Sonnet 5 Drops, Fable 5 Will Return & Fusion’s First Plant Gets Licensed W/ Philip Johnston | #268
In a Nutshell
Sonnet 5 was released as a stopgap after the US government pulled Anthropic's Fable 5 flagship model for national security reasons, while fusion startup Helion secured regulatory approval for its first plant supplying Microsoft. The discussion centers on accelerating humanoid robotics scaling toward commodity pricing, massive orbital compute deployment via StarCloud's space-based data centers, and energy abundance through both fission and fusion. Key themes include experts consistently underestimating exponential growth in solar, EVs, batteries, and now robots, with incentive prizes and vertical integration (SpaceX, Rocket Lab) emerging as dominant forces across AI, energy, and space infrastructure.
These notes were generated by AI and may contain inaccuracies.
Sonnet 5 is positioned as a temporary solution to fill the gap until Fable 5 returns. It offers mediocre capability at a high price point, yet demand persists. Anthropic's flagship model Fable 5 has been offline for 15 days after being pulled by the US government. Axios reports it may return within days. Historians will view this period as marking the middle or endgame of recursive self-improvement. Helion has cleared the required Washington state regulatory approvals for its Orion Fusion power plant, signaling that fusion power is finally arriving.
If the right long-term metric or figure of merit is tracked, the timing of major technological breakthroughs can be predicted with accuracy. The discussion shifts to data centers and space topics.
Peter D. Mandis hosts the Moonshots podcast, focused on the coming singularity and age of abundance. The panel includes Dave London (managing partner of Link Exponential Ventures and top funder of MIT and Harvard AI startups), Sel (global trader and CEO of Open Exo), and AWG (in-house ASI). Philip Johnston joins as a guest.
Sel dropped his son off at camp near San Sebastian, Spain, and is currently in Barretts. He has visited five countries in the last three days. Alex reports he is GDP maxing. Peter focuses on happiness maxing and gratitude maxing. Dave is at Link Studio, noting increased activity with Northeastern and Princeton alongside MIT and Harvard.
TechCrunch is active in San Francisco. Liputan and Andrew Feldman (number two at Nvidia) spoke to teams on Friday. The team previously collaborated with Eric Schmidt, Eric Bolson, and Daniel Larus.
Twenty stories across six fronts will be covered, including robotics, energy, and data centers. Capital continues to flow rapidly. The mission is to keep viewers informed and optimistic about the future.
The previous episode with Eman was described as exceptionally strong. EMOD released new model combinations in the last 48 hours. The first topic is robotics, specifically predictions about the number of robots that will exist on Earth.
Rames Nom, an early Singularity University faculty member and futurist, shared a chart demonstrating how experts consistently underestimate exponential growth. The left chart shows solar growth: the yellow exponential line represents actual growth while expert predictions remain flat and horizontal each year. The center chart shows the same pattern for EV growth, and the right chart shows it for battery sales. Experts remain sublinear in their estimates despite living through exponential growth.
Sel notes that every presentation he gives includes slides on this phenomenon. A notable example is the 2013 prediction that Moore's law would end by 2022. Technology press articles making similar claims have appeared every two years for 60 years. Experts excel at measuring current technology but fail to account for the compounding ecosystem around it. The original book included the headline "beware the experts."
Alex defines an expert as someone who can tell you exactly how something cannot happen. Experts are often ingrained in the past; disruption threatens their status. This observation connects directly to the robotics discussion.
Morgan Stanley originally predicted 14,000 Chinese robots, later raised the figure to 26,000, and has now increased it to 50,000 with a projection of 500,000 robots by 2030. There are currently 140 humanoid robot companies developing hardware in China. Elon Musk projects tens of millions to 50 million robots by 2030, with billions invested into the early 2030s.
A chart from Andrea Harowitz projects $16 billion in hardware investments in Q1 of 2026, with the US finally catching up. Dave's investment thesis shifts from an AI-centric entrepreneurial ecosystem to a hardware-centric ecosystem.
Alex states that super intelligence will spill out of data centers into the streets, primarily through autonomous vehicles and humanoid or near-humanoid robots. At low numbers of humanoid robots per capita, applications remain industrial and logistical. At approximately one humanoid robot per capita, domestic robots become widespread, resembling an iRobot-style regime with domestic staff available to everyone.
When robot density reaches 10 or 100 humanoid robots per capita, the humanoid form may no longer be necessary. Microbots and nanobots become viable, enabling solutions to grand physical-world challenges that would be uneconomical at one robot per capita. Humanity will rapidly scale through low, one-per-capita, and many-per-capita regimes.
Dave emphasizes thinking in 10-year investment themes, noting that 10 years in the AI age equals roughly 100 years in a normal world. The next one to two years remain dominated by white-collar automation, AI algorithms, chip design, AI-designed chips, and early space data centers. Beyond two years, automation of data center construction, biotech experiments, chemical mixing, and gel reading will all be robotic.
The US gutter-cleaning industry alone is a $1 billion annual market. A robot that cleans gutters and washes windows would address a $20 billion global market. Construction involves thousands of individual tasks. Kitchen automation is already being pursued, with fast-food restaurants able to buy at scale. Travis Clutnick's company Adams is developing fully roboticized kitchens.
The $16.5 billion invested in the last quarter came through 500 deals. Peter notes that Star Wars featured hundreds of droid variations; the current proliferation of specialized robotics companies suggests this vision is materializing.
Peter observes that robots were largely absent from Star Trek except for early characters like Data. Gene Roddenberry focused on societal implications of technology and humanistic themes. The computer character (played by Roddenberry's wife Christine) was roboticized but lacked the empathic voices of modern models. Rod Roddenberry will appear at the Moonshot Summit as a judge for the future vision X-Prize.
The show's emphasis on infinite diversity and infinite creativity prioritized human and humanoid-alien interpersonal interactions. Recent series have added robots through retroactive continuity, but nothing approaches the density seen in Star Wars.
Notable US companies include Tesla (Optimus), Figure, and 1X. In China, Unitree has received significant publicity; one of its founders appeared at the Abundance Summit. The latest Unitree R1 video has 11 million views. The R1 sells for $4,900, comparable to a cheap used car.
The R1 is extremely capable, but the critical work lies in software layers. Buyers will likely purchase algorithms separately to enable advanced behaviors. At $5,000, these robots become accessible to individuals with reasonable incomes, triggering an explosion of experimentation.
Alex identifies the critical threshold: humanoid robots assembling other humanoid robots. Once assembly costs approach zero, only raw materials and energy costs remain, driving unit costs toward near-zero. This enables physical labor too cheap to meter. Approximately two-thirds of the service economy involves physical labor.
At the Gigafactory, automation already exists in CNC milling machines and auto lathes that manufacture parts. The remaining human task is moving parts between machines and performing final assembly. Humanoid robots need only replicate this transfer and assembly function, making the problem significantly easier than literal robot manufacturing from raw materials.
At $4,900, robots move toward commodity pricing. The value layer shifts to software. Entrepreneurs can purchase robots and develop applications on top of them without corporate budgets or permissions. This mirrors the Raspberry Pi moment, enabling a new revenue engine through software applications.
Industrial use cases and dull, dirty, dangerous jobs will be automated first. This phase is expected to take roughly a decade before domestic applications such as gardening become common.
Export controls on frontier models can be viewed as immigration-style regulations on importing or exporting super intelligence. As robotic embodiment costs drop (primarily from China), similar import controls or national-security restrictions may emerge in the West, motivated by concerns beyond simple dumping.
Blitzy provides autonomous software development using thousands of specialized AI agents that analyze enterprise-scale codebases. The platform generates and pre-compiles code, delivering 80% or more of development work autonomously. Enterprises report 5x engineering velocity increases.
The Trump administration has expressed interest in investing in the robotics industry. Tesla and Figure's Brett Adcock have received major funding. The next topic covers robots in law enforcement.
The Orlando Police Department began using drones as first responders on June 17th. The system includes nine docks at different locations and 11 Skyo network drones. Skyo is the current US manufacturer; DJI drones from China face import controls for security reasons and cost three to ten times less than Skyo units.
Rick Smith, CEO of Axon (taser and body-cam manufacturer), holds the contract. A single-drone trial showed the drone reaching scenes faster than patrol officers about one-third of the time and providing useful information in 97% of cases. Drones will be stationed on buildings throughout cities for near-instant response.
Concerns about privacy and overreach exist. The system is designed not to look in windows or spy on residents. With good policy and governance regarding data retention, access, and use, the technology can save lives by enabling faster assessment of emergency situations.
In 2015, the largest insurance company in Mexico (30 million users) deployed drones because they reached accident scenes faster than ambulances. Drones scanned and mapped areas before vehicles were moved, providing complete information.
As robotic embodiment costs trend toward zero, the West may encounter regimes already familiar in China due to demographics. The Chinese Communist Party maintains officers on a per-block basis. In the US, first responders are distributed on a per-municipality or precinct basis. Drones enable a new model: geographic density comparable to fire hydrants, allowing remote activation for any block-level incident.
A suspect in Sacramento was holding a knife and police deployed a drone with a magnet attached to disarm the individual and make the scene safe. The drone pulled the knife from the suspect's hand while he appeared to be sleeping. This represents a niche example of drones being used for de-escalation in law enforcement scenarios.
During the Abundance Summit, Rick Smith from Axon demonstrated a taser-equipped drone, tasing a person on stage who was wearing a protective suit. The global drone market is currently valued at approximately $100 billion and is expected to grow significantly. Drones are being deployed extensively in Ukraine, with Eric Schmidt funding drone companies to support Ukrainian independence efforts.
Scientific American published research showing that crime rates in the US have decreased by half due to the deployment of first responders in the right place at the right time, largely enabled by GPS technology. Drone footage provides even greater accuracy for positioning responders. The video surveillance capabilities of drones are considered immediately impactful, with resolution continuing to improve.
Physical interventions like disarming individuals remain niche applications, but video capabilities are available now. Drones could become as abundant as fire hydrants at minimal cost. The concept was predicted in Minority Report, where police deployed spider-like robots to search buildings. China is currently deploying drones for first response more frequently than the US or the West.
Future developments may include new form factors such as spider drones and swarms of drones rather than individual units. As costs decrease, swarms will become more common. Public reaction will depend on how drones are armed and the guidelines governing their use.
Drones can rapidly deliver medical equipment to accident sites, including defibrillators in traffic-jammed locations. Electric vertical takeoff and landing aircraft could transport ambulance personnel. A representative from the Dutch police force described training hawks to drop mesh wires onto drones used by drug dealers, wrapping and tangling them.
Below a certain size, birds present a challenge as they attack drones. In Ukraine, strands of optical fiber from drone attacks create a significant environmental issue on the ground.
The most promising near-term application discussed is wildfire response, where drones could reach fires at inception and extinguish them before they cause hundreds of millions or billions in damage.
Flying cars exist but remain relatively sparse. Skies are expected to become densely filled with drones in coming years. Drone ambulances are viewed as a particularly compelling use case because no one would oppose technology designed to save lives. This could unlock regulatory barriers, airspace access, and prove overall efficacy.
A substack article discussed the trajectory toward knowing anything, anytime, anywhere through orbital satellites, 200 satellites from Will Marshall at Planet, aviation-layer imaging from flying cars and drones at centimeter and sub-centimeter resolution, and autonomous vehicles gathering terabytes of road data.
The comprehensive imaging capability raises privacy concerns but also offers the positive outcome that crime could be eliminated. When people know they are being observed, they behave differently. A desk bot with a camera causes people to act differently. The Lindbergh Foundation flew drones over elephant herds, and poachers stayed away when drones were overhead.
A story from China illustrated the difference in entrepreneurial culture. A person claimed that building a successful company has nothing to do with team or business plan, but rather depends entirely on what the government needs next. This suggests that loss of freedom and privacy may correlate with reduced innovation and entrepreneurship.
Switzerland voted to lift its ban on nuclear plants. After Fukushima in 2017, the country had phased out nuclear power completely, but is now reversing course. Switzerland currently has four aging reactors supplying 40% of its power, which were scheduled for shutdown but will instead be upgraded.
France has 57 operating reactors and succeeded by mass-producing a single reactor design rather than starting from zero with each project. The UK has nine reactors and Spain has seven. Ramez Naam Salem, described as one of the most extraordinary thinkers in energy, explained that France's success came from standardized reactor designs.
Europe faces an energy crisis after losing access to cheap Russian oil. Most countries underinvested in nuclear energy compared to France. Europe is not a major producer of solar PV and has a cultural orientation toward energy scarcity. Rising global temperatures and increasing power consumption are forcing a reevaluation of nuclear energy.
Europe will need to radically increase power consumption and production per capita. Nuclear energy offers an attractive path forward. Without increased energy capacity, Europe risks suffering under heat domes. Thousands of Europeans die annually from heat overexposure, a preventable outcome with better air conditioning and higher energy availability per person.
AI demand is transforming energy from an environmental issue into a capacity and commercial profit-driven issue. Every country must deliver sufficient energy capacity. In the short to medium term, fission through small modular reactors and generation 3 plants will provide base load until fusion becomes available.
The longstanding joke that fusion is always 50 years away no longer applies. Approximately 50 privately funded fusion companies have raised about $6 billion. Two US companies lead: Helion and Commonwealth Fusion.
Commonwealth Fusion, led by CEO Bob Mumgaard, is building a tokamak design and expects to construct its first 400-megawatt plant around 2032. Bob Mumgaard appeared on stage at the Abundance Summit.
Helion, backed by Sam Altman as the largest early investor since March 2012, had Altman serving as executive chairman until approximately two months ago when he stepped down to enable large-scale partnerships with OpenAI. On June 16, Helion received required Washington state regulatory approvals for its Orion fusion power plant, intended to supply Microsoft with 50 megawatts of power starting in 2028.
This would represent the first fusion plant coming online, albeit at a relatively small scale of 50 megawatts compared to gigawatt-level plants. Helion has raised approximately $1 billion at a $5.4 billion valuation.
The triple product metric, defined as the product of plasma density, confinement time, and plasma temperature, has shown steady progress over the past half century toward self-sustaining and net-positive fusion reactions. Progress has been continuous rather than the result of a single recent breakthrough, though high-temperature superconducting ribbon has provided economic advantages for certain reactor architectures.
A parallel exists between fusion and AI/ASI: both involve compressing matter or knowledge into a small volume to achieve a phase transition. Watching the triple product provides a figure of merit that could predict fusion's arrival from decades away, similar to how the Hutter Prize tracked compression of English Wikipedia and could have predicted large language models.
Clean, cheap, and dense energy is foundational to abundance because the cost of computation, desalination, transportation, manufacturing, and agriculture would reduce to essentially the cost of materials. Energy abundance would cause other scarcities to become negotiable. Energy correlates directly with GDP, health, and education outcomes.
Fission plants including small modular reactors and generation 3 plants are not expected online until the early to mid-2030s. Fusion plants at the 400-megawatt scale, whether from Commonwealth Fusion or scaled Helion systems, face similar timelines. Romez Naam Salem's company Aentik addresses near-term grid constraints by charging batteries between 1:00 a.m. and 6:00 a.m. and discharging during peak periods.
Space propulsion represents a potential killer app for compact fusion reactors. With abundant fusion energy, efficiency becomes less critical for storage mechanisms. Any clean storage method becomes viable even if inefficient, because the cost of electricity would be negligible. The same principle applies to rocket launches and powering space stations or moon bases.
Helion's pulse fusion device directly recovers energy to generate zero-carbon electricity. The process begins with deuterium and helium-3 fuel injected as gas into a formation chamber where it becomes superheated plasma. Capacitors charge magnets that invert the plasma's magnetic field into a toroidal shape.
Magnets fire sequentially, accelerating plasmas toward each other at over 1 million miles per hour. They collide and merge into one hot dense plasma. The magnetic field increases rapidly, compressing the plasma at over 10 Tesla. Fusion reactions convert matter into energy, strengthening the plasma's magnetic field.
The plasma's magnetic field pushes back on the machine's field, changing magnetic flux. Per Faraday's law, this induces current in the coils, which is directly recaptured as electricity and returned to the capacitors.
The architecture skips multiple steps found in conventional tokamak or inertial confinement designs that heat water to produce steam for turbines. Direct magnetic recovery potentially enables greater efficiency and more compact designs.
Both Helion and Commonwealth Fusion raised approximately $1 billion each. When the hosts attended MIT, fusion research budgets were in the tens of millions. The shift to commercial sector investment in hard science is considered unprecedented.
Nat Friedman, former CEO of GitHub, and Daniel Gross founded a $1.8 million incentive prize. Nat Friedman announced the Vesuvius Challenge on stage at the Abundance Summit in March 2023. The challenge addressed scrolls buried and burnt under Mount Vesuvius in 79 AD that were fully carbonized and unreadable without destruction.
The solution used CT scans to capture data and AI to read the ancient Greek text. After nearly 2,000 years, the scrolls have been read for the first time. The AI linearized and laid out 22 columns of ancient Greek text. Hundreds of additional scrolls remain to be read.
Alex described computational archaeology powered by AI as potentially transformative. The killer app of the singularity is seen as supercharging computational archaeology. Nikolai Fiorov, one of the fathers of cosmism, is cited for the philosophy that humankind's common task is to resurrect every human who's ever lived using technology. The Vuvius challenge represents the beginning of a larger arc of technological progress that may require completion of the singularity.
A thought experiment was presented: imagine scanning the Earth to gain fine spatial and temporal precision for every atom, then feeding the entire Earth state into an AI to recover Earth's history. Environmental DNA was mentioned as an example—DNA aerosolized from animals alive and dead, and DNA with a surprisingly long half-life in soil near cemeteries. With strong enough AI and scanning technology, large fractions of our past light cone could potentially be recreated.
Colossal's work extracting DNA from fossils to bring back the direwolf and woolly mammoth was referenced as a limited version of this approach.
AI excels at interpolating fragments where no human could reverse engineer original characters, such as in the Vuvius scrolls. This capability is not from regular LLM AI like Anthropic or OpenAI transformers. David Seagull, founder of Two Sigma and MIT board member, funds Project Open Athena, which provides AI compute resources for alternate AI versions not based on transformers that excel at these fragment interpolation problems.
Use cases include reverse engineering fragments of DNA to determine what happened in a room, or piecing together traces from world events where only one historical interpretation fits the scattered evidence. Core neural network AI built from the ground up is required rather than LLM approaches.
Incentive prizes can aggregate the best experts globally to solve problems even for non-technical founders. The $1.8 million Vuvius prize reportedly brought tens of millions of dollars or more in genius-level effort. This follows the X-Prize principle, where 30x the prize money is cumulatively spent to solve challenges.
Selim emphasized the prize model's power and the abundance thesis: challenges previously thought unsolvable are falling due to emerging technologies. The before-and-after images showing data recovered from previously unreadable scrolls were described as mindboggling.
Nick Leonard, CEO and co-founder of Voice Run, described the platform as a runtime and development platform for voice agents. Agents are built and configured in code rather than no-code platforms, giving developers total control and enterprises extensibility. Voice CLI was built first with cloud code, codecs, and open claw in mind. The platform supports building, deploying, testing, simulating, and analyzing voice agents at scale.
Grok 4.5 is based on a 1.5 trillion parameter V9 foundation model. Elon Musk claims he will iterate and release a new model every month for the rest of the year and start pre-training every month—not just distillation or post-training cycles, but entirely new pre-training runs from scratch.
Alex's earlier comments about Grok being put on life support drew spicy feedback. The goal is a competitive frontier beyond the current OpenAI-Anthropic duopoly, with Chinese open-weight models trailing. xAI's historical strategy has been brute force, and with SpaceX bringing more compute online, this approach combined with off-the-shelf algorithms could succeed.
Grok 4.5 combined with Cursor represents early movement. Cursor acquisition currently involves post-training existing models, but Elon has promised Cursor will become part of the pre-training recipe. If successful, Grok could reach the frontier through brute force compute.
xAI has tens of billions in dry powder and is bringing top talent from SpaceX and Tesla. The goal is to be number one and create Grok as the ultimate seeker of truth.
Elon does not see the race as two-horse between OpenAI and Anthropic. He views Google as the primary competitor due to their vertical monopoly and custom silicon providing 10-100x unlock. Google has already integrated their stack, while Anthropic is not perceived as close.
Verticalization is winning across industries including space and AI. Cursor data enables training on actual human-machine work processes rather than just human artifacts, creating a flywheel of better models, more usage, richer workflows, and further improvement.
The Cursor acquisition was viewed as a brain transplant for Grok's trajectory, analogous to ChatGPT incorporating Codex as its mainline model. The Cursor dataset provides reasoning traces from developers focused on code generation, enabling a leap toward the near frontier in code generation. However, xAI must still achieve its own recursive self-improvement loop, as post-training on developer traces has limits. Anthropic has historically been strong at recursive self-improvement while xAI has been weaker, though brute force compute may help close the gap.
Anthropic's flagship model Fable 5 has been offline for 15 days after the US government pulled it for national security concerns. Axios reports it may return within days. Secretary Lutnik credited Anthropic for working with the Pentagon and NSA on risks, though they have not yet signed off. Stripe tested Fable 5 by overhauling 50 million lines of codebase in a single day—work that would have taken engineers many months. The government is treating Fable 5 as a commercial AI model like a controlled munition.
If Fable 5 returns, the shutdown period will have denied Chinese organizations access to Anthropic's frontier or near-frontier models for reasoning trace distillation. This represents approximately one month of denied catch-up time. A less generous interpretation is that this marks movement toward the middle or endgame of recursive self-improvement, where frontier intelligence access becomes a regulated zone requiring US person status and strict export controls.
The shutdown period allowed critical systems to safeguard themselves against Fable 5. Three or four independent nonprofit and for-profit organizations stood up bulk vulnerability scanning using Fable 5. The pause did not decelerate AI overall and may have been a net accelerant, as Chinese models like GLM 5.2 gained traction among those seeking near-frontier capabilities, creating pressure on Chinese labs to catch up.
Regulatory risk is now a first-order variable for investors. Technical founders should avoid building products on a single model and ensure model swap capability.
Sonnet 5 was released to fill the gap while Fable 5 remains offline. AI demand vastly exceeds chip supply, driving Anthropic revenues higher. Only top use cases gain access, and preferential routing is occurring. Sonnet 5 offers mediocre capability at a high price point due to compute costs, yet demand remains strong.
The Sonnet 5 announcement appears bizarre. Traditionally, the Sonnet series represents distillation of Opus, with Haiku as further distillation of Sonnet, showing optimal price-performance tradeoffs. However, cost versus performance curves for agentic tasks show Sonnet 5 as a Pareto improvement over Sonnet 4.6, yet Opus 4.8 remains superior on a cost-performance basis.
Frameworks now exist to bounce between models while maintaining context and prompt history. The all-Claude stack offers convenience with MCP wrappers and connectors already built in, similar to Apple's integrated approach. Cost-effective alternatives require third-party context management platforms. This positions Anthropic as becoming the Apple of AI.
There may be branding strategy at play, with Fable and Mythos as new high-end models and Sonnet potentially repositioned as the new low-end or Haiku-level offering while Fable 5 remains inaccessible.
AI impacts health through prevention of heart disease. Dr. Don Musalem, Chief Medical Officer at Fountain Life, described heart disease as personal after losing her husband to sudden cardiac death when their daughter was five. Fifty percent of people die of heart attacks with no warning signs. AI-powered CT angiography with analytics finds detectable coronary artery disease in 88% of people tested, with 23% showing soft plaque invisible on traditional calcium score CTs. Multimodal testing including diagnostic laboratory studies paired with lifestyle recommendations enables early intervention.
Philip Johnston, co-founder and CEO of StarCloud, joined the discussion. StarCloud has raised about $200 million with its last round over a billion dollars. The company launched StarCloud 1, the first NVIDIA H100 GPU in orbit, and trained its first LLM in space. StarCloud launched in November 2025 on Falcon 9.
StarCloud started about two and a half years ago in January 2024 with a team of about 20 engineers based in Redmond, Washington, near the Starlink manufacturing facility. About half the team came from SpaceX and the rest from data center companies including AWS and Azure.
StarCloud 1 launched with five GPUs—two from Marani and three from NVIDIA, including the H100. The team trained nanoGPT, a very tiny model from Karpathy, and ran Gemma, DeepMind's cut-down version of Gemini, in space. Additional work is ongoing.
The team has moved beyond basic demonstrations to useful workloads. High power inference is being performed on SAR (synthetic aperture radar) data in collaboration with various Department of Defense entities. Doom was also run on the space platform.
The guests are recognized for being ahead of the orbital compute trend. Alex, Dave, and Peter have been vocal supporters, including Peter in his interview with Elon. The goal is building the Dyson swarm for future generations, with a preference for multiple Dyson swarms to avoid solar system-scale monopoly. The concept of Matryoshka swarms—swarms inside swarms inside swarms—was raised, noting Jupiter and Saturn have plenty of atoms available for disassembly.
Three launches are booked for the following year. StarCloud 2 is scheduled for January and will deliver approximately 100 times the power generation of StarCloud 1. It will feature the largest commercial deployable radiator in space. The spacecraft will carry H100s, Nvidia's black chip, Bitcoin mining ASICs, and an AWS Outpost on-premises server blade to run EC2 instances on orbit for customers.
StarCloud 3 is planned as the next major step—a 200 kW, 3-ton spacecraft designed to fit the Starship head dispenser form factor. A welded chassis version measuring about 6 meters long has been built. The design includes 100-meter deployables. Approximately 50 of these spacecraft can fit per Starship launch, providing roughly 10 megawatts of new compute capacity per launch. Frequent Starship launches are anticipated given the two Starship Gigafactories targeting three Starships per day.
When addressing investor concerns about competing with Elon and Google, the strategy is to operate more as an energy and infrastructure provider than a traditional cloud company. An agreement with Crusoe allows them to provide a box with power, cooling, and connectivity; Crusoe finances the chips and sells to any customers at any price while paying a rental fee similar to Equinix. In the early phase, SpaceX is expected to primarily serve its own (Cursor, Grok, xAI) workloads before potentially offering cloud services to companies like Anthropic. Pure infrastructure provision is viewed as further out for SpaceX.
Because StarCloud does not own launch, it will have a higher cost base than SpaceX but aims for a lower cost base than other hyperscalers. If OpenAI needs space compute, options include paying xAI/SpaceX, building their own satellites (likely far behind), or turning to the most advanced third-party provider once Starship cadence increases.
The company began with space-based solar power in mid-2023. A trip to Starbase, Texas before the first launch highlighted the scale of Starship development. Concepts from Asimov in the 1940s regarding space-based solar were seen as becoming reality. The break-even launch cost for space-based solar was calculated at around $50 per kilogram. However, 90-95% energy loss in transmission from space to ground made the economics challenging.
Noting that most new U.S. energy projects were being built to power data centers, the team shifted to bringing the data center to space. Recalculating for orbital compute produced a break-even launch cost of approximately $500 per kilogram. This formed the basis of a white paper released in summer 2024 and became the foundation of the company.
Launch capacity is currently the primary constraint. Bookings are being attempted on Relativity Space's first launch in January. Relativity Space is roughly the size of New Glenn and half the size of Starship. Two form factors may be needed: one for Starship and another compatible with Stoke Space, which is developing a reusable upper stage.
The business has two phases. While launch costs remain relatively high, operations focus on Falcon 9 and similar vehicles to provide edge and cloud services for other spacecraft, particularly DoD and Earth observation constellations. In a three-to-four-year timeframe as Starship production and cadence ramp, the focus shifts to competing directly with terrestrial data centers on energy cost. Three launches are booked for the next year, but Falcon 9 availability for 2028 is essentially gone due to government bookings of 20 launches.
The core near-term use case involves receiving raw satellite imagery (hyperspectral, SAR, etc.) and processing it on orbit rather than downlinking massive datasets. In a recent demonstration, SAR data was processed to identify tank coordinates, which were then downlinked. This avoids multi-day waits for ground station passes to downlink 100 gigabytes of data.
In 10-20 years, significant compute is expected in sun-synchronous orbits, beginning with Lagrange points or distinct solar orbits that trail or lead Earth. Dawn-dusk sun-synchronous orbit could accommodate roughly 10 terawatts of compute—about 20 times the entire U.S. power grid. Lower orbits (LEO) suffer 45 minutes of eclipse every 90-minute orbit, requiring batteries and increasing cost.
SSO is expected to become crowded. The ideal altitude for continuous sunlight is approximately 1,200 km. Most new compute capacity is projected to move to space within 10 years, though total space compute may still be under 5% at that point. In 20 years, space compute is unlikely to exceed 50% of global capacity, but a 50-year horizon could see 99% of compute in space.
Serious discussions are underway with various launch providers. If SpaceX cannot provide sufficient capacity, alternative solutions will be needed. The assessment is that SpaceX will hold a launch monopoly in five years but not in ten years.
Elon discussed direct-to-cell service at the All-In Summit, with phones using acquired spectrum expected to ship in around two years alongside new satellites. The goal is high-bandwidth connectivity allowing video streaming anywhere. SpaceX controls the full stack: launch, satellites, spectrum, and increasingly compute. It was noted that SpaceX may eventually manufacture its own phones rather than relying solely on partners. Rumors of a Tesla phone and potential acquisitions of T-Mobile or Charter Communications were mentioned.
Starlink is projected to generate hundreds of billions in free cash flow over the next five to ten years and could make SpaceX the world's most valuable company. Valuations of $10 trillion by 2030 scaling toward $100 trillion were discussed.
Starlink satellites operate at approximately 460 km in 50-degree inclination orbits that do not reach the poles. AI/data center satellites are better suited to higher altitudes (ideally 1,200 km) for continuous sunlight. The 400-500 km band will become extremely crowded, but AI satellites will occupy different real estate.
At 1,000 km, most debris would de-orbit within about 50 years. Historical context was provided regarding the 1970 U.S. government dispersal of 400 million needles at 3,000 km altitude to create an artificial ionosphere for radio reflection; all have since de-orbited due to solar and lunar gravitational interactions.
Cooling uses liquid through aluminum deployable radiators. The International Space Station has proven radiator technology for 20 years, but those systems are expensive and heavy. The company's design achieves 10 times less mass per watt of dissipation and approximately 100 times lower cost per watt than ISS radiators. The first large commercial deployable radiator will fly in January.
Masayoshi Son's statement that energy represents only 7% of compute cost was addressed. While chip costs currently dominate, energy plus infrastructure together represent at least 30% of total cost. A prior post noted that Son sold all Nvidia stock in 2019.
Building new terrestrial energy projects faces 5- to 10-year lead times due to permitting requirements. Space-based deployment can occur much more rapidly. Even at energy and infrastructure break-even, the approach makes economic sense because it delivers approximately 10 times lower costs on both energy and infrastructure.
Space systems eliminate the need for batteries, cooling towers, big chillers, and backup power. The only infrastructure required is a dirt-cheap radiator. The remaining cost factor is launch, which continues trending rapidly lower.
Conservative projections 20 to 30 years out point to exponential growth on the lunar surface. Elon Musk's statement that Optimus represents the von Neumann probe concept is viewed as accurate. Placing 100,000 Optimus units on the lunar surface to establish a self-replicating factory would trigger hyper-exponential development curves.
Mass drivers on the moon are expected within roughly 20 years. Timelines significantly shorter or longer than 20 years would both be surprising.
Rocket Lab, founded by Sir Peter Beck, has reached a $64 billion valuation. The company built Electron, a smaller launcher with 91 successful flights and high reliability. Its next vehicle, Neutron, is approximately half the size of Falcon 9, features a reusable first stage, and targets first launch by the end of the year.
Rocket Lab is acquiring Iridium, a 66-satellite constellation operating at 780 km altitude that began commercial service in November 1997. The original design planned 77 satellites (atomic number of iridium); the name was retained after reduction to 66 satellites rather than changing to dysprosium.
Iridium holds 10.5 megahertz of globally coordinated L-band spectrum licensed through the ITU. This spectrum access across all countries represents the core strategic value. Vertical integration—combining launch, satellite manufacturing, spectrum, and operations—follows the SpaceX model and is becoming the dominant structure in the new space economy. The acquisition will likely be completed using Rocket Lab stock, capitalizing on its high revenue multiple while adding cash-generative, profitable assets.
Laser links are expected to become the primary communications method within 5 to 10 years, though radio will retain niche utility due to lower cost and lack of gimbal requirements. Current hardware already incorporates laser terminals: three gimbal lasers fly on the second satellite launching in January. The next 25 satellites will carry two Starlink PLA plug-and-play laser terminals plus one SDA-compliant laser for government satellite connectivity.
Laser communications offer the additional advantage of being unregulated.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
