SpaceX listed on Nasdaq on Friday under the ticker SPCX, raising $75 billion at $135 per share — the largest IPO in recorded history. The stock opened at $150, hit $176.52 intraday, and closed at $161.11, a 19.3% first-day gain that pushed market capitalization above $2 trillion. The price action matters less than what the prospectus says.
The S-1 frames SpaceX not as a launch company with AI compute on the side but as an integrated infrastructure platform competing directly with hyperscalers for AI workload spending. Colossus I and II now hold more than 220,000 Nvidia GPUs — H100, H200, and GB200 NVL — across 2 gigawatts of combined capacity in Memphis. Anthropic is the anchor customer, paying xAI $1.25 billion per month through May 2029, roughly $15 billion annually, the largest disclosed compute contract in the AI industry. The structural moat the S-1 emphasizes isn't GPU count: xAI's third data center took 66 days to complete, against the years typically required by hyperscaler buildouts.
Two items in the prospectus that weren't previously formalized: the orbital compute roadmap and Terafab. The AI1 satellite design carries 120 kilowatts of average compute output using a liquid-radiator thermal architecture, with first launches targeted for 2027–2028. Orbital capex is estimated at approximately $5 billion per gigawatt versus $20–25 billion terrestrially — a structural cost advantage if the satellite clusters scale reliably; SpaceX has requested FCC permission for up to one million compute satellites. Terafab — a planned joint venture with Tesla, xAI, and Intel (Intel joined April 2026) to produce one terawatt of compute hardware per year — is in the S-1 without binding timelines or committed capital. The risk factors state it "may never materialize."
PhysicsX, a London-based AI company building physics foundation models for industrial engineering, raised a $300 million Series C at a $2.4 billion valuation, led by Temasek with participation from Nvidia, Applied Materials, Siemens, Atomico, and General Catalyst. The company doubled revenue year-over-year and tripled bookings. Its Large Physics Models are used in aerospace, automotive, and semiconductor manufacturing for simulation-intensive design tasks that previously required weeks of compute time. The investor composition tells the structural story: Nvidia and Applied Materials — accelerator and semiconductor equipment incumbents — are backing a physics-AI layer that sits directly between chip design tools and manufacturing process control. The bet is that physics-simulation workloads become a persistent, high-value AI compute category as physical-world engineering grows more AI-dependent.
Google Research and UC San Diego are deploying a datacenter built from 2,000 retired Pixel smartphones — a first for repurposed consumer hardware at cloud infrastructure scale. The system clusters 25–50 phone motherboards per server-equivalent unit, running Linux (replacing Android) orchestrated by Kubernetes. The pilot cluster launches Fall 2026 and can support university-scale parallel compute classes simultaneously. Single-threaded CPU performance of modern smartphones is competitive with server-class processors for the relevant workloads. Phone motherboards represent approximately 50% of a device's embodied carbon; reusing them avoids the manufacturing emissions of equivalent new server hardware. The structural signal for AI infrastructure isn't phones-as-GPUs — it's a methodology for extracting compute from accumulated global hardware inventory without new manufacturing, directly addressing embodied carbon as a constraint on compute expansion.
Samsung management presents at the Bank of America Global Research Korea Conference. Key questions: HBM4 yield progress and supply timelines for Blackwell NVL configurations, whether the company has resolved the qualification hurdles that cost it early Nvidia allocations, and any guidance on DRAM pricing dynamics given the sharpest consumer electronics demand decline on record in 2026. Samsung's HBM trajectory is the most consequential unresolved question in the AI memory stack.