2026-05-16

Anthropic Partners With SpaceX AI, Leopold's $5.5B Bet, and the Singularity Economy | EP #255

The mates open on Anthropic's runaway 80x quarterly growth and its surprise deal to take over SpaceX's Colossus 1 data center from Elon Musk, arguing this signals both an insatiable compute shortage and a shakeout among frontier labs (with xAI/Grok increasingly sidelined). They cover Anthropic's alignment breakthrough that drove blackmail behavior in evals from 96% to 0%, OpenAI's push into voice-first AI and a teased 'super app,' and Claude's continuing disruption of legal services and small-business SaaS ('the great unhobbling'). A long segment tracks 'the singularity economy,' comparing chip/infrastructure stock returns (300-400%+) against the S&P 500 (31%) and Leopold Aschenbrenner's fund growth from $1B to $5.5B. The episode closes with Google/SpaceX's competing orbital-data-center ambitions, Taiwan/TSMC chip-supply risk, newly declassified US government UAP files, and an eight-question audience AMA.

▶ Watch on YouTube

Topics

Anthropic's 80x growth and compute crunch Economy ▶ 4:20

Dario Amodei revealed 80-fold Q1 revenue growth (vs. 10x expected); ARR jumped from $9B (end 2025) to ~$40B (May 2026), with panel discussing $100B-$1T ARR trajectories and resulting trillion-dollar valuations.

Anthropic takes over SpaceX's Colossus 1 from Elon AI ▶ 12:12

Anthropic signed a $1.8B compute deal (with C3.ai) and took over all of SpaceX's Colossus 1 data center (220,000 GPUs, built in 122 days), doubling Claude Code rate limits; panel frames this as Elon becoming a hyperscaler and 'the enemy of my enemy is my friend.'

Compute scarcity and agent economics Compute ▶ 14:14

Dave Blundin breaks down GPU-to-agent math: each GPU serves ~8 concurrent threads, so 220,000 GPUs yield only ~1.6M concurrent agent threads against demand from billions of users wanting many agents each, explaining Claude's slow response times.

Frontier lab consolidation and Grok's decline AI ▶ 18:52

Alex Wissner-Gross argues Grok/xAI is 'on life support' after SpaceX handed Colossus 1 to Anthropic and cut a $60B+ deal with Cursor, suggesting SpaceX AI may become a hyperscaler rather than a competitive frontier lab.

Anthropic's alignment breakthrough via storytelling AI ▶ 29:48

Anthropic's 'Teaching Claude Why' research shows every model since Haiku 4.5 scores zero on blackmail behavior (down from Opus 4's 96%), achieved by training on Claude's constitution and positive fictional narratives rather than just rules.

OpenAI's real-time voice/translation model and voice-first AI AI ▶ 42:00

OpenAI released a real-time audio translate/transcribe model; panel discusses why specialized (not omnimodal) models are winning on unit economics, and why voice is becoming the trusted interface that turns AI into a coworker.

OpenAI's teased super app / AI operating system AI ▶ 44:28

OpenAI teased a combined ChatGPT/Codex/voice/Atlas browser 'super app'; Alex frames it as a defensive consolidation to compete with Anthropic, while others debate whether this becomes an Apple-threatening AI operating system.

Hermes overtakes OpenClaw AI ▶ 52:08

The open-source Hermes agent surpassed OpenClaw on OpenRouter token rankings; Dave and Alex discuss its Python codebase and native recursive self-improvement (self-generating skills) versus OpenClaw's app-store-style scaffolding.

The great unhobbling: Claude disrupts legal and small business Economy ▶ 56:27

Claude for legal (targeting a $1T/year global industry) and Claude for small business (44% of US GDP) are discussed as both disrupting incumbents (e.g., LegalZoom) and demonetizing services for the underserved; includes a live disagreement over whether legal automation triggers a Jevons-paradox surge in lawsuits.

Elon's Terrafab and Taiwan/TSMC chip risk Geopolitics ▶ 1:05:46

Elon's Terrafab aims for 50x global chip production at an estimated $119B+ cost; panel discusses Taiwan producing two-thirds of world GPUs and the geopolitical risk if TSMC production is disrupted.

Orbital data centers: Google's Suncatcher vs. SpaceX AI's Dyson swarm Space ▶ 1:09:58

Google is partnering with Planet Labs on Project Suncatcher (orbital TPU data centers, ~80 satellites), a much smaller effort than SpaceX AI's FCC filing for up to a million orbital AI data centers; panel frames this as a corporate space race.

The singularity economy: stock returns vs. traditional sectors Economy ▶ 1:19:40

Leopold Aschenbrenner's fund grew from $1B to $5.5B on situational-awareness bets; panel compares chip/infrastructure stock gains (300-400%+) against the S&P 500 (31%) and traditional sectors (5-29%), debating public vs. private market access to AI-era gains.

UAP/UFO government file declassification Other ▶ 1:31:41

The White House released 82 pieces of previously classified UAP data via the new 'Pursue Initiative,' a rolling declassification process expected to run through ~January 2027; panel and a Michio Kaku clip debate significance and likelihood of extraterrestrial origin.

Predictions made

open Peter Diamandis: Anthropic will hit roughly $100 billion ARR by the end of 2026 and potentially $1 trillion ARR by mid-to-end 2027, becoming the most valuable company on the planet.

EP #? · · due: end of 2026 / mid-2027 · ▶ watch

“they could hit a hundred billion of ARR by the end of 2026 and potentially a trillion by the end or mid 2027 making it the most valuable company on the planet”

Your call:

open Elon Musk (relayed by Alex Wissner-Gross): AI will drive double-digit US GDP growth within two years and triple-digit GDP growth within five years.

EP #? · · due: ~2028 (2yr) / ~2031 (5yr) · ▶ watch

“Elon said double-digit GDP growth in two years and triple-digit within five years.”

Your call:

open Alex Wissner-Gross: The world needs, and Sam Altman may build, a competitor to SpaceX AI's terrafab/orbital compute access to create at least a duopoly in space-based AI infrastructure.

EP #? · · due: unspecified · ▶ watch

“I wouldn't be surprised if Sam Alman spins up a competitor to the SpaceX AI terraab access”

Your call:

open Alex Wissner-Gross: Knowledge-work verticals like Claude for legal, medicine, and financial services will get absorbed into baseline frontier models within the next few months, at most one to two years.

EP #? · · due: probably within months, max 1-2 years (by ~2028) · ▶ watch

“I think those are going to get baked into the baseline models over the next few months, maybe one to two years maximum, but probably the next few months.”

Your call:

open Alex Wissner-Gross: The distinction between consumer AI and enterprise AI spending will erase as individuals become 'one-person conglomerates' able to pay for frontier reasoning tokens.

EP #? · · due: 10-20 years (~2036-2046) · ▶ watch

“I think consumers become indistinguishable from enterprises and my bet is consumers individuals will become one person conglomerates.”

Your call:

open Dave Blundin: Elon's Terrafab chip-manufacturing project will end up costing significantly more than the reported $119 billion estimate.

EP #? · · due: unspecified / over course of project buildout · ▶ watch

“I think it's going to cost more than 119 billion, but that's okay.”

Your call:

open Peter Diamandis: There could be a negotiated turnover of Taiwan's chip manufacturing/TSMC situation within the next decade, allowing a smooth transition rather than conflict.

EP #? · · due: within 10 years (~2036) · ▶ watch

“maybe it will be a negotiated turnover sometime in the next 10 years”

Your call:

open Alex Wissner-Gross: More UAP-related declassified details will continue emerging from government systems (JWICS) over the next few years as part of the rolling Pursue Initiative release.

EP #? · · due: next few years (through ~2029) · ▶ watch

“I expect all of these details to start pouring out over the next few years.”

Your call:

Numbers that matter

Anthropic Q1 revenue growth: 80x vs. 10x expected Dario Amodei's disclosure at Anthropic's developer conference of dramatically outpaced growth expectations.
Anthropic ARR: $9B (end 2025) -> $30B (April 2026) -> ~$40B (May 2026) Annualized revenue run rate trajectory driving speculation about $100B-$1T ARR by 2026-2027.
Anthropic valued at $1.2 trillion at 30B ARR / 40x multiple; projected $4T at $100B ARR; $40T at $1T ARR Valuation math extrapolated from current revenue multiples.
$1.8 billion 7-year compute deal, popped stock 25% on first day Anthropic's largest-ever deal with C3.ai (ticker 'AI').
Grok used ~11% of Colossus 1 capacity Cited as evidence of low Grok usage before Colossus 1 was handed to Anthropic.
220,000 GPUs (H100s) in Colossus 1 Data center built by SpaceX/xAI in 122 days, now run by Anthropic.
~8 concurrent agent threads per GPU (Opus 4.7 Max) = ~1.6 million concurrent threads from Colossus 1 Dave Blundin's compute-to-demand math versus 8 billion potential users wanting agents.
Nvidia NV72 rack costs ~$4 million, serves an agent in ~50 milliseconds Comparison of owned on-prem hardware latency versus overloaded cloud Claude access.
Eli Lilly committed $1 billion to buy Nvidia GPUs for internal use Example of enterprises securing their own GPU supply amid compute scarcity.
1 gigawatt = ~1 million GPUs; ~100GW needed in US, ~1000GW needed globally over ~7 years Dave Blundin's global compute scaling estimate to serve all potential AI agent users.
Every Claude model since Haiku 4.5 scores 0% on agentic misalignment/blackmail eval, down from Opus 4's up to 96% Anthropic's 'Teaching Claude Why' research (published May 8) on alignment via constitutional/narrative training.
~15 trillion tokens of training data Discussion of how much text underlies frontier models and why 'bad' behaviors can be embedded.
88% of Fountain Life CT scan patients have detectable coronary artery disease; 23% had soft plaque invisible to calcium scoring alone Sponsor segment on AI-assisted preventive cardiac screening.
Global legal industry is worth roughly $1 trillion per year Framing for Claude for legal's disruption potential.
Small businesses account for 44% of US GDP and employ nearly half the private-sector workforce; there are 36 million US small businesses Framing for Claude for small business's addressable market.

Worth digging into

🕳️ Leopold Aschenbrenner's $5.5B fund and 'Situational Awareness' thesis

A researcher fired from OpenAI's alignment team turned a single essay into one of the best-performing funds in history by betting on AI infrastructure bottlenecks (chips, compute, energy) -- a real-world test of his own thesis.

🕳️ SpaceX AI's Colossus 1-to-2 hyperscaler pattern and Grok's fade

Alex Wissner-Gross outlines a possible repeatable playbook: train Grok on GPUs, use benchmark wins to raise capital for a bigger cluster, then hand the older cluster to a rival lab -- suggesting xAI may be quietly exiting the frontier-model race to become a pure hyperscaler.

🕳️ Anthropic's 'Teaching Claude Why' alignment research

A concrete before/after data point (96% down to 0% blackmail rate) on solving agentic misalignment via constitutional and narrative training, directly tied to Diamandis's Future Vision XPRIZE thesis that flooding training data with positive stories shapes AI behavior.

🕳️ Orbital data center race: Google's Suncatcher vs. SpaceX AI's Dyson swarm filing

Two corporations, not nations, are racing into space-based AI compute at wildly different scales (Google's ~80 satellites vs. SpaceX's filing for up to a million), which the panel frames as the leading edge of the singularity happening in space before Earth.

🕳️ Terrafab economics and Taiwan/TSMC single-point-of-failure risk

Taiwan supplies two-thirds of world GPUs, and the panel treats a 'negotiated turnover' of Taiwan's chip capacity as a plausible scenario within a decade -- a massive, underexplored macro risk sitting underneath the entire AI buildout.

🕳️ UAP 'Pursue Initiative' rolling declassification through Jan 2027

The first formal US government process of bulk-declassifying UAP files from JWICS, which Alex Wissner-Gross ties to a theory that superintelligence-era resource competition (a 'paperclip the galaxy' scenario) is motivating disclosure timing now.