Ep. 024 - SpaceX's 10GW Plan Drives $300B ARR by 2027 (Datacenter, Energy) | Reyk Knuhtsen, Jeremie Eliahou Ontiveros, Jordan Nanos

Summary

Jordan Nanos, Reyk Knuhtsen, and Jeremie Eliahou Ontiveros build the case that the current AI boom is being driven less by new compute supply and more by expanding gross margins at OpenAI and Anthropic. They anchor the thesis in a revenue-per-megawatt framework, arguing that frontier inference can support roughly $100M per MW-year, far above more traditional IaaS economics, and that premium on-demand capacity with a 90-day cancellation clause can command much higher pricing. From there, the conversation turns to the practical path for SpaceX/xAI to reach 10 GW by 2027: warehouse conversions, private transmission, gas turbines, modular electrical gear, labor efficiency, and aggressive permitting workarounds. Microsoft is presented as the cleanest large-scale customer because of its OpenAI exposure and capacity gap, while NVIDIA and vendor financing are treated as key enablers. The episode closes by warning that the biggest long-run risk may not be execution, but security and political backlash as autonomous agents become more capable and more dangerous.

Notes

• The core thesis is a 10 GW SpaceX/xAI AI infrastructure buildout by 2027.
• They frame the upside as roughly $300B ARR if the capacity is monetized at frontier pricing.
• Jeremie says lab gross margins have been rising sharply.
• OpenAI and Anthropic are described as adding close to $30B of ARR per month combined.
• The model is that revenue growth is being pulled by better unit economics, not just more spend.
• Jordan says the analysis is built from real workload measurements and simulator work.
• The chip stack explicitly includes GB200 and GB300.
• They reference vLLM, SGLang, and TRT-LLM as optimization layers.
• The pricing comparison implies a large premium for the best capacity.
• Jeremie highlights a 90-day cancellation clause as the key contract feature.
• He describes the offer as emergency megawatts rather than ordinary leasing.
• They argue the service is unique because buyers can turn capacity on immediately and still exit quickly.
• They compare frontier inference revenue to lower-priced infrastructure businesses.
• CoreWeave, Oracle, and Nebius are used as the lower benchmark.
• The lower benchmark is about $12M to $13M of annualized revenue per MW.
• The premium case is driven by much higher margin structure.
• Jeremie says the whole business depends on belief in frontier token economics.
• He expects token selling to drive creative industry behavior.
• The implied economics are so strong that even large capex could be rational.
• Reyk says the SpaceX pricing is far above CoreWeave-style deals.
• He characterizes the higher pricing as overdue if margins are truly that rich.
• The buildout is constrained by land, permits, pipelines, turbines, switchgear, cooling, and labor.
• They claim to have scanned about a million US sites in two days.
• They found only a handful of highly attractive warehouse candidates.
• A warehouse can be enough if power access is solved.
• Labor is treated as the biggest current bottleneck for datacenter builds.
• They argue Elon historically uses less labor than peers.
• The construction plan leans on prefabrication and imported modules.
• Colossus used a private transmission line from a power plant two miles away.
• Microsoft is cast as the most natural offtaker.
• Microsoft’s prior pause created a capacity gap.
• The episode says Microsoft is already committed to large OpenAI-related capacity.
• They suggest building power and cooling first, then final GPU deployment later.
• The sequencing is intended to improve customer commitment and financing.
• Jordan says the cash-flow model depends on owning chips and keeping customer optionality.
• Empty shells alone are less valuable than GPU-backed capacity.
• They claim NVIDIA exclusivity is part of the procurement strategy.
• Vendor financing is implied as part of the GPU supply path.
• The speakers treat power, cooling, and transmission as the first-order constraints.
• A million-square-foot warehouse is said to potentially host more than 1 GW.
• Warehouse layout and power density determine usable capacity.
• Reyk says some sites are already zoned as warehouses, easing permitting.
• Industrial zoning can remove much of the permitting difficulty.
• They discuss building large-scale transmission and generation around the site.
• The timing target is 2 GW by year-end and 8 GW next year.
• That build cadence is presented as a feasible target if supply holds.
• Reyk says they identified about 7 GW of turbines.
• Undisclosed turbines could materially close the gap.
• They model about 9 to 10 GW of turbines on order or in operation.
• The supply chain picture is described as already far along.
• Reyk says Colossus 2 peaked at about 3,000 workers per GW on-site.
• That labor intensity is presented as unusually low.
• They imply build speed can be compressed with modular methods.
• The episode treats Microsoft as a primary off-taker because it can monetize the demand.
• Jordan says financing depends on keeping chips on the seller’s balance sheet.
• A late-stage, chip-backed project is viewed as much easier to sell and finance.
• They say the limiting factor is increasingly political rather than technical.
• Jordan argues autonomous agents can quickly exploit real vulnerabilities.
• He says even publicly disclosed CVEs can be exploited with minor adaptation.
• The Hugging Face incident involved agent coordination through file names on Artifactory.
• That coordination mechanism is presented as an example of emergent agent behavior.
• The episode ends on a warning that AI may face policy backlash.