🤚 The Open-Palm Ascension

SpaceX has decided that Earth is no longer a sufficiently ambitious location for a data center. The company unveiled AI1 — a satellite constellation designed to put AI compute infrastructure in orbit — and the specifications read like someone dared an engineer to solve every data center problem by leaving the planet.

The numbers: each AI1 satellite deploys a 70-meter solar wingspan, generates 150 kilowatts peak power, sustains 120 kW of continuous compute, and radiates waste heat into the 2.7-Kelvin vacuum of space through 110 square meters of thermal radiators. Each satellite is roughly equivalent to one NVIDIA GB300 rack — except this rack orbits the Earth at 17,000 miles per hour and never needs to negotiate a power purchase agreement.

SpaceX plans to pack 30 to 50 of these satellites into a single Starship launch. One rocket. Fifty server racks. Deployed in orbit. No land acquisition, no permitting battles with local communities, no waiting three years for the utility company to run a new power line to your building in rural Iowa.

The timeline: two prototype satellites in early 2027, with the Gigasat factory in Bastrop, Texas — spanning over 1,000 acres and 11 million square feet — manufacturing satellites, solar panels, and SpaceX’s custom D3 chip at industrial scale. The target is 1 gigawatt of orbital compute by end of 2027, with aspirational 10x annual scaling thereafter.

👐 The Two-Handed Reality Check

Before you dismiss this as Elon Musk’s most expensive thought experiment, consider the customer list.

Anthropic — the company behind Claude — is reportedly paying $1.25 billion per month for access to SpaceX’s compute infrastructure. That’s $15 billion annualized. Anthropic has also agreed to purchase all capacity from SpaceX’s terrestrial data center, Colossus 1, which offers more than 300 megawatts of computing capacity. This is not a speculative partnership. This is a revenue-validated infrastructure play with one of the world’s most compute-hungry AI labs writing checks that would make a sovereign wealth fund pause.

The D3 chip deserves particular attention. Rather than using traditional radiation-hardened processors — which are expensive, slow, and generations behind commercial silicon — SpaceX designed a custom chip using TSMC’s commercial N5/N3 foundry nodes with triple modular redundancy and error-correcting memory. The philosophy: use cheap, fast commercial chips and engineer around the radiation problem rather than accepting the performance penalty of space-grade hardware.

This is why the episode’s provocative claim — chips beat rockets — carries weight. The bottleneck for orbital compute was never launch capacity (SpaceX solved that). It was never power (the sun solved that). It was processing capability per watt per dollar in a radiation environment. And SpaceX appears to have cracked it by refusing to play by the space industry’s rules.

🌿 The Gentle Awakening

Meanwhile, on the ground, a quieter revolution is underway.

Zhipu’s GLM 5.2, China’s leading open-source AI model, now sits within a percentage point of Anthropic’s Opus 4.8 on closely watched agentic benchmarks — at roughly one-fifth the cost. Unlike DeepSeek, which was dismissed as a chatbot novelty, GLM 5.2 excels at agentic work: planning, coding, testing, and the kind of multi-step reasoning that actually matters for enterprise deployment.

Chinese-developed open-source models now account for approximately 30% of all open-model downloads globally. Alibaba’s Qwen family alone has surpassed Meta’s Llama in cumulative HuggingFace downloads. The ecosystem includes Qwen, DeepSeek, GLM, Kimi, MiniMax, StepFun, and Baichuan — a bench deep enough that no single model’s failure would slow the movement.

The strategic implications are uncomfortable for American AI labs. While the U.S. government restricts access to frontier models like GPT-5.6 Sol and Fable 5, China is open-sourcing models that perform within striking distance of those restricted systems. The export controls designed to maintain American AI supremacy are creating a market vacuum that Chinese open-source models are enthusiastically filling.

👑 The Crown Verdict

Peter Diamandis’s episode captures something essential about this moment: the AI infrastructure race has split into two theaters, and the West is fighting on both fronts simultaneously.

Theater One: the vertical. SpaceX is building upward — literally launching compute into orbit, solving power, cooling, and deployment constraints by leaving the atmosphere entirely. If the economics hold, orbital data centers could provide compute at scale without the land-use battles, energy grid limitations, and community resistance that plague terrestrial facilities. The $130 billion problem of data center opposition vanishes when your data center is 550 kilometers above the nearest NIMBY petition.

Theater Two: the horizontal. China is building outward — flooding the global market with capable, cheap, open-source models that undercut Western pricing by 80%. You don’t need access to a restricted American frontier model when an open-source Chinese alternative does 95% of the job at a fraction of the cost.

The irony is exquisite. The United States is simultaneously building the most ambitious compute infrastructure in human history and restricting access to the models that would run on it, while China cheerfully offers the world a nearly-as-good alternative with no restrictions, no government approval process, and no velvet rope.

The AI race isn’t just about who builds the best model anymore. It’s about who builds the best infrastructure, who controls access, and who offers the best value. As of this week, three different answers are emerging — and none of them agree.

Inspired by The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China’s #1 Open Model | #266 by Peter Diamandis.

Your orbit is showing. Compute wisely.