Sonnet 5 Drops, Fable 5 Will Return & Fusion's First Plant Gets Licensed w/ Philip Johnston | #268

Anthropic's Claude Opus 5 was pulled offline for approximately 15 days after the US government intervened over national security concerns, with Secretary Lutnick crediting Anthropic for working through the Pentagon, though the NSA had not yet signed off. Axios reported the model could return within days. Stripe reportedly used Claude Opus 5 to overhaul 50 million lines of code in a single day, work that would have taken engineers many months. Alex argued the shutdown may have denied Chinese organizations access to frontier reasoning traces for roughly one month and suggested the episode could mark a phase change toward permitting regimes for frontier AI analogous to Cold War bloc systems, though he caveated it likely did not materially decelerate AI development overall. Anthropic released Sonnet 5 during the outage period. Alex assessed it as a Pareto improvement over Sonnet 4.6 but inferior to Opus 4.8 on a cost-performance basis, and Dave characterized it as mediocre capability at a high price point. Alex assessed the AI frontier as effectively a duopoly between OpenAI and Anthropic, with Grok 4.5 combined with Cursor showing early signs of XAI competing at the frontier. Musk has argued the primary race is against Google because Google's custom silicon provides a 10 to 100x training advantage and Google holds a vertical monopoly across AI hardware and software.

Helion received regulatory approval from Washington state on June 16 for its Orion fusion power plant, which is contracted to supply Microsoft with 50 megawatts of power starting in 2028. The company has raised approximately one billion dollars at a 5.4 billion dollar valuation. Helion's architecture accelerates magnetically propelled plasmas toward each other at over one million miles per hour under a magnetic field exceeding 10 Tesla, and its key differentiator is direct electrical recovery from the fusion plasma via Faraday's law, bypassing the conventional steam turbine chain. Alex described this as making Helion the closest fusion startup to a radical compactification model, potentially enabling deployment at the township or city level. Sam Altman stepped down as executive chairman roughly two months ago, clearing the way for Helion to pursue large-scale partnerships with OpenAI. The 50 megawatt output is modest compared to gigawatt-scale plants, and small modular reactors and Gen 3 fission plants are not expected online until the early to mid 2030s. Approximately 50 privately funded fusion companies have raised about 6 billion dollars collectively.

Morgan Stanley raised its projection for Chinese humanoid robots from 14,000 to 26,000 to 50,000, with a further projection of 500,000 by 2030. Elon Musk projects tens of millions to 50 million robots by 2030 and billions in the early 2030s. The Unitree R1 sells for 4,900 dollars. Alex argued that approximately two thirds of the service economy constitutes physical labor and that physical labor costs will eventually be driven to near zero once robots can assemble other robots. He also warned that Western organizations are woefully behind Chinese organizations in driving down unit costs of general purpose robotic embodiments and that import controls on Chinese robotic hardware are likely to emerge as costs fall.

Philip Johnston, co-founder and CEO of Star Cloud, described the company as a roughly 20-person team based in Redmond, Washington, founded in January 2024, that has raised approximately 200 million dollars at a valuation exceeding one billion dollars. The company launched Star Cloud One on a Falcon 9 in November 2025 carrying five GPUs including the first Nvidia H100 placed in orbit, and subsequently trained what Johnston described as the first LLM in space using Andrej Karpathy's nano GPT. Star Cloud Two is booked for January and will carry H100s, Nvidia Blackwell chips, Bitcoin mining ASICs, and an AWS Outpost blade, generating approximately 100 times the power of Star Cloud One. Star Cloud Three is a 200 kilowatt spacecraft roughly six meters long, and approximately 50 units can fit per Starship, equating to about 10 megawatts of new compute capacity per Starship launch.

Johnston's business model positions Star Cloud as an energy and infrastructure provider rather than its own cloud, with Crusoe paying a rental fee and financing the chips itself. He targets approximately 10 times cheaper energy and infrastructure costs compared to terrestrial data centers, arguing that energy and infrastructure together represent at least 30 percent of total data center cost and that orbital compute eliminates batteries, cooling towers, large chillers, and backup power. New energy projects on Earth face five to ten year permitting lead times, which Johnston cited as a structural advantage for orbital deployment. Johnston's radiator technology achieves 10 times less mass per watt and approximately 100 times less cost per watt compared to the ISS radiator, has passed thermal vacuum testing, and is scheduled to fly in January. A contract was signed with SpaceX to put two Starlink laser terminals and one SDA-compliant government laser terminal on each of the next 25 Star Cloud satellites.

Johnston's long-range predictions are that in 10 years most new compute capacity will be deployed in space but total space compute will still be under 5 percent, and at roughly 50 years 99 percent of all compute will be in space contingent on a Dyson Swarm-type buildout. AI compute satellites would ideally orbit at 1200 kilometers, the lowest altitude with no solar blackout throughout the year, while Starlink operates at 400 to 460 kilometers, meaning the two use cases do not compete for the same orbital band.