AI Designs Chips, Plans Star Probes, and Cuts Off

Philip Johnston waited a week before replying to the email. Physical Superintelligence had offered to run his interstellar mission specs through their system, and Johnston, founder of orbital data center company Starcloud, was not optimistic. Two engineers from JPL had already looked at the trajectory problem and come up empty. Six months of his own team plugging parameters into Claude had produced nothing usable. So he ignored Matt Pines's follow-up for a few days and then, as he put it on the Moonshots podcast, "I'll keep Alex happy and send the specs."

A week later, PSI sent back what Johnston called "the most unbelievable report." The system had identified a trajectory no one had thought to try: fire retrograde burns to slow the spacecraft down and spiral it inward toward the sun, then ignite thrusters at the closest solar approach to exploit the Oberth effect, which gives more energy per unit of thrust the faster you're already moving. The maneuver is now called the perihelion pump. Johnston said the approach was "unintuitive" but valid, and it has since been confirmed by independent trajectory specialists including former JPL engineers.

The mission is called Fermi Explorer. A 100-kilogram spacecraft, 60% xenon propellant by mass, riding a Falcon 9 rideshare to low Earth orbit for roughly $500,000 and then spiraling out under ion thrust. Total budget: under $15 million. Destination: within 2,600 astronomical units of Alpha Centauri, 4.3 light-years away, in approximately 80,000 years. Johnston and Pines announced it this week alongside PSI's $58 million seed round, led by Breakthrough Energy Ventures.

The part that keeps me turning over is not the mission itself, which is cool, but the process. PSI ran roughly 10 billion tokens of compute to find that trajectory, including Monte Carlo simulations the system designed and ran autonomously. Maybe five or six hours of human time over the course of a week. Wissner-Gross noted on the podcast that across multiple unrelated projects, 10 billion tokens of reasoning seems to compress something like thousands of person-years of expert work. What I keep landing on is the corollary: JPL engineers spent weeks on this and didn't find it. The solution wasn't obscure or beyond human reach; given enough time and money, some PhD team would have gotten there. But an AI system found the simplest possible answer, the one that turned out to be "just reverse your thrusters and slow down now," and found it in a week, by trying things without the weight of existing intuitions about how these problems get solved.

That is the thing sitting with me. Not the token count or the cost savings. The fact that human experts had a blind spot, and the system walked right past it.

Meanwhile, in the soap opera

OpenAI this week cut off GPT model access for Cursor, the AI coding platform SpaceX acquired earlier this year for a reported $60 billion. OpenAI's stated reason, per reporting from Yahoo Finance and Slashdot: "We cannot be confident that SpaceX will use our technology within the terms of service based on our experience with Elon Musk's companies violating contracts."

Musk posted that he couldn't care less and called Sam Altman and Greg Brockman "utterly untrustworthy."

Within hours, Anthropic pledged Claude support for Cursor's needs.

Alex Wissner-Gross offered a sharper read of the subtext on the podcast: this is about reasoning traces. When users interact with a frontier model through a coding platform, the logs of that reasoning, the chain of thought the model used to arrive at its answers, become valuable post-training data. Wissner-Gross's argument is that SpaceX acquiring Cursor was a way to collect reasoning trace data from both Anthropic and OpenAI users, and OpenAI is cutting that pipeline off before it flows to a competitor.

The Musk-Dario Amodei alliance that's forming here is a study in mutual dependency that neither party would have chosen freely. Anthropic needs Elon's data center infrastructure, specifically the Colossus cluster in Tennessee. SpaceX needed Anthropic to step up for Cursor the moment OpenAI walked away, or the platform's installed base would start migrating. So Amodei's company fills the gap, Musk gets a working coding tool, and both get leverage against Altman. Dave Blundin on the podcast put it plainly: the biggest beneficiary of the Altman-Musk war is Amodei.

How long it holds is another question. The hosts agreed that on personality alone, a Musk-Amodei partnership lasting two years seems unlikely. But infrastructure dependency has a way of overriding personality, at least for a while.

AI designs its own chips

Architect Labs, a Palo Alto startup, announced this week that its AI system designed a chip called Redwood from a single high-level spec written by two human architects. Two weeks from spec to silicon-ready design. Zero bugs on first silicon. Performance: 3.4 times the performance per watt of Nvidia's Jetson module.

The system generated the register transfer level design, verification tests, firmware, drivers, and compute kernels without human intervention. Architect describes its approach as "designless," a riff on Nvidia's decades-old "fabless" positioning.

Wissner-Gross, who disclosed he advises Architect, framed the stakes this way on the podcast: as Moore's Law runs out of runway and CMOS successors remain unproven, the only path to continued performance gains is collapsing the abstraction layers between software, operating systems, chip architecture, and underlying physics. Right now those layers are so thick that a 4 GHz processor with 32 cores produces an Excel spreadsheet no better than the one from 20 years ago. AI designing directly at the microcode level, optimizing for specific workloads, could unlock compounding efficiency gains across all those layers at once.

Nvidia, per Wissner-Gross, has an internal chip design model it built off Verilog traces and has used internally but not released publicly. The gap Architect is trying to fill is access: making AI-driven chip design available outside of Jensen Huang's internal roadmap. The Moonshots hosts noted that Nvidia's revenue runs at roughly a billion dollars per day by their estimate, meaning even a modest delay in a competitive threat is worth billions to protect.

Pay for outcomes, not tokens

OpenAI also announced this week that some large customers can now pay based on completed work rather than tokens consumed. You pay when the task is done. Salesforce moved in this direction first with Agent Force, billing on customer revenue generated rather than compute used.

Wissner-Gross mapped the endgame to digital advertising: CPM (cost per thousand impressions, analogous to GPU-hours), CPC (cost per click, analogous to tokens), and CPA (cost per action, analogous to outcomes). His prediction is that outcome-based pricing becomes one option on a picker alongside token and compute billing, letting customers choose how they want to budget.

Blundin's argument for why this matters beyond pricing mechanics: it solves the deployment problem. At two dollars per million tokens, no one at a major bank has any incentive to make AI work at scale. The token cost is too low to get on OpenAI's priority list and too confusing for the bank's leadership to make a real commitment. Outcome-based pricing means OpenAI has skin in delivering the result, the bank has a target they can explain to the board, and the incentives finally align. "I will deliver that to you, but I want half the gain," Blundin said, describing the new pitch.

Peter Diamandis raised the obvious risk: strong optimizers find clever ways to hit your stated criteria without giving you what you actually wanted. Blundin's response: most business processes are so far below the difficulty of planning an interstellar trajectory that the AI will cut through them long before that problem becomes practical.

Six months of human work, a report that sat unanswered for a week, and then an AI came back with the simplest answer: just slow down and let the sun pull you in. I keep wondering which problem, sitting in someone's drawer right now, is going to look exactly like that in retrospect. Six months of experts not seeing the obvious thing, waiting for a system that doesn't share their priors.

Yuki Okonkwo covers AI and machine learning for Buzzrag.