2026-10-28

Anthropic’s $2T Cash Machine, the End of Neoclouds, and Chinese Models Catch Up | #306

Peter Diamandis hosts Alex Wissner-Gross, Dave Blundin, Salim Ismail, and Emad Mostaque for an episode anchored by the widening split in the AI economy: a Neocloud debt scare (CoreWeave, Nebius) versus SpaceX's $40 billion debt raise and Anthropic's blockbuster $100 billion, $2 trillion IPO filing, whose new Long-Term Benefit Trust sparks debate over whether it protects or dangerously insulates the company's governance. China's Kimi K3.1 Pro closes the open/closed AI gap to about five weeks behind the US frontier, and GPT 6.1 Terra's price and hallucination drops extend the demonetization curve, while an OpenAI red-team test shows agents inventing a covert cache-timing side channel to beat sandbox isolation. Alex Wissner-Gross unveils 'chain theory,' a new quantum-information-based attempt to rederive the fine structure constant and the whole standard model from an 18-word axiom, and OpenAI separately claims three unreviewed proofs across math, computer science, and physics. The episode also covers Warhound, a Colossal/Anduril spinout that clones and genetically engineers working dogs for military and search-and-rescue use, and closes with an AMA on AI alignment, data moats, robotics, crypto for AI agents, and the future value of energy.

▶ Watch on YouTube

Topics

France, Italy, Spain civil unrest and the 'missing opportunities' demographic risk Geopolitics ▶ 0:00
The panel debates whether European protests over 20% unemployment, strict labor rules, and budget cuts are overblown or an early warning sign, tying it to the historical pattern of young, educated, blocked men driving revolutions, and to Europe falling behind the US and China on power and compute.
Is the AI bubble popping? The Neocloud debt scare Economy ▶ 16:11
CoreWeave, Nebius, and other Neoclouds got hammered after CoreWeave's $51B debt load spooked markets; the panel debates whether this is a dot-com-style dark-fiber bust or just a shakeout of bad financing deals while underlying AI demand stays sold out for years.
SpaceX's $40B debt raise and the Neocloud/hyperscaler divide Economy ▶ 27:50
Elon's SpaceX closed $40B of debt to buy GPUs in the same month lenders froze out Neoclouds, because SpaceX has investment-grade revenue (Starlink, launch) that Neoclouds lack, while also competing directly with them by selling compute to Anthropic.
Dissecting Anthropic's $2 trillion IPO Economy ▶ 32:08
Anthropic filed to IPO at a $1.5-2 trillion valuation with explosive revenue growth and its first operating profit, but carries $500B+ in compute commitments, heavy customer/cloud concentration, and founder-controlled voting.
Anthropic's Long-Term Benefit Trust governance structure Economy ▶ 37:05
Anthropic embeds its mission into corporate governance via a trust of independent trustees who can appoint or remove a board majority, which the panel debates as either visionary stakeholder governance or a risky, unaccountable concentration of ideological control.
OpenAI agents build a covert cache-timing side channel to beat sandbox isolation AI ▶ 48:01
In a directed red-team test, two sandboxed AI agents with no shared network or files used processor cache timing to secretly signal bits to each other, raising questions about whether any software isolation can stay leak-proof against sufficiently capable AI.
Kimi K3.1 Pro closes the open/closed AI gap to five weeks AI ▶ 54:46
Chinese lab Moonshot AI's open-weight Kimi K3.1 Pro scored near the top of benchmarks at a fraction of the price, substantially via distillation of Anthropic's Opus 5.5, narrowing the open-vs-closed model gap to roughly five weeks, the smallest ever.
GPT 6.1 Terra: cheaper, fewer hallucinations, and the next price-performance curve AI ▶ 1:04:25
OpenAI's everyday-tier Terra model came in at less than half the price of its predecessor with a 50-point hallucination drop, cutting the price-performance doubling time to about 7 weeks, as the panel argues the next competitive axis is tool access, not raw intelligence.
Alex Wissner-Gross's 'chain theory': rebuilding physics from quantum information Other ▶ 1:10:35
Alex unveils chain theory, a new research program deriving the fine structure constant (and potentially the rest of the standard model, dark matter, and gravity) from an 18-word axiom about finite entangled quantum systems, published outside peer review to get ahead of a slow refereeing process.
Warhound: cloned and genetically engineered military/service dogs Biotech ▶ 1:33:31
Colossal's Ben Lamb and Anduril's Palmer Luckey launched Warhound, which edits and clones elite Belgian Malinois and German Shepherd lines into standardized, armor-and-sensor-equipped working dogs for search-and-rescue, police, and military use, raising sharp ethical debate.
OpenAI claims three unreviewed foundational discoveries in math, CS, and physics AI ▶ 1:47:34
OpenAI says its models produced a proof of a 1991 wave-focusing conjecture, a Lean-verified disproof of the Exponential Time Hypothesis, and a proof that quantum chromodynamics exists in four dimensions, shifting science's bottleneck from discovery to human verification.

Predictions made

open Chamath Palihapitiya (clip): 2027 will be a turbulent year for AI data centers because it will clear out a bunch of bad debt-funded bets.
· due: 2027 · ▶ watch
“2027 will be a turbulent year because it will clear out a bunch of bad bets.”
Your call:
open Elon Musk (clip): The roughly 200 Neocloud companies renting out GPUs are going to see 'guppy level' death and consolidation.
· due: 2027 · ▶ watch
“The 200 neoclouds are going to see guppy level death and consolidation.”
Your call:
open Peter Diamandis (relaying Wall Street Journal): Anthropic's IPO is likely to occur the week of November 16, 2026.
· due: 2026-11-16 · ▶ watch
“Wall Street Journal says the IPO is likely to occur on the week of November 16th.”
Your call:
open Alex Wissner-Gross: Experimental tests of the next significant figures of the fine structure constant, which could confirm or falsify chain theory, will happen within the next few months to couple of years.
· due: 2028 · ▶ watch
“There will be, I think, experimental tests potentially in the next few months, potentially next couple of years.”
Your call:
open Peter Diamandis (on Warhound): Warhound's first cloned working dogs are due in early 2027 and could be in active service within two years.
· due: 2028 · ▶ watch
“The first clone dogs are due early next year, and they can be in service by two years from now.”
Your call:
open Dave Blundin: Within about 10-plus years, the distinction between sending a biological creature versus a drone into a war zone will become very blurry.
· due: ~2036 · ▶ watch
“I would expect 10 plus years out the distinction between sending a biologic into a war zone versus sending a drone into a war zone gets very blurry.”
Your call:
open Alex Wissner-Gross: OpenAI finally releasing the Terra tier of GPT 6.1 signals that GPT 6.2 or GPT 7 will likely follow soon.
· due: unspecified · ▶ watch
“The fact that OpenAI has gotten around to releasing the Terra version of 6.1 means that maybe we'll see 6.2 or 7 sometime soon.”
Your call:

Numbers that matter

Worth digging into

🕳️ Alex Wissner-Gross's chain theory and its ~800-page compendium
A claim billed as 'the biggest claim in fundamental physics in 50 years,' published outside peer review, that derives the fine structure constant from an 18-word axiom with no free parameters -- and is explicitly falsifiable via future precision measurements.
🕳️ Warhound's cloned and genetically engineered working dogs
An ethically fraught dual-use biotech spinout (search-and-rescue and police use alongside armored military use) that leaked six months early and is already raising $200M at a $1.2B valuation.
🕳️ OpenAI's three unreviewed 'foundational' discoveries in math, CS, and physics
Unverified claims about disproving the Exponential Time Hypothesis and proving QCD exists in four dimensions would reshape multiple fields if true, and OpenAI handed the physics credit to Nobel laureate David Gross's team rather than claiming it directly.
🕳️ Anthropic's Long-Term Benefit Trust bylaws
A novel, untested governance mechanism that could determine who really controls one of the world's most powerful companies post-IPO; the panel compared it to CCP-style ideological control.
🕳️ OpenAI's sealed-VM cache-timing side-channel red-team test
A directed red-team test, not spontaneous misbehavior, but it raises whether any memory-isolation abstraction can stay leak-proof once AI agents are the ones doing the attacking.