiAnalysis debates Google’s AI exodus, compute bets, and agent-built software

  • Agents
  • AI Infrastructure, Compute, Chips, And Energy
Thumbnail for SemiAnalysis debates Google’s AI exodus, compute bets, and agent-built software
Image: Ep. 23 - Everyone Leaves Google, Elon Forecasts 1T ARR, Reflecting On GPT-5 | Jon from Asianometry

IntroductionSection 01

Published August 7, 2026, this SemiAnalysis discussion brings together Jon Y of Asianometry, Doug O’Laughlin, and Jordan Nanos to examine where practical AI advantage is accumulating. They use senior departures from Google DeepMind to debate whether a giant research organization can keep turning breakthroughs into products when rare technical leaders leave. The conversation then follows the money into SpaceX and the biggest cloud companies’ compute plans, before showing how coding agents let Jon build personal software and why Nvidia’s software support keeps outside teams on its GPUs.

Senior AI Departures Are Reshaping Google DeepMindSource12:06

The prototype seems to be that these guys leave and then they raise money from a bunch of investors including Google Ventures. And then go buy a bunch of GPUs and do whatever they want.

The panel treats Jeff Dean’s exit and the departure of several Gemini leads as a broad transition inside Google DeepMind. Predictions that Demis Hassabis could become Alphabet CEO remain speculation.

  • Hassabis is taking the chief scientist role while continuing to lead Isomorphic Labs, according to the discussion.
  • Dean helped create foundational Google infrastructure including MapReduce, Spanner, Bigtable, TensorFlow, and TPUs.
  • Prominent researchers can leave Google, raise outside capital, buy GPUs, and pursue their ideas independently.
  • The loss of several less visible Gemini leads suggests wider organizational change, rather than one famous departure.

Google's Research Culture May Not Turn Breakthroughs Into Winning ProductsSource17:28

We're going to become an infrastructure company. We're going to do good enough. We're going to have a good enough product on the transformer but they're probably not going to stay for the next and the next and the next.

The panel debates whether Google produces exceptional research but struggles to turn it into winning products. Jeff Dean’s departure sharpens the disagreement over whether large teams can replace rare, system-wide technical judgment.

  • Google’s broad bureaucracy may be less focused than an AI startup, although startups can also become bureaucratic.
  • Maps, Google Cloud, Kubernetes, and TPUs challenge the claim that Google cannot execute.
  • A critic argues that Google still failed to capitalize fully on its early cloud advantage.
  • Financial incentives could push Google toward infrastructure and a good-enough AI product while still rewarding shareholders.
  • TSMC shows the strength of distributed expertise, but Dean’s breadth across hardware and software may be unusually difficult to replace.

SpaceX's Compute Forecast Pulls a Trillion-Dollar Bet ForwardSource28:49

They're going to move their 1 gawatt forecast to 10 gawatt by the end of 2027 and 20 gawatt after that. which means that they are now he believes bringing in their $1 trillion forecast from 2031 to 2030.

The panel says SpaceX plans to move from 1 gigawatt to 10 gigawatts of compute by the end of 2027, then to 20 gigawatts. That reported acceleration underpins an earlier $1 trillion forecast.

  • The host attributes the 10- and 20-gigawatt figures to Musk’s comments to SpaceX investors.
  • He moves his $1 trillion forecast from 2031 to 2030 because of the faster buildout.
  • The panel asks what happens if about $1 trillion of industry spending grows toward $10 trillion over five years.
  • Microsoft, Amazon, Google, and Meta are also spending heavily on chips, increasing the chance of abundant compute.

Coding Agents Let Jon Build Software Around His Own WorkflowSource35:58

And I have my own video editor and I've integrated Wikipdia comments search into it. I've integrated a video search into it. Like it dropped it in with the correct attributions.

Jon used coding agents to build a compact video editor and other personal tools without understanding their implementation. Jon made the fourth version work by narrowing the editor to one job, writing a detailed specification, and testing the result.

  • Jon built a roughly 40-megabyte editor around his keyboard-heavy workflow after rejecting conventional editors.
  • Three failed versions led him to reduce the editor to one core job.
  • Claude helped produce a large specification, which he gave to Codex for a deliberately simple build.
  • A newer model completed a difficult timeline feature after asking roughly 25 requirements questions.
  • A few hundred dollars in subscriptions produced attributed search, faster editing, and a separate audio-processing app.

Nvidia's Software Support Keeps External AI Teams on GPUsSource42:52

Yeah, I think TPUs are great, but I think that the support you get outside of Google on the software side is zero. That's what Fable's for, bro. You're going to roll your own software.

The panel describes memory scarcity as a possible opening for a SpaceX chip business. It then argues that outside AI teams choose Nvidia because Google provides too little software and customer support for TPU users.

  • Memory shortages are constraining products in Taiwan and giving suppliers unusual pricing power.
  • CXMT reportedly refused to discount a large Apple order, showing the strength of current demand.
  • SpaceX and former Google leaders are cited as examples of outside teams choosing Nvidia systems.
  • TPUs may train models well, but outside users would need to recreate tools, software, consoles, and support.

Claims & connections

Tags

  • AI Infrastructure
  • Coding Agents