EWS/SpaceX is framed as a multi-engine compounding machine: satellite compute, Starlink connectivity, launch, defense, AI, and direct-to-cell could turn space infrastructure into a trillion-dollar-plus revenue stack if the execution moat holds. The core thesis is that space data centers are technically hard but not physics-limited, with cooling, solar, and debris management the main hurdles, while demand from inference, agents, robots, and sovereign AI keeps rising. The model work pushes valuation into 2028-2030 scenarios, with sandbagged outputs from $527 by 2028 to $926 by 2030, and aggressive upside tied to Starship launch economics and Starlink penetration. The biggest risks and open questions are whether space compute can scale reliably by 2027, whether debris and thermal systems can be solved, and whether supply locked up in Vera Rubin and Starlink is enough to stay ahead of competitors.
EWS was said to be on pace for 2 GW all-in by year-end, with current expectation closer to 2.5 GW in a best case; if all deployed compute were rented out, the run-rate would be north of $100B in ARR by year-end.
On the Tennessee buildout, a “Minihard” building was described as taking a country’s worth of GPUs in a week, with GB300 monetization modeled at 25-50M per rack at 50% utilization, versus 100-200B if the same compute is used internally.
The EWS mix was put at 10% training and 90% inference, with inference described as the rented product; Google and others were said to have paid 30-50B per GW, and the next big EWS customer was floated as Microsoft at north of $1B/month.
Meta’s Muse was cited as having 1M users and already tapped out on compute, used as evidence that inference demand is not collapsing; Vera Rubin was said to raise token density so the same space and power produce far more tokens.
Elon was said to have secured 30-40% of Vera Rubin supply, which was framed as enough to make EWS the top hyperscaler on the planet because rivals cannot catch up if they cannot source the chips.
The rationale for Nvidia supplying EWS heavily was that high utilization across racks and GPUs raises monetization across the stack; the target was closer to 40%-50% utilization, not 10%-30%.
SpaceX was framed as having four engines: rocket launches, global comms, cell-phone connectivity, and AI compute, with Starlink and Star Shield also tied to private government and military communications spending.
Space-based compute was argued to be engineering-hard but not physics-limited; the main bottleneck was cooling via radiators, not solar power, with radiators and power generation identified as the two hardest problems.
AI1 was described as materially larger than a V3 satellite, roughly twice the weight while fitting in the same area compact, with big radiators as the key addition and long solar strips as the biggest component.
Space solar was said to generate about 5x the energy for the same panel capex because it avoids atmosphere and can point at the sun all the time; in orbit, solar was described as scaling from tens of kW to 100-200 kW.
Kesler syndrome was raised as the debris-risk objection: collisions with old satellites or natural objects can create a domino effect of more collisions, though active debris avoidance was said to be already routine, sometimes three times per day per satellite.
The AI/robotics use case was pushed hard: Optimus tied to Starlink could send live video to a brain in the sky, implying hundreds of thousands to millions of tokens versus roughly 100 tokens for a normal LLM request; a 20 ms delay over a robot was called far better than routing through a ground data center.
MCP was defined as model context protocol, likened to plugins/connectors that let agents pull, scan, digest, and act on data without manual API work; less than 1% of newsletter users had signed up, but MCP users were said to generate 15x more usage per user.
The valuation work ran from crude models to Monte Carlo: an early-2026 model with EWS, Starlink, Rocket Launch, XAI, and data centers mapped to a $2.8T result and a $210 stock price, while a broader model excluding government business mapped to $1,383 by 2030.
Morgan Stanley’s $3.5T revenue call for SpaceX in 2040 was contrasted with an Elon target of $3.5T in 2033; separate launch math claimed $800k to build a rocket versus $13.3M lifetime return, and a 200-launch Starship model implied about $999B in lifetime revenue, roughly $1,500 per share on the notes’ framing.
A Monte Carlo over the whole stack produced a sandbagged $527 target by 2028, or 3.61x from today, and $926 by 2030, or 6.18x, with a stated total TAM of $83.5T after adding humanoids, SpaceX, robo-taxis, communications, launch, and military.
The final stance on buying SpaceX was that 1.75T-2T is bond-like exposure to Starlink and the rest is upside optionality; Starlink alone was placed around 170B by 2030, while direct-to-cell at 30%-40% global penetration was framed as a $5T market.