EP 16 - Google's Dream Team Just Quit, and Nobody Can Find the AI Bear Case

Four of Google's top AI researchers walked out the same day, on top of others who left earlier this summer, including Jeff Dean after 25 years at the company. Anthony Batt and Harry DeMott open by asking what that exodus actually signals to a business owner or investor: not that Google is finished, but that its best talent would rather build AI somewhere else. From there they work through what they call the theater of the AI industry: Anthropic and OpenAI's steady drumbeat of existential warnings, which they read as marketing and regulatory positioning as much as genuine caution, Mark Zuckerberg's new manifesto on invention over automation, and Elon Musk's comparatively quiet rollout of Grok 4.6 and Grokbot, priced at half of what OpenAI and Anthropic charge at the frontier.

The second half of the episode turns to what they call the receipts: the actual demand data behind the AI buildout. They walk through Gavin Baker's search for a single bear-case data point that came up empty, hyperscaler cash flow up 28 to 32% year over year, GPU rental rates roughly doubling in seven months, and NVIDIA's $500 billion consortium to turn compute into a tradable asset. They also get into Michael Burry's depreciation argument against the AI trade, why NVIDIA effectively controls its own depreciation schedule, and Kieran Goodwin's "nobody knows" thesis on where this all lands. They close on a practical note for anyone running a business right now: hardware prices are climbing, and the real bottleneck isn't chips, it's finding people who already know how to use AI well.

What we get into:

00:01 Welcome to Episode 16

00:50 The plan for today: theater, concrete moves, and the receipts

01:23 Why mainstream AI coverage is mostly headlines, not analysis

02:00 Google's AI product problem and four top researchers walking out the same day

03:43 Jeff Dean leaving after 25 years, and what that signals

05:01 Demis Hassabis stepping back as Sergey Brin returns to run AI day to day

06:09 Founder energy vs. management energy, and the Apple comparison

07:46 Apple's real innovation is chips, not developer experience

08:06 Is Google finished, or just a sleeping giant

09:03 Not just the starting five walked out, it was the whole all-star squad

09:50 Has search peaked? Do google rankings even matter anymore

10:52 Newsletters, podcasts, and YouTube as the new discovery layer

11:13 Anthropic's fear-based messaging as a business strategy

12:35 Elon calling Anthropic "misanthropic," and the safety-marketing debate

14:09 If the model is really that dangerous, why not just shut it down

14:39 The cynical read: regulatory capture dressed up as safety

16:29 Manufacture danger, then ask Washington to regulate you

18:12 Where real caution is warranted versus where it's just marketing

19:31 Ranking the trust levels: Google, then Sam Altman and Dario Amodei, then Zuckerberg

20:50 What AI actually changes about how people think, versus how search worked

23:29 The literacy gap: most people don't know how to ask AI the right questions

25:00 Zuckerberg's new manifesto on invention over automation

26:19 A great founder with products that are bad for society, and why that's still worth admiring

28:37 Facebook's history of buying and copying instead of inventing

29:29 Elon Musk, SpaceX, xAI, Colossus, and Cursor

29:54 Why Microsoft barely comes up in this conversation at all

32:46 Grok 4.6 and Grokbot ship quietly, at half the price of the frontier labs

34:13 Why Grok plus Cursor's coding harness could be a serious combination

35:07 The Texas gigafactory and Elon's fully vertical AI stack

37:32 Grok isn't the best model yet, but the gap looks closable

38:14 Owning atoms versus renting compute: Elon's structural edge over OpenAI and Anthropic

39:05 Cursor's session data as an unintentional recruiting tool

41:41 Turning to the receipts: is AI demand actually real

42:16 Gavin Baker's search for a bear case that came up empty

43:52 GPU rental rates roughly doubling in seven months

44:07 The addressable market: half a million agentic AI users out of 8 billion people

44:29 The data center PR problem, granola versus water usage

46:46 No one paying for AI today is going to stop using it

48:51 CoreWeave's earnings as clean proof of accelerating compute demand

49:20 NVIDIA's $500 billion consortium to turn compute into a tradable asset

50:21 Five-year-old chips already sold out through the end of the decade

51:40 Michael Burry's depreciation argument, and why NVIDIA controls its own schedule

52:46 Kieran Goodwin's "nobody knows" thesis on where this all lands

56:35 Rising compute costs as the new cost of doing business

57:49 A loaded gaming PC now costs $4,500, and a Mac Studio runs $8,300 with a six-week wait

1:02:32 The real bottleneck isn't chips, it's talent

1:03:14 Why the AI talent gap will take roughly nine years to close

1:05:58 Using AI to plan a real vacation instead of a Google search

1:07:39 Zero times zero equals zero: what you bring to AI determines what you get out

Future-Proof is a weekly conversation between Anthony Batt and Harry DeMott on where AI, developer tools, and the businesses building on top of them are headed next.

Subscribe for a new episode every week.