Open source is coming for the frontier labs, but will Anthropic's IPO still be one of the best-received listings in a decade? — In this discussion Aaron Ross and Aaron Dillon go over why those two things are not in conflict.
On this episode of the Cap Table Podcast, Aaron Ross and Aaron Dillon break down a week that touched every layer of the AI and defense stack: Anduril's Thunder reveal one week after the AWS battlefield-datacenter partnership, and why our $100B implied valuation still leaves room — Raytheon carries a $261B market cap on $88B of 2025 revenue, Boeing $166B, Lockheed $118B, General Dynamics $100B, Northrop $74B, L3Harris $52B — putting a reasonable post-IPO resting place at $225–275B if Anduril stays sovereign-only, and materially higher if they open the commercial door and re-rate toward a Palantir-style multiple. Our house view on the single-product defense names (drone defense, autonomous surface vessels, one-platform plays) is that they're acquisition candidates, not compounders: a 3x in three years to a prime is a fine outcome, but we'd be honest that the exit is the thesis. On the model layer, Kimi K3's open-weight release is the real story of the week, and we'd start by killing the myth — running a Chinese open-weight model on your own hardware sends nothing to Beijing. The Chinese labs are winning open source precisely because they're GPU-constrained, forcing software-side innovation like multi-agent orchestration and resource-allocated sub-models that showed up in Chinese research quarters before Western adoption. We'd put 80–85% of global AI compute spend on open weights within five to ten years, taking the closed frontier labs from roughly 75–80% share today to 15–20% — while the pie grows enough that OpenAI and Anthropic still book ~10x current revenue. Which is why we're not bearish on the Anthropic IPO, reportedly as soon as October: the S-1 will show roughly $4B as of December against a run-rate now rumored near $100B, printing something like $40–50B on a GAAP basis, and public markets have never seen a pure-play frontier model provider before. The Figma episode is the cautionary tale on the other side — embed in a platform and the platform eventually ships your product — and we'd expect Anthropic to behave exactly like Google has for 25 years, because that is the correct thing to do for shareholders. Our entry-point discipline stays where it's been: we're not recommending open-source model companies to clients, we're recommending infrastructure. Inference providers, model routers, token routers, semiconductor and datacenter names — Baseten, Fireworks AI, Factory AI — mostly $5–15B valuations, effectively mid-caps, and almost entirely private. Everyone's models run on that layer regardless of who wins above it, which is the best risk-adjusted return in AI right now. We close on OpenAI's portable screenless device and where personal hardware actually ends: audio recording as default within 6–12 months, video shortly after, businesses adopting before consumers, and eventual brain-computer symbiosis — Neuralink is in human patients today. Whoever holds that data, our bet is Apple or Google, not OpenAI. Build across stages, size your entries, and use time diversification — tranche annually, recycle liquidity in year five, and don't concentrate into a market moving this fast.
Not investment advice.
00:00 Three stories: Anduril's Thunder, Anthropic's IPO, OpenAI's device
01:18 Anduril vs. the defense primes: the $261B ceiling
04:35 Hardware margins, revenue multiples, and the Palantir comp
06:03 Why we're cautious on single-product defense startups
09:41 Anduril goes global: Israel, Ukraine, and frontier-tech buyers
10:50 Kimi K3 and the open-weight threat to Anthropic
11:34 No, running a Chinese model doesn't send data to China
12:52 Why China is winning open source
15:53 80% of AI compute will run on open weights
19:23 Anthropic's IPO math: $4B to $50B in three quarters
22:02 Figma, Claude Design, and the 800-pound gorilla problem
26:32 Where we'd actually put money: AI infrastructure
29:18 OpenAI's screenless device and the Apple fight
30:19 "In 6 to 12 months, everything will be recorded"
34:35 Who gets your personal data: Apple, Google, or OpenAI?
36:22 The case that privacy is already gone
38:11 The endgame: Neuralink and full symbiosis
48:58 Digital Aaron: why businesses adopt first
51:48 How to invest: entry price, tranches, time diversification