Ep. 024 - SpaceX's 10GW Plan Drives $300B ARR by 2027 (Datacenter, Energy)

Summary

The episode makes the case that rising AI lab gross margins are driving higher revenue per watt, which in turn supports premium pricing for scarce compute and underpins SpaceX’s proposed 10 GW datacenter plan. The speakers say their model combines real inference workloads on GB200/GB300 hardware with a simulator that tracks GEMMs, collectives, KV cache, bandwidth, and future Rubin-class hardware. They argue the market already supports very high economics—on the order of $100M per MW-year at the top end—with standard IaaS GPU pricing far lower unless buyers pay for immediacy or flexibility. The second half shifts to execution: warehouse conversion, nearby gas pipelines, turbines, mobile power, permitting workarounds, labor shortages, and sourcing from China are all framed as the practical bottlenecks. Microsoft is presented as the clearest near-term customer path for OpenAI-linked demand, while low-risk 90-day cancellation terms and vendor financing are described as making the contracts financeable enough to scale rapidly.

Notes

• The core thesis is that AI/datacenter economics, not just engineering, drive the SpaceX opportunity.
• The speakers say current inference economics can support about $100M per MW-year.
• Five-year average IaaS GPU pricing is described as roughly $12–13M per MW-year in baseline form.
• Premiums appear when capacity is on-demand or contractually scarce.
• A 90-day cancellation clause is treated as a major buyer-risk reducer.
• The technical forecast is grounded in real workload testing and a simulator.
• The model uses GB200 and GB300 testing plus op-by-op simulation.
• Long context windows are a memory-pressure problem because they inflate KV cache.
• The simulator tracks GEMMs, collectives, and future bandwidth/flops changes on Rubin.
• Rubin is expected to shift the performance and power envelope materially next year.
• The speakers say Frontier/open-source scale and long contexts make efficiency and KV offloading central.
• The argument is that compute supply remains constrained because capital must be deployed well ahead of revenue.
• Build lead times are typically 12 to 18 months, which keeps supply behind demand.
• The debate narrows to whether SpaceX can build 10 GW quickly enough.
• SpaceX’s scaling question is execution, not abstract pricing power.
• The site-selection problem is framed as finding warehouses, power, and people.
• The team reportedly scanned about a million US sites in roughly two days.
• The practical template is warehouse plus gas pipeline.
• They found several million-square-foot warehouse candidates.
• The power picture includes large turbine availability, including undisclosed units.
• One estimate says Elon already has about 9 to 10 GW of turbines in the fleet.
• Labor is presented as a massive bottleneck alongside switchgear and cooling gear.
• The build strategy is expected to lean on Chinese-made electrical and mechanical equipment.
• Preassembled Chinese modules are viewed as speed-friendly, especially for on-demand capacity.
• Google signing with SpaceX is treated as market validation.
• Elon’s historical pattern is to build with less labor than peers.
• Colossus 2 reportedly peaked around 3,000 workers per day, about 3x below top peers.
• The supply-chain team reportedly saw 5 to 10 GW of next-year orders.
• Mississippi is used as a precedent for sidestepping Tennessee permitting.
• The Mississippi build started with a 1.2 GW power-plant concept and then added mobile turbines.
• The final Mississippi setup is described as 69 turbines and DOJ intervention.
• Warehouse zoning is cited as a permitting shortcut versus powered land.
• In some cases the only additional permit needed is air.
• One described build uses a private transmission line from a power plant about 2 miles away.
• The strategy is to find warehouses and then locate nearby power generation.
• The conversation says these sites are not optimized for top-tier SLA customers.
• Anthropic-style self-build concepts are being discussed with 99.7% uptime.
• Customers are said to accept lower SLAs if speed to market improves.
• Tesla batteries are mentioned as a redundancy aid.
• Google and Anthropic deals are used as proof there are real customers.
• Microsoft is presented as the clearest customer route for OpenAI-linked demand.
• Microsoft paused data-center expansion in late 2024 and early 2025.
• The speakers estimate Microsoft’s OpenAI off-take at about 7 GW.
• Some Microsoft capacity is framed as serving OpenAI via infrastructure-as-a-service.
• The Microsoft/OpenAI route is said to price closer to $12M per MW-year than $100M.
• Microsoft and Meta are described as the most active pre-leasers this year.
• Chevron/Pecos County is cited as a new behind-the-meter 2.7 GW agreement.
• The capacity gap is said to hit late 2026 and H1 2027, with many leases not arriving until late 2027–2028.
• Cancelable contracts are described as low-risk because downside is limited to three months of payments.
• The low-risk structure is said to make CFO approval straightforward.
• The speakers claim 10 GW of contracts implies over $300B of contractual value.
• They contrast balance-sheet-heavy infrastructure with more flexible on-demand compute.
• Nvidia is expected to finance part of the build because payback is under a year.
• Vendor financing could make the build nearly cash-neutral.
• SpaceX and xAI are described as Nvidia-exclusive at the frontier.
• The speakers say Google is not likely to compete at the frontier, while Microsoft is.