Deep Dive
AI capex with sub-one-year returns flips the script
The street fixated on SpaceX's AI infrastructure spending, but CFO Brett revealed the economics that matter: less than a one-year payback period. That reframes capex from risky bet to straight cost of goods sold. ARK's Daniel emphasized this unlocks the case for aggressive reinvestment—you're not speculating on AI, you're deploying proven infrastructure with immediate returns. The real number is monetization: SpaceX quoted $30-50B per gigawatt, roughly double the $15B industry average. That spread exists because hyperscalers are desperate for any compute capacity they can access, and SpaceX has launch cadence and power density competitors can't match. Elon signaled this wasn't short-term either—the $100B annual recurring revenue target by year-end signals they're leaning into the business model hard, contradicting earlier comments suggesting terrestrial compute was a temporary play.
Starship's upper-stage catch changes the unit economics forever
Flight Test 14 hitting orbit next month isn't just a engineering milestone—it's the linchpin for SpaceX's entire $28.5 trillion addressable market. The mission includes three firsts: reaching orbit, deploying V3 commercial satellites with significantly more bandwidth, and attempting to catch the upper stage mid-air. That third piece matters because full reusability is what drives launch costs from today's ~$1,000 per kilogram down to sub-$100 per kilogram long-term. Without catching the booster, you lose half the rocket. With it, you unlock orbital data centers—the whole AI infrastructure thesis breaks without that cost curve. On Starlink specifically, V3 satellites enable a 20x bandwidth improvement per Starship flight compared to Falcon 9 runs. At scale they plan 60 V3 sats per launch; that's an 8% bump to all bandwidth SpaceX has ever launched, from a single mission. Starlink's already got 12 million subscribers growing rapidly—V3 capacity could accelerate adoption significantly.
Power becomes the hard constraint—20 gigawatts in 24 months
Scaling AI infrastructure hits a wall every hyperscaler knows: power access. SpaceX committed to 2 gigawatts by end of 2025 and 5-10 gigawatts (closer to 10, per Elon) by end of 2026. Building the compute infrastructure itself is the easy part—engineers who've built rockets laugh at data centers. The bottleneck is power generation: 10 gigawatts of compute requires roughly 20 gigawatts of electrical capacity—20 large nuclear reactors' worth. CoreWeave, the incumbent, is targeting 8 gigawatts by 2030. SpaceX wants to do that in under two years. Elon personally acquired a gas turbine company weeks ago, signaling the strategy: near-term power from mobile generation while securing long-term grid access. The question Daniel surfaced: how do they actually secure that much power in that timeframe when it's scarce globally? The fact they're committing publicly suggests they either have deals locked or the engineering confidence to move faster than anyone expects.
Starlink's direct-to-cell gets distributed infrastructure from customers
Starlink's direct-to-cell service—internet from satellite straight to phone—works great in open sky but fails in buildings and dense urban areas. SpaceX's solution: when customers buy Starlink broadband and install a dish, bundle it with hardware that acts as a localized cell tower. Customers fund the infrastructure rollout; SpaceX gets the coverage density. This is Tesla's FSD playbook copy-pasted into satellite infrastructure. Tesla gathered billions of miles from customer vehicles to train its AI system, building a data advantage competitors can't replicate. SpaceX is doing the same: every Starlink dish becomes a potential mobile cell node. The spectrum SpaceX acquired from Echostar will fuel this network. It's elegant because it solves the chicken-and-egg problem of mobile coverage—customers pay for broadband anyway, the cell tower capability is incremental hardware cost. As the network densifies, service improves, adoption accelerates.