TBPN Podcast Summary

Episode: Will AWS Buy Google’s TPUs, Remembering Claude The Gator, Ricursive Raises $35M | Diet TBPN
Date: December 4, 2025
Hosts: John Coogan & Jordi Hays

Overview

This episode of TBPN features a lively discussion between John Coogan and Jordi Hays about the rapidly evolving landscape of AI accelerators, notable news around Amazon’s Trainium 3 chip, Google’s TPUs, competitive dynamics with Nvidia, and the broader implications for cloud providers and AI companies. The hosts weave in real-world examples from startups harnessing these technologies and digress into Silicon Valley stories, fond memories of San Francisco’s albino alligator Claude, and the mechanics (and pitfalls) of secondary venture deals and new AI chip startups.

Key Discussions and Insights

1. AI Chips Wars: Amazon Trainium 3 vs. Google’s TPU vs. Nvidia

• Amazon’s Trainium 3 Launch

  • [00:42] Amazon Web Services announced the public launch of Trainium 3, a custom AI chip claimed to be 4x faster than its predecessor, produced by AWS’s Annapurna Labs.

  • [01:08] Annapurna Labs, acquired for $350M, has been developing AWS’s custom silicon for a decade, with ambitions to reduce AI model training and operation costs by up to 50% compared to equivalent GPUs.

  • [02:06] Real-world example: Dean Leiters, cofounder at Descartes, achieved a breakthrough in real-time AI video generation using Trainium 3, which his team developed during an intense coding sprint.

    “The moment that I saw it worked, I saw four people just start jumping up and down.” — Dean Leiters, relayed by John Coogan [02:29]

• Market Competition & Customer Choice

  • [04:12] Rumors that major AI firm Anthropic may favor TPUs over Trainium, even though Amazon remains an Anthropic shareholder.

  • [04:38] Interview snippet: AWS CEO Matt Garman states commitment to supporting customer choices and continuing GPU partnerships (especially Nvidia), echoing Jeff Bezos’s ethos of being “customer obsessed.”

    “If the customer says ... ‘I want you to give me an Nvidia GPU ... or a TPU in your server,’ they might do that because that's actually in Amazon's DNA.” — John Coogan [06:24]

  • [05:35] Trainium 3 may find its niche in specialized workloads (e.g., real-time video for Twitch streamers, which Amazon owns).

  • [07:10] Discussion of AWS and Google Cloud’s new partnership—to enable high-speed cross-cloud links—illustrates evolving, pragmatic alliances.

• Competition with Nvidia

  • [08:52] Nvidia’s astronomical financial growth: revenue ramping from $27B (2023) to $130B (2025), net margins swelling to 56%.

  • [08:52] This empowers Nvidia, but also incentivizes other hyperscalers and OpenAI to form an “anti-Nvidia alliance” to drive down costs.

    “56% net profit margins on $130 billion of revenue. ... That's a lot of acquisition.” — John Coogan [09:45]

2. Technical Comparisons: Chips and Capabilities

• Trainium 3 vs. TPU & Nvidia
  • [10:22] The “semianalysis crew” offers critical comparison:
    • Trainium 3: “New and underpowered ... just 667 TFlops. The gap between Nvidia and Trainium is actually increasing rather than decreasing.” — John Coogan [10:26]
    • TPU v6e is competitive with Nvidia’s A100s.
    • Ironwood (Nvidia) is expected to quickly gain market share and is competitive with Blackwell.
    • Potential: “Maybe they come from behind and they just destroy TPU and we're all talking about Trainium next year.” — John Coogan [11:40]

3. Ecosystem Shifts and Cloud Partnerships

• [07:39] As AI workloads become more central, next-generation companies may build entire infrastructure strategies around the best available AI platforms, rather than treating AI as a bolt-on service.
• [08:10] Interoperability and cross-cloud communication (AWS-GCP high-speed linking) become vital for customers needing access to different strengths across providers.

4. Remembering Claude, San Francisco’s Albino Alligator

• [11:58] News segment interrupts technical talk:

  “San Francisco's beloved albino alligator has passed away at age 30.” — John Coogan [12:00]

• [12:21] Commentary on social media speculation, joking about Anthropic’s sponsorship and conspiracy theories around Claude’s demise.

5. Government Investment in Chip Startups & New AI Design Firms

• [13:08] Trump administration invests $150M (Chips Act) in XLite, a lithography startup; further $35M raised by Recursive Intelligence, an AI-driven chip design firm targeting more automated approaches to custom silicon.

  “A startup they hope will remake the $800 billion chip industry ... recently raised $35 million to kickstart Recursive Intelligence.” — John Coogan [14:07]

• [14:51] Recursive Intelligence reportedly raised at a $750M valuation for only 5% dilution, defying expectations about the capital intensity of chip companies.

6. AI Capabilities: Economic Diffusion & the Limits of Today’s Models

• [15:30+] Discussion on RL environments, AI model capabilities, and why economic “diffusion lag” (the slow spread of technology into real-world use) is often overblown.

  “If these models were actually like humans on a server, they'd diffuse incredibly quickly.” — John Coogan (quoting Dwarkesh) [19:39]

• [18:05] Even top podcasters struggle to automate content-clip discovery—a task still dependent on human judgment, emotional nuance, and identification of humor.

  "One of the seeming missing capabilities is the ability to identify humor or even ... something almost emotional." — Jordi Hays [18:19]

• The hosts critique the idea that "diffusion lag" is the main bottleneck for AI’s commercial impact, pointing out real gaps in model capability.

7. Silicon Valley Saga: Secondary Markets & Startup Economics

• [23:02] A $1.5B judgment against Anthropic for training AI on copyrighted books; one author skipping the payout, wittily admitting to principle “but too lazy to be highly principled.”

• [23:36] Exposé on dodgy practices in secondary venture deals, especially involving Anduril shares — including double-layer fees, lack of transparency, and explicit contradictions of company bylaws.

  “A double layered SPV with ... 8% upfront fee, 3% annual fee for two years, 20% carried interest, and the craziest part, an implied price per share that is completely insane.” — John Coogan [24:12]

  “Don't do this. Instead, why don’t you start a company and apply to Y Combinator, build an actual business instead of going around hustling SPVs in companies that don't want to sell shares.” — John Coogan [26:48]

Notable Quotes & Memorable Moments

• “Amazon’s so bad at hype. Trainium is used by 500 million people through Bedrock, but their marketing team just can't.” — John Coogan [05:00]
• “Nvidia has an insane amount of power right now ... that's one of the greatest revenue ramps at scale in history.” — John Coogan [08:52]
• “Some of the clips that they make are so sloppy ... and the fact that we're still paying humans to do that, still, I mean, it just feels notable.” — Jordi Hays [19:14]
• “I'm principled, but too lazy to be highly principled.” — Read from an author for TBPN [23:29]

Timestamps by Segment

| Timestamp | Topic | |-----------|------------------------------------------------------------------------------------------| | 00:00 | Will AWS buy Google’s TPUs? Amazon launches Trainium 3 | | 01:08 | Annapurna Labs history & custom chip design | | 02:06 | Startup Descartes’ breakthrough with Trainium 3 | | 04:12 | Anthropic’s mixed results and chip supplier competition | | 04:38 | AWS CEO Matt Garman on customer choice and chips | | 05:35 | Real-time video AI and Twitch connection | | 07:10 | AWS-Google Cloud partnership: high-speed linking | | 08:52 | Nvidia’s dominance and anti-Nvidia alliances | | 10:22 | SemiAnalysis review: Trainium 3 vs. Nvidia/TPU specs | | 11:58 | Remembering Claude the alligator | | 13:08 | Government & VC investment in next-gen chip startups (XLite, Recursive Intelligence) | | 15:30 | AI progress, economic diffusion lag, missing capabilities | | 18:05 | Podcast content curation, limitations of current AI models | | 23:02 | Anthropic lawsuit, author’s witty response | | 23:36 | Anduril secondary market drama; warning against predatory SPV structures |

Tone & Style Notes

• The hosts’ banter is irreverent, quick-witted, and blends deep technical insight with Valley gossip and humor.
• They oscillate seamlessly from technical analysis to storytelling, speculation, and even affectionate obituaries (Claude the gator).
• Industry critiques are frank and occasionally biting, especially around venture and startup shenanigans.

In Summary

This episode delivers a condensed but information-packed tour of the cloud AI chip wars (Amazon, Google, Nvidia); real-world applications and startup stories; thoughtful and skeptical takes on diffusion and AI capabilities; and sharp commentary on Silicon Valley’s financial ecosystem—punctuated with humor and the human touch (RIP Claude). It’s a must-listen for those tracking the intersection of cloud, chips, AI, and venture.