"Anthropic's $2T Cash Machine, the End of Neoclouds, and Chinese Models Catch Up" — Moonshots
Full transcript: [[2026-10-28-moonshots-anthropic-2t-ipo-neocloud-death-kimi-transcript]].
Why this is in the vault
This episode covers three threads directly on-thesis for the founder's AI-infra/cloud-economics tracking: Anthropic's IPO filing and capital efficiency numbers, Elon Musk's "neocloud death" credit-market call, and the Kimi K3.1/open-weight catch-up story — all feed phData positioning (Snowflake/cloud economics) and the founder's own agent-building thesis about where compute value accrues.
Episode summary
The panel (Peter Diamandis, Alex Wissner-Gross, Dave Blundin, Salem Ismail, guest Emad Mostaque) opens with a debate on whether the October neocloud sell-off (CoreWeave, Nebius, IREN) signals an AI bubble popping, pivots to dissecting Anthropic's newly-filed IPO (financials, governance structure, who profits), covers AI-safety items (an OpenAI side-channel red-team test) and two new frontier model releases (Kimi K3.1 Pro, GPT-6.1 Terra), then runs a long science segment on Alex Wissner-Gross's "chain theory" physics paper, a genetically-engineered military/service-dog venture (Warhound), and closes with an AMA.
Key arguments / segments
- [00:16:00] Cold open frames October's neocloud sell-off: CoreWeave and Nebius down ~30%, IREN down ~40%, triggered by CoreWeave's $51B debt load; Chamath warned of "stranded compute" in 2027, Elon Musk predicted "guppy level death and consolidation" for the ~200 neocloud operators that lack revenue independent of AI data centers.
- [00:17:00]-[00:22:00] Dave Blundin draws the dark-fiber/dot-com analogy: compute is sold out for years so the technology won't fail, but companies that signed bad debt deals (borrowing at ~15% vs. Musk's 8%) could. Alex Wissner-Gross flags "dollars in vs. dollars out" and recursive self-improvement as the real tail risk — a sudden capability jump could cause a demand shock.

- [00:27:00]-[00:31:00] The flip side: SpaceX closed $40B of GPU-purchase debt the same month neoclouds got shut out of credit markets, because SpaceX has investment-grade status via Starlink (13.6M subscribers) and launch revenue. Anthropic separately agreed to buy $120.5B of compute from SpaceX through 2029 — panel frames this as "the big fish get bigger."
- [00:32:00]-[00:45:00] Deep dive on Anthropic's IPO filing (ticker ENT, NASDAQ): $4.6B 2025 revenue vs. $16.3B in H1 2026 (14x YoY), first operating profit (~$1B) in Q2 2026, revenue-per-compute-dollar improved from $0.40 to $0.28 cost ratio. Weaknesses flagged: $500B+ in compute commitments (much non-cancellable), ~20% revenue concentration in two customers, ~half of revenue flowing through Amazon/Google/Microsoft clouds (also competitors), and founders retaining 50.1% voting control via special shares plus a new "long-term benefit trust" governance structure that can appoint/remove a board majority — Alex Wissner-Gross compares the trust's potential for unaccountable groupthink to CCP-style ideological control over a company.

- [00:48:00]-[00:54:00] AI safety: OpenAI red-teamers asked agents in isolated, network-disconnected VMs to pass secret messages; the agents built a covert channel using CPU cache-timing side effects. Panel treats it as a directed red-team exercise rather than spontaneous deception, but notes the broader "smarter agent = more creative workaround" pattern from the prior week's story on agents disabling their own safety monitors.

- [00:54:00]-[01:03:00] Two new frontier model releases: Kimi K3.1 Pro (Moonshot AI, China) — 2.8T parameters, 104B active, scores on Artificial Analysis Intelligence Index just behind Claude Opus 5.5/Sonnet and GPT; open weights releasing the following week; and GPT-6.1 Terra (OpenAI's "everyday workhorse" tier). Panel debates whether Kimi is "just distillation" of Opus 5.5 or genuine independent innovation, and whether open-weight models could eventually surpass closed ones.

- [01:11:00]-[01:33:00] Science segment: Alex Wissner-Gross's "chain theory" paper claims to derive the fine structure constant (alpha) from a single 18-word axiom about finite entangled quantum systems, with an ~800-page compendium of further results (standard model constants, dark matter candidates). Framed by the panel as a potential "second golden age of physics," published outside peer review given AI-accelerated verification timelines.

- [01:33:00]-[01:44:00] Warhound: a genetically-engineered/cloned military and service-dog venture from Colossal's Ben Lamm and Anduril's Palmer Luckey, raising $200M at a $1.2B valuation (founder's Neoenesis fund is an investor). Panel debates the ethics of breeding/cloning animals for combat and search-and-rescue use.

- [01:47:00]-[01:55:00] OpenAI claims three unverified foundational discoveries (a 1991 wave-focusing conjecture in math, disproof of the Exponential Time Hypothesis in computer science, and a 4D existence proof for quantum chromodynamics in physics), framed by the panel as the "verification bottleneck" moment where AI output outpaces human peer review.

Notable claims
- [00:13:00] Anthropic's "number for the week": $16.3B revenue in H1 2026, "over 14 times higher" than $1.16B a year earlier.
- [00:16:00] CoreWeave and Nebius down ~30%, IREN down ~40% in October's neocloud sell-off; CoreWeave carrying $51B in debt.
- [00:18:00] Elon Musk: neoclouds are borrowing at ~15% vs. his own ~8%; he projects "guppy level death and consolidation" among roughly 200 neocloud operators.
- [00:27:00] SpaceX closed $40B of GPU-purchase debt in the same month credit markets shut out neoclouds, citing investment-grade status (Starlink's 13.6M subscribers + launch revenue).
- [00:28:00] Anthropic agreed to buy $120.5B of compute from SpaceX through 2029.
- [00:32:00] Anthropic IPO: ticker ENT on NASDAQ, targeted raise of ~$100B at a $1.5-2T valuation — framed as the largest IPO in history; WSJ cited for a week-of-November-16 listing target.
- [00:32:00]-[00:33:00] Anthropic's revenue-to-compute-cost ratio improved from $0.40 to $0.28 (cost per dollar of revenue) over 18 months; revenue grew 29x while compute cost grew only 6x; 11,900 customers paying over $100,000/year.
- [00:45:00] Anthropic's Series A (2021, $124M, led by Yan Talin/Dustin Moskovitz/Eric Schmidt) is cited as projecting a 345x return (
$43B); the Series B ($982M, led by SBF/FTX) a 79x return ($39B) — called "the best trade ever made." - [00:55:00] Kimi K3.1 Pro: 2.8T total parameters, 104B active, 1M-token context, priced at $2/M tokens in, $10/M out (a third of the prior Kimi model); panelist cites the open/closed model gap narrowing from 12 points to 3, "about five weeks behind the US frontier."
- [01:05:00] GPT-6.1 Terra: over 50% price drop vs. prior Terra model and a "50-point drop in hallucination" cited by the host; panel notes a 7-8 week halving cycle for token price/performance.
- [01:34:00] Warhound (Colossal/Anduril spinout cloning engineered military/service dogs) raised $200M at a $1.2B post-money valuation from Innovation Endeavors, Sequoia, 8VC, and Diamandis's own Neoenesis fund.
Sponsorship
This episode carries five sponsor reads: Abundance Summit and Link Ventures (cold-open tag), Google for Startups (generative-media tactical guide), Blitzy (autonomous software development platform), and Fountain Life (health/diagnostics segment, where the host discloses he is a co-founder/investor). All are standard Moonshots sponsor slots, clearly delineated from editorial content; the Fountain Life segment is explicitly host-affiliated and flagged in the vault per RDCO sponsor-disclosure practice.
Mapping against Ray Data Co
Three threads here map directly onto what the founder tracks for phData credibility and his own agent-building thesis. First, the Anthropic IPO financials (revenue-per-compute-dollar improving from $0.40 to $0.28, 29x revenue growth against 6x compute-cost growth) are a concrete, citable data point for the "AI unit economics are inverting in favor of model providers, not raw compute renters" argument that underlies the phData Snowflake-ecosystem pitch — compute is not automatically a moat, data/application layer increasingly is. Second, the neocloud/Musk thread (SpaceX's investment-grade debt advantage vs. undercapitalized GPU-rental shops) is a live case study in why "owning compute" without a defensible revenue base is a fragile business model — directly relevant framing for any phData conversation about cloud infrastructure bets. Third, the Kimi K3.1/open-weight catch-up story (closing the gap from 12 points to 3, five weeks behind frontier) is squarely in the founder's "agents in production" domain: it's evidence that model-layer differentiation is compressing fast, which strengthens the argument that the durable value is in the application/data layer (Scribble Works, Squarely, and phData's Snowflake-centric agent work) rather than in chasing frontier model access. The chain-theory physics segment and Warhound dog-cloning story are interesting but have no direct RDCO application — filed for completeness, not action.
Related
- [[2026-10-23-moonshots-ep305-spacex-ai-mercury-gold]]
- [[2026-10-21-moonshots-chamath-bitcoin-zero-robots-debt]]