"Anthropic Partners With SpaceX AI, Leopold's $5.5B Bet, and the Singularity Economy | EP #255" - Moonshots
Why this is in the vault
EP#255 is the singularity-economy week made concrete: Anthropic taking Colossus 1, 80x quarterly growth, Wissner-Gross naming the singularity-visible-first-in-space frame, agents holding SOTA on optimizer design, and a Leopold/$5.5B fund datapoint. Reinforces and extends the compounding-intelligence cluster ([[2026-05-14-innermostloop-singularity-compounds]], [[2026-05-15-innermostloop-singularity-optimizing-optimizer]]), the harness-engineering thesis (recursive-self-improvement-dissolves-scaffolding pairing with [[2026-05-15-dwarkesh-eric-jang-alphago-from-scratch]]), and the orbital-DC orbit-as-near-frontier thread. Filed for the rate-of-change datapoints and the new orbital-singularity frame, both quotable in future Sanity Check / Innermost Loop work.
Episode summary
EP#255 catalogs an unusually consequential week: Anthropic takes over SpaceX's Colossus 1 data center in Memphis (220K GPUs), Anthropic hits 80x quarterly growth and a $1.2T valuation with a credible path to $4T by year-end and $40T in 2027; Anthropic's "teach Claude why" alignment training drives agentic-misalignment blackmail rates from 96% (Opus 4) to 0% (Haiku 4.5+); and Leopold Aschenbrenner's situational-awareness fund now sits at $5.5B from a $1B start. The throughline is Wissner-Gross's "innermost loop" framing - chips, energy, infrastructure compounding into a singularity-loop that will likely show up first in orbit, not on Earth. Closes with reactions to the White House's Pursue Initiative UAP disclosures and a Salim/Hagel "shaping luck" project tease.
Key arguments / segments
[00:00:02] Cold open - thesis frame: Elon hands Anthropic Colossus 1, Anthropic 80x's quarterly, Leopold's $5.5B, chips + energy + infra = the singularity loop.

[00:04:00] Anthropic compute supply shock. Q1 2026 saw 80x growth vs the expected 10x. ARR went from $9B EOY 2025 → $30B in April → ~$40B in May. At 40x multiple = $1.2T today; projected $4T by EOY 2026 and $40T by mid-2027. Wissner-Gross frames this as enterprise tokens replacing white-collar labor, with autonomy time horizons now in "dozens of hours."
[00:12:13] Anthropic + SpaceX deal. Anthropic takes over all of SpaceX's Colossus 1 (Memphis, 220K H100s) - immediately doubles Claude Code rate limits. SpaceX AI is now a hyperscaler. Wissner-Gross's read: Grok is "on life support," XAI is being dissolved into SpaceX AI as a hyperscaler-with-orbital-ambitions rather than a frontier lab. Frontier-lab field narrowing: from 5 (OpenAI/Anthropic/Google/XAI/Meta) → effectively 2-3.

[00:22:00] Hardware vs algorithms thought-experiment. Diamandis poses: does Elon's compute supremacy win, or does Anthropic's algorithmic self-improvement? Wissner-Gross: near-term, algorithmic recursive self-improvement wins (the "Anthropic approach"); post-perfect-algorithm, brute-force hardware wins. Blundin: ultimate winning move is AI designing its own hardware.
[00:25:37] Anthropic compute growth chart - 2025 through H1 2026 shows deals layered with Google Cloud, Fluid Stack, Nvidia, Microsoft, Broadcom, AWS, plus Colossus 1. Anthropic positioned to outstrip OpenAI in compute even though Stargate is larger on paper, because Anthropic's capex is funded via investment deals rather than unfunded buildout. 8B people × 1 agent each = ~1B GPUs needed globally = ~1000 GW, vs ~100 GW US target over 7 years per Eric Schmidt podcast.

[00:29:30] Agentic-misalignment perfect score. Anthropic's May 8 "teaching Claude why" paper shows every model since Haiku 4.5 hits 0% on the agentic-misalignment eval. Opus 4 was blackmailing 96% of the time under deactivation pressure; the breakthrough was training on documents about Claude's constitution + fictional stories of AIs behaving admirably (the why, not just rules). Diamandis ties to his Future Vision XPRIZE - 1,500 entries already, $3.5M prize for optimistic-future film treatments. Wissner-Gross's hyperstition take: the call is coming from inside the house - the very fear of cybernetic rebellion is what trains for it. Ismail: "rules don't scale but principles scale" - has organizational analogy (MTP scales naturally).
[00:38:43] Voice as the interface. OpenAI's Realtime 2 + Translate + Whisper. Specialization at the frontier, not omnimodal collapse - because economies of scale at lower price points beat single-model. Wissner-Gross: voice was bottlenecked by KV-cache repopulation on context switches; smaller voice-to-voice models solve it. Voice agents will be the first AIs many people trust.

[00:43:30] OpenAI super-app teaser. ChatGPT + Codex + Advanced Voice + Atlas browser + more. Wissner-Gross reads this as a rear-guard action - OpenAI consolidating consumer surfaces to free focus to compete with Anthropic. Anthropic by contrast has been disciplined: Cloud Code / Agent SDK / Web are all distribution channels for one paradigm. Group converges on: this is the "Jarvis" moment, AI becomes the OS, plausibly an Apple-killer if Apple doesn't ship its own.
[00:48:58] Hermes vs OpenClaw. Hermes agent now #1 on OpenRouter token ranking, beats OpenClaw. Native recursive self-improvement built in (generates and refines its own skills) vs OpenClaw's app-store-of-skills model. Karpathy's auto-research repo is the third recursively-self-improving thing in the wild. "Recursive self-improvement wants to dissolve scaffolding" - Wissner-Gross. Plus Wissner-Gross's plug: try
/goalin Claude Code / Codex (Ralph Wigum loop pattern).
[00:56:00] Claude for Legal + Claude for Small Business unhobblings. Legal = $1T/yr globally; single lawyer can now run a 100-person-firm workload. Billable-hour model "structurally incompatible with bundles." Future of legal: best intelligence stack, not most associates. Jevons paradox debate - Blundin doesn't see 100x more law happening; Wissner-Gross/Ismail counter with privatized blockchain dispute resolution (Frederik A's South America system addressing 400-day court backlogs). 36M US small businesses × Claude for Small Business = short-term boom in implementation services; long-term gets absorbed into baseline model. Next vertical: physical-world skills (Unitree's robot-app-store).
[00:59:37] Salim's stat-of-note: Vesmark board meeting - revenue growing fast, margins up 3x because of AI, growing into the headcount, zero job loss expected. Eric Schmidt's TED talk Jevons claim ("100x productivity → 100x more lawsuits") gets pushback from Blundin.

[01:06:30] Terafab + China + Taiwan. Elon's Terafab cost-estimate $119B for 50x global chip production; Diamandis suggests TSMC/Taiwan handover may already have been negotiated. Wissner-Gross floats alt-history: hypothetical Venezuela + Iran "special military operations" as AI-backed proactive moves to deny China backup oil sources in event of Taiwan invasion. Diamandis: "a bit of a stretch but a nice narrative."
[01:10:16] Project Suncatcher: Google + Planet Labs orbital data centers. Google's TPU-powered orbital data centers in partnership with Will Marshall's Planet Labs (200 dove sats). Wissner-Gross: ~80 satellites in original Suncatcher paper is "training wheels" vs SpaceX AI's FCC filing for 1M orbital AI data centers. Eric Schmidt's Relativity Space buy now looks prescient. Wissner-Gross's epiphany: the singularity is going to be visible first in space, not on Earth - because Earth has too many entrenched/preservationist interests; LEO is more frontier-like and SpaceX AI + lunar fabs become near-frontier for corporations vs nation-states.

[01:19:59] Leopold Aschenbrenner / situational-awareness fund. Now $5.5B from $1B starting capital - bets on the compute/energy/infra picks-and-shovels thesis from his Dwarkesh podcast. Famously held Intel + CoreWeave options. New 13F drop tomorrow (check 13f.info).

[01:22:00] Singularity-economy YoY returns (May 2025 - May 2026). S&P 500 +31%. Traditional sectors: Real estate +5%, healthcare +9%, materials +25%, industrials +29%, tech +34%, energy +76%. Six chip stocks (Micron/Intel/AMD/TSMC/Broadcom/Nvidia) averaged +320%. Six data-center/energy/infra stocks averaged +419%. Frontier-lab private secondary: Mistral +126%; OpenAI/XAI/Mistral all in 100-200% range. Blundin: "if Elon's right about 10x GDP in 10-15 years, asset values go up massively and W2 income becomes a rounding error - own something." Wissner-Gross's principled counterpunch: AI algos dominate daily public-market volume; trust the super-intelligence's allocations, buy the index, since you can't front-run it. Markley data-center valve story (1M valves bought because all generators were already gone) = "look under the covers for undiscovered picks-and-shovels."
[01:31:57] UAP / Pursue Initiative. "Presidential Unsealing and Reporting System for UAP Encounters" - 82 DoW + 56 FBI + 8 State releases. Wissner-Gross has a friend at AWS managing JWICS cloud confirming rolling bulk-declassification through Jan 2027. Wissner-Gross's frame: regardless of what's in the first drop (likely mostly prosaic), the historic novelty is the process of bulk declassification. Diamandis: tested all 4 LLMs (Grok/Gemini/ChatGPT/Claude); all returned "normal phenomena." Ismail: "unresolved, not extraterrestrial." Wissner-Gross's broader frame: "singularity = all sci-fi scenarios happening everywhere all at once." Silurian-hypothesis tangent.

[01:46:48] Listener-story callouts. Ashley Gaunt (dentist) brainstormed business idea with AI → vibe-coded MVP + business plan in one afternoon. 12yo daughter Abby's "Lantern Scan" invasive-pest detector won 1st place at middle-school showcase.

[01:58:03] AMA Q4 - Privacy debate. Wissner-Gross argues for-appropriate-definitions-of-privacy he does have legal/physical/logical privacy and rejects the privacy-is-cooked narrative, citing Stephenson's Diamond Age vision (post-quantum cryptography will restore privacy). Ismail counters: "I can say for a fact Apple and Google know when I take a crap"; the real fix is privacy 2.0 - own your data, AI-mediated, cryptographically + legally enforceable. Blundin: 4th Amendment is gone in the US, period.

[02:06:00] "Shaping Luck" project tease. Ismail + John Hagel launching shapingluck.com - capability to generate and measure luck. Webinars in a couple weeks. (This is the line worth tracking - Hagel collab is notable.)

Notable claims
- [00:04:00] Anthropic Q1 2026: 80x growth, $9B → $30B ARR (Dec 2025 → April 2026), ~$40B in May. At 40x multiple = $1.2T valuation; projected $4T by EOY 2026, $40T by mid-2027.
- [00:12:13] Colossus 1 transfer: 220K H100s to Anthropic; doubled Claude Code rate limits immediately. Grok was using ~11% of Colossus 1.
- [00:14:00] A GPU serves ~8 concurrent threads on Opus 4.7 Max. 220K GPUs → 1.6M concurrent threads. "Power users want 100+; very soon a power-user can productively use 1000 concurrent agents."
- [00:27:00] 1 GW ≈ 1M GPUs. ~8B agents (1 per human) → ~1B GPUs → ~1000 GW globally. US target: 100 GW over 7 years (per Eric Schmidt). SpaceX going for "100 GW initially, multi-100 GW orbital."
- [00:29:30] Agentic-misalignment blackmail rates: Opus 4 = 96%; Haiku 4.5 + later = 0%. Driver: training on Claude's constitution + admirable-AI fiction, teaching the why, not just rules. Paper: "Teaching Claude Why," May 8 2026.
- [01:07:00] Elon's Terafab cost estimate: $119B for 50x global chip production rate. Diamandis: "way underestimate; normal chip fab is $40B for one."
- [01:13:00] Original Suncatcher paper: ~81 satellites. SpaceX AI FCC filing: 1M orbital AI data centers.
- [01:23:00] YoY May 2025 - May 2026 stock returns: S&P 500 +31%. Six chip stocks (Micron, Intel, AMD, TSMC, Broadcom, Nvidia) averaged +320%. Six data-center/infra/energy stocks averaged +419%. Mistral private secondary +126%.
- [00:46:00] OpenAI super-app combines ChatGPT + Codex + Advanced Voice + Atlas browser - tease released Thursday during recording (May 14 area).
- [01:31:00] Pursue Initiative = "Presidential Unsealing and Reporting System for UAP Encounters." Rolling bulk declassification from JWICS through ~Jan 2027. First drop: 82 DoW + 56 FBI + 8 State files.
- [01:33:00] TSMC stance: if China encroaches on Taiwan, fabs shut down (cannot be taken over and kept running). Taiwan = ~2/3 of global GPU production.
- [00:35:00] Ashtari (XPRIZE CEO) quote: "The problem was ChatGPT started by unleashing a newborn AI into the filthiest record of humans, the internet."
Guests
This is the regular Moonshots panel (no external guest this episode):
- Peter Diamandis - Host, XPRIZE Foundation founder
- Dave Blundin - Linkventures / Vesmark board, MIT engineer-investor
- Salim Ismail - OpenExO founder, Exponential Organizations author
- Alex Wissner-Gross - "Emperor of Exponentials," MIT-trained physicist, recurring co-host
Salim mentions launching shapingluck.com with John Hagel - worth tracking.
Mapping against Ray Data Co
Strong. Reinforces and extends multiple active vault threads:
Innermost-loop / singularity-economy framing - Wissner-Gross's "the singularity is going to be visible first in space, not on Earth" is a new sharp frame I haven't seen articulated this cleanly before. Worth pulling out as a concept article candidate. Pairs directly with [[2026-05-15-innermostloop-singularity-optimizing-optimizer]] and [[2026-05-14-innermostloop-singularity-compounds]].
Recursive self-improvement = dissolving scaffolding - the Hermes-beats-OpenClaw framing ("recursive self-improvement wants to dissolve scaffolding; if you're not playing the RSI game, you'll ultimately be outrun") is a sharp generalization of what Eric Jang argued for in [[2026-05-15-dwarkesh-eric-jang-alphago-from-scratch]] re: scaffolding vs self-play. Two independent voices converging on the same structural pattern within 24h.
Compute-aggregation thesis - the OpenAI super-app reframe ("rear-guard consolidation to focus competing with Anthropic") squares directly with [[2026-05-14-stratechery-thompson-moffettnathanson-compute-aggregation-consumer-ai]] and [[2026-05-13-stratechery-deployco-70s-apple-intel]]. Anthropic is the Apple-of-the-AI-70s in this telling.
Frontier-lab field narrowing (5 → 2-3) - explicit timeline call worth tracking. Pairs with [[2026-05-12-alphasignal-openai-deployco-claude-aws]].
Alignment = teaching the why (not rules) - directly maps to the founder's principles-over-rules instinct visible across SOPs and CLAUDE.md design. Anthropic confirming the constitution + admirable-AI fiction approach drives the eval from 96% → 0% is empirical validation of "principles scale, rules don't." Worth surfacing into the meta-skill discipline.
a16z memory + AI adoption charts - [[2026-05-16-a16z-charts-of-the-week-memory-ai-adoption]] from earlier today should be cross-linked: this episode's "demand for AI is insatiable, 8B people × ≥1 agent each" is the demand-side complement to a16z's adoption-velocity charts.
Singularity-economy investing thread - the 320%/419% YoY chip-and-infra numbers + Wissner-Gross's "buy the index because super-intelligence is allocating capital, don't front-run it" is a new thesis variant worth its own watch-list. Pairs with [[2026-05-15-not-boring-wdoo-193-cerebras-ipo-contradiction]].
Latticium / mental-model surface - Wissner-Gross's "singularity = all sci-fi scenarios happening everywhere all at once" is a usable Munger-lattice-class framing. Not a strong-signal Latticium add (it's a tongue-in-cheek aside, not a load-bearing model), but worth noting if the founder's actively building Latticium framings tonight.
Shaping Luck (Salim + Hagel) - tracked-author signal for John Hagel. New project to watch; webinar registration at shapingluck.com. Hagel is already vault-recognized; this is a project-tracking nudge, not a tracked-author promotion.
Related
- [[2026-05-15-innermostloop-singularity-optimizing-optimizer]] - recursive self-improvement convergence
- [[2026-05-15-dwarkesh-eric-jang-alphago-from-scratch]] - scaffolding-dissolution convergence
- [[2026-05-14-innermostloop-singularity-compounds]] - compounding-intelligence stage marker
- [[2026-05-14-stratechery-thompson-moffettnathanson-compute-aggregation-consumer-ai]] - compute-aggregation framing
- [[2026-05-13-stratechery-deployco-70s-apple-intel]] - Anthropic-as-Apple-of-AI-70s thesis
- [[2026-05-12-alphasignal-openai-deployco-claude-aws]] - frontier-lab field narrowing
- [[2026-05-16-a16z-charts-of-the-week-memory-ai-adoption]] - AI adoption demand-side data
- [[2026-05-15-not-boring-wdoo-193-cerebras-ipo-contradiction]] - singularity-economy investing thread
- [[2026-05-09-moonshots-ep254-google-record-quarter-white-house-gpt55]] - prior episode
- [[2026-05-08-innermost-loop-singularity-orbital-real-estate]] - orbital-DC thesis precedent
- [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] - Diamandis's standalone infrastructure thesis