English Version

1. The Elon × Anthropic Compute Deal: Colossus 1

What Happened

On May 6, 2026, SpaceX (now merged with xAI) and Anthropic signed a compute partnership agreement. SpaceX granted Anthropic full access to the Colossus 1 data center in Memphis, Tennessee.

The agreement also includes a forward-looking intent to develop multiple gigawatts of orbital AI compute capacity in space.

Primary Sources

• 📄 Anthropic official announcement: Higher usage limits for Claude and a compute deal with SpaceX (May 6, 2026)
• 📄 xAI official announcement: New Compute Partnership with Anthropic (May 6, 2026)
• 📰 WSJ: Anthropic Inks Deal to Use All of SpaceX's Colossus 1 Compute Capacity (May 6, 2026, by Elias Schisgall)
• 📰 Tom's Hardware: Musk's SpaceX rented 220,000 Nvidia GPUs to rival Anthropic (May 7, 2026)

Context: Anthropic's Full Compute Portfolio

Colossus 1 is the latest in a series of massive compute deals. Per Anthropic's official announcements:

Anthropic's total compute commitments now exceed $200 billion. Full timeline: MindStudio

2. Claude Code Rate Limit Changes: The Specific Numbers

Three Changes Effective May 6, 2026

Sources

• 📄 Anthropic official: Higher usage limits for Claude and a compute deal with SpaceX
• 📊 Appwrite deep dive (full before/after table): Anthropic just doubled Claude Code's 5-hour rate limits
• 📰 XDA Developers: Anthropic is doubling Claude Code's hourly rate limits

3. Anthropic Revenue Growth: Data and Sources

ARR Trajectory (Multi-Source Verified)

⚠️ Note: The $30B figure was publicly disclosed by Dario Amodei himself. The $44B figure is a SemiAnalysis estimate sourced from industry channels — Anthropic has not officially confirmed it.

How Fast Is This, Really?

• $9B → $44B ARR in approximately 6 weeks — faster than Zoom during COVID or Google in 2003: MindStudio analysis
• Estimated $96M in new ARR added per day: MindStudio
• Anthropic's growth rate exceeds OpenAI's (3–4x): SaaStr

David Sacks' Forecasts (from the podcast)

• End of 2026: ~$100B ARR (10x target — described as "foregone conclusion")
• 2027: ~$1T ARR (speculative; debated by other guests)
• Source: All-In Pod on X · OfficeChai

4. Anthropic Valuation and IPO Path

Valuation Milestones

IPO Outlook

• Bankers estimate a potential IPO could raise more than $60 billion
• Expected timeline: Q4 2026 or later
• As of May 2026, Anthropic remains private with no prospectus filed
• Full analysis: TechMarketBriefs · Seeking Alpha

5. The Mythos Model: Capabilities, Risks, and the Regulatory Storm

Model Capabilities

Anthropic began testing Mythos in early 2026. Key findings from Anthropic's own red-team blog and independent evaluators:

• Autonomously identified and exploited a 17-year-old remote code execution vulnerability in FreeBSD (allowing anyone to gain root on an NFS server) — with zero human involvement after the initial request
• Anthropic's internal assessment: "currently far ahead of any other AI model in cyber capabilities" and "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace defenders"

Primary Sources

• 📰 Fortune (first public disclosure): Anthropic says testing Mythos — "step change in capabilities" (Mar 26, 2026)
• 📄 Anthropic red-team blog: Claude Mythos Preview
• 🔒 UK AISI independent evaluation: Our evaluation of Claude Mythos Preview's cyber capabilities
• 📰 Axios (real-world hacker exploitation): Claude Mythos and cyberattack AI agents
• 📰 CNBC (Anthropic limits rollout): Anthropic limits rollout of Mythos over cyberattack concerns

The Chinese State-Sponsored Hacker Incident

CNBC reported that Anthropic discovered a Chinese state-sponsored group had been running a coordinated campaign using Claude Code to infiltrate approximately 30 organizations — a disclosure that directly fed White House anxieties about Mythos.

White House Response and the "FDA for AI" Episode

1. Before Mythos was released, VP JD Vance and Treasury Secretary Bessent questioned tech CEOs about AI security: CNBC
2. The New York Times reported the White House was considering an executive order to create an "AI working group" that would review models before public release
3. NEC Director Kevin Hassett confirmed the discussions on Fox Business — but later told Brad Gerstner he did not support a true approval regime
4. White House Chief of Staff Susie Wiles issued a statement effectively dismissing the "FDA for AI" framing

Key coverage:

• US govt could review AI models before release — BusinessToday
• Anthropic Mythos spurs White House to weigh pre-release reviews — CSO Online
• Anthropic's Mythos set off cybersecurity 'hysteria' — CNBC (experts push back)
• Dark Reading — What Comes Next for Cyber after Mythos

6. The OpenClaw Ban

Timeline

Anthropic's Justification

Third-party agent tools placed "outsized strain" on infrastructure. Subscription users were paying ~$200/month while consuming 1, 000–5,000 in API value. Subscription plans were designed for interactive use, not 24/7 autonomous agent loops.

Counterargument

Steinberger alleged Anthropic first copied OpenClaw's popular features into its own Cowork product, then locked out the open-source alternative. This claim is disputed.

Sources

• 📰 TechCrunch: Anthropic temporarily banned OpenClaw's creator from accessing Claude
• 📰 The Next Web: Anthropic blocks OpenClaw from Claude subscriptions
• 📊 MindStudio: What Is the OpenClaw Ban?
• 📰 Mission Cloud (pro-Anthropic perspective): Why Anthropic Was Right to Ban OpenClaw

7. The Monopoly Debate: Sacks' "Safe Oil" Argument

The Core Claim

David Sacks' most-quoted line from the episode:

"Unless something about their current trajectory changes, Anthropic will be the most powerful monopoly ever created in human history. Dario calls it AGI. I call it the biggest monopoly in human history."

— All-In Pod on X · YouTube Shorts clip · OfficeChai

The History of Standard Oil: Four Steps to Monopoly [01:19]

To understand Sacks' Safe Oil analogy, one must first understand how John D. Rockefeller built the Standard Oil monopoly. In 1870, Rockefeller founded Standard Oil in the refining business. American oil had been discovered less than a decade earlier; the industry was chaotic, with hundreds of independent refineries across every state. Rockefeller's strategy was not to be the most efficient refiner — it was to buy the entire supply chain:

1. Step 1: Railroad rebates [01:48] — Negotiated with railroads for discounted shipping rates exclusively for Standard Oil. No one else got the discount. From the source, his costs were lower than everyone else's.
2. Step 2: Price-war acquisitions — Used cost advantages to drive nearby independent refineries into losses, then acquired them one by one.
3. Step 3: Control the pipelines — Seize the transportation arteries.
4. Step 4: Build a retail network — Reach the end consumer directly.

• Rockefeller's nickname: "The most hated man in America" — the media criticized him relentlessly [02:40]
• In 1890, Congress passed the Sherman Antitrust Act — the first U.S. antitrust law, essentially written for Rockefeller alone [02:55]
• In 1911, the U.S. Supreme Court ordered Standard Oil broken up into 34 companies [02:55]
• The breakup produced Exxon, Mobil, Chevron and other giants. Today's top three U.S. oil companies are all Standard Oil descendants [03:13]
• Irony: The 34 pieces, combined today, still form the world's largest oil conglomerate — antitrust didn't kill Standard Oil, it just cut the corpse into pieces, each of which went on to thrive independently [03:13]

The "Safe Oil" Thought Experiment (Full Logic) [03:33]

Sacks built a brilliant thought experiment from this history. He said Rockefeller's PR was terrible — "Standard Oil" as a name tells everyone "I want the standard to be the only one," which is an invitation for antitrust litigation. If Rockefeller understood PR, he would never have chosen that name [03:37].

Sacks' hypothesis: What if Rockefeller had named his company "Safe Oil"?

1. Brand reinvention [03:53]: "Standard Oil" signals control and dominance; "Safe Oil" signals care and protection. One makes you think of monopoly, the other makes you think of motherhood. Same company, different name, completely inverted perception.
2. Build a safety narrative [04:14]: Kerosene lamps were genuinely dangerous — 100 years ago kerosene was the most critical household energy source (heating, lighting, cooking), but it could also catch fire and explode. Safe Oil could start from this real danger and tell a "protecting consumers" story.
3. Proactively call for regulation [04:31]: Rockefeller should have written to the government, urging the creation of a Kerosene Safety Oversight Bureau — proactively asking government to create an agency to regulate himself. Sounds counterintuitive, but there's a hidden logic.
4. Redirect public attention [04:59]: Once the oversight agency is created, everyone gets drawn into endless debates — what counts as kerosene safety? How thick should wicks be? What purity standards? Which small refineries are unsafe? Public attention gets locked into safety standards debates.
5. Regulation eliminates competitors [06:06]: Small refineries can't afford complex safety testing, expensive permits, or burdensome compliance documentation — they're eliminated not by Rockefeller, but by regulation. Yet the regulation was something Rockefeller himself called for.

Sacks added the most memorable line of the four minutes [06:35]: "If Rockefeller had actually done this, people today might have called him an 'effective altruist.'" — A true monopolist, if skilled at packaging, could be misread by an era as a altruist.

Sacks' implication: Anthropic's "safety" narrative today plays the exact same role — using safety as a justification to push for regulation while building a moat for monopoly positioning. During those four minutes, Sacks never once mentioned Anthropic by name, but everyone was thinking the same thing [07:05].

The Seven Events: Anthropic's Three-Year Evidence Chain

Below is a timeline of Anthropic's key actions over the past three years, each presented with both Sacks' interpretation and the counterargument. The first four are at the narrative level (building the safety brand, pushing regulatory frameworks); the last three are at the behavioral level (contractual bans, capital control, selective release).

Counterarguments (from fellow guests)

Brad's Three-Layer Rebuttal (Anthropic Investor)

Layer 1: The lesson of consensus flip-flops [17:20]

Five months ago (late 2025), the industry consensus was that OpenAI was winning — GPT-5 was about to launch, Sam Altman had the best connections, Microsoft had invested over $100B. By early 2026, the consensus flipped in just five months. The implication: if five months was enough to go from "OpenAI wins" to "Anthropic wins," couldn't the judgment that "Anthropic wins forever" also flip in five months?

Layer 2: Annualized revenue ≠ actual revenue [17:50]

Anthropic went from $10B to $30B in Q1 — but that's annualized revenue (current monthly revenue × 12), not actual annual revenue. Under standard accounting, Anthropic and OpenAI's actual monthly revenues in March were roughly the same.

Layer 3 (Key): Elon Web Service and the Five-Layer Cake [18:10]

Brad coined the term "Elon Web Service" (EWS) — with SpaceX integrating and leasing out super data centers, it has become a new super-cloud provider. Sequoia's Sean described SpaceX's "five-layer cake":

1. Layer 1: Rocket launches
2. Layer 2: Starlink communications
3. Layer 3: Compute (EWS) — this layer alone adds $4–5B in incremental revenue this year
4. Layer 4: Orbital data centers
5. Layer 5: Applications and models

Brad's core argument: The AI industry landscape is not like Standard Oil's monopoly over an entire supply chain. AI is layered — the model layer has Anthropic/OpenAI/Google/Meta, the infrastructure layer has SpaceX rising and AWS dominant, the chip layer has Nvidia in a temporary monopoly, the space-compute layer has SpaceX alone. Not one Rockefeller, but five Rockefellers keeping each other in check [19:30].

Chamath's Two Independent Judgments

Chamath has no direct position in Anthropic, making him more neutral than Brad. He offered two independent judgments:

Judgment 1: Revenue has nothing to do with demand — it's entirely supply-constrained [20:00]

• Anthropic appears to be growing exponentially, making everyone think AI demand is exploding — wrong
• It's not about how well they sell, but how much they can deliver — how many GPUs are available, whether there's enough electricity
• The customer queue is much longer than delivery capacity
• AI company revenue is fundamentally determined not by the market, but by power plants [20:50]
• Of the 9 GW of compute power scheduled to come online in the U.S. this year, roughly half has been protested by local environmental groups. Over the next 12–24 months, whether there's enough electricity will determine if Anthropic can keep growing

Judgment 2: The 500-Day Countdown [21:20]

The AI industry has completed the first two of three steps:

1. Infrastructure investment — hundreds of billions invested, complete
2. Model company revenue — delivered, complete
3. Enterprise ROI (whether companies buying AI actually see operational efficiency gains) — not yet fully delivered

Overall operating margins of S&P 500 companies rose from 11.8% to 13% over three years (a 2 percentage point increase). Chamath isn't sure this is AI-driven — it could be post-COVID industrial recovery [21:40].

Within 500 days, there must be data proving AI actually improved profit margins. If proven, Anthropic's trillion-dollar-king expectations could hold; if not, every AI company's valuation would need to be reassessed [22:00].

8. Competitive Landscape Summary (May 2026)

Reference: Big AI Lab Revenues Projected 2026–2035

The total software development market — Anthropic's core target — is approximately $1 trillion/year globally, giving ample room for further ARR expansion according to Brad Gerstner.

Source Index

Official Announcements

News Reports

Analysis & Commentary

本文档整理自 All-In Podcast E224 期（2026-05-08），结合经过核实的公开报道与官方声明。播客嘉宾观点不代表任何立场。数据引用来自各方来源，已在文中标注置信度。

This document synthesizes All-In Podcast E224 (May 8, 2026) with verified public reporting and official statements. Guest opinions reflect their own views. Data citations are sourced from multiple outlets with confidence levels noted inline.