"OpenAI Cuts Off Elon's Cursor, Humanity's First Star Probe, and Trump's Nuclear Mars Ship" — Moonshots EP #285

Full transcript: [[2026-09-02-moonshots-ep285-openai-cursor-star-probe-mars-ship-transcript]].

Why this is in the vault

Weekly Moonshots news-roundup with two threads directly useful to RDCO: a live-fire case study of an AI system (Physical Super Intelligence) autonomously solving a hard, novel engineering problem in a week that would normally take a specialist team months, and a frank panel discussion of outcome-based AI pricing as OpenAI's answer to Anthropic's enterprise lead — both bear on how RDCO thinks about agent capability ceilings and its own services/pricing model. Also captures the OpenAI-Anthropic-Elon vertical-integration drama with concrete detail (reasoning-trace access as the real motive) worth tracking against the existing [[2026-05-12-stratechery-spacex-anthropic-xai-musks-two-companies|Stratechery]]/SpaceX-IPO threads already in the vault.

Episode summary

Guests Philip Johnston (StarCloud) and Matt Pines (Physical Super Intelligence) join the Moonshot mates to announce the Fermi Explorer mission — a $15M, sub-100kg spacecraft aimed at Alpha Centauri, with its full trajectory (including a novel "perihelion pump maneuver") designed autonomously by an AI physics system in about a week and ~10 billion tokens. The rest of the episode covers OpenAI pulling GPT model access from SpaceX-owned Cursor amid the Musk-Altman feud (with Anthropic stepping in), Sam Altman's four-month AGI timeline claim tied to OpenAI's unreleased Astra model, a Palo Alto startup's fully AI-designed chip (Architect Labs' "Redwood," reportedly 3.4x Nvidia Jetson's performance-per-watt), Elon Musk's satellite-based planetary geoengineering pitch, and Trump/NASA's announcement of a nuclear-powered Mars spacecraft targeting a 2028 launch.

Key arguments / segments

[00:00:08] Cold open teasing all three "moonshot" threads: the interstellar probe, the Architect Labs chip, and Musk's satellite geoengineering pitch.
[00:08:00] Philip Johnston and Matt Pines announce the Fermi Explorer mission: a ~100kg smallsat, ~$15M all-in budget, launching within 3 years, targeting 99% of the distance to Alpha Centauri within 80,000 years using ion thrusters and a 5-year spiral-out/retrograde-burn trajectory.
[00:11:02] Pines describes how PSI's AI system found the "perihelion pump maneuver" — an unintuitive trajectory (retrograde burns near aphelion, then thrust at perihelion to exploit the Oberth effect) — after human JPL-adjacent engineers spent 6 months without a workable answer. Total AI compute: ~10 billion tokens, ~5-6 hours of human steering time.
[00:21:11] Panel debates the Fermi paradox as motivation for the mission — Wissner-Gross frames three resolutions (humanity is first / a "great filter" exists / a "prime directive" zoo scenario) and argues the zoo scenario is likeliest.
[00:30:45] Discussion of PSI's $58M seed round (led by Breakthrough Energy Ventures) coming out of stealth the same day as the mission announcement; framed as proof-of-concept for AI-physics systems tackling problems previously bottlenecked by trained human specialists.
[00:35:00] OpenAI cuts GPT-model access to Cursor (SpaceX-owned, acquired for ~$60B) over concerns it can't ensure SpaceX honors its terms of service; Anthropic immediately steps in with Claude support. Panel's read: Elon needs Anthropic now that OpenAI is gone, and the real prize in play is reasoning-trace data, not just coding-agent market share.
[00:46:00] Sam Altman tells Time he expects an internal AGI-level system by year-end (~4 months out), tied to unreleased model "Astra," which reportedly coordinated 16 agents on a research-math proof and can independently run experiments inside OpenAI's own codebase.
[00:50:51] Wissner-Gross speculates Astra's real breakthrough is a fix for the finite-context-window problem — describing current multi-agent teams as an "oral history" band-aid where agents die around ~1M tokens and pass down lossy summaries to successors.
[00:52:01] Panel covers the shift toward outcome-based AI pricing (Salesforce Agentforce, then OpenAI) — charging per completed outcome rather than per token/API call — with Blundin arguing this is the real unlock for enterprise AI adoption because it changes vendor incentives from "sell compute" to "sell results."
[01:03:28] Architect Labs (Palo Alto) announces "Redwood," a chip it says was designed end-to-end by AI from a single human-written spec in two weeks — zero bugs on first silicon, claimed 3.4x Nvidia Jetson's performance-per-watt. Panel (several of whom are investors/advisors in the company) frames it as a genuine threat to Nvidia's design moat.
[01:15:32] Promo segment for the September 25 Moonshots Live event in LA and a CBS Star Trek 60th-anniversary documentary premiere tie-in, plus a tangent on where Star Trek's imagined future under- and over-shot reality (weak on AI/biotech, strong on faster-than-light travel).
[01:23:00] Elon Musk's tweet-thread on the 2027 AI power crunch (15 GW of compute that can't be turned on due to grid/transformer/cooling bottlenecks, not chip supply) and his plan to vertically integrate natural-gas turbine manufacturing at SpaceX/Tesla to bootstrap solar buildout.
[01:28:25] Musk's separate pitch for satellite-based planetary geoengineering ("sentient satellites" between Earth and sun to titrate solar flux) as a hedge against periodic extinction-level events; panel is broadly supportive, flags governance/tragedy-of-the-commons risk (who decides the planet's thermostat).
[01:37:04] Fountain Life health-segment sponsor read (Diamandis's own preventive-health company) on CT angiography findings — 88% of a sampled patient cohort had detectable coronary artery disease, 23% had non-calcified "soft plaque" invisible to standard calcium scoring.
[01:44:53] Trump/NASA announce a nuclear-powered interplanetary spacecraft targeting a 2028 Mars launch (via NASA administrator Jared Isaacman), cutting transit time from ~7 months to 1-2 months; panel frames this as an implicit NASA-vs-SpaceX-Starship race to Mars, alongside a newly chartered U.S. Space Academy modeled on West Point/Annapolis.
[01:56:09] Closing banter and sign-off.

Notable claims

Fermi Explorer: ~$15M total mission cost, ~100kg spacecraft (60% xenon propellant), targeting 99% of the distance to Alpha Centauri within 80,000 years, launch within 3 years [00:09:00].
PSI's AI system found the mission trajectory using ~10 billion total tokens and an estimated 5-6 hours of human steering time over one week [00:28:00] — offered by the panel as a general "reference class for problem difficulty" (10^9-10^10 tokens = roughly a human-team-decade of specialist effort).
Sam Altman: internal AGI-level system expected within ~4 months (by year-end 2026), per Time magazine, tied to unreleased model "Astra" [00:46:00]. Panel is skeptical of the specific "AGI" framing given Altman's prior (deleted) 2022-era "AGI achieved internally" claim, but treats the underlying capability claims (16-agent coordinated math proof, autonomous experiment-running) as credible.
Architect Labs' "Redwood" chip: designed end-to-end by AI from a single spec in 2 weeks, zero bugs on first silicon, claimed 3.4x Nvidia Jetson's performance-per-watt [01:03:28-01:04:00]. Note: several panelists (Wissner-Gross, Diamandis) disclosed they are investors/advisors in Architect Labs — treat the performance claim as company-sourced and unverified by a third party.
Elon Musk: ~15 GW of 2027 AI compute capacity will sit idle due to power-delivery infrastructure bottlenecks (transformers, cooling, wiring), not chip supply or raw electricity generation [01:23:00].
NASA nuclear Mars spacecraft: targeting 2028 launch, cutting Mars transit time from ~7 months to 1-2 months [01:44:53].

Guests

Philip Johnston — Founder/CEO, StarCloud (orbital data centers); co-founder of Physical Super Intelligence; second Moonshots appearance.
Matt Pines — Co-founder, Physical Super Intelligence (AI-for-physics lab); first Moonshots appearance.
Hosts: Peter Diamandis, Dave Blundin, Alex Wissner-Gross (regular Moonshot mates; a fourth host, "Salem," was absent this episode).

Sponsorship

Rotating sponsor reads, standard for this show:

Google for Startups — pitched as access to Google DeepMind generative-media models via a production-deployment technical guide [00:22:01].
Blitzy — autonomous software-development platform claiming 80%+ of a dev sprint automated, positioned against Nvidia/enterprise coding-agent competitors [00:45:00].
VoiceRun — code-first voice-agent development/runtime platform [01:10:00].
Fountain Life — Diamandis's own longevity/preventive-health clinic company; the health segment functions as a house promo (self-referential — Diamandis is a Fountain Life co-founder/investor), not a pure third-party sponsor [01:36:00].

Mapping against Ray Data Co

The Fermi Explorer trajectory-design story is the most concrete data point in this episode for RDCO's own agent-capability tracking: an AI-physics system solved a genuinely novel, high-constraint optimization problem (not a lookup or interpolation task) faster than a team of human specialists, using a workflow of "human writes tight spec, AI does the design work, human reviews the output" — structurally similar to RDCO's own agent-delegation posture (station-spec-author / station-code-author pattern) but at a much harder problem class. Worth filing as a positive data point for the thesis that agent capability continues to widen the set of tasks safely delegable to AI. The context-window "oral history" framing [00:50:51] is also directly relevant — RDCO's own CLAUDE.md hard rule #4 (route long artifacts through subagents to avoid [[context-rot]]) is effectively the same band-aid Wissner-Gross describes at a larger scale; worth revisiting if Astra or a similar model actually ships a fix, since it would change the calculus on when sub-agent fan-out is still necessary. The outcome-based-pricing discussion [00:52:01] is a useful reference point if RDCO ever reconsiders its own services pricing (e.g., phData engagement structure or a future AI-consulting offering) — pay-per-outcome vs. pay-per-token is a live debate in the same market RDCO operates adjacent to. The OpenAI/Cursor/Anthropic reasoning-trace framing is a good complement to the existing [[2026-05-12-stratechery-spacex-anthropic-xai-musks-two-companies|Stratechery SpaceX-Anthropic-xAI vertical-integration note]] already in the vault. Mars/nuclear-propulsion and satellite-geoengineering threads are lower relevance to RDCO directly — filed for vault completeness on the tracked Moonshots source and general "singularity is accelerating" trend-tracking.

[[2026-08-29-moonshots-ep284-nvidia-96b-quarter-openai-chip]] — immediately prior episode; continues the OpenAI-chip and AI-infra-capex thread.
[[2026-08-27-moonshots-ep283-sam-altman-singularity-slowdown]] — prior episode's AGI-timeline and agent-swarm-operator threads, direct continuity with this episode's Astra/AGI claims.
[[2026-05-12-stratechery-spacex-anthropic-xai-musks-two-companies]] — deeper treatment of the Musk/Anthropic/OpenAI vertical-integration dynamics referenced in the Cursor-cutoff segment.
[[2026-08-13-moonshots-200gw-grid-sodium-batteries-wave-datacenters]] — companion energy/grid-bottleneck thread relevant to the 15GW idle-compute claim.

moonshots ep285 openai cursor star probe mars ship