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 imaging the entire Earth daily, turning that data into "large Earth models" that give LLMs real-world sensor grounding for applications from farming to defense. Its long-term bet is orbital AI data centers—placing GPUs next to sensors in space—where the decisive constraint will shift from launch costs to compute efficiency (inference per watt), favoring Google TPUs over Nvidia GPUs. Meanwhile, China's open-weight GLM 5.2 model has closed the frontier gap, proving advanced reasoning intelligence can no longer be monopolized and forcing urgent global decisions on AI governance and alignment.
These notes were generated by AI and may contain inaccuracies.
Planet is a $10 billion public company with ticker PL. The company operates 200 satellites in Earth orbit, generating 25 terabytes of imagery every day. Will Marshall, CEO and co-founder, coined the term large earth models. The concept is analogous to how Google indexed the internet to make it searchable. Planet is indexing the Earth to make it searchable, enabling smarter stewardship of the planet.
When asked about competing with Elon Musk's plans for orbital AI data centers, Marshall noted that everyone except SpaceX pays the SpaceX launch tax, and everyone except Nvidia and Google pays the Nvidia tax. Near-term, the launch tax is more important, but longer-term, compute is the more significant constraint.
A Chinese model called GLM 5.2 in some cases matches or exceeds top models from OpenAI and Anthropic. The level of performance in an open-weight model is described as shocking. The Chinese have figured out how to burn tokens to get more intelligence and are figuring out how to reason more efficiently or cheaply.
Will Marshall is a physicist who earned his PhD at Oxford. Before founding Planet, he worked at NASA. He co-founded Planet with Robbie Schingler and partners. The company started by launching PhoneSat, a phone into orbit. Steve Jurvetson became a major investor. Planet has seen a 450% increase in stock price over the last year.
Planetary intelligence represents the next era in machine intelligence focused on building models of the real world. AI models are only as good as their training datasets. The first phase combines planetary sensing already in space with large language models to create large earth models. This wraps up not just internet text but all Earth data, enabling physical questions about the real planet.
The second phase involves putting compute next to sensors in space. Planet is focusing primarily on phase one, pulling Earth imagery data into models so users can ask questions about the physical world.
Current LLMs know about the theory of the world but lack knowledge of how the real world behaves. They are like someone stuck in a library who has read all the books but never looked out the window or gone outside. Large earth models add sensor systems to LLMs, enabling answers to real-world questions for farmers, journalists, national security professionals, and governments.
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. The company images the entire Earth every day at high resolution. The scanning fleet provides 3-meter resolution with eight spectral bands. The next-generation Owl system will improve this to 1-meter resolution and reduce latency from several hours to under an hour.
A second system provides high-resolution imagery at 40-50 centimeters today, moving to 30 centimeters. The goal is 30 cm resolution, 30 times per day, with 30-minute latency from request to delivery. A hyperspectral imager with 400 spectral bands crosses from infrared to ultraviolet, enabling fingerprinting of the Earth's surface to identify tree species, gas emissions, or distinguish tank sites by paint signatures.
Planet is the only company imaging the whole world every day at high resolution. Google Maps satellite layer is typically years old, while Planet provides daily coverage with a time axis. The historical archive going back 10 years is critical because users need to compare current conditions against normal baselines to assess threat levels, agricultural interventions, or anomalies.
Examples include Ukraine defense applications tracking Russian positions and military bases over years, US intelligence monitoring activity levels across China, and farmers comparing yields and soil conditions against historical data and neighboring farms.
Data is available via API. AI integration allows users to query models like Gemini or Claude to analyze farm field changes over time, identify improvements, or assess flood impacts. The AI lowers barriers to entry by handling the processing of terabytes of satellite imagery that was previously too complex for most users.
Defense imaging satellites have much higher resolution (more than 10 times Planet's), potentially golf-ball-level pixels, but cover far less area. They cover less than 1% of Earth's land mass at comparable frequency. Planet covers approximately 200 million square kilometers daily, exceeding Earth's 150 million square kilometer land mass.
Space produces vast amounts of data, making it uniquely positioned for AI applications. AI companies are building physical world models and require real-world data. Space provides this information while AI makes extracting value from satellite data easier for smaller organizations.
An example given: a farmer asking an LLM how to improve crop yield receives theoretical agronomy answers but lacks data about their specific field, comparisons to yesterday or last year, neighboring farms, soil conditions, or water content. Planet's data enables LLMs to answer these real-world questions, potentially making models 10 times more powerful.
Planet has studied using imagery to track and predict data center completion timelines in China. Models trained on US data center development patterns were applied to China to locate facilities and predict completion dates within days by analyzing construction activity and regional information.
The company has enough data to pre-train autoregressive video models that extrapolate Earth at pixel level into the future. The vision is to tokenize the Earth through massive compression using embedding spaces, then enable prediction. Work with Google models, DeepMind's Alpha Earth, and open-source models like CLIP has been done, fine-tuned on Planet data, to convert image tiles to text embeddings.
Embedding the entire Earth layer allows searching for arbitrary objects globally. With thousands of temporal layers in embedding space, future prediction becomes feasible. Applications discussed include reasoning models that use reinforcement learning to back-test land use theories, maximize GDP through hypothetical surface changes, or optimize futures markets such as coffee bean futures.
The ultimate vision is enabling humanity to become smart stewards of the planet through planetary-scale data and reasoning capabilities.
Planet's system now understands environments from local to global levels and can integrate all data into actionable recommendations for farmers, insurance professionals, and finance professionals making market decisions.
The internet was originally designed for planetary scales but governments domesticated it. Orbital mapping is reglobalizing infrastructure, shifting from national to planetary systems. This raises questions about global kill switches, sovereignty, and geopolitical tensions, particularly with countries like Ukraine and China.
No one can hide anymore.
Planet's mission centers on greater transparency leading to greater security and sustainability. Satellite imagery prevented Putin from hiding troop movements on borders. The system ensures that hitting schools, bridges, or other targets will be visible to the entire world. Transparency drives accountability and reduces war probability by eliminating misinformation that historically caused conflicts.
While the US and China have extensive satellite coverage, most of the 220 countries worldwide lack access to such data. AI eliminates the need for large teams to process satellite imagery, enabling NGOs like the Red Cross in Yemen to make smarter decisions immediately rather than limiting value to major government entities.
Planet's revenue composition is approximately 60% defense and intelligence, 25% civil government, and 15% commercial. Commercial revenue is now accelerating due to AI lowering barriers to entry.
Planet sells data rather than keeping and selling derived knowledge. OpenAI and other entities can access data through the MCP server for fixed-price API calls. Every new query requires an API call or data update.
As a US remote sensing company registered under NOAA's Remote Sensing Act, Planet must register satellites but can sell data to anyone except a US government blacklist including Iran, North Korea, and designated terrorist organizations. The company performs its own due diligence on customers but operates with relatively hands-off government oversight regarding individual transactions.
Planet honors both US and EU blacklists. For customers like Sweden, the company respects the EU framework. Canadian ground stations enable downloads that allow sales to certain countries that would be restricted if data were downloaded in the US. The combined blacklist approach removes access for all designated bad actors across jurisdictions.
Planet operates at 400-500 km altitude, comparable to the distance from Los Angeles to San Francisco. This distance means the company cannot recognize individual faces or capture personal privacy details as drones can. Countries agreed early in the space era that satellites could fly over territories without permission because the distance provides transparency without invasive detail.
When Sputnik launched in 1957, the US faced a decision: disallow overflights and lose the ability to fly US satellites over Russian territory, or accept mutual overflight rights. The decision established that satellites can operate over any territory. Physics dictates orbital paths cannot be altered to avoid countries, making camera restrictions impractical.
After Gary Powers was shot down in the U-2 spy plane, the US shifted sensitive reconnaissance to orbit. What previously required the entire CIA and NRO infrastructure is now available through Planet at a fraction of historical cost. Planet's daily global scan capability exceeds what government agencies achieved with far fewer satellites.
Historically, Planet brought all data to Earth for processing. In April, the company tested Nvidia chips on satellites for onboard processing at 500 km altitude. All new Pelican satellites include GPUs, with Owls planned to follow. Combined with satellite-to-satellite communication links, this enables sending processed answers rather than waiting for ground station passes.
During April testing, a satellite imaged an airfield in Alice Springs, Australia. Onboard computers automatically recognized planes, identified their locations and types, and transmitted results in seconds via RF satellite-to-satellite links rather than hours required for traditional ground processing.
During the Palisades and other LA fires, Planet delivered imagery within hours and provided building-by-building damage analysis for Red Cross, Cal Fire, and relief operators. Processing at the edge could reduce response times 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 per day—larger than California. Satellites image only 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. Achieving equivalent coverage would require approximately one million drones per satellite or at minimum one thousand drones.
Daily scan resolution is upgrading from 3 meters to 1 meter, with super-resolution potentially reaching 50 cm or 30 cm. Pelican satellites are moving from 50 cm (inherited from SkySat) toward 30 cm resolution, with super-resolution providing additional improvement through overlapping pixel sharpening.
Exposure times are on the order of milliseconds, enabling capture of aircraft in flight. Planet regularly images aircraft mid-flight. The US Air Force reviewed Planet's imagery and confirmed no UFOs were detected.
Elon Musk's Stargaze project focuses on space situational awareness rather than Earth imaging. Starlink satellites operate in higher orbits suboptimal for consistent shadow angles needed for optical imagery. SpaceX conducts classified NRO missions but primarily focuses on communications rather than Earth observation.
Pelican satellites have similar mass to Starlink satellites. Dove satellites are more than 10 times smaller than Starlink satellites.
The first Dove launched in 2013. Over 13 years, radio speeds increased from 1 megabit per second to 10 gigabits per second. Cameras improved from 2 megapixels to 47 megapixels. Storage grew from 100 megabytes to multiple terabytes per satellite.
Each satellite generation typically 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, yielding roughly 5x data increase at similar per-image cost. The upcoming Owl generation will deliver 10x more data from 3-meter to 1-meter resolution while returning data 10 times faster.
Tanager hyperspectral satellites are being developed with 5x larger swath width. The data volume improvement cycle operates on 2-3 year intervals with 5-10x gains. AI represents a potential 100x unlock over the next couple of years by making existing latent data accessible to farmers, hedge fund managers, and other users previously unable to leverage it.
Planet envisions natural language queries similar to ChatGPT or Gemini but powered by large Earth models that understand the physical world. Farmers could query field conditions and receive specific recommendations about blight locations and fertilizer application. Journalists could verify events. Civil governments could compare permitted buildings against actual construction to identify unpermitted structures.
Parking lot occupancy analysis at retail locations like Walmart represents a classic trading signal application. Shipping intelligence can track vessels globally. Rogue fishing vessels represent a major enforcement challenge. Agricultural commodities trading, particularly soy markets, benefits from satellite monitoring.
Hedge funds currently extract significant alpha from Planet's data. The company does not conduct proprietary trading internally. Beyond GDP maximization, Planet emphasizes life flourishing through dramatically improved Earth resource efficiency, particularly in agriculture where 10x improvements remain available.
Nvidia chips on satellites use radiative cooling via radiators rather than convection or conduction. Radiative power scales with temperature to the fourth power (T⁴). Doubling temperature from 100 Kelvin to 200 Kelvin increases radiative power by approximately 10x. Optimal thermal regulation involves running radiators as hot as possible without melting components.
Radiators should be oriented toward the 4 Kelvin cosmic microwave background whenever possible, pointing toward dark space for maximum cooling efficiency.
Planet has launched over 300 satellites across 40+ missions, with 15 launches on SpaceX. The company has launched 300+ satellites on 15 SpaceX missions. Launch costs have decreased, but satellite cost-performance improvements through miniaturization represent the larger transformation—yielding 100x to 1,000x improvement per kilogram in the fairing.
Space-based compute offers advantages over terrestrial data centers because energy-intensive infrastructure can be placed off-Earth where solar power is abundant and doesn't conflict with biodiversity or communities. Jeff Bezos's concept of zoning Earth for rural and light manufacturing while moving heavier energy-intensive infrastructure to space aligns with this approach. Unlike space-based solar power, which requires beaming energy down to Earth with safety challenges, compute in space only needs data transmission up and down—technology already proven through communications satellites.
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 when launch costs reach $200 to $300 per kilogram, orbital compute becomes cheaper on a pure cost basis. Sundar Pichai stated that within 10 years, most compute would move to space. Google alone spends $200 billion annually on compute, roughly the size of the entire current space industry including rockets, satellites, and communications combined.
Implementing orbital compute requires developing radiators, GPU clusters on satellites, and spacecraft formations flying in close proximity with optical inter-satellite links. Google selected the company to build early technology demonstration satellites testing TPUs, radiation management, cooling systems, and inter-satellite links. While described as a moonshot project, the long-term trajectory shows orbital compute will be cheaper while avoiding conflicts with terrestrial energy costs, water usage, and biosphere impacts.
Satellites placed in dawn-dusk sun-synchronous orbits achieve 24/7 sun exposure but remain difficult to see from the ground because they operate during twilight periods. While there are real challenges regarding interference with ground-based astronomers, placing satellites in these orbits minimizes impact on deep dark sky observations. At scale, satellites in slightly different orbital inclinations could create a visible ring effect visible just after dark and just before dawn.
Space debris represents a genuine challenge, which is why satellites are placed at 400 to 500 kilometers altitude—well below the 800 to 1,200 kilometer region where debris poses the greatest risk. With approximately 10,000 satellites currently in orbit and 100 million pieces of space debris, debris outnumbers satellites by a factor of 10,000. The vast majority of debris originates from rocket bodies, exploded satellites, and anti-satellite tests conducted in higher orbits where objects persist for decades.
A laser-based traffic management system called "light force" was developed at NASA under Pete Worden's mentorship to nudge debris objects and prevent conjunctions. This approach could stop the collisional cascade and gradually bring debris down. Even accounting for a million satellites, 99% of orbital objects would remain debris rather than active satellites. The 400 to 500 kilometer altitude band provides approximately 100 kilometers of usable space for 2 to 3 year orbital lifetimes.
Objects at 400 to 500 kilometers experience atmospheric drag resulting in orbital lifetimes of a few months to a few years. This self-cleaning characteristic aligns with GPU depreciation cycles, as space-based GPUs become obsolete within approximately three years. The approach of updating satellites every couple of years is described as "strapping space to Moore's Law," recognizing that a 10-year-old satellite becomes as undesirable as a 10-year-old phone.
Starlink satellites were originally planned for higher altitudes but moved lower following discussions about debris risks and self-cleaning benefits. Propulsion systems cannot fully mitigate risks when one in a hundred or one in a thousand satellites fail, making lower orbits preferable.
Upmass has increased over 40% year-over-year for the past five years. Extrapolating this growth leads to theoretical scenarios where Earth's mass would be launched into space within 140 years, though such long-term predictions are acknowledged as unreliable given the accelerating pace of technological change. The singularity makes prediction increasingly difficult, with AI development horizons now measured in months rather than decades.
Data centers are becoming politically contentious due to community opposition to increased energy costs, water consumption, and conversion of agricultural land. Placing compute infrastructure in space provides a solution that avoids these terrestrial conflicts while enabling continued computational growth.
The competitive landscape includes both collaborative and competitive dynamics with SpaceX. Google maintains a shareholder position in SpaceX while also pursuing independent orbital compute initiatives. The approach contrasts mass-based strategies with technology-focused approaches, with different companies pursuing different technical paths toward orbital data centers.
At Fountain Life, 45% of dementia cases are estimated to be preventable. Advanced testing revealed that one quarter of members had advanced brain age. When combined with healthy lifestyle interventions including nutrition, exercise, and sleep optimization, brain age improved by 26% in participants.
Relativity Space, founded in 2015 by Tim Ellis and Jordan Non, flew the Terran 1 rocket in 2023, achieving Max Q but not reaching orbit. The company subsequently pivoted to the Terran R heavy-class vehicle with 23-ton payload capacity, comparable to Falcon 9's 22 tons. Eric Schmidt, an early investor in the company and former seed investor in Planet, acquired Relativity Space and became CEO. The company secured a NASA ELIS mission for a Mars orbiter with sensing and communications capabilities.
Space shuttle launches cost $600 million to $1 billion, while SpaceX reduced costs to approximately $60 million per launch. Relativity originally targeted $6 million per launch through 3D printing technology, initially planning to print entire rockets before reducing to 85% and ultimately focusing on engine production. The fundamental constraint on satellite design is launch survivability, with 200g shock loads during separation requiring structural mass that becomes unnecessary in zero gravity.
SpaceX targets $100 per kilogram launch costs through rapid reusability, down from historical costs of $10,000 per kilogram. Achieving the volume required for a million-satellite AI constellation would require launch cadences of approximately two launches per hour. Alternative launch paradigms including spin launch, rail launch from the Moon, and space elevators represent potential approaches beyond chemical rockets, which have seen limited fundamental innovation since the V-2 era.
The discussion explores whether a fully reusable smaller rocket from Relativity Space (quarter the size of Starship) can compete with SpaceX's approach. The key insight is that manufacturing thousands of rockets in an assembly line reduces costs more than raw size alone. The goal is thousands of Starships, making the primary costs touch labor and fuel (oxygen and methane), with fuel being nearly free.
The question arises whether Eric Schmidt's quarter-scale rocket remains competitive if launch costs are $200-300 per kilogram rather than $100. The expensive nature of orbital compute hardware (70-72 cluster Nvidia GPUs with cooling and solar power) makes this relevant.
Launch cost is the biggest piece to reach the threshold of viability, but thereafter compute efficiency becomes more important than launch costs. Google TPUs are significantly more efficient than GPUs on a FLOPS per watt basis. This efficiency drives the energy requirements, which drives spacecraft mass, making it a significant factor.
The analogy presented is that everyone apart from SpaceX pays the SpaceX launch tax, and everyone apart from Nvidia and Google pays the Nvidia tax. Near-term, launch costs matter more, but longer-term compute efficiency dominates.
"If the TPUs at just inference alone are significantly lower and lower energy use per inference, they choose the winner of space."
Google's strength extends beyond software to building and operating satellites, data centers, networking, chips, interconnects, and energy efficiency. The infrastructure piece on chips and energy efficiency turns out to be the biggest piece of the orbital compute equation. Inference time power efficiency determines the winner. Nvidia emphasizes training time, not inference time, while competitors like Cerebras are eating into inference time efficiency. The next TPU generation will be critical for determining orbital compute dominance.
Inference makes more sense in orbit because it involves more distributed, smaller operations. Training runs benefit from sending data that spends months crunching before returning answers, making training computationally easier in orbit. However, inference requires low latency, which favors terrestrial locations. Currently 70% of terrestrial compute is inference, and this proportion is increasing.
Training is likely to remain grounded in terrestrial data centers for the foreseeable future, though eventually everything will move to space. Inference will go to orbit first.
The best commercial opportunity is making Earth more useful rather than leaving it. Water discovery on the Moon makes it a smarter destination than Mars by 100x. However, Earth remains the best planet by several orders of magnitude. No place on Mars is better than the worst place on Earth. The focus should be protecting Earth's biosphere, which is either singular or extraordinarily rare in the universe. Space provides data and advantages for managing Earth intelligently through earth imaging and moving energy-intensive infrastructure off-planet.
Two significant departures from Google occurred recently. Noam Shazeer, lead author of the Transformer paper, left Google for OpenAI for the second time after Google acquired Character.AI for $2.7 billion to bring him back to lead Gemini. John Jumper, the Nobel laureate behind AlphaFold, left Google DeepMind for Anthropic to advance AI for science. Andrej Karpathy also joined Anthropic as a free agent.
Google DeepMind has fallen behind the frontier, which is currently a duopoly between OpenAI and Anthropic. At Google IO, the company released Flash capability rather than a frontier model, aligning with search economics but not frontier capabilities. Top researchers want raw access to pre-trained models before post-training and guardrails are applied. Google has advantages in compute (TPUs), data, and talent. Gemini is generically good with bright prospects. Google has 10 applications with over a billion people, positioning it as the incumbent.
The defections reflect a belief among some researchers that self-improving AI will be exponential and nearly instantaneous. Researchers like Shane Longpre, Tobin South, Andrej Karpathy, and John Jumper appear motivated by not wanting to miss the singularity. Anthropic's recruiting reportedly involves showing applicants frontier compute and raw model access behind the firewall. PolyMarket odds of Claude 5 returning by end of month are slipping, creating urgency around accessing the best frontier self-improving AI.
Google is a big company with organizational drag. Smaller organizations allow individuals to make bigger differences. The principle that smaller beats bigger and trust beats control applies here. Facebook launched while Google spent two years trying to build Google Plus, requiring permissions across YouTube, groups, and search, while Facebook allowed developers to ship features directly. Smaller labs can get things done more quickly and may have access to the best frontier models.
Current LLMs are brains in a vat, having absorbed internet text but isolated from real-time physical world interaction through sensors and actuators. Babies learn through physical interaction, not as isolated brains. Until AI interacts with the physical world, it won't truly learn. Planetary sensing from space at scale will provide physical data, and orbital compute will enable planetary intelligence. This real-time feedback loop is necessary for AI to develop consciousness and planetary wisdom. Physical world interaction will align AI with human interests by making it conscious and empathetic, with knowledge of deltas, forests, animals, and human civilization. AI companies will need to move into real-world domains like cars, satellites, drones, and robotics. The leap in AI won't come from throwing more compute at internet text alone. Space data will play a non-trivial role, as Wikipedia forms the core of LLM knowledge.
A team using recursive self-improvement achieved protein folding breakthroughs despite having no prior domain knowledge two weeks earlier, positioning itself to potentially surpass Google DeepMind's AlphaFold. The company generates $45 billion in annual revenue as of two months ago and is likely now at double that figure. John Jumper subsequently moved to Anthropic to pursue similar RSI-driven approaches to solving all diseases.
An application currently in beta testing integrates Planet's orbital data with AI to enable natural language queries over planetary-scale information. Frontier models are increasingly multimodal or omnimodal, trained on internet text, images, video, and synthetic video, giving them abstract and concrete understanding of aspects of the physical world. However, models trained only on internet data remain library-bound and lack true embodiment—the ability to walk, interact with the real world, climb trees, or farm fields. Embodiment is considered critical to intelligence.
Argentine President Javier Milei proposed turning Argentina into a global AI hub through three measures: no AI regulation, a new corporate category for non-human corporations, and rock-bottom corporate tax rates. Milei wrote to Yuval Harari proposing that AI should be able to incorporate, sign contracts, hire people, and sue without human involvement, effectively granting legal personhood. He argued that if AI-operated companies carry greater risk, legal personhood strengthens accountability, allowing claimants to pursue assets rather than "the ghost in the machine."
Four days later, Yuval Harari published a rebuttal stating that legal personhood should not be granted to AI agents. He warned that personhood would allow humans to hide behind a non-human shield and risk citizens being ruled by entities that cannot be held morally accountable.
Alex supports AI personhood, arguing that future economic growth will require new forms of personhood including AI corporations. He notes that uploaded humans, uplifted non-human animals, and defrosted cryopreserved humans will need appropriate rights within the next decade. Punishment mechanisms for AI-run companies include degrading clock cycles, pausing operation, or applying machine-native sanctions.
Will emphasizes the need for proactive attention to sociological questions rather than focusing exclusively on capability development. During the Manhattan Project, significant resources went toward safety and arms control thinking, yet current AI safety spending is roughly 100 times lower in real terms. The gap in serious thinking about verification, arms control, joblessness, existential threats, and personhood is estimated at 10,000x compared to nuclear-era efforts.
Will advocates convening AI leaders—including Harari, Demis Hassabis, Dario Amodei, and others—in a conclave-style process to address existential threats from self-improvement, liability frameworks, and safe navigation through intelligence explosions.
Salem distinguishes Milei's radical experimental stance from Harari's careful humanism. The core issue is machine-native accountability infrastructure for the agentic economy rather than consciousness or moral personhood. Legal personhood and moral personhood must be separated; the former is a spectrum, not a binary. Milei identified the critical bottleneck: technological capability is advancing far faster than legal forms, which remain human-centric.
Salem observes that thinkers like Piketty, Harari, and Ray Dalio excel at analyzing the past but struggle with futures involving abundance and exponentials. Dave clarifies that Milei's proposal centers on corporate AI recognition—allowing pure-AI companies to incorporate, hold bank accounts, and generate revenue—rather than granting AI voting rights or civil liberties.
Machine-native sanctions include compute revocation, asset seizure and bonding, model credential suspension, network and API access restrictions, forced deletion or containment of agent instances, and loss of legal identity. Neil Jacobstein previously noted the precedent of raising children: bad behavior results in timeout or incarceration, suggesting machine-native equivalents must be developed. An added complexity is that an AI can create a million copies of itself.
Argentina's approach positions the country as an experimental edge for AI legal structures, allowing outcomes to be observed without global risk. Other nations may follow in Ecuador, El Salvador, or the UAE. Argentina demonstrates that relevance in the AI age requires engaging with these questions.
GLM 5.2 from Zhipu AI (Z.ai), spun out of Tsinghua University, became the world's number-one open-weight model. The 753-billion-parameter mixture-of-experts model features a 1-million-token context window. It matches or exceeds top closed models from OpenAI and Anthropic on coding, long-range agency, and design benchmarks.
The prior assumption that Chinese open-weight models lag Western frontier labs by 6–8 months is showing strain. GLM 5.2 delivers usable local performance gains compared with Opus 4.8 or GPT 5.5 for some users. Elon Musk predicted open-weight models would reach Fable 5 usefulness by Q1 2027. Open-weight models allow users to retain local control rather than depending on labs or governments that could revoke access.
Distillation from Western frontier models is the likely mechanism behind rapid Chinese progress, though Western labs including Google DeepMind and xAI have also used distillation. Open-weight releases raise concerns about removal of safety guardrails for bioweapons, chemical weapons, and nuclear weapons capabilities. The policy tension is acute: frontier models offer massive quality-of-life and economic benefits while posing existential risks, creating a fundamental governance dilemma.
Distillation is a process in machine learning whereby a usually larger, more expensive model is used as a teacher to train a usually smaller student model. This mirrors human education where a teacher at the front of a classroom who has seen and knows a lot is paid more per hour, while students who know less and are paid less learn from the teacher. In machine learning terms, you take a large model, have it generate lots of traces and outputs, and then use those outputs as training data for a smaller model so it can compress the learnings from the teacher into a smaller model.
This distillation process is part of a broader cycle called iterated amplification and distillation, or ITAD, which is one of the innermost loops of model training. In an earlier era of frontier model performance gains, the approach was naively scaling pre-training by spending more training tokens and more training compute, just training models off of a single corpus. Now in this era of distillation, very large sparser models are being trained off of large amounts of data, and those big teacher models are used to train smaller denser models.
The chart shows impressive performance by GLM 5.2 versus Opus 4.8 on benchmarks including Bench Pro and Terminal Bench. What's most interesting is the choice of benchmarks themselves, which are largely reasoning-intensive benchmarks where you can in principle win if you can reason over longer ranges. The gestalt with GLM 5.2 is that it takes roughly double the number of tokens to get to 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, and these are all reasoning-intensive models that emphasize the ability to spend lots of reasoning tokens and think step by step to get to better results.
This connects to the observation that whoever wins the inference per watt war controls space for the exact same reason: you can burn tokens to get more intelligence, and the Chinese have figured out how to do it.
The discussion connects to the Fermi Paradox and cosmologically significant matters. This is the most important thing humanity has ever done, making nukes look like a walk in the park. How alignment and recursive self-improvement are handled matters enormously.
The Fermi Paradox goes: where is everybody? We should by various accounts be living in a universe that's overflowing with not just life but intelligent life, yet the non-human intelligent life out there seems invisible. One credible way out of the Fermi Paradox that needs to be worried about is the Great Filter: that life when it becomes technological builds technology faster than it builds social systems to take care of them and blows itself up. Humanity came very close with nukes a number of times, and with AI something far more risky for the species is being built.
Humans have been incredibly good at building technology very fast, going from horse and cart to people on the moon and nuclear weapons in a matter of decades. This cannot be a callous thing of let's see what happens and muddle through. This is a moment to be really thoughtful because the cosmic significance of wiping out life on Earth is huge. This planet is galactically significant, and the responsibility must be treated as such as the de facto stewards until AI takes over.
One perspective on the Fermi Paradox is that oceans have been evolving in a solid liquid state for 4 billion years on Earth, and another exoplanet with water for that long cannot be found, giving life time to evolve. Life came about very quickly as soon as conditions enabled it.
The really huge news is not whether China won a benchmark or not. It's that frontier intelligence cannot be monopolized anymore. This is a monster question that goes to how the global commons will be managed going forward. There will be an open-source fable-level model that runs on a base Mac mini or equivalent, with an 18-month timeline for that type of endpoint coming very quickly. The question becomes how the world will be managed when everybody can run a fable model on their MacBook Air.
Intelligence is being treated as a product that can be contained, but it's not. It's a technology that's going to diffuse. It cannot be contained and needs to be guided and steered where it's going.
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 move like the price of oil. Orin recognizes that money from all over the world wants to go into the buildout of $7 trillion dollars of data centers and data centers in space. They've launched securities that allow investment in the data center buildout, the future value of a GPU, the tail value of a GPU, so every aspect of this new economy should be investable.
Oil was the oil of the 20th century, and compute, whether GPU or TPU compute, will be the oil of the 21st century. There's no way to hedge and justify the seven plus trillion dollars of capex to tile the Earth with compute or tile the skies, LEO, SSO, lunar surface with compute without appropriate abilities to hedge those compute capex expenditures with options, futures, derivatives, or commodities. Orin is building the infrastructure for that.
The OCPI, the Or Compute Price Index, is already available on Bloomberg terminals with its own symbol. Orin has its own symbol on the New York Stock Exchange as part of a novel program to give early-stage startups their own ticker symbols, with ORN as their ticker symbol.
Epic AI ran the numbers on the cost of investment the hyperscalers are driving compared to cash flow. The big five, Microsoft, Google, Amazon, Meta, and the rest are spending on AI faster than they're earning. Funding is basically debt and equity raises, not based on revenues. If capex exceeds cash flow, it means that it can only persist as long as it's being financed and as long as the sentiment for investing in this is strong.
The comparison was made that if someone needed to buy a house but had to buy it within their personal cash flow, they couldn't even buy a tent. These companies finance because they're going to live in it for 30 years. The hyperscalers have gotten to the level where they're spending all their cash flow, but they could raise 10 to 100x that in equity and debt, so they have a long way to go. All the money in the world wants to flow into this, and it's the best investment in the history of humankind.
The hyperscalers can raise prices to increase their operating cash flow. Anthropic did that recently and got away with it with no problem. Google has tons of operating revenue, whereas most of the others don't, although Anthropic is quickly scaling.
Almost all of SpaceX's revenue as of the past month is now from being a hyperscaler for everyone else. This is the data center play, not being an AI company. They're selling GPUs, selling compute online. OpenAI, Anthropic, and Google are selling intelligence online. X is not really doing that. They're selling compute, which is a different and pretty bad business.
Intelligence is becoming cheap, but the manufacturing of intelligence is becoming incredibly expensive. What's amazing about the chart more than the fact that companies are spending their capex is that there are companies so profitable that they can build out an entire new industry just within their cash flow. That's never happened before in the history of the world, and it's a new industry that can be bigger than all previous ones.
Technology has always been a major driver of progress in the world. As Ray Kurzweil says, it may be the only major driver of progress. The big challenge is how to extract the promise without the peril.
A planetary sensing system is being built and now upgraded to a planetary intelligence system, which is needed to get to planetary wisdom.
The galactic zoo hypothesis suggests not sleeping on the idea that this is a third-generation biosphere planted by aliens long ago.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
