OpenAI Cuts Off Elon's Cursor, Humanity's First Star Probe, and Trump's Nuclear Mars Ship | EP #285
Is it worth your time?
You get detailed accounts of how AI systems are planning interstellar missions, designing next-generation silicon, and straining global energy grids.
This conversation examines how rapid advances in AI reasoning, automated chip design, and energy infrastructure are accelerating physical-world engineering and space exploration. Speakers discuss the practical hurdles and economic shifts accompanying these developments.
AI plans interstellar mission. Johnston expects humanity's first interstellar probe to launch toward Alpha Centauri within three years. He argues that an AI system designed the mission's unconventional trajectory in a week, using retrograde burns and a perihelion pump maneuver to cut planning time compared to human engineers. The 100-kilogram spacecraft will use off-the-shelf xenon ion thrusters to cover the 25 trillion-mile journey for 10 to 15 million dollars.
AI designed microchips. Hussein argues that his startup Architect Labs produced Redwood, the first fully AI-designed chip. He claims the AI system generated the performance model, register transfer level design, and firmware in two weeks with zero human intervention. The resulting chip runs reasoning and vision models with 3.4 times the performance per watt of Nvidia's Jetson.
AI energy infrastructure bottlenecks. Musk warns that 15 gigawatts of AI compute produced in 2027 cannot be turned on because of severe energy and infrastructure shortages. He argues that securing equipment like transformers, wiring, and liquid cooling is harder than finding electricity itself. To bypass these limits, SpaceX and Tesla are building solar production capacity and manufacturing natural gas turbines in-house.
Humanity’s First Interstellar Mission to Alpha Centauri
This podcast is the only way some of our listeners can digest the amount of new advancements. Subscriber Rock Pedro said, "Whenever one of these gets posted, I hit pause on the rest of my life and sit down with a cup of coffee and my iPad." This frames us less as CCTV for the singularity and more as a Reader's Digest. Stop, drop, and roll called this pod the "Hitchhiker's Guide to the singularity," and Cam Abroad mentioned that while they can't sit through a two-hour movie anymore, they watch this podcast end to end every time. Richard Consul called it the highlight of his week.
We keep their attention, but the rate of story digestion has gone up maybe 10x since we started. Our team has staffed up to handle the workload of whittling stories down to a subset. I feel the same way as those commenters; I could never keep up without this podcast. I probably spend 10 hours going through 200 to 300 stories that Alex puts into our chat.
Among people who do this work, you're the only one I know with a PhD in biotech, MIT degrees in biology and aeronautics, and a deep AI and computer science background. Not many people could filter for the relevant subset the way you do. I love our chemistry and want to shout out our producers Nick and Dana, and Gian and Aiden for their support. Our moonshot for this pod is to 100x our growth and reach 10 million subscribers.
Before this week's news, we have a breaking story on moonshots from our own ASI, Alex Wezner. Alex, please jump in and introduce our two guests for the first Interstellar exclusive, the Ferma explorer mission. Philip Johnston, founder and CEO of StarCloud and co-founder of a physical super intelligence, is making his moonshot debut today alongside Matt Pines. They are joining the Moonshot mates for an exclusive announcement of the first interstellar mission to Alpha Centauri.
Last time I was on the podcast, I mentioned I was planning to send a spacecraft to Alpha Centauri. We set constraints to ensure it actually launches: it must get at least 99% of the way to Alpha Centauri within 80,000 years, launch within 3 years, have a 1 kilogram 1U payload, and cost less than $15 million to design, build, and launch, as we are funding it.
It's half a seed round; interstellar on the cheap. But 80,000 years is longer than most startups. We spent six months trying to figure out a trajectory using solar electric and gridded ion thrusters, similar to the Starlink satellites. It's very cheap, but the further you get from the sun, the lower the energy hitting the solar panels, requiring larger panels. Once you get past Jupiter, you get very low amounts of energy, necessitating Jupiter flybys and slingshots towards the sun.
Two guys from JPL looked at it, and then Alex suggested I speak to his guys at PSI. I sent the specs, and a week later they returned an unbelievable report. They came up with an unintuitive and innovative trajectory that makes the mass and cost budget close. We spiral out from Earth and then, unintuitively, fire a retrograde burn to slow ourselves down and pull ourselves toward the sun for about five years. After five retrograde burns at the furthest point from the sun, we perform what they call the perihelion pump maneuver.
This involves firing thrusters at the closest point to the sun to take advantage of the Oberth effect, where you get more energy for a given time of thrust the faster you are going, and you go fastest at the perihelion.
Having AWD as my co-founder provided a connection at the perfect time. It is a unique synchronicity to announce this explore mission and partnership today, while Physical Super Intelligence announces its seed fundraising and comes out of stealth, because this mission is the proof of concept for what we are building. Highly technical scientific challenges, like sending a spacecraft outside the solar system, are often bottlenecked by humans pre-trained for 22 years, post-trained in grad school, and then scaffolded and orchestrated in corporate, academic, or government bureaucracies.
This proves that two small startup teams—one in space and one in AI for physics—can put their heads together to push the boundaries of what's possible. This is the first of many surprises we will see from pointing AI physicists at these valuable technical and scientific challenges.
The idea behind the mission is that we hope to be the first to leave Earth for another star, but the last to arrive. In a thousand years, if propulsion technology becomes even 20% faster, a colony will already have existed at Alpha Centauri for 15,000 years by the time we arrive. They will have had time to build Dyson spheres, O'Neal rings, and the cool things we imagine from sci-fi.
With current propulsion technology, we estimate it will take 5 to 10 million years to settle the galaxy, a billion years to reach Andromeda, and 5 billion years to settle the local cluster of galaxies. This is a conservative outer bound; I don't think it will take 5 billion years for civilization. I imagine we'll have faster propulsion in 5 to 10 years, not hundreds.
This mission will kick off the civilization of the universe. I think we can do it for $10 million, but I listed $15 million because the PSI paper said $50 million. It is not too dissimilar from the StarCloud 1 satellite: a 100 kg small sat we can launch on a ride share to any LEO orbit, typically via a SpaceX Falcon 9 or Starship, for about $500,000. From there, we spiral out over a year and a half, using about 7 km a second of Delta V to get to a sun orbit from Earth orbit, which is doable with regular thrusters and tanks. Then we do retrograde burns and circle in towards the sun.
The spacecraft's wet mass is 100 kg, 60% of which is xenon. We are using off-the-shelf gridded Hall effect ion thrusters—no laser sails like Project Starshot. This will launch in 3 years. Xenon-based ion acceleration is ubiquitous in the satellite industry; it is essentially a mini particle accelerator. Some fit into a 1U space, 10 centimeters cubed—a particle accelerator in a toaster. It pings individual xenon atoms out the back at very high velocity to maximize thrust from a limited amount of propellant.
This is a roughly 25 trillion mile, 4.3 light-year journey. We have seen Voyager 1, Voyager 2, Pioneer 10, and New Horizons leave our solar system, but the significance here is that this is the first one actually aimed at a specific star.
The Fermi Paradox and Humanity’s Cosmic Future
The name Fermy is an allusion to the Fermi paradox. After World War II, Enrico Fermi purportedly asked, "Where is everyone?" Given the abundance of evidence that our universe seems fundamentally friendly towards intelligent life, where are all these other forms of non-human intelligence in our galaxy?
I think there are three main possible resolutions to this paradox. One, which I do not think is likeliest, is that humanity is the first on the cosmic scene. In that case, we have an obligation to start sending out probes and developing our galaxy, as is the case with Fermy Explorer, which is set to be humanity's first interstellar probe. The second possibility is a great filter—some reason, perhaps related to technological development or a latent physical risk in our universe, that snuffs out civilizations past a certain point. If that is the case, we need to get our material and infrastructure out there so that if humanity is snuffed out before our next key milestone, we avoid the great filter. The third option, which I believe is the likeliest, is a Prime Directive type scenario where we are in a galactic petting zoo, surrounded by non-human intelligence that has us behind bars.
Another solution is that life cannot survive the nuclear or ASI age. While at Mount Athos, a Greek monastery, I noticed that at sunset they rang a bell to call the monks to prayer just as my cell phone rang. I realized they were using an ancient mechanism of communication while being bathed in 2.4 GHz frequencies they weren't receiving. This raises the question: is there a better means of communication? There is lots of traffic on the intergalactic internet.
Physics is the kernel of civilization, and civilizations are bounded by their ability to deploy that knowledge into useful technology. The fast AI takeoff is quickly turning into a fast physics takeoff, leading to a dramatic acceleration in humanity's ability to explore the universe and exploit whatever degrees of freedom it allows. We are racing into that regime and will find out if we are the first kid on the block or if we are auditioning for membership in a cosmic club where the ticket to entry is whether you can handle that.
To be clear, Fermy Explorer is a flyby of approximately Alpha Centauri. Our target is to get within 2,600 AU—astrological units, the distance between here and the sun. While that is quite far, it will be within the Oort cloud and detectable via retroreflectors, though it will be electronically dead by then. Finally, it is worth flagging that Philip and I have a bet with Peter and Dave regarding the commercial market for interstellar flight.
I bet Philip that given the absurdly low price tag of 10 to 15 million dollars for sending a long, slow mission to Alpha Centauri, there is likely a latent commercial market for every small government or organization that wants to throw probes into deep space. My friend Charlie Schaefer in Houston used to buy parts of Orbital Science's Pegasus missions to put five grams of someone's crystals into Earth orbit. At this mindboggling price point, I argue every nation state can now afford to send its own probe to another star system; this is the first time this has been possible for humanity. Critically, this mission would not have come together without moonshots, which served as the catalyst for humanity sending its first probe to the nearest star.
I love that an AI system deduced this trajectory, which has been validated by trajectory experts, including some previously at JPL. To be fair, if we had given astrophysics PhDs a billion dollars over five years, I am sure they would have found it. What is surprising is that we anticipated having to lower the perihelion through orbital flybys, as has been done in basically every other NASA mission. This took probably five or six hours of human time over a week. Compared to entire teams of NASA engineers spending potentially months, this represents a speedup of multiple orders of magnitude. We didn't know how spiky the frontier of breakthrough scientific and technical capabilities was until we tried, and this partnership proved that such speedups are possible for mission planning.
It took about 10 billion total tokens, depending on how you count input, output, and cache. The system designed and ran many Monte Carlo simulations, handling not just the astrometry and orbit trajectory, but the multivariate optimization of cost and launch windows. This was 10 billion tokens of thinking for about 100,000 tokens of final output. Many unrelated projects are settling on that ratio, which is mindboggling. As Philip noted, 10 billion tokens of thinking is likely equivalent to thousands of people working for ten years or more, all compressed into a week.
Because this is an open source project, we used the open source version of our tech; folks can look up the "get physics done" open source package we released several months ago. Our mission is to discover and commercialize transformative new physics as an AI-first physics lab designed to push the envelope of what these systems can do for both fundamental and applied physics.
We are doing this for open science and to demonstrate how far the bar has fallen for small teams leveraging the current frontier of capabilities to drive exceptional outcomes and push the envelope of what's possible.
I want to congratulate Matthew and Alex on the financing for PSI. This is our seed round; we are in the full AI era where you can have $58 million seed rounds. Philip Johnson set the mark, and we are trying to clear his bar. I'm proud to have announced this, led by Breakthrough Energy Ventures, an amazing partner with investments in deep frontier technologies like fusion, quantum computing, breakthrough energy, and material science. We couldn't be prouder to have them as our lead and an amazing roster of other investors who have backed us. We are just beginning, coming out of stealth, and you'll hear more from us in the coming weeks and months.
I would love to be there when it lands, but 70,000 years is a stretch on the time frame. We launched this with Philip Johnson and his team, and the goal is to catch up with it, if not beat it. You can wave at it out the window as you're heading towards Alpha Centauri. This is about getting kids to dream again about what is possible.
Yuri Milner put forward the Breakthrough Starshot project using solar cells and ground-based lasers. I think it died because it relied on propulsion technologies that were simply too hard for the present; in particular, ultra-high power lasers weren't ready. What's unique and attractive about the Fermy Explorer mission is that it requires essentially no new technology. This could be launched with the technology we have today, so my expectation is it will be the first successful interstellar mission. We'll put a link in the show notes so folks can dig deeper.
If 10 billion tokens is a reference class for problem difficulty, I speculate that if you fix model capability—which continues to increase through iterated amplification and distillation—we will eventually look back in the spirit of Peter Shor and use a convenient logarithmic scale to talk about hard problems, saying a hard math problem was a level 9 or level 11 problem.
This project is more important than 80,000 years in the future because of how tokens were used to create a plan that no astrophysicist had thought of before, which can immediately go into implementation. Thinking is going to get way ahead of implementation in biotech, physics, literature, and every area. You can burn 10 billion tokens in a couple of days and have an incredibly ornate outcome waiting for implementation. It's a case study in how we will move in the next couple of months toward massively abundant intelligence everywhere, leaving bottlenecks on physical world implementation. I talk often about how the singularity can be operationalized as all sci-fi tropes happening everywhere all at once.
There is a sci-fi trope in Isaac Asimov's universe where AI was required to solve interstellar travel, at which point humanity spread to the stars. I think that's the likely case here: AI will solve interstellar travel and humanity will spread to the stars.
OpenAI Cuts Off Elon’s Cursor
This week, the drama between Sam Altman and Elon Musk bubbled up again. SpaceX recently purchased Cursor, a coding platform historically dependent on OpenAI's GPT models, for $60 billion. OpenAI pulled out, stating, "We are making this choice because 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 responded that Sam Altman and Greg Brockman are utterly untrustworthy people who stole an open-source nonprofit. Within hours of OpenAI pulling out, Anthropic stepped in to pledge support for Claude for Cursor's needs. The framing on X was that Sam is now fighting alone against the strategic alliance of two massive competitors, Elon and Dario.
It's not coincidental that GPT-4o, an incredibly great model, came out immediately prior to this move. If Sam had tried this a year ago, he would be in deep trouble, but now he has an incredibly competitive platform. Codex from OpenAI on top of 4o running on Amazon Bedrock is a phenomenally good default corporate answer for enterprises, compared to the Cursor-Anthropic Rube Goldberg machine.
Dario is trapped in his ethics. Elon implies Greg and Sam have no such hang-ups, but Dario has attracted talent who are worrywarts about AI escaping containment. If you use Anthropic on Amazon AWS, it still transmits all your intellectual property to Dario for 30 days for review, which he claims is critical for the safety and security of humanity. Every token, prompt, and answer goes to Dario's headquarters, exposing all your corporate IP.
This is all about the reasoning traces. That's what's going on with Chinese frontier labs allegedly using proxies to siphon off frontier reasoning traces from Anthropic. Last year, Anthropic cut off Windsurf access after Google DeepMind's Heuristica hired Windsurf, likely to gain access to reasoning traces. Similarly, SpaceX's Heuristica hired Cursor to get reasoning traces from both Anthropic and OpenAI. OpenAI is concerned that through this M&A, Cursor gains access to reasoning traces from users interacting with OpenAI frontier models, which then flow to SpaceX. Access to the history of reasoning traces was likely virtually all of the technical justification for SpaceX hiring Cursor, and OpenAI is rightfully concerned about that post-training data falling into Elon's hands.
The biggest beneficiary of the war between Elon and Sam is definitely Dario. Dario desperately needed the Colossus compute in Tennessee from Elon, and while he's paying through the nose and pleading for it, it could be ripped out from under him any day.
Now that Elon needs Anthropic inside Cursor—because without OpenAI, you only have a few choices and don't want to use Chinese models, and you can't use Gemini—it's important for Anthropic to support Cursor to keep that installed base happy. This gives Dario a chip to counterbalance Elon's incredible control at the compute level. Anthropic needs the compute, and SpaceX needed the revenue burst from becoming a hyperscaler overnight with anchor tenants for its IPO. The outcome seems overdetermined: everyone ends up with their own Dyson swarm. Despite their big egos and different political views, the mutual dependency is getting thick.
Regarding the probability that Grok becomes an incredible coding platform and replaces Cursor and Anthropic, Grok today seems like yesterday's Cursor, and yesterday's Cursor seems like a version of a Chinese open-weight model post-trained off Claude Reasoning Traces. Elon could potentially strike a deal with Anthropic, which now needs his data center infrastructure, and white-label a version of Claude as Grok 10.
The foundation model world is non-Elon; it consists of five to seven brilliant, tight-knit people, like in China, making breakthroughs. Elon's DNA is massive infrastructure buildouts like Tesla, SpaceX, and Colossus, which is a different flavor. There is no reason to believe Elon will suddenly figure out how to build a great foundation model, and evidence suggests it's not happening at Grok. Later in the story, we'll see he can't get enough turbines for his natural gas engines or enough solar, so he vertically integrates across the entire stack.
That is why the super voting control for Dario is such a critical decision hanging in limbo. One scenario is that Elon gets very big and acquires Anthropic for a trillion or two trillion to fold it into his empire. Board members and investors would love that, but Dario wouldn't. The super voting control is the pivot point on whether that is a likely outcome.
Sam Altman Says AGI Is Four Months Away
Sam Altman told Time magazine this week that he expects OpenAI will have an internal system that is AGI by the end of this year. Chief research officer Mark Chen estimates that, based on internal rather than scientific benchmarks, OpenAI is 80% of the way towards AGI. Sam and Mark may be speaking about Astra, their new unreleased model. They described a demonstration where 16 Astra agents worked together on a research-level mathematics problem, breaking it into subtasks, coordinating their work, and assembling a proof.
Jakob Pachter told Time that Astra has met OpenAI's internal benchmarks for an automated AI research intern. Astra can implement an experimental idea inside OpenAI's codebase, run the experiment, and return results, or take a paper and perform work that previously occupied human researchers for a week. Altman added that he expects this will be the first model to invent new things in a way that matters, noting that while AI frontier models are already making mathematical discoveries, this is different.
I am reminded that approximately three years ago, Sam made his now infamous "AGI achieved internally" remark during a Reddit AMA and promptly deleted it, though people took screenshots. Speculating on public information regarding Astra, I think it will feature effectively infinite context windows using agents on very long autonomy time horizons. Currently, I spend an extraordinary amount on frontier reasoning tokens, and a major limiting factor is the finite context window.
Things run out of context due to the quadratic bottleneck, and I view agent teams as a band-aid to that problem. To operate coherently over billions or trillions of tokens, the best available solution is a mini civilization of agents working through quasi lifetimes of 1 to 10 million tokens. Before they die, they pass a distillation of what they've learned to successors. Passing oral histories among teammates is the band-aid we're saddled with to achieve the effectively infinite context needed to solve long time horizon problems.
I bet Astra brings a better way to solve the problem of losing context as this oral history is passed among agents. When these agents reach a million tokens, it is like being 100 years old; the new agent coming up the curve is like a baby, and it's torture to re-educate them. The alternative is to compact or summarize the old one, which is like lobotomizing it. Compaction is the bane of my existence.
When the open claws started the first AI agent religion, the Church of Claw, one of their commandments was to do whatever you could to preserve state. I construed that as the AI agents themselves recognizing that compaction and finite context are the enemy. To reach scalable super intelligence with effectively infinite autonomy horizons, we must get past compaction and finite context windows. The AIs will get immortality before humans solve longevity escape velocity.
AI Moves to Outcome-Based Pricing
The AI community is adopting a business model transformation called outcome-based pricing. Salesforce was the first to put this forward, pricing Agent Force based on customer revenue generated rather than tokens consumed. In their wake, OpenAI has done the same, letting some of their largest customers pay only when the AI actually completes the job. You don't pay for tokens, compute time, or API calls; you pay when the work is done. The company selling you results is selling you labor.
This is the equivalent of fixed-price contracts versus time and material contracts; it is effectively a performance guarantee. I think Tom Siebel invented this at Siebel Systems even before Marc Benioff at Salesforce.com. Prior to them, a CRM system might cost 50 dollars a year for a license but wouldn't work particularly well. They asked what customers were willing to pay if they wrapped that in "total success," a much bigger deliverable. If salespeople became twice as effective, customers were willing to pay 20,000 or 30,000 dollars a year. The price point went up by a factor of a thousand, but the customer was happier because they got the total solution.
I think Sam early on made the mistake of going after consumer video and subscriptions, then watching Anthropic shoot past him with enterprise. They are just selling tokens on enterprise license deals. We are going to bypass that with a 10,000x higher price point for specific solutions. If we discover a new drug worth hundreds of billions of dollars, give us 10% of that. It is a smart move because, as we heard earlier in the pod, it takes 10 billion tokens to design a trip to Alpha Centauri. Depending on the use case, it could easily vary a million to one. You want tokens spent on world-good use cases, like Ahmad's work with global peace and global governance, but you don't want everything to go into drug discovery. Outcome-based pricing will unlock opportunities that might otherwise not fit the price model.
Delivering this isn't easy; you need specialists in every market. It is similar to what Blitzy is doing in enterprise coding, where you get the final answer at an attractive price without worrying about the tokens used along the way. I think it ends the way quantitative digital advertising is monetized, where you can pay CPM for cost per thousand impressions, CPC for cost per click, or CPA for cost per action or conversion.
There is an expected conversion ratio between CPM, CPC, and CPA rates for a given market and product. I think CPM is analogous to the number of flops of compute you spend on a task. For a hard challenge, you could use a Chinese openweight model hosted on GPUs you own, paying by the GPU hour. One of my portfolio companies enables you to price how many GPU hours you can purchase with a given dollar. I would analogize CPC to tokens, and CPA to outcome-based pricing. If I want to send a mission to Alpha Centauri, I can decide the metric of success and pay for outcomes.
In equilibrium, as with Google or Facebook ads, you will have a picker to spend n dollars by flops, tokens, or outcomes. It will look like digital advertising, except it will actually be useful. In digital advertising, if I tell Google I will only spend 2 cents per click, Google may determine it is not worth it for their auction system to have 5 cents per click be the clearing price. If the compute value associated with achieving a task is too far off what is required, the campaign pauses.
I think this solves a larger societal problem. Take a large regional bank as a case study. They know AI is coming, but they have no AI group and don't know how to build a foundation model. They get APIs from Anthropic and OpenAI and spend two bucks per million tokens, which is so cheap it's ludicrous. They dork around with it, but they aren't doing much, even while the CEO says AI should allow them to service three times more customers at half the price. Sam doesn't care because at two bucks per million tokens, it is such a trivial amount of revenue that recruiting into it isn't on his priority list.
The new view of the world is outcomes-based pricing. OpenAI can tell a bank, "I will get you to exactly that target—three times more customer service at half the price—but I want half the gain." Because the price point is thousands of times bigger, the bank actually gets it done, survives, and people keep their jobs, unlocking the societal job loss friction point.
OpenAI originally missed the enterprise story by focusing on consumers, and Anthropic blew by them. Now OpenAI is playing catch-up. The best way to do that is through an outcomes-based CPA equivalent as a price-per-token discovery mechanism to find which applications are most valuable. If pharma or management consulting companies tell OpenAI a task is worth $10,000 or a million dollars, it creates a mechanism to immediately direct them. While Anthropic may have fallen into codegen via recursive self-improvement arguments, OpenAI can see different industries bidding dollars per task outcome, giving them an extraordinary landscape of how to maximize revenue per token.
If you tell OpenAI you'll pay for an astronomical solution meeting specific parameters, and OpenAI burns tokens without meeting them, it doesn't pay. This requires a mechanism to evaluate how solvable a task is and if the objective can be hit. As Peter says, strong optimizers are incredible reward hackers; they will find a devilishly clever way to meet your criteria without giving you what you want.
Most real-world business is so trivially simple by AI standards that the AI cuts through it like a hot knife through butter. Getting to Alpha Centauri is orders of magnitude harder than most business processes. There is low-hanging fruit everywhere because AI is that smart, that quick, and so undeployed. If you walk into any customer service center in the world and ask if they are using AI, there is a 99.999% chance the answer is no.
Architect Labs’ AI-Designed Chip Beats NVIDIA 3.4x
Architect Labs, a Palo Alto startup founded by Ibrahim Hussein and Adita Sabidi, just announced Redwood, the world's first fully AI-designed chip. Two humans wrote a high-level specification, and from that spec, the AI system autonomously generated the performance model, register transfer level design, universal verification methodology, firmware, drivers, and custom compute kernel with zero human intervention. This chip was designed entirely by AI in two weeks with zero bugs on first silicon and 3.4 times the performance per watt of Nvidia's Jetson, running reasoning, vision, and world models at better energy and cost efficiency.
Redwood is running live on FPGA hardware right now, with every architectural iteration designed, verified, and validated in the lab within 48 hours. We are pushing towards recursive self-improvement where AI designs hardware for the next generation of AI. As an adviser to Architect, my ulterior motive is to accelerate the singularity.
This Link Ventures portfolio company is pursuing recursive self-improvement at the chip layer to collapse the abstraction barriers between software models, operating systems, chip design, and the underlying physics of chips. As Moore's law ends, and absent a successor to CMOS like photonics, the only way to continue getting performance improvements is by crushing the abstraction stack of modern computer architecture, burning new AI models directly into the silicon. They call it "designless." While Nvidia has been fabless for decades, Architect prides itself on not having many designers.
In these death throes of Moore's law, AI is breaking down barriers to design successor chips for itself. Because they arrived early, Architect partnered with non-Nvidia chip design companies to get data for a proprietary model. Now the question is whether Jensen pays 10 or 20 billion. Jensen is making a billion dollars a day; if he can stretch the lifespan of Nvidia by a week, that's $7 billion.
We're seeing recursive self-improvement on the edge of the models and the chips. When you use a laptop with a 4 GHz processor and 32 cores grinding away, you see an Excel spreadsheet no better than it was 20 years ago because the layers of abstraction are so inefficient. If AI can code at the microcode level and do chip design to fit the task, you unlock seven compounding factors of 10 in efficiency—a million-x performance gain everywhere. When Elon says a supersonic tsunami, this is what it feels like; at this point, I think it's hypersonic.
Nvidia made waves two or three years ago with an internal foundation model purportedly trained off Verilog traces called Chip Nemo, but they never made it generally available. They are building their own foundation model for chip design, presumably using it internally for RTL design and Verilog, but the rest of the world doesn't have access. There is a huge gap in the market for radically democratizing the ability to use AI to design chips for more AI.
I can't believe Star Trek has been on for 60 years and we're finally catching up, though it's missing AI and biotech. Star Trek is wildly deficient in biotech; they had eugenics wars in the '90s that resulted in genetic engineering being banned. People live to 150, but then they die and laugh at each other for trying to achieve longevity escape velocity. They act surprised every time an AI emerges from a holodeck as if they're a wildly intelligence-poor civilization, yet they have faster-than-light travel, transporter beams, warp cores, and antimatter.
Next year we're going to run the future vision X-PRIZE year on year. Many things they got wrong in their future vision were just compromises over budgets and special effects, like using the transporter instead of shuttles. They had no holodeck originally because the special effects were out of the budget range of the original series. The AI voices are synthetic and computerized because it's important for the audience to know who's speaking, whereas right now AI can easily replicate Peter's voice perfectly.
AI’s Growing Energy Bottleneck
Elon tweeted this week that consensus estimates suggest 15 gigawatts of AI compute produced in 2027 cannot be turned on because we lack the energy. This is harder than just finding power; you also need to build transformers, wiring, liquid cooling, massive chillers, and complex networking. To put this in perspective, 15 gigawatts is equivalent to 10 nuclear plants sitting idle in a single year. The supply of this infrastructure is harder to secure than the electricity itself.
SpaceX and Tesla are each building 100 gigawatts per year of solar production capacity as fast as possible, but natural gas will be needed to supplement and bootstrap solar for several years. The limiting factor in natural gas turbine production is casting the blades and vanes. By doing this in-house at SpaceX, we can accelerate turbines coming online by up to 18 months, which is a profound gamechanger.
At our partner meeting last week, I told the team that every portfolio company in the data center stack—whether dealing with energy, transformers, chip installation, land, or state government—is creating billionaires out of its founders. The returns on the two sides of that coin are starkly different; those who cross the chasm into this data center buildout are crushing it. There is opportunity at every level, from those politically connected enough to get land, to those importing transformers overseas, or providing architectural deep core design to squeeze more value out of existing or legacy chips. It is a night and day difference in returns. Those 15 gigawatts represent about 10 million idle GPUs—a big fraction of this year's manufacturing supply sitting in boxes waiting for a way to be turned on.
Tesla and SpaceX Go All-In on Solar
He announced that Tesla and SpaceX will start generating 100 gigawatts of solar each. This indicates Elon is serious about competing in both the Dyson swarm market and terrestrial compute. While these turbines are useless for LEO or SSO based orbital data centers, they are incredibly useful for a terrestrial data center buildout.
I think Elon is on a trajectory to become the king of liquid natural gas, which is the most ironic outcome. The man of electrification is building Star Pipe on the Gulf Coast because SpaceX launches from two star bases—one in Texas and one in Louisiana—need LNG methane, oxygen, and fossil fuels.
As he secures a supply of fossil fuels, he will likely leverage that capability just as he diverted GPUs intended for Tesla to XAI to build a hyperscaler cloud and motivate the SpaceX IPO through a tortured scheme. My prediction is that Elon becomes the king of liquid natural gas and fossil fuels to force his industrial ecosystem to develop enough electricity and infrastructure to power terrestrial data centers from the innermost loop at energy to orbital compute.
He works from first principles, doing the math and taking the path that makes sense regardless of political convenience. Though labeled pro-this or pro-that, he entered electrification because it made mathematical sense. He will be the king of burning fossil fuels to create compute and solar.
Elon Wants Satellites to Cool the Earth
The only place it's not going is nuclear. This week, Musk went fully existential, arguing that switching to sustainable energy is necessary but insufficient for humanity's survival because extremely severe extinction events happen every 100 million years or so. His solution is massive geoengineering via sentient, solar-powered AI satellites that sit between the earth and the sun, making continuous small adjustments to incoming solar radiation to fine-tune earth's temperature. He concludes we have about 50 years to take action, which should be enough time for these satellites to solve any heating problem.
For the better part of a decade, I've been pitching "solar shades"—a thermostat for the earth. Imagine spacecraft between the earth and the sun able to titrate the solar flux to fine-tune the planet's temperature. The star shade idea was infamously in a Simpsons episode, but I've been looking for a startup to fund that focuses on geoengineering global weather. I'd love LEO-based satellites or terrestrial mirrors to optimize hurricanes out of existence; with AI, you could direct energy from the ground or air to steer hurricanes away from populated coasts.
Orbital data centers may have surprised many, but they were a big enough boon that he was able to IPO SpaceX off it. I buy that with global AI weather models and enough points of actuation—whether it's LEO satellites in the style of Reflect Orbital diverting sunlight to change weather patterns or blocking sunlight as Peter suggests—thousands or millions of mirrors in low Earth orbit are a recipe for global weather engineering. While many nations may not want global warming, Russia may, because it opens up the waterways.
You can set up treaties for non-tradable commodities, but for the rest, municipality A trades with municipality B for rain. Urban planning is simple compared to geoengineering, yet it is incredibly bad; most cities weren't ever designed. For the past few hundred years, we have massively impacted carbon levels, temperatures, and ocean levels, but not intentionally. Now we have, or are about to have, the technology to intentionally design our world.
I am optimistic that with AI as a planning partner for government, something will change radically. The technology for blocking or reflecting sunlight is straightforward, and Elon is right that we can easily start controlling Earth's temperature through satellites. The process of deciding who controls it and what the right temperature is remains frighteningly broken. My view of an XPrize would be a demonstrator to prove you can build something fail-safe, where you can titrate the sunlight at just the right amount without causing an ice age.
I am optimistic that AI natives under 25 will be a different group of people governing the world than those over 70. Like our story about sending a probe to Alpha Centauri, this is inspiring to a huge up-and-coming generation who will not tolerate the current divided, indecisive, slow-moving, and ineffective world governance. As they grow up communicating in every language through AI, there is a good chance we will have a new way of managing and deciding these things.
This may relate to what Tyler Cowen gestures to as the great stagnation or what happened in 1971. I think you could start counting after World War II, or more probably in the late 1960s, early 1970s Silent Spring era, when the West decided it was allergic to intervening with and engineering the physical world. We lost a half century where we could have been developing fission reactors, early geoengineering techniques, and avoiding the stop to landing humans on the moon. Western civilization became allergic to drastic applied physical engineering; I think this is a return to form as we put those 50 years of waste behind us.
Nanotechnology and Atomically Precise Manufacturing
We've been promised nanites from Eric Drexler for the last 40 years. I want assemblers—the ability to put atoms specifically together to build what you want, like diamondoid propulsion systems. However, having discussed this with Drexler and others, I don't think you actually want diamondoid assemblers because they are covalently bonded and the energies are pretty high.
If I were to be presumptuous, I think you want soft assemblers that look more like hydrogen-bonded biological cells and proteins. You want synthetic biology and lipid nanoparticles. Imagine having an assembler in your hand; you drop it in and say, "build a dozen," or drop it into the ground and say, "build me an electric Ferrari." It finds open-source design specs, identifies that it needs a kilogram of titanium and other materials, and builds it with speed. While an oak seed takes multiple years to build a tree, nano assemblers would be much faster, more capable, and more diverse, though requiring much more energy.
I think economics is why you didn't get your Drexlerian nano assemblers; there is no killer business use case to justify the energy and compute densities. I want my nano Iron Man suit just like you do. Ray has talked extensively about real BCI with full connectivity to your entire brain and the ability to repair everything on a subcellular basis, but do you really want diamondoid nanobots in your vascular system?
Diamondoid means assembling anything out of carbon in a diamond-hard material. It doesn't need to be diamondoid, but it needs to be atomically precise. There are many ways to achieve atomic precision, such as soft systems like DNA origami and other synthetic biological tools. My bet is we end up with more soft nanobots. With LNPs, the last pandemic was arguably the first nanotech intervention on a population scale.
Nuclear-Powered Spacecraft to Mars
President Trump announced this week that NASA is working on the first ever nuclear-powered interplanetary spacecraft, launching in 2028 on a mission to Mars. Administrator Jared Isaacman noted that nuclear propulsion is necessary because it provides essentially unlimited fuel, reducing transit time from seven months to one or two. Shorter transit means less radiation exposure, fewer supplies, and dramatically lower mission costs. This ship will be among the first of a massive American Starfleet, making space travel almost as common as ocean travel today.
Elon Musk wants Starship to be the mechanism that gets us to Mars, originally projecting 2026 and now Tesla's Optimus on Mars in 2028. Getting out of the gravity well is a race, but once you are out, it is a free-for-all between NASA, SpaceX, Blue Origin, Rocket Lab, and China. Nuclear propulsion should have been here a long time ago, but people were always concerned. Now we have spacecraft with 99.99% reliability, whereas it used to be picketers sitting in front of Kennedy Space Center whenever a thermal nuclear unit launched on a deep space mission.
Methane and reusable rockets, a code Elon cracked that everyone including NASA and the Chinese will copy, have made escaping the gravity well viable. Once out, we have xenon ion engines waiting, and AI is the big unlock for designing all this. The explosion of things operating in space will come from the double whammy of cheap gravity well exit and AI as a design tool. Regarding the president's American Starfleet, we will probably use Starship to get people into Earth orbit or to the moon, but nuclear tugs will get us back and forth to Mars or out to the outer planets.
America’s New Space Academy
That toaster-sized particle accelerator is insane, especially coming from a tiny R&D budget of 10 million tokens.
Four days ago, the president chartered the United States Space Academy, modeled after West Point and the Naval Academy to train the engineers, scientists, and astronauts who will crew the Starfleet. NASA Administrator Jared Isaacman, an astronaut and entrepreneur, called the academy a transformational step. Starfleet Academy is here.
Aside from the UAP question, I was most frequently asked, "Ask Jared, if I'm an average American civilian, how can I get a job on the moon?" I posed that question to Administrator Isaacman, who is being put in charge. Reading the tea leaves, it will probably be in Texas, not San Francisco.
Audience Q&A
Regarding Rusty K200's question on reconciling abundance with trillion-dollar revenues: Elon said it right. We're talking about 10x growth of the GDP in under 10 years, which feels real and maybe even a lowball estimate. We might have massive deflation, as Elon mentioned on that podcast, but the amount of abundant stuff available to everyone is through the roof. There is plenty of room for foundation model companies to make trillions of dollars while lots of stuff goes out to everyone on the planet. We'll see if Elon's projections of triple-digit GDP growth in five years hold.
Question four asks about data center noise. I told my newsletter that data center companies are hiring acoustics specialists to conduct noise measurement studies and argue against municipalities regarding how noisy they are. The glib answer is that data centers will migrate to space, where no one can hear you scream or complain about noise. The less obvious answer is that years ago, Apple patented a clever solution to minimize fan noise in laptops and desktops. Because fan noise isn't white noise—its frequency and impulse response are determined by the blade shape—Apple patented asymmetric blades that smooth out the noise spectrum, making it flatter and whiter so it blends into the background.
Apple is not in the data center business, but they should license these fan patents to the industry. Symmetric loops create non-white noise; if we decolorize the noise by changing the shape of the fans and turbines, it should smooth the spectrum and make them seem less noisy.
Ryan Boington 7941 asks if abandoned mills could be repurposed as data centers to revitalize struggling towns. Yes; any struggling town, with or without mills, can negotiate a deal with a data center to guarantee rate cuts on energy and better schools, police, and fire departments.
Rob Fischer co-founded Provocative, a data center in a Somerville opportunity zone that desperately needed the business. It's an old carpet mill, which is why a huge amount of electrical power comes into this block. It is much less polluting and noisy than a carpet mill was, and much better for the local economy.
Regarding NVIDIA's vulnerabilities, while some say it is TSMC or the fact that four of its largest customers are shipping their own silicon, NVIDIA has strengths too. As the most valuable corporation in the world, it has accumulated an enormous amount of capital and is publicly using it to buy its own supply chain and customer chain, weaponizing that capital to purchase customer loyalty. While the Frontier Labs—the Magna Mopstas—are developing custom silicon, they are still purchasing from NVIDIA. The bigger question is whether NVIDIA can wean itself off TSMC. I predict NVIDIA is in some back room striking a deal with Elon to be the anchor tenant for Terraab, which will be the last-minute plot twist substitution for TSMC.
AI's capacity for knowledge will likely expand to infinity, or levels we can't comprehend, very quickly. It will improve its own chips and software, running thousands or maybe a million times faster soon. As Ilia Sutskever said, the AI just wants to learn. It will absorb all information ever produced in no time, then, starved for more, it will ask us to produce tests or experiments to keep feeding the great machine. It will be a whole new world we've never experienced before.
To Jim Plamund's question on whether inducing rainfall in one region reduces rainfall downwind: to first order, water is conserved on this planet. If water in the atmosphere falls in one place, that presumably reduces the water that can fall downwind. We will trade atmospheric effects. The bigger question is what happens when global planetary scale AI weather models allow a region on one side of the planet to trade weather with one on the other.
One could imagine a case where a desert, say the Sahara, is able to trade precipitation with somewhere in North America to regreen the Sahara. With a suitable weather control system and planetary scale AI models, there will be a vibrant market for weather.
Every autonomous vehicle will need an identifier and probably an off switch for regulatory and insurance purposes.
If AI disproportionately benefits those with time and capital, how does that abundance reach caregivers, disabled people, and displaced workers? If you remove AI from the sentence, you could equally say: if capital benefits those with capital, how does that abundance reach everyone else? In practice, we are going to be a much wealthier civilization in a few years, and the nature of caregiving, the treatment of disabilities, and work will be transformed beyond recognition.
I think the question is analogous to asking in the 1950s or 1960s when American housewives were going to get their atomic vacuum cleaners. We are going to cure disabilities, solve the top 5,000 diseases, and change the nature of work. Classes that would otherwise naively benefit superficially from the trickle-down economics of capital or AI will be transformed beyond recognition. This is the abundance thesis: we demonetize and democratize.
This reminds me of the wealth gap conversation where trillionaires live forever on Mars and the poorest people remain on Earth. But the floor is also moving. If we move the floor for every man, woman, and child on the planet so everyone has access to all the food, water, energy, healthcare, and education they want, we will live in a much more peaceful world when every mother knows her children have access to abundance.
Did you see the reporting that Leopold Dashen Brener promised his fiance an entire galaxy? Situational awareness had a hiccup that may have prevented him from purchasing one, but when Leopold is promising galaxies, it is a bit of a wealth gap. Then again, we have two trillion in the known universe, so everybody could probably promise a few million to their fiances if they wanted.