Uploaded July 2026 | Updated September 2026, 6 hours ago
SpaceX/xAI has become the only place hyperscalers can rent instantly usable GPU capacity. Microsoft claims ~$650bn in signed AI demand it can't serve for 1–2 years due to power/datacenter constraints; Google needed "bridge capacity" for the same reason. xAI already has three leasing deals — Google (~$1bn/month, fully ramped by October), Anthropic + Reflection AI (~$1.4bn/month combined, already ramped) — and the host predicts 1–2 more deals before year-end, likely from Google again, Anthropic again, or Microsoft (not OpenAI, for obvious Elon reasons).
The moat is power, not chips. Citing Epoch AI aerial analysis: ~440,000 working chips at Colossus 2 (330k B300s + 110k B200s), possibly 45–70% of all active premium Nvidia chips on Earth. Energy comes from ~60 mobile gas turbines (~1.65 GW, heading to 2 GW), the APR Energy acquisition, a Caterpillar/Solar Turbines JV with a $3bn order book, and 12 Doosan 380 MW turbines on order — potentially 6 additional GW over three years while everyone else is stuck in utility queues.
The financials: 2026 AI leasing revenue ~$47bn, Q4 exit run rate $72–82bn/year; 2027 bull case ~$90bn with a $120bn exit run rate. Q4 2026 base case: ~$18.4bn consolidated quarterly revenue (neck-and-neck with Tesla), $4.4bn operating income ($2.3bn bear / $6.8bn bull), near-breakeven FCF despite $13bn/quarter capex — the trick being Amazon-style depreciation (chips over 5–6 years) so only ~20% of capex hits earnings annually. Demand evidence: everyone hitting session limits on Claude/Grok/GPT means inference supply, not demand
Follow Brian Wang on NextBigFuture.com
Follow Brian Wang on X.com/@NextBigFuture
Patreon - Next Big Future
This channel brings you the most important science and technology breakthroughs and analysis at least once per week.
SpaceX/xAI has become the only place hyperscalers can rent instantly usable GPU capacity. Microsoft claims ~$650bn in signed AI demand it can't serve for 1–2 years due to power/datacenter constraints; Google needed "bridge capacity" for the same reason. xAI already has three leasing deals — Google (~$1bn/month, fully ramped by October), Anthropic + Reflection AI (~$1.4bn/month combined, already ramped) — and the host predicts 1–2 more deals before year-end, likely from Google again, Anthropic again, or Microsoft (not OpenAI, for obvious Elon reasons).
The moat is power, not chips. Citing Epoch AI aerial analysis: ~440,000 working chips at Colossus 2 (330k B300s + 110k B200s), possibly 45–70% of all active premium Nvidia chips on Earth. Energy comes from ~60 mobile gas turbines (~1.65 GW, heading to 2 GW), the APR Energy acquisition, a Caterpillar/Solar Turbines JV with a $3bn order book, and 12 Doosan 380 MW turbines on order — potentially 6 additional GW over three years while everyone else is stuck in utility queues.
The financials: 2026 AI leasing revenue ~$47bn, Q4 exit run rate $72–82bn/year; 2027 bull case ~$90bn with a $120bn exit run rate. Q4 2026 base case: ~$18.4bn consolidated quarterly revenue (neck-and-neck with Tesla), $4.4bn operating income ($2.3bn bear / $6.8bn bull), near-breakeven FCF despite $13bn/quarter capex — the trick being Amazon-style depreciation (chips over 5–6 years) so only ~20% of capex hits earnings annually. Demand evidence: everyone hitting session limits on Claude/Grok/GPT means inference supply, not demand
Follow Brian Wang on NextBigFuture.com
Follow Brian Wang on X.com/@NextBigFuture
Patreon - Next Big Future
This channel brings you the most important science and technology breakthroughs and analysis at least once per week.
