All-In PodcastWhy Anthropic's best model is locked up despite a $30B ramp
Mythos scored so well on cyber offense that Anthropic delayed its release; a 100-day hardening window did not stop Claude Code from driving a $30B run rate.
CHAPTERS
Brad Gerstner returns, “retard maxing,” and show banter sets the tone
The episode opens with Brad Gerstner subbing in as the “fifth bestie,” plus extended banter about moderation, therapy/rumination, and internet culture. They tee up a packed agenda focused on AI model releases, AI coding agents, and geopolitical risk.
- •Brad Gerstner joins in place of Friedberg
- •Running jokes and community callbacks (moderation, “retard maxing,” etc.)
- •A quick riff on productivity/“detachment” as a life philosophy
- •Transition into the first major topic: Anthropic withholding a new model
Anthropic withholds “Mythos” for cyber risk: what they claim it can do
Jason introduces Anthropic’s decision to delay release of its newest model, Mythos, citing serious security concerns. The hosts recount Anthropic’s claims: the model finds long-dormant vulnerabilities across operating systems, browsers, and core libraries and can chain multiple flaws into sophisticated exploits.
- •Anthropic says Mythos autonomously discovered thousands of vulnerabilities
- •Examples cited: OpenBSD, FFmpeg, Linux kernel, major browsers/OSes
- •Claimed capability: chaining 3–5 vulnerabilities into an exploit path
- •Anthropic releases a dramatic safety/hype video featuring Dario and team
Self-regulation vs marketing stunt: Brad, Sacks, Chamath debate Anthropic’s motives
Brad argues Anthropic deserves credit for sandboxing the model and coordinating a 100-day defensive push rather than “move fast and break the internet.” Sacks highlights Anthropic’s history of fear-based marketing but says cyber risk is plausibly real; Chamath calls it mostly theater and questions what can be fixed quickly.
- •Brad: proactive industry coalition (“Project Glasswing”) and controlled release is responsible
- •Sacks: Anthropic often uses scare studies (e.g., past ‘blackmail’ narrative), but cyber risk here is credible
- •Chamath: compares to GPT‑2’s staged rollout; argues advanced hackers can already do much of this
- •Argument over effectiveness of a 100-day patch window vs structural internet-scale tech debt
National security and coordination: CIA/FBI, “Team America,” and the patch window
The conversation turns to game theory: whether adversaries already have comparable capabilities and what role government should play. They discuss industry coordination (Frontier Model Forum), the need to treat even ‘performative’ risk as real, and the practical reality that AI coding gains directly raise cyber offense and defense stakes.
- •Jason pushes for serious coordination with intelligence and law enforcement agencies
- •Sacks: treat it seriously regardless of marketing; use the window to patch vulnerabilities
- •Brad: coordination among frontier labs already exists; prefers market-led action over moratoriums
- •Shared view: AI coding advances inevitably amplify cyber capabilities on both sides
OpenClaw ‘ankled’: subscription cutoffs, API pricing, and first-party agent competition
Sacks and Jason unpack OpenClaw’s complaint that Anthropic restricted access via the $200 subscription plan, forcing heavy users onto metered API pricing. They debate whether this is rational pricing (ending an implicit subsidy) or an anti-competitive move—especially as Anthropic launches an in-house agent framework soon after.
- •Anthropic limits OpenClaw usage via flat subscription; pushes power users to API pricing
- •Jason frames it as cutting off oxygen; Brad frames it as stopping ‘dollars for 10 cents’
- •Sacks raises bundling/price-discrimination concerns if Anthropic’s first-party agent is subsidized
- •Enterprise security/compliance needs cited as a reason firms prefer first-party tooling
Is Anthropic dominant in AI coding? Market share vs TAM arguments
The hosts argue about what “dominance” means in an early market. Jason and Chamath suggest Anthropic may control a majority of coding tokens today, while Brad emphasizes the overall software TAM is enormous and fast-changing—making monopoly claims premature.
- •Sacks presses for a yes/no on dominance; discussion centers on coding-token share
- •Jason: Anthropic likely >50% of coding-token usage; Chamath: ~50–60% with Codex also strong
- •Brad: even if token share is high, overall market penetration remains small relative to total software spend
- •Consensus: the market moves quickly; leadership can flip within months
Open source agents and distributed training: Bittensor, “Bridges,” and enterprise realism
Jason argues open source agent tooling could undercut frontier labs, citing crypto-incentivized development and rapid iteration. Chamath separates two ideas: distributed/open training may be disruptive, but enterprises won’t outsource sensitive production code transformation to open communities; Brad and others note open source still penetrates enterprises heavily (Linux/K8s precedent).
- •Jason highlights Bittensor subnet projects (e.g., ‘Bridges’) as rapid-improvement flywheels
- •Chamath: open source pretraining/orchestration could win if capital markets dry up, but enterprises won’t open-source production code rewrites
- •Historical open source adoption in enterprises (Linux, Kubernetes, Postgres) used as counterpoint
- •Broader theme: agents will be the next battleground; trust/security boundaries matter
Anthropic’s $30B run-rate shock: fastest revenue ramp and the ‘TAM for intelligence’
They review reports that Anthropic has surged to roughly a $30B annualized run rate, fueled by Claude Code and enterprise demand. Brad frames it as proof that AI revenues can scale exponentially, that compute constraints are currently the limiter, and that the addressable market is fundamentally the ‘TAM for intelligence.’
- •Timeline discussed: 2023 monetization → 2024 ~$1B run rate → 2025 acceleration → ~30B run rate claim
- •Enterprise adoption cited: 1,000+ customers reportedly paying $1M+/year
- •Brad: demand-pull, not just go-to-market; AI shifts from IT budget to labor augmentation/replacement
- •Compute constraints: limited GW capacity caps growth; more capacity could accelerate revenues
Profitability and margins: gross vs net revenue confusion, inference cost declines
Chamath cautions the discussion is still at the “revenue talk” stage and notes public confusion about gross vs net reporting. Brad argues margins should be improving quickly because fixed compute is being utilized harder and inference costs are falling sharply; they debate how fast these companies can become sustainably profitable given massive capex needs.
- •Chamath: unclear how ‘gross’ vs ‘net’ is being reported; outsiders lack margin transparency
- •Brad: major cost is compute; utilization + declining inference costs can drive rapid gross-margin improvement
- •Discussion of ‘accidental profitability’ due to compute buildout lag vs demand surge
- •Sacks: early evidence ROI exists, rebutting ‘AI bubble’ critiques tied to hyperscaler capex
Vibe shift: Anthropic ripping, OpenAI reloading—competition is not zero-sum
Brad credits Anthropic’s focus (coding/agents) for out-executing OpenAI recently, but warns against counting OpenAI out, citing an imminent model release and Codex momentum. The group frames the market as enormous and multi-winner, with Meta and Google also positioned to compete—while emphasizing the need for “clean” competitive behavior to avoid antitrust backlash.
- •Brad: Anthropic was ‘counted out’ last year; now outpacing OpenAI in recent months
- •OpenAI’s next major model teased as competitive with Anthropic’s newest capabilities
- •Sacks: coding leadership can create a data flywheel into agents; firms should avoid discriminatory tactics
- •Meta/Google cited as capital/compute heavyweights that can pressure frontier labs
Community/business interlude: Liquidity Summit updates and speaker announcements
A lighter segment updates listeners on the sold-out Liquidity event and new speaker announcements. The hosts plug upcoming events and joke about last-minute ticket demands and podcast-related social dynamics.
- •Liquidity is sold out; waitlist and capacity constraints mentioned
- •Newly announced speakers include Brad Gerstner and Thomas Lafont (Coatue) plus more
- •Mention of the All-In Summit in Los Angeles and application link
- •Running jokes and side banter about X, subscribers, and show logistics
Iran war ceasefire and market reaction: escalation traps, off-ramps, and diplomacy
The final major segment turns to the Iran conflict: a two-week ceasefire, diplomatic travel, and Trump’s public statements. Sacks is careful to speak personally (not for the White House), endorsing the ceasefire as critical to avoiding escalation; Brad ties market resilience to expectations of limited entanglement and outlines potential upside if multiple geopolitical “planes land.”
- •Ceasefire described along with ongoing diplomacy and regional meetings
- •Sacks: wars escalate; pulling out is hard; ceasefire is a major achievement
- •Brad: markets priced conflict as limited; upside if Iran/Lebanon/Ukraine tracks stabilize
- •Discussion of oil/strait risk, regional stability, and investor sentiment
Israel’s influence and U.S. politics: coalition strains, popularity decline, and next steps
They debate why the U.S. entered the conflict, referencing reporting on Netanyahu’s pitch and internal U.S. deliberations. Chamath stresses the President ultimately decides; he warns Israel risks losing predictable U.S. support if it doesn’t help create an off-ramp, while Sacks cites Israeli political concerns about declining U.S. public approval.
- •Debate over whether U.S. policy is overly influenced by Netanyahu/Israel vs U.S. sovereign decision-making
- •Concerns about domestic coalition fractures and public opinion shifts
- •Sacks references Naftali Bennett highlighting Israel’s popularity problem in U.S. polling
- •General agreement that backing ceasefire/off-ramp efforts could improve political sustainability
