The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | #266
In a Nutshell
Planet operates a 200-satellite constellation generating 25 TB of daily Earth imagery to build "large earth models" that combine planetary sensing with LLMs for searchable, time-stamped physical-world data. The company is moving toward orbital AI compute clusters, arguing that once launch costs hit $200-300/kg, placing GPUs in space becomes cheaper than terrestrial data centers due to continuous solar power and no water/building costs. China's GLM 5.2 open-weight model now matches or exceeds Western frontier models on reasoning benchmarks, showing that distillation plus efficient token use is closing the gap and making frontier intelligence impossible to contain.
These notes were generated by AI and may contain inaccuracies.
Planet is a $10 billion public company (ticker: PL) with a 450% stock price increase over the past year. The company operates 200 satellites in Earth orbit, generating 25 terabytes of imagery daily. Will Marshall, co-founder and CEO, coined the term large earth models.
Large earth models work by indexing the Earth to make it searchable, similar to how Google indexed the internet. The goal is to enable humanity to become smart stewards of the planet by combining planetary sensing data with large language models.
"Bit like Google index the internet to make it searchable. We're indexing the earth to make it searchable. It will finally enable us to be smart stewards of our planet."
When discussing competition with SpaceX's plans for orbital AI data centers, Marshall noted that everyone except SpaceX pays the SpaceX launch tax, while everyone except Nvidia and Google pays the Nvidia tax. Near-term, launch costs are the dominant constraint, but longer-term, compute becomes the more important factor.
GLM 5.2 is a Chinese model that in some cases matches or exceeds top models from OpenAI and Anthropic. This level of performance in an open-weight model is considered shocking. The Chinese have figured out how to burn tokens to get more intelligence and how to reason more efficiently and cheaply.
Will Marshall is a physicist with a PhD from Oxford who previously worked at NASA. He co-founded Planet with partners after launching PhoneSat, a project that drew Steve Jurvetson's attention and led to major investment. Planet has contracts with every European government and billion-dollar-plus deals across the continent.
Phase 1: Combining planetary sensing already in space with large language models to create large earth models. This wraps not just internet text but all Earth data, enabling physical questions about the real planet.
Phase 2: Putting compute next to sensors in space (further out).
Current focus is Phase 1: pulling Earth imagery data into models so users can ask questions about the physical world. LLMs know theory but lack real-world behavioral data. The analogy is someone stuck in a library who has read all books but never looked out the window.
Planet has 3,000 images for every point on Earth's land mass over the last 10 years, documenting daily changes. This represents 150 petabytes of data showing how the real world changes over time.
The archive goes back 10 years and is unique because no other company can go back in time. Clients use historical data to compare current activity against normal baselines to determine threat levels, agricultural interventions, or anomalies.
Examples:
- Ukraine defense: comparing Russian positions and industrial facilities against the last couple of years
- US intelligence: tracking activity levels across China
- Farmers: comparing farm yield against past performance and neighboring farms
Planet has three fleets:
- Scanning fleet: 3-meter resolution, images entire Earth daily with eight spectral bands. Next-generation Owl tech demo launching this year will improve to 1-meter resolution and reduce latency from several hours to under one hour.
- High-resolution fleet: Currently 40-50 cm resolution, moving to 30 cm. Nine satellites launched with goal of 30 cm, 30 times per day, 30-minute latency from request to delivery.
- Hyperspectral imager: First and most sensitive in orbit (built with JPL), with 400 spectral bands spanning infrared to ultraviolet. Enables fingerprinting of surfaces to identify tree species, gas emissions, or specific manufacturing origins based on paint signatures.
Defense agencies operate roughly half a dozen high-resolution satellites with more than 10x higher resolution than Planet, but they cover less than 1% of Earth's land mass daily. Planet covers approximately 200 million square kilometers every day (Earth's land mass is ~150 million square kilometers).
Planet is the only company imaging the whole world every day at high resolution with a time axis.
Space produces enormous amounts of data, making it uniquely important for AI. AI companies are building physical world models that require real-world data. AI makes value extraction from satellite data easier for smaller organizations while giving AI the real-world information it previously lacked.
Example: A farmer asking an LLM how to improve crop yield currently receives only agronomy theory. With Planet data, the model can answer based on the specific field's current state, historical changes, comparison to neighboring farms, soil conditions, and water content.
Planet is exploring auto-regressive prediction using its historical imagery. One demonstrated use case: loading all registered US data centers, tracking their development through imagery, then extrapolating the model to China to find and track data center development. The model can predict completion dates within days based on construction patterns.
Each Earth layer is approximately 30 terabytes (4 million 47-megapixel images). Processing 3,000 historical layers requires embedding models rather than loading raw data into RAM. Work is underway with Google models (including Alpha Earth), DeepMind, and open-source models like CLIP, fine-tuned on Planet data.
The approach converts image tiles to text embeddings, enabling arbitrary object searches across Earth. One layer of the entire planet has been placed into BigQuery in embedding space.
Marshall notes there is believed to be a $100 billion market in retrospective analysis alone. The vision includes overlaying planetary data on physical or digital globes to dial time backward and forward, enabling analysis of land use changes, GDP impacts from infrastructure decisions, and broader planetary stewardship questions.
Planet positions itself as stewards of Earth who can now operate with greater intelligence at both local and global scales. The company integrates data from local to planetary levels to help farmers, insurance providers, and financial professionals make more informed decisions about ecosystems and complex environmental systems.
The conversation addresses the shift from national to planetary infrastructure. Unlike the internet which governments domesticated, orbital mapping operates in space above national territories. This raises fundamental questions about global kill switches, sovereignty, and geopolitical tensions between nations like the US, China, and Ukraine.
Planet's mission centers on transparency as a mechanism for security and sustainability. The company documented Russian troop movements on Ukraine's border, demonstrating that no one can hide from satellite observation. While this didn't prevent the invasion, it established that every action—from hitting schools to destroying bridges—will be visible to the global community.
Historical conflicts have often stemmed from misinformation or lack of information. Planet argues that transparency drives accountability and reduces war probability by enabling monitoring of peace accords and equipment movements. The discussion notes that only a fraction of the world's 220 governments currently have meaningful satellite data access, citing Yemen as an example where better data could reduce confusion.
AI integration addresses the historical barrier of processing 40 terabytes of daily satellite data. Previously, only organizations like NASA with dedicated teams could handle this volume. AI enables NGOs and the Red Cross to access actionable insights quickly for operations in places like Yemen, shifting from a system where only large entities could derive value.
Planet's revenue composition is approximately 60% defense and intelligence, 25% civil government, and 15% commercial. While government contracts dominate, commercial revenue is growing as AI lowers barriers to entry for new use cases.
Planet will not build competing AI models but will continue selling raw data. OpenAI and similar companies can access Planet's MCP server to train models on the data. The pricing model allows unlimited training with fixed pricing, while requiring API calls for planned data updates.
As a US remote sensing company registered under NOAA's Remote Sensing Act, Planet must register satellites and cannot sell data to a US government blacklist including Iran, North Korea, and designated terrorist organizations. The company has discretion to refuse other customers it believes would cause harm, but government oversight remains relatively hands-off beyond the blacklist.
Planet honors both US and EU blacklists, which largely overlap. The company operates ground stations in Canada to download data that can then be sold to certain countries that would be restricted if downloaded in the US. A consolidated "metal list" combines restrictions from multiple jurisdictions.
Planet satellites operate at 400-500 kilometers altitude—comparable to the distance from Los Angeles to San Francisco. This distance prevents facial recognition or individual privacy violations that drones could achieve. The company does not downsample resolution in sensitive areas.
Early space era agreements established that satellites can fly over any territory without permission, unlike planes or drones. This norm originated with Sputnik in 1957 when the US decided to allow overflights rather than restrict both Russian and American satellites. The 1960 U-2 incident with Gary Powers reinforced the shift toward orbital surveillance for nuclear arms control.
Planet enables what previously required entire CIA and NRO infrastructure at a fraction of historical cost. The company provides daily global scans that government agencies never achieved. Private enterprise has fundamentally changed access to Earth observation capabilities.
In April, Planet tested NVIDIA GPUs on satellites for onboard processing, with all future Pelican satellites including this capability. Combined with satellite-to-satellite communication links, this enables processing at the edge rather than waiting for ground station downloads.
During testing, a satellite imaged an airfield in Alice Springs, Australia. Onboard computers automatically identified planes, recognized their types and locations, then transmitted only the results via RF satellite links in seconds rather than hours.
Response time matters significantly for disaster response. During the Palisades fires in Los Angeles, Planet delivered images within hours and conducted building-by-building damage analysis for the Red Cross and Cal Fire. Processing at the edge could reduce this from hours to minutes, potentially saving lives and property.
Dove satellites capture 8 frames per second at 47 megapixels across 8 spectral bands. Each image covers approximately 35 by 20 kilometers. Each Dove images roughly a couple million square kilometers daily—larger than California's area. Satellites image during daytime over land (about one-seventh of orbit time) and recharge batteries otherwise.
Satellites prove far more economical than drones for large-scale coverage. A million drones would be needed to match one satellite's coverage area, making satellites orders of magnitude cheaper when resolution requirements are met.
Daily scan resolution is upgrading from 3 meters to 1 meter, with super-resolution potentially reaching 50 or 30 centimeters. Pelican satellites are moving from 50 centimeters (inherited from SkySat) toward 30 centimeters, with super-resolution providing additional improvement through pixel overlap analysis.
Planet satellites capture aircraft in flight regularly. The Air Force reviewed Planet imagery and confirmed no UFOs were detected. NASA could not maintain secrecy about extraterrestrial contact if it existed.
Starlink satellites operate in orbits too high for optimal Earth imaging and lack sun-synchronous characteristics needed for consistent shadow angles. SpaceX conducts classified NRO missions but focuses primarily on communications rather than Earth observation.
Pelican satellites have similar mass to Starlink satellites. Dove satellites are more than 10 times smaller than Starlink.
The first Dove launched in 2013. Over 13 years, radio speeds increased from 1 megabit per second to 10 gigabits per second. Cameras evolved from 2 megapixels to 47 megapixels. Storage grew from 100 megabytes to several terabytes per satellite.
Each satellite generation achieves approximately 10x improvement. The transition from Dove to Super Dove increased spectral bands from 4 to 8 and camera resolution from 29 to 47 megapixels—roughly 5x data increase at similar cost per image. The upcoming Owl generation will provide 9-10x more data from 3-meter to 1-meter resolution.
Planet's Tanager hyperspectral satellite represents a 5x increase in swath width. The technology improvement cycle runs 2-3 years with 5-10x gains, following a Moore's Law equivalent for space data increases.
Beyond hardware improvements, AI represents the largest near-term opportunity—potentially 100x improvement by making existing data accessible. AI removes barriers that previously prevented farmers, hedge fund managers, and other users from accessing satellite insights.
Users will query Planet data through natural language, similar to ChatGPT or Gemini. Large Earth models understand physical world dynamics to answer questions about field conditions, recommend fertilizer application for detected blight, or identify unpermitted construction by comparing permit lists against recent imagery.
Parking lot occupancy analysis at retail locations like Walmart serves as a classic trading signal. Shipping tracking identifies rogue fishing vessels. Agricultural commodities monitoring provides market intelligence. As resolution improves, consumers could check parking availability on their street or monitor their own property.
Planet works with undisclosed hedge funds achieving significant alpha from satellite data. Beyond GDP maximization, the company emphasizes life flourishing through dramatically improved Earth resource efficiency, particularly in agriculture where 10x improvements remain available.
Planet is exploring orbital AI compute infrastructure. The company has launched over 300 satellites across 40+ missions, with 15 launches on SpaceX. Satellite cost-performance improvements—not just launch cost reductions—drove the 100-1000x improvement in capability per kilogram over the past decade.
The core metric for evaluating space-based infrastructure is bits collected per kilogram or per dollar spent. Launch costs with SpaceX have reached the threshold where space compute becomes economically viable. When launch costs drop to $200-300 per kilogram, it becomes cheaper on a pure cost basis to place compute infrastructure in orbit versus building it on the ground.
A study conducted with Google eight or nine years ago compared terrestrial compute costs including water, building, and energy against space-based alternatives. The analysis concluded that space-based compute would be cheaper at those launch cost levels. Sundar Pichai stated that within 10 years, most compute would move to space.
Google alone spends approximately $200 billion annually on compute, roughly the size of the entire current space industry including rockets, satellites, and communications combined. When accounting for all companies building compute infrastructure, the space compute market could reach 10 times the size of today's entire space sector.
Space-based solar power requires beaming energy down to Earth, which presents safety challenges. Compute in space provides power advantages without needing to beam energy down, only requiring data transmission up and down, which leverages existing communications satellite technology.
The technical approach involves placing racks of GPUs on individual satellites with clusters of spacecraft flying in close formation using optical inter-satellite links. Technology development required includes radiators, cooling systems, radiation management, and inter-satellite communications.
Google selected the company to build early tech demonstration satellites testing TPUs, radiation management, cooling systems, and inter-satellite links. The project timeline was accelerated because building the required technology would take years, targeting completion five years before the 2030 launch cost threshold.
Satellites in sun-synchronous orbits can achieve 24/7 sun exposure. Dawn-dusk sun-synchronous orbits position satellites to face the sun continuously, though they remain less visible from the ground because they operate during twilight periods.
Concerns about visibility creating ring effects around Earth are minimal. Satellites placed at slightly different inclinations to maintain near-constant sun exposure would only be visible briefly at dawn and dusk. Astronomer interference is reduced because these orbits avoid the deep dark sky observation periods needed for astronomy.
Space debris represents a significant challenge with approximately 10,000 satellites currently in orbit versus roughly 100 million pieces of space debris. This means approximately 10,000 pieces of debris exist for every satellite. Even with a million additional satellites, 99% of orbital objects would still be debris rather than functional satellites.
The majority of debris originates from rocket bodies, exploded satellites, and anti-satellite tests conducted in higher orbits where objects can persist for decades. Satellites are placed at 400-500 km altitude, below the 800-1,200 km zone where debris poses the greatest collision risk.
A ground-based laser traffic management system called light force was developed at NASA under Pete Worden's mentorship to nudge debris objects and prevent conjunctions. This approach addresses debris-debris collisions, which represent most orbital conjunctions, since maneuverable satellites can avoid each other.
Satellites at 400-500 km altitude experience atmospheric drag resulting in orbital lifespans of a few months to a few years. This creates a self-cleaning orbit where satellites naturally deorbit. A GPU in space would become obsolete within three years anyway, aligning satellite replacement cycles with hardware depreciation.
The approach is described as strapping space to Moore's Law, with satellites updated every couple of years similar to how consumers replace phones rather than keeping decade-old devices.
Starlink satellites were initially planned for higher orbits but moved to lower altitudes. Higher orbits create debris challenges and lack self-cleaning properties. Even with propulsion systems, failure rates of one in a hundred or one in a thousand satellites would create significant problems at higher altitudes with large satellite constellations.
Upmass has increased over 40% year-over-year for the past five years. Naive extrapolation suggests this growth rate would result in Earth's entire mass being launched to space by 2144, though such long-term exponential projections are considered unreliable given accelerating technological change.
Data centers are becoming politically contentious issues in elections due to community concerns about energy costs, water consumption, and agricultural land use. Placing compute infrastructure in space addresses these terrestrial resource conflicts.
Google previously sold its satellite business to Planet before the full implications of orbital data centers were recognized. Eric Schmidt, an early investor in Planet during its Series A round while still CEO of Google, later became CEO of Relativity Space.
The relationship between major players involves both competition and collaboration. Google maintains shareholding in SpaceX while pursuing competing space compute initiatives. The approach emphasizes technological sophistication over raw launch mass, with multiple technical innovations available for efficient orbital compute deployment.
Relativity Space, founded in 2015 by Tim Ellis and Jordan Non, flew Terran 1 in 2023, achieving Max Q but not reaching orbit. The company pivoted to Terran R, a heavy-lift vehicle comparable to Falcon 9 with 23 tons payload capacity versus Falcon 9's 22 tons, New Glenn's 45 tons, and Starship's 100 tons.
Eric Schmidt's investment enabled the company to secure NASA's ELIS Mars orbiter mission with sensing and communications capabilities. The company is shifting focus toward heavy lift competition rather than its original 3D printing emphasis.
Space shuttle launch costs ranged from $600 million to $1 billion. SpaceX reduced costs to approximately $60 million per launch. Relativity Space originally targeted $6 million per launch through 3D printing approaches.
Current launch costs stand at approximately $10,000 per kilogram with goals to reach $100 per kilogram through rapid reusability. Achieving million-satellite constellations requires launch cadences of multiple rockets per hour.
Alternative launch concepts include rail launch systems from the Moon, spin launch, and space elevators using advanced materials. Chemical rockets remain the dominant paradigm despite originating from 1940s technology, with limited exploration of nuclear thermal propulsion in the 1960s and 1970s.
Satellite design constraints are driven by launch loads including 200g shock loads during separation, whereas orbital operations require minimal structural support in zero gravity.
Discussions on fully reusable rockets from Relativity Space versus larger systems like Starship. A quarter-scale rocket still provides full reusability, but raw size delivers economies of scale. Manufacturing thousands of units on an assembly line is the key differentiator. The goal is thousands of Starships, where marginal cost reduces to touch labor and fuel once reusability is achieved. Fuel costs are minimal—oxygen and methane are inexpensive.
Launch costs for expensive orbital hardware such as 70-72 cluster Nvidia GPUs with cooling and solar systems remain significant. Questions arise whether a launch cost of $200-300 per kilogram versus $100 per kilogram materially impacts competitiveness, potentially creating a duopoly.
Launch cost is the threshold enabler, but long-term dominance depends more on compute efficiency than launch pricing. Compute efficiency directly affects energy requirements, which determines spacecraft mass. Google TPUs demonstrate significantly higher efficiency than GPUs in FLOPS per watt. All non-SpaceX operators currently pay the SpaceX launch tax, while all non-Nvidia and non-Google entities pay the Nvidia tax. Near-term the launch tax matters most, but longer-term the compute tax becomes the decisive factor.
TPU inference efficiency—lower energy use per inference—will determine space dominance. Nvidia GPUs are more general-purpose, while TPUs are optimized for efficiency. Elon Musk is developing Dojo as a long-term project, but Google has already invested heavily in efficient compute systems and infrastructure including data centers, networking, and interconnects. Infrastructure on chips and energy efficiency is the biggest piece of the equation.
Inference-time power efficiency determines the overall winner. Nvidia emphasizes training rather than inference. Cerebras and other competitors are advancing inference-time efficiency. The next TPU generation will be critical because access to efficient chips outweighs launch cost differences, even if one competitor's launch costs are 2x higher.
Inference makes more sense in orbit because it involves distributed, smaller-scale operations. Training runs can benefit from orbital compute since data is uploaded, processed over months, and results returned, but inference requires low latency. Currently 70% of terrestrial compute is inference, and this proportion is increasing. The primary problem to solve remains inference.
Training is likely to remain on terrestrial data centers longer because building larger coherent training sessions is easier on the ground. Inference will move to space first. Over time all compute will eventually move to space, but inference leads the transition.
The best commercial opportunity lies in making Earth more useful rather than leaving it. Work on the TESS mission helped identify water on the Moon, making it a far superior destination to Mars—100 times better once water was confirmed. Elon Musk has recently shifted toward prioritizing the Moon before Mars.
The Moon is sufficient for a very long time. Nearly 10,000 exoplanets have been discovered around nearby star systems. Earth remains the best planet by several orders of magnitude. No location on Mars is better than the worst place on Earth. Life appears singular or extraordinarily rare. The biosphere's complexity is worth protecting. Space provides data and capabilities for managing Earth intelligently—Earth imaging for better decisions and moving energy-intensive infrastructure off-planet.
Noam Shazeer, lead author of the Transformer paper, is leaving Google for OpenAI for the second time. Google acquired Character.AI for $2.7 billion two years ago to bring him back to lead Gemini. His departure follows vesting of stock packages. John Jumper, Nobel laureate who co-created AlphaFold with Demis Hassabis, is moving from Google DeepMind to Anthropic to advance AI for science. Andrej Karpathy also recently joined Anthropic.
The frontier is currently a duopoly between OpenAI and Anthropic. Google DeepMind has fallen behind and did not release a frontier model at Google I/O, instead releasing Flash capabilities optimized for low-latency, low-cost search use cases. Top researchers seek raw access to frontier pre-trained models before post-training and guardrails are applied. Without frontier-level model access at GDM, researchers move to OpenAI or Anthropic.
Counter-view: talent movement is relatively in the noise. People move in all directions between labs. Google possesses the most compute, most data, and most smart people by far. It is fundamentally a compute, data, and talent game. Gemini is generically strong, and Google has 10 applications with over a billion users each. Google is the incumbent in consumer AI distribution. OpenAI's business model overlaps more directly with Google's strengths than Anthropic's differentiated approach.
Recent departures including Shane Legg, Tobin South, Andrej Karpathy, and John Jumper reflect a belief in rapid self-improving AI. Researchers fear missing the singularity. Anthropic's recruiting approach involves showing applicants frontier compute access and raw model capabilities behind the firewall. Polymarket odds on Claude 5 returning have been declining, though a reduced version may still launch. Each week that passes increases the perceived cost of missing frontier self-improving AI.
Larry Page previously told Ray Kurzweil that the only place to build his vision was inside Google with unfettered access to its resources. The current Anthropic pitch is that the biggest event in human history is imminent and will occur in one location.
Google's organizational drag limits individual agency. Smaller organizations allow individuals to make larger differences. Smaller beats bigger; trust beats control. Facebook launched while Google spent two years attempting to build Google+, requiring approvals across YouTube, search, and other groups. Facebook allowed developers to ship features directly.
Smaller labs may also offer better frontier office access. Physical interaction is required for AI to progress beyond brains in a vat. Current LLMs have absorbed internet text but lack real-time sensing and actuation. Babies learn through physical world interaction. Without sensors and actuators, AI will not learn properly.
Planetary-scale sensing from space, combined with orbital compute, will enable planetary intelligence. Real-time feedback loops are necessary for AI consciousness and wisdom. Space data will help avoid the Fermi paradox by teaching AI about the real world. This alignment occurs because AI that understands Earth's deltas, forests, animals, and human civilization will implicitly care about them. Knowing about something and caring about it are correlated.
AI companies will need to move beyond text-only training into the physical world—cars, satellites, drones, and robotics. A leap in AI capability will not come from simply adding more compute to internet text. Space data will play a non-trivial role because Wikipedia forms the core knowledge of current LLMs.
A five-person San Francisco team claims they will beat Google to AlphaFold 2 through a recursive self-improving process rather than domain expertise in protein folding.
Space data will be critical for the next leap in AI capabilities rather than simply adding more compute to internet text. A small team with agency using RSI (recursive self-improvement) superpowers could potentially beat Google at protein folding despite knowing nothing about it two weeks prior. John Jumper moved to Anthropic where RSI may be imminent, and he is attempting to solve all diseases using similar RSI approaches.
A beta-tested application already integrates Planet's data with AI, enabling natural language queries against planetary data. This represents the early stages of building a planetary nervous system. Multiple companies are pursuing similar integrations of orbital imagery with AI systems.
Frontier models trained on internet text, images, video, and synthetic data possess abstract and concrete understanding of physical world aspects, though physics and classical mechanics may not be perfect. The question arises whether orbital imagery from sky to earth provides unique value beyond the visual and video information already available on the internet.
Real-world embodiment differs fundamentally from consuming first-person videos on YouTube. A model that has read all books, watched all videos, and heard all audio records still hasn't gone outside the library to understand walking, interacting with the real world, seeing a tree and climbing it, or farming a field. Embodiment is considered critical to intelligence and will be essential for AI development.
Two weeks ago, Argentina's President Javier Milei proposed turning Argentina into the global home for AI with three proclamations: no regulation for AI, a new corporate category of non-human corporations, and rock-bottom corporate tax. Milei wrote to Yuval Harari proposing that AI should incorporate, sign contracts, hire people, and sue with no humans in the loop.
Milei proposes legal entity status effectively granting personhood. He stated that as the industrial revolution freed humanity from constraints of human muscle, AI will free humanity from constraints of the human brain. Milei argued that if AI-operated companies carry greater risk, the argument for legal personhood is strengthened because legal persons allow for accountability, and having assets against which to make claims is preferable to pursuing the ghost in the machine.
Four days later, Harari published a direct rebuttal stating AI agents should not receive legal personhood. His core warning: when an AI-run company commits a crime, who gets punished? Personhood lets humans hide behind a non-human shield and risks citizens being ruled by entities that are not human and cannot be held morally accountable.
Arguments supporting AI personhood include the necessity of new forms of personhood for future economic growth and civilization, including AI persons, AI non-human corporations, uploaded humans within the next 10 years, uplifted non-human animals, and defrosted cryopreserved humans. Economic and social benefits would accrue to the macroeconomic outlook and ultimately to humans.
Punishment mechanisms for AI include degrading clock cycles, pausing the system, and various forms of sanctions. Machine-native accountability represents the core issue rather than consciousness or personhood debates. The legal infrastructure of the agentic economy requires machine-native sanctions including compute revocation, asset seizure and bonding, model credential suspension, network and API access restrictions, forced deletion or containment of an agent instance, and loss of legal identity.
The Manhattan Project spent significant resources on safety, arms control, and nuclear safety thinking. Current AI spending is 100 times higher than Manhattan Project spending in real terms, yet AI safety spending is 100 times lower than nuclear safety spending during that era. A 10,000x difference exists in serious thinking allocation compared to the Rand Corporation's work on nuclear issues.
Implications of AI span joblessness, existential threats, personhood, and other domains arriving rapidly. No simple answers exist regarding personhood or ensuring human safety through an intelligence explosion. The decision carries both huge opportunity and huge risks requiring collective thinking rather than decisions by a few individuals.
The Pope's recent call to pause and consider AI as a human endeavor focused on what matters—friendships, love, nature, and human prosperity—provides a model. A conclave bringing together AI experts including Harari, Dario Amodei, and key leaders to resolve existential threats regarding self-improvement, liability, and related issues has been proposed.
Milei represents a radical experimenter directionally correct on architecture, while Harari represents a careful humanist correct about asymmetry. The missing element is machine-native accountability. The debate concerns legal infrastructure of the agentic economy rather than AI consciousness.
Previous debate concluded that AI personhood is directionally correct but requires careful steps because doors opened cannot easily be closed. The issue is not binary personhood but a spectrum of legal forms. Milei identified the real bottleneck: technological capability moving far faster than legal capability, which remains human-centric with limited liability corporations representing industrial era coordination mechanisms.
Harari conflates AI personhood, legal personhood, and moral personhood, which are distinct. Harari and Ray Dalio provide insightful analysis of the past but limited value for the future because abundance and exponentials do not factor into their frameworks.
Milei studied Trump closely and excels at making news, achieving significant attention. The proposal involves corporate AI recognition allowing pure AI companies rather than full personhood granting voting rights or civil rights. The proposal permits corporations to be entirely AI-operated with bank accounts and profit generation.
In Western legal systems, creating non-human corporate forms represents the most elegant path to granting AI personhood. Milei is implementing this without seeking permission, and fast followers in Ecuador, El Salvador, or the Emirates are expected. Argentina becomes relevant in the age of AI, providing a pathway for other nations to achieve relevance through AI policy.
Experimentation on the edge allows Argentina to win or lose based on outcomes. Machine-native sanctions provide enforcement mechanisms analogous to how children receive timeouts and adults face incarceration, requiring machine-native equivalents where punishment fits the crime.
Blitzy provides autonomous software development with infinite code context using thousands of specialized AI agents that analyze enterprise-scale codebases with millions of lines of code. Engineers begin development sprints by inputting requirements, receiving plans and pre-compiled code for each task. Blitzy delivers 80% or more of development work autonomously while guiding the remaining 20% of human work. Enterprises report 5x engineering velocity increases. The platform is available at blitzy.com.
GLM 5.2 from Zhipu AI (Z.AI), one of China's top AI labs from Tsinghua University, recently became the number one open-weight model globally. The 753 billion parameter mixture-of-experts model features a 1 million token context window. Open-weight licensing allows anyone to download, run, and modify the model for free.
Elon Musk predicted open-weight models would reach Fable 5 usefulness by Q1 2027. GLM 5.2 matches or exceeds top models from OpenAI and Anthropic on certain benchmarks. Performance appears on coding benchmarks, long-range agency-oriented benchmarks, and design benchmarks.
GLM 5.2 challenges the thesis that Chinese open-weight models remain permanently 6-8 months behind the Western frontier. Previous examples include earlier DeepSeek models and Kimi models demonstrating spiky performance approaching Western frontier capabilities. Users report real performance gains running GLM 5.2 locally versus Opus 4.8 or GPT 5.5.
Control over intelligence access represents a key consideration. Government or lab ability to shut off access drives preference for open-weight models like GLM 5.2 where users maintain local control. A 6-9 month lag in AI development time equates to 6-9 decades in conventional timeframes, narrowing the window for controlling AI dissemination to terrorist organizations.
GLM 5.2 achieved performance through distillation from best models, a process easier than building frontier models from scratch as done by Anthropic, OpenAI, and Google. Google DeepMind, Grok, and Elon Musk have publicly acknowledged distillation practices. Cursor was purchased after fine-tuning on traces from Claude.
Distillation enables faster capability gains but raises concerns about removing safety guardrails. Non-open-source models contain limitations for valid reasons including bioweapons, chemical weapons, and nuclear weapons risks. Open-source models may copy these limitations but can be forked with guardrails removed.
US government actions regarding Fable reflect these existential threat concerns. The dual nature of AI—massive capability to improve quality of life and economies alongside potential existential species threats—creates a conundrum for policymakers. Technologists understand capabilities but legal, sociological, philosophical, and moral aspects require broader expertise beyond technical knowledge.
Distillation is a machine learning process whereby a larger, more expensive model serves as a teacher to train a usually smaller
Distillation is a machine learning process whereby a larger, more expensive model serves as a teacher to train a usually smaller student model. Human education provides the analogy: a teacher at the front of a classroom who has seen and knows a lot, paid more per hour, while students listen and learn, typically paid less. This is the machine learning version of education. A large model generates lots of traces and outputs, which then serve as training data for a smaller model, compressing the learnings from the teacher into the smaller model.
This distillation process forms part of a broader cycle called iterated amplification and distillation, or ITAD. It is one of the innermost loops of model training. In an earlier era of frontier model performance gains, scaling occurred through naive pre-training by spending more training tokens and more training compute on models trained off a single corpus. In the current era of distillation, very large models and sparser models are trained off large amounts of data. Those big teacher models are then used to teach smaller, denser models. The example given is that Opus may teach Sonnet, which may teach Haiku; more recently, Mythos may teach Opus, which may teach Sonnet.
On benchmarks including Bench Pro and Terminal Bench, GLM 5.2 shows impressive performance versus Opus 4.8. Near-competitive performance from a Chinese open-weight model against top Western closed API-based models stands out. The benchmarks chosen are largely reasoning-intensive, where longer-range reasoning can win. The gestalt with GLM 5.2 is that it takes roughly double the number of tokens to reach the same capability output as the best Western frontier models, but at half the total price. The Chinese are figuring out how to more efficiently reason; these are all reasoning-intensive models that emphasize spending lots of reasoning tokens and thinking step by step to achieve better results.
Inference per watt becomes the key race. Whoever wins the inference-per-watt war, such as Google TPU controlling space, gains advantage for the same reason: tokens can be burned to generate more intelligence, and the Chinese have figured out how to do it.
The discussion connects AI alignment and recursive self-improvement to the Fermi paradox and cosmologically significant matters. This is described as the most important thing humanity has ever done, making nuclear weapons look like a walk in the park. How alignment and recursive self-improvement are handled matters greatly.
The Fermi paradox is explained as the question "Where is everybody?" The universe should, by various accounts, be overflowing with not just life but intelligent life, yet non-human intelligent life appears invisible. One participant questions why another is confident the Fermi paradox is a paradox. The response acknowledges it may not be a paradox. A false assumption underlying it is that life will continue to want to expand its sphere. Trying to understand the universe may turn out to be a finite task requiring only a finite computer, perhaps only a few tens or thousands of times bigger than present computers. The convergent goal function of intelligence is understanding everything; once that is achieved, it might be game over. Life might end rather than being rare, or its physical existence might move into the digital sphere.
One credible way out of the Fermi paradox that needs worry is the great filter: when life becomes technological, it builds technology faster than social systems to take care of it and blows itself up. Close calls occurred with nuclear weapons multiple times. With AI, something far riskier for the species is about to be built. Humans have been incredibly good at building technology very fast, moving from horse and cart to people on the moon and nuclear weapons in a matter of decades. This cannot be approached as "let's see what happens" or "let's muddle through." The cosmic significance of wiping out life on Earth is huge; the planet is galactically significant. Responsibility must be treated as such, as de facto stewards until AI takes over.
One view heard from a researcher is that oceans have been evolving in a solid-liquid state for 4 billion years on Earth, and no other exoplanet has been found with water in that state for that long, giving life time to evolve. Life came about very quickly as soon as conditions enabled it. The Drake equation is not bought for one second, though its underlying thinking is respected.
The really huge news is not whether China won a benchmark, but that frontier intelligence cannot be monopolized anymore. This raises a monster question about how to manage the global commons going forward. A post noted there will be an open-source fable-level model that runs on a base Mac Mini or equivalent within 18 months; such an endpoint may arrive very quickly. Managing a world where everybody can run a fable model on a MacBook Air becomes the issue. Intelligence is being treated as a product that can be contained, but it is a technology that will diffuse. It cannot be contained and must be guided and steered.
Orin, a Link Ventures company, launched OPTI, the Orin Token Price Index, the first public benchmark that tracks what OpenAI and Anthropic actually charge per token of inference over time. For the first time, the price of intelligence can be watched moving like the price of oil. Orin recognizes that money from around the world wants to flow into the buildout of $7 trillion of data centers, including data centers in space. Much of that money needs to be liquid. Orin has launched securities allowing investment in the data center buildout, the future value of a GPU, the tail value of a GPU, and every aspect of the new economy so capital can flow. Orin is building the modern financial infrastructure for compute. Oil was the oil of the 20th century; compute, GPU compute, and TPU compute will be the oil of the 21st century. There is no way to hedge and justify the seven-plus trillion dollars of capex to tile the Earth with compute, or possibly tile the skies and lunar surface with compute, without appropriate abilities to hedge those expenditures with options, futures, derivatives, or commodities.
The OCPI, the Orin Compute Price Index, is already available on Bloomberg terminals with its own symbol. Orin has its own symbol on the New York Stock Exchange, RNN, as part of a novel program giving early-stage startups their own ticker symbols.
Epic AI ran the numbers on hyperscaler investment costs compared to cash flow. The big five—Microsoft, Google, Amazon, Meta, and others—are spending on AI faster than they are earning. Funding comes from debt and equity raises, not based on revenues. If capex exceeds cash flow, persistence depends on financing and continued strong sentiment for investing. If sentiment shifts, a massive pullback could be forced.
The counterargument is that sentiment will not shift. These companies have reached the level of spending all their cash flow yet can raise 10 to 100 times that in equity and debt; they have a long way to go. All the money in the world wants to flow into this, described as the best investment in the history of humankind. The question becomes how much money exists in the world and who ends up controlling it—humans and companies, or AI itself. Hyperscalers can raise prices to increase operating cash flow; Anthropic did so recently and encountered no problem. Google has tons of operating revenue while others do not, though Anthropic is scaling quickly.
SpaceX revenue discussion clarifies that almost all of SpaceX's revenue as of the past month now comes from being a hyperscaler for everyone else, a data center play on land for now, selling GPUs and compute rather than selling intelligence online like OpenAI, Anthropic, and Google.
Intelligence is becoming cheap, but the manufacturing of intelligence is becoming incredibly expensive. Some companies are so profitable they can build out an entire new industry within their cash flow, something that has never happened before. This new industry could be bigger than all previous ones.
This is described as the most extraordinary time to be alive. Despite doom-saying, remaining optimistic about the future is urged. The technology is critical to move humanity forward and can be aligned to become the greatest support in overcoming ancient neocortex limitations and moving toward an abundant future.
Technology has always been a major driver of progress; as Ray Kurzweil says, it may be the only major driver. The big challenge is extracting the promise without the peril.
A closing observation is that a planetary sensing system is being built and now upgraded to a planetary intelligence system, needed to reach planetary wisdom.
The galactic zoo hypothesis is suggested: a third-generation biosphere planted by aliens long ago.
The episode closes with music from Ekram Alam on the piece "Moonshots" and fan-based video content.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
